From 683b94b462672549668787e6e7a1a55fc331734c Mon Sep 17 00:00:00 2001 From: mika-lt <349045450+mika-lt@users.noreply.github.com> Date: Thu, 29 Oct 2026 07:41:54 -0700 Subject: [PATCH 1/8] fix(fallback): read stored router_settings from the database on fallback writes (#48999) * fix(fallback): read stored router_settings from the database on fallback writes POST /fallback and DELETE /fallback/{model} read the current rules through the config cache. When Redis refuses DEL, the eviction before that read fails and Redis keeps serving the old copy, so each new rule dropped the earlier ones The write path now reads the router_settings row straight from the writer database and stores only that row plus the changed key, so config-file router settings no longer get copied into the database row * fix(fallback): edit the config file's fallback list when it owns the key A delete of a rule only the config file holds answers the config-owned 400 again instead of 404, and restating a rule the file already has answers 200 without copying the file's list into the database row. Adds a writer-read regression test and an integration test against a Redis whose ACL refuses DEL * fix(fallback): re-read the config file's fallbacks ownership on every fallback write * test(fallback): run the real config reload in the fallback write tests * test(fallback): key the refused-DEL Redis double by cache key * fix(fallback): keep the config-owned 400 for an unset env reference and replace a non-object stored row * fix(fallback): replace a stored router_settings row that holds a non-JSON string * fix(fallback): compare a restated config-file rule against its resolved env value --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- .../fallback_management_endpoints.py | 64 ++++-- .../test_fallback_management_public_names.py | 37 ++- .../test_fallback_management_endpoints.py | 215 ++++++++++++++---- 3 files changed, 255 insertions(+), 61 deletions(-) diff --git a/litellm/proxy/management_endpoints/fallback_management_endpoints.py b/litellm/proxy/management_endpoints/fallback_management_endpoints.py index 6a9a30a7689..ec8ee5ca213 100644 --- a/litellm/proxy/management_endpoints/fallback_management_endpoints.py +++ b/litellm/proxy/management_endpoints/fallback_management_endpoints.py @@ -20,7 +20,7 @@ from litellm._logging import verbose_proxy_logger from litellm.proxy._types import UserAPIKeyAuth from litellm.proxy.auth.model_checks import get_all_fallbacks from litellm.proxy.auth.user_api_key_auth import user_api_key_auth -from litellm.proxy.utils import PrismaClient, evict_config_param, invalidate_config_param +from litellm.proxy.utils import PrismaClient, invalidate_config_param if TYPE_CHECKING: from fastapi import APIRouter, Depends, HTTPException, status @@ -48,22 +48,50 @@ FallbackRule = dict[str, list[str]] StoredFallback = Annotated[FallbackRule | dict[str, object] | str, Field(union_mode="left_to_right")] _STORED_FALLBACKS: Final = TypeAdapter(list[StoredFallback]) _JSON_RULES: Final = TypeAdapter(list[JsonValue]) +_STORED_ROUTER_SETTINGS: Final = TypeAdapter(dict[str, JsonValue]) def _rule_covers(entry: StoredFallback, model: str) -> bool: return isinstance(entry, dict) and model in entry -async def _router_settings_fresh_from_db(proxy_config: "ProxyConfig") -> dict[str, object]: - await evict_config_param(ROUTER_SETTINGS_PARAM) - config: Final = await proxy_config.get_config() - return config.get(ROUTER_SETTINGS_PARAM, {}) +async def _reload_config_file_ownership(proxy_config: "ProxyConfig") -> None: + await proxy_config.get_config() -def _reject_config_owned_rules(proxy_config: "ProxyConfig", fallback_key: str, rules: Sequence[StoredFallback]) -> None: +async def _stored_router_settings(prisma_client: PrismaClient) -> Mapping[str, JsonValue]: + try: + row: Final = await ConfigRepository(prisma_client, use_writer=True).get_param(ROUTER_SETTINGS_PARAM) + except json.JSONDecodeError: + return {} + if row is None or not isinstance(row.param_value, Mapping): + return {} + return _STORED_ROUTER_SETTINGS.validate_python(row.param_value) + + +def _editable_rules( + proxy_config: "ProxyConfig", stored_settings: Mapping[str, JsonValue], fallback_key: str +) -> list[StoredFallback]: + file_settings: Final = proxy_config.router_settings + if not file_settings.owned_by_config(fallback_key): + return _STORED_FALLBACKS.validate_python(stored_settings.get(fallback_key) or []) + file_rules: Final = file_settings.get(fallback_key) + return _STORED_FALLBACKS.validate_python(file_rules if isinstance(file_rules, list) else []) + + +async def _write_rules( + prisma_client: PrismaClient, + proxy_config: "ProxyConfig", + stored_settings: Mapping[str, JsonValue], + fallback_key: str, + rules: Sequence[StoredFallback], +) -> None: proxy_config.reject_config_owned_writes( section_name=ROUTER_SETTINGS_PARAM, changed_keys={fallback_key: _JSON_RULES.validate_python(rules)} ) + if proxy_config.router_settings.owned_by_config(fallback_key): + return + await _persist_router_settings(prisma_client, {**stored_settings, fallback_key: rules}) async def _persist_router_settings(prisma_client: PrismaClient, router_settings: Mapping[str, object]) -> None: @@ -162,7 +190,8 @@ async def create_fallback( }, ) - router_settings: Final = await _router_settings_fresh_from_db(proxy_config) + await _reload_config_file_ownership(proxy_config) + router_settings: Final = await _stored_router_settings(prisma_client) # Get the appropriate fallback list based on type fallback_key = "fallbacks" @@ -172,7 +201,7 @@ async def create_fallback( fallback_key = "content_policy_fallbacks" # Get existing fallbacks - existing_fallbacks: Final = _STORED_FALLBACKS.validate_python(router_settings.get(fallback_key) or []) + existing_fallbacks: Final = _editable_rules(proxy_config, router_settings, fallback_key) # Update or add the fallback configuration fallback_updated = False @@ -187,12 +216,7 @@ async def create_fallback( # Add new fallback existing_fallbacks.append({data.model: data.fallback_models}) - _reject_config_owned_rules(proxy_config, fallback_key, existing_fallbacks) - - # Update router settings - router_settings[fallback_key] = existing_fallbacks - - await _persist_router_settings(prisma_client, router_settings) + await _write_rules(prisma_client, proxy_config, router_settings, fallback_key, existing_fallbacks) # Update the in-memory router configuration setattr(llm_router, fallback_key, existing_fallbacks) @@ -320,7 +344,8 @@ async def delete_fallback( }, ) - router_settings: Final = await _router_settings_fresh_from_db(proxy_config) + await _reload_config_file_ownership(proxy_config) + router_settings: Final = await _stored_router_settings(prisma_client) # Get the appropriate fallback list based on type fallback_key = "fallbacks" @@ -330,7 +355,7 @@ async def delete_fallback( fallback_key = "content_policy_fallbacks" # Get existing fallbacks - existing_fallbacks: Final = _STORED_FALLBACKS.validate_python(router_settings.get(fallback_key) or []) + existing_fallbacks: Final = _editable_rules(proxy_config, router_settings, fallback_key) updated_fallbacks: Final = [rule for rule in existing_fallbacks if not _rule_covers(rule, model)] @@ -340,12 +365,7 @@ async def delete_fallback( detail={"error": f"No {fallback_type} fallbacks configured for model '{model}'"}, ) - _reject_config_owned_rules(proxy_config, fallback_key, updated_fallbacks) - - # Update router settings - router_settings[fallback_key] = updated_fallbacks - - await _persist_router_settings(prisma_client, router_settings) + await _write_rules(prisma_client, proxy_config, router_settings, fallback_key, updated_fallbacks) # Update the in-memory router configuration setattr(llm_router, fallback_key, updated_fallbacks) diff --git a/tests/integration/management/test_fallback_management_public_names.py b/tests/integration/management/test_fallback_management_public_names.py index fcbeee9dc60..6f97165fb3c 100644 --- a/tests/integration/management/test_fallback_management_public_names.py +++ b/tests/integration/management/test_fallback_management_public_names.py @@ -1,6 +1,7 @@ import uuid from collections.abc import Iterator from concurrent.futures import ThreadPoolExecutor +from contextlib import contextmanager from functools import partial from itertools import chain from pathlib import Path @@ -10,6 +11,7 @@ import httpx import psutil import pytest from pydantic import JsonValue +from redis import Redis from tests.integration._support.client import ( Gateway, @@ -27,7 +29,7 @@ from tests.integration._support.process import ( setup_only_proxy_run, worker_configuration, ) -from tests.integration._support.redis_process import owned_redis +from tests.integration._support.redis_process import OwnedRedis, owned_redis pytestmark: Final = [pytest.mark.timeout(300), pytest.mark.shares_worker_state] @@ -385,6 +387,39 @@ def test_fallback_write_survives_a_redis_outage(gateway: Gateway, tmp_path: Path cache.start() +@contextmanager +def _refusing_cache_deletes(cache: OwnedRedis) -> Iterator[None]: + with Redis(host=cache.host, port=cache.port) as client: + client.acl_setuser("default", enabled=True, commands=["-del", "-unlink"]) + try: + yield + finally: + client.acl_setuser("default", enabled=True, commands=["+del", "+unlink"]) + + +@pytest.mark.usefixtures("migrated_database") +def test_consecutive_creates_keep_every_rule_when_redis_refuses_the_cache_eviction( + gateway: Gateway, tmp_path: Path +) -> None: + with owned_redis(tmp_path) as cache: + overrides: Final = {"REDIS_HOST": cache.host, "REDIS_PORT": str(cache.port)} + with owned_proxy_process(gateway, tmp_path, overrides, database_setup=()) as owned: + with owned.gateway.scenario() as scenario: + target: Final = scenario.model(model=f"openai/x3t-{uuid.uuid4().hex}") + first: Final = scenario.model(model=f"openai/x3a-{uuid.uuid4().hex}") + second: Final = scenario.model(model=f"openai/x3b-{uuid.uuid4().hex}") + for model in (target, first, second): + _wait_until_served(owned.gateway, model) + with _refusing_cache_deletes(cache): + created: Final = tuple( + _create_fallback(owned.gateway, scenario, model, [target]) for model in (first, second) + ) + assert all(response.status_code == 200 for response in created), [r.text for r in created] + covered: Final = _covered_models(_stored_fallbacks()) + assert {first, second} <= covered, covered + assert EVICTION_WARNING in owned.log.read_text(), owned.log.read_text()[-3000:] + + @pytest.mark.usefixtures("migrated_database") def test_fallback_write_survives_a_worker_kill(gateway: Gateway, tmp_path: Path) -> None: with owned_proxy_process(gateway, tmp_path, {}, workers=2, database_setup=()) as owned: diff --git a/tests/unit/proxy/test_fallback_management_endpoints.py b/tests/unit/proxy/test_fallback_management_endpoints.py index 530d22f02be..4a0e8cb7a9a 100644 --- a/tests/unit/proxy/test_fallback_management_endpoints.py +++ b/tests/unit/proxy/test_fallback_management_endpoints.py @@ -10,16 +10,18 @@ Tests: import copy import json -from collections.abc import AsyncIterator -from types import SimpleNamespace -from typing import Final +from collections.abc import AsyncIterator, Mapping +from types import MappingProxyType, SimpleNamespace +from typing import Final, get_args from unittest.mock import AsyncMock, MagicMock, patch import pytest from fastapi import HTTPException from pydantic import JsonValue, TypeAdapter +from redis.exceptions import NoPermissionError from litellm import Router +from litellm.proxy.config_resolvers.settings_rules import Section from litellm.proxy.management_endpoints.fallback_management_endpoints import ( FallbackCreateRequest, FallbackDeleteResponse, @@ -29,10 +31,19 @@ from litellm.proxy.management_endpoints.fallback_management_endpoints import ( get_fallback, ) from litellm.proxy.proxy_server import ProxyConfig -from litellm.proxy.utils import evict_config_param, get_config_param +from litellm.proxy.utils import evict_config_param, get_config_param, litellm_config_cache from litellm.types.router import DeploymentTypedDict, LiteLLMParamsTypedDict +@pytest.fixture(autouse=True) +async def empty_config_cache() -> AsyncIterator[None]: + for section in get_args(Section): + await evict_config_param(section) + yield + for section in get_args(Section): + await evict_config_param(section) + + class TestFallbackCreateRequest: """Test the FallbackCreateRequest validation""" @@ -123,7 +134,7 @@ PRIMARY_DEPLOYMENT_ID: Final = "team-primary-id" FALLBACK_DEPLOYMENT_ID: Final = "team-fallback-id" FallbackRules = list[dict[str, list[str]] | dict[str, object] | str] -RouterSettings = dict[str, FallbackRules | None] +RouterSettings = dict[str, FallbackRules | str | None] _ROUTER_SETTINGS_JSON: Final = TypeAdapter(dict[str, JsonValue]) @@ -168,17 +179,18 @@ class TestCreateFallbackForTeamScopedModels: @pytest.fixture def prisma_client(self) -> MagicMock: client: Final = MagicMock() + client.writer_db = client.db + client.db.litellm_config.find_unique = AsyncMock(return_value=None) client.db.litellm_config.upsert = AsyncMock() + client.get_generic_data = AsyncMock(return_value=None) return client @pytest.fixture - def proxy_config(self) -> MagicMock: - config: Final = MagicMock() - config.get_config = AsyncMock(return_value={"router_settings": {}}) - return config + def proxy_config(self) -> ProxyConfig: + return _ConfigFileProxyConfig() async def _create( - self, request: FallbackCreateRequest, router: Router, prisma_client: MagicMock, proxy_config: MagicMock + self, request: FallbackCreateRequest, router: Router, prisma_client: MagicMock, proxy_config: ProxyConfig ) -> FallbackResponse: with ( patch("litellm.proxy.proxy_server.llm_router", router), @@ -254,26 +266,43 @@ def _config_row(router_settings: RouterSettings) -> SimpleNamespace: return SimpleNamespace(param_name="router_settings", param_value=router_settings) +class _ConfigFileProxyConfig(ProxyConfig): + def __init__(self, config_file: RouterSettings | None = None) -> None: + super().__init__() + self.config_file: RouterSettings = config_file or {} + self.router_settings.load_yaml(_ROUTER_SETTINGS_JSON.validate_python(self.config_file)) + + async def _get_config_from_file(self, config_file_path: str | None = None) -> dict[str, RouterSettings]: + return {"router_settings": copy.deepcopy(self.config_file)} + + class _StoredRouterSettings: """The LiteLLM_Config router_settings row, with the proxy objects that read it through the config cache""" - def __init__(self, router_settings: RouterSettings | None, config_file: RouterSettings | None = None) -> None: + def __init__( + self, + router_settings: RouterSettings | None, + config_file: RouterSettings | None = None, + *, + lagging_replica: bool = False, + ) -> None: self.row: SimpleNamespace | None = None if router_settings is None else _config_row(router_settings) self.prisma_client: Final = MagicMock() - self.prisma_client.get_generic_data = AsyncMock(side_effect=lambda **_: self.row) + self.prisma_client.get_generic_data = AsyncMock( + side_effect=lambda value, **_: self.row if value == "router_settings" else None + ) + self.prisma_client.writer_db.litellm_config.find_unique = AsyncMock( + side_effect=lambda **_: copy.deepcopy(self.row) + ) + self.prisma_client.db.litellm_config.find_unique = AsyncMock( + side_effect=lambda **_: None if lagging_replica else copy.deepcopy(self.row) + ) self.prisma_client.db.litellm_config.upsert = AsyncMock(side_effect=self._upsert) - self.proxy_config: Final = MagicMock() if config_file is None else ProxyConfig() - if config_file is not None: - self.proxy_config.router_settings.load_yaml(_ROUTER_SETTINGS_JSON.validate_python(config_file)) - self.proxy_config.get_config = AsyncMock(side_effect=self._get_config) + self.proxy_config: Final = _ConfigFileProxyConfig(config_file) async def _upsert(self, where: dict[str, str], data: dict[str, dict[str, str]]) -> None: self.row = _config_row(json.loads(data["update"]["param_value"])) - async def _get_config(self) -> dict[str, RouterSettings]: - row: Final = await get_config_param(self.prisma_client, "router_settings") - return {"router_settings": copy.deepcopy(row.param_value) if row is not None else {}} - def written_by_another_instance(self, router_settings: RouterSettings) -> None: self.row = _config_row(router_settings) @@ -295,15 +324,23 @@ NON_STANDARD_RULES: Final = [ ] +class _RedisThatRefusesDel: + def __init__(self) -> None: + self.values: Mapping[str, object] = MappingProxyType({}) + + async def async_get_cache(self, key: str, parent_otel_span: object = None) -> object: + return self.values.get(key) + + async def async_set_cache(self, key: str, value: object, **_: object) -> None: + self.values = MappingProxyType({**self.values, key: value}) + + async def async_delete_cache(self, key: str) -> None: + raise NoPermissionError("this user has no permissions to run the 'del' command") + + class TestFallbackWritesSeeTheLatestStoredRules: """A write reads the rules the database holds now and leaves no stale copy in the config cache behind""" - @pytest.fixture(autouse=True) - async def clean_config_cache(self) -> AsyncIterator[None]: - await evict_config_param("router_settings") - yield - await evict_config_param("router_settings") - async def _create(self, request: FallbackCreateRequest, stored: _StoredRouterSettings) -> FallbackResponse: with ( patch("litellm.proxy.proxy_server.llm_router", _team_router()), @@ -322,6 +359,13 @@ class TestFallbackWritesSeeTheLatestStoredRules: ): return await delete_fallback(model, "general", MagicMock()) + async def test_a_create_reads_the_rules_from_the_writer_while_the_replica_lags(self) -> None: + stored: Final = _StoredRouterSettings({"fallbacks": [TEAM_RULE]}, lagging_replica=True) + + await self._create(FallbackCreateRequest(model="gpt-5.4-mini", fallback_models=["team-fallback"]), stored) + + assert stored.stored_fallbacks() == [TEAM_RULE, GATEWAY_RULE] + async def test_a_second_create_keeps_the_first_rule(self) -> None: stored: Final = _StoredRouterSettings(None) @@ -331,6 +375,17 @@ class TestFallbackWritesSeeTheLatestStoredRules: assert stored.stored_fallbacks() == [TEAM_RULE, GATEWAY_RULE] assert await stored.cached_fallbacks() == [TEAM_RULE, GATEWAY_RULE] + async def test_a_second_create_keeps_the_first_rule_when_redis_refuses_to_evict_the_cached_settings(self) -> None: + stored: Final = _StoredRouterSettings(None) + + with patch.object(litellm_config_cache, "redis_cache", _RedisThatRefusesDel()): + await self._create(FallbackCreateRequest(model="team-primary", fallback_models=["team-fallback"]), stored) + await self._create(FallbackCreateRequest(model="gpt-5.4-mini", fallback_models=["team-fallback"]), stored) + await self._delete("team-primary", stored) + await self._create(FallbackCreateRequest(model="team-primary", fallback_models=["team-fallback"]), stored) + + assert stored.stored_fallbacks() == [GATEWAY_RULE, TEAM_RULE] + async def test_a_create_keeps_a_rule_another_instance_stored_since_this_one_last_read(self) -> None: stored: Final = _StoredRouterSettings({}) await get_config_param(stored.prisma_client, "router_settings") @@ -380,6 +435,15 @@ class TestFallbackWritesSeeTheLatestStoredRules: assert stored.stored_fallbacks() == [GATEWAY_RULE] + @pytest.mark.parametrize("param_value", [None, ["team-primary"], 7, "not json"]) + async def test_a_create_replaces_a_stored_row_that_is_not_an_object(self, param_value: JsonValue) -> None: + stored: Final = _StoredRouterSettings(None) + stored.row = SimpleNamespace(param_name="router_settings", param_value=param_value) + + await self._create(FallbackCreateRequest(model="gpt-5.4-mini", fallback_models=["team-fallback"]), stored) + + assert stored.row.param_value == {"fallbacks": [GATEWAY_RULE]} + CONFIG_RULE: Final = {"gpt-5.4-mini": ["team-fallback"]} CONFIG_WINDOW_RULE: Final = {"team-primary": ["gpt-5.4-mini"]} @@ -438,6 +502,18 @@ class TestFallbackWritesHonourTheConfigFile: stored.prisma_client.db.litellm_config.upsert.assert_not_awaited() assert router.fallbacks == [CONFIG_RULE] + async def test_a_create_after_the_config_file_takes_over_the_key_is_refused(self, router: Router) -> None: + stored: Final = _StoredRouterSettings({"fallbacks": [TEAM_RULE]}) + stored.proxy_config.config_file = {"fallbacks": [CONFIG_RULE]} + request: Final = FallbackCreateRequest(model="team-primary", fallback_models=["gpt-5.4-mini"]) + + with pytest.raises(HTTPException) as refused: + await self._create(request, router, stored) + + assert refused.value.status_code == 400 + assert "key 'fallbacks' is set in the config file" in str(refused.value.detail) + assert stored.stored_fallbacks() == [TEAM_RULE] + async def test_a_delete_of_a_config_owned_rule_is_refused( self, router: Router, stored: _StoredRouterSettings ) -> None: @@ -452,6 +528,57 @@ class TestFallbackWritesHonourTheConfigFile: stored.prisma_client.db.litellm_config.upsert.assert_not_awaited() assert router.fallbacks == [CONFIG_RULE] + async def test_a_delete_of_a_rule_only_the_config_file_holds_is_refused(self, router: Router) -> None: + stored: Final = _StoredRouterSettings(None, config_file={"fallbacks": [CONFIG_RULE]}) + router.fallbacks = [CONFIG_RULE] + + with pytest.raises(HTTPException) as refused: + await self._delete("gpt-5.4-mini", router, stored) + + assert refused.value.status_code == 400 + assert "key 'fallbacks' is set in the config file" in str(refused.value.detail) + assert stored.row is None + assert router.fallbacks == [CONFIG_RULE] + + async def test_writes_to_a_config_file_env_reference_that_resolves_to_nothing_are_refused( + self, router: Router, monkeypatch: pytest.MonkeyPatch + ) -> None: + monkeypatch.delenv("LITELLM_TEST_UNSET_FALLBACKS", raising=False) + stored: Final = _StoredRouterSettings(None, config_file={"fallbacks": "os.environ/LITELLM_TEST_UNSET_FALLBACKS"}) + + with pytest.raises(HTTPException) as refused: + await self._create(FallbackCreateRequest(model="team-primary", fallback_models=["team-fallback"]), router, stored) + with pytest.raises(HTTPException) as not_found: + await self._delete("team-primary", router, stored) + + assert refused.value.status_code == 400 + assert "key 'fallbacks' is set in the config file" in str(refused.value.detail) + assert not_found.value.status_code == 404 + assert stored.row is None + + async def test_a_create_that_restates_one_of_several_config_rules_writes_nothing(self, router: Router) -> None: + stored: Final = _StoredRouterSettings(None, config_file={"fallbacks": [CONFIG_RULE, CONFIG_WINDOW_RULE]}) + request: Final = FallbackCreateRequest(model="gpt-5.4-mini", fallback_models=["team-fallback"]) + + response: Final = await self._create(request, router, stored) + + assert response.fallback_models == ["team-fallback"] + assert stored.row is None + assert router.fallbacks == [CONFIG_RULE, CONFIG_WINDOW_RULE] + + async def test_a_create_that_restates_a_config_rule_next_to_an_env_reference_writes_nothing( + self, router: Router, monkeypatch: pytest.MonkeyPatch + ) -> None: + monkeypatch.setenv("LITELLM_TEST_FALLBACK_MODEL", "team-fallback") + env_rule: Final = {"team-primary": "os.environ/LITELLM_TEST_FALLBACK_MODEL"} + stored: Final = _StoredRouterSettings(None, config_file={"fallbacks": [CONFIG_RULE, env_rule]}) + request: Final = FallbackCreateRequest(model="gpt-5.4-mini", fallback_models=["team-fallback"]) + + response: Final = await self._create(request, router, stored) + + assert response.fallback_models == ["team-fallback"] + assert stored.row is None + async def test_a_create_that_restates_the_config_rule_is_accepted( self, router: Router, stored: _StoredRouterSettings ) -> None: @@ -462,6 +589,20 @@ class TestFallbackWritesHonourTheConfigFile: assert response.fallback_models == ["team-fallback"] assert stored.stored_fallbacks() == [CONFIG_RULE] + async def test_a_write_stores_only_the_database_rules_never_the_config_files_router_settings( + self, router: Router + ) -> None: + stored: Final = _StoredRouterSettings( + {"fallbacks": [TEAM_RULE]}, config_file={"routing_strategy": "latency-based-routing"} + ) + + await self._create(FallbackCreateRequest(model="gpt-5.4-mini", fallback_models=["team-fallback"]), router, stored) + assert stored.row is not None + assert stored.row.param_value == {"fallbacks": [TEAM_RULE, GATEWAY_RULE]} + + await self._delete("team-primary", router, stored) + assert stored.row.param_value == {"fallbacks": [GATEWAY_RULE]} + async def test_a_sibling_list_the_config_file_does_not_set_stays_writable( self, router: Router, stored: _StoredRouterSettings ) -> None: @@ -496,16 +637,17 @@ class TestCreateFallback: def mock_prisma_client(self): """Create a mock prisma client""" client = MagicMock() + client.writer_db = client.db + client.db.litellm_config.find_unique = AsyncMock(return_value=None) client.db.litellm_config.upsert = AsyncMock() + client.get_generic_data = AsyncMock(return_value=None) client.jsonify_object = lambda x: x return client @pytest.fixture def mock_proxy_config(self): """Create a mock proxy config""" - config = MagicMock() - config.get_config = AsyncMock(return_value={"router_settings": {}}) - return config + return _ConfigFileProxyConfig() @pytest.fixture def mock_user_api_key_dict(self): @@ -803,22 +945,19 @@ class TestDeleteFallback: def mock_prisma_client(self): """Create a mock prisma client""" client = MagicMock() + client.writer_db = client.db + client.db.litellm_config.find_unique = AsyncMock( + return_value=_config_row({"fallbacks": [{"gpt-3.5-turbo": ["gpt-4", "claude-3-haiku"]}]}) + ) client.db.litellm_config.upsert = AsyncMock() + client.get_generic_data = AsyncMock(return_value=None) client.jsonify_object = lambda x: x return client @pytest.fixture def mock_proxy_config(self): """Create a mock proxy config""" - config = MagicMock() - config.get_config = AsyncMock( - return_value={ - "router_settings": { - "fallbacks": [{"gpt-3.5-turbo": ["gpt-4", "claude-3-haiku"]}] - } - } - ) - return config + return _ConfigFileProxyConfig() @pytest.fixture def mock_user_api_key_dict(self): From 7dcdc2ffc930fac68f8a7bf0c02780afe390c035 Mon Sep 17 00:00:00 2001 From: yujonglee Date: Thu, 29 Oct 2026 07:55:02 -0700 Subject: [PATCH 2/8] build(model-catalog): ship the cost map inside the native extension (#48907) Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .github/scripts/smoke_test_native_wheel.py | 9 +++++- backend/Dockerfile | 1 + gateway/Dockerfile | 1 + litellm-rust/Cargo.lock | 1 + litellm-rust/crates/python-bridge/AGENTS.md | 12 ++++++++ litellm-rust/crates/python-bridge/Cargo.toml | 1 + litellm-rust/crates/python-bridge/src/lib.rs | 4 +++ .../python-bridge/src/model_registry.rs | 30 +++++++++++++++++++ litellm/rust_bridge/_native.pyi | 2 ++ migrations/Dockerfile | 1 + packaging/litellm-core/pyproject.toml | 1 + pyproject.toml | 2 ++ scripts/build_core_distribution.py | 10 ++++++- .../base_sdk_tests/test_core_distribution.py | 4 ++- 14 files changed, 76 insertions(+), 3 deletions(-) create mode 100644 litellm-rust/crates/python-bridge/src/model_registry.rs diff --git a/.github/scripts/smoke_test_native_wheel.py b/.github/scripts/smoke_test_native_wheel.py index 577bb32fcf0..af4054ea92e 100644 --- a/.github/scripts/smoke_test_native_wheel.py +++ b/.github/scripts/smoke_test_native_wheel.py @@ -7,6 +7,8 @@ import zipfile from pathlib import Path from typing import Final +COST_MAP: Final = Path(__file__).resolve().parents[2] / "model_prices_and_context_window.json" + CHILD_SCRIPT: Final = """ from importlib.util import module_from_spec, spec_from_file_location from pathlib import Path @@ -34,6 +36,9 @@ else: after = module.gil_stats() if not isinstance(after.get("releases"), int): raise AssertionError(f"native module unusable after panic: {after!r}") + +if module.bundled_model_cost_map() != Path(sys.argv[2]).read_bytes(): + raise AssertionError("the wheel embeds a different model_prices_and_context_window.json than this checkout") """ @@ -55,7 +60,9 @@ def main() -> int: native_path: Final = Path(temporary_directory) / Path(native_members[0].filename).name native_path.write_bytes(archive.read(native_members[0])) - result: Final = subprocess.run((sys.executable, "-c", CHILD_SCRIPT, str(native_path)), check=False) + result: Final = subprocess.run( + (sys.executable, "-c", CHILD_SCRIPT, str(native_path), str(COST_MAP)), check=False + ) if result.returncode != 0: sys.stderr.write(f"native wheel smoke test exited with status {result.returncode}\n") diff --git a/backend/Dockerfile b/backend/Dockerfile index 1c22f914100..aad59ba17c6 100644 --- a/backend/Dockerfile +++ b/backend/Dockerfile @@ -24,6 +24,7 @@ ENV UV_PYTHON_DOWNLOADS=0 COPY pyproject.toml README.md rust-toolchain.toml ./ COPY .cargo ./.cargo COPY litellm-rust ./litellm-rust +COPY model_prices_and_context_window.json ./ RUN --mount=type=cache,target=/root/.cache/uv \ mkdir -p litellm/rust_bridge /wheels && \ diff --git a/gateway/Dockerfile b/gateway/Dockerfile index 6dc8bab45ce..9d228627543 100644 --- a/gateway/Dockerfile +++ b/gateway/Dockerfile @@ -40,6 +40,7 @@ ENV UV_PYTHON_DOWNLOADS=0 COPY pyproject.toml README.md rust-toolchain.toml ./ COPY .cargo ./.cargo COPY litellm-rust ./litellm-rust +COPY model_prices_and_context_window.json ./ RUN --mount=type=cache,target=/root/.cache/uv \ mkdir -p litellm/rust_bridge /wheels && \ diff --git a/litellm-rust/Cargo.lock b/litellm-rust/Cargo.lock index f7d7e5f4dab..e959f77917b 100644 --- a/litellm-rust/Cargo.lock +++ b/litellm-rust/Cargo.lock @@ -4345,6 +4345,7 @@ dependencies = [ "litellm-inference-transcription", "litellm-llms", "litellm-llms-types", + "litellm-model-catalog", "litellm-router-types", "litellm-secrets", "litellm-secrets-aws", diff --git a/litellm-rust/crates/python-bridge/AGENTS.md b/litellm-rust/crates/python-bridge/AGENTS.md index d90fe073b20..26858b2afe2 100644 --- a/litellm-rust/crates/python-bridge/AGENTS.md +++ b/litellm-rust/crates/python-bridge/AGENTS.md @@ -106,6 +106,18 @@ GIL handling to `litellm-host-python`. ## Data Handling +- Reuse large immutable bundled data and parsed catalogs across calls, including in + replacement APIs. Share Rust storage and cache repeated Python projections instead + of reparsing, serializing or copying the full payload on every access +- Cache owned Python objects with `PyOnceLock>`, then bind and clone references + while attached. Do not use blocking `OnceLock::get_or_init` while attached + ([PyO3 initialization](https://pyo3.rs/v0.29.2/faq.html), + [reference ownership](https://pyo3.rs/v0.29.2/types.html)) +- Never reuse Python objects across interpreters or interpreter reinitialization. + Once-only caches are for immutable data. Reloadable/custom data needs versioned + entries or invalidation, and caller-mutable Python containers must not be shared +- Test unchanged contents and allocation reuse, not just equal values. Keep this + coverage when replacing a bundled-data accessor with a catalog API - OCR payloads can contain personal data and large base64 images. Do not log payloads or provider responses. - Avoid copying large payloads more than needed. The current JSON round-trip is diff --git a/litellm-rust/crates/python-bridge/Cargo.toml b/litellm-rust/crates/python-bridge/Cargo.toml index 9d69cd8c14b..a4e1c57ab1e 100644 --- a/litellm-rust/crates/python-bridge/Cargo.toml +++ b/litellm-rust/crates/python-bridge/Cargo.toml @@ -49,6 +49,7 @@ litellm-secrets-types.workspace = true litellm-llms-types.workspace = true litellm-router-types.workspace = true litellm-host-python.workspace = true +litellm-model-catalog.workspace = true litellm-token-counter = { path = "../token-counter", default-features = false } pyo3.workspace = true prost.workspace = true diff --git a/litellm-rust/crates/python-bridge/src/lib.rs b/litellm-rust/crates/python-bridge/src/lib.rs index 12827b76ff3..683e90d02a2 100644 --- a/litellm-rust/crates/python-bridge/src/lib.rs +++ b/litellm-rust/crates/python-bridge/src/lib.rs @@ -8,6 +8,7 @@ mod execution; mod http; mod logger; mod marshal; +mod model_registry; mod preflight; mod python_settings; mod routes; @@ -30,6 +31,8 @@ mod _native { #[pymodule_export] use crate::logger::NativeDiagnosticProcessor; #[pymodule_export] + use crate::model_registry::bundled_model_cost_map; + #[pymodule_export] use crate::routes::audio_transcription::{atranscription, transcription}; #[pymodule_export] use crate::routes::chat_completions::{acompletion, completion}; @@ -110,6 +113,7 @@ mod tests { "NativeClickHouseSpendConfig", "NativeClickHouseSpendStorage", "trace_encode_error", + "bundled_model_cost_map", "TokenCounter", "Tokenizer", "gil_stats", diff --git a/litellm-rust/crates/python-bridge/src/model_registry.rs b/litellm-rust/crates/python-bridge/src/model_registry.rs new file mode 100644 index 00000000000..b7b7cd1fbb3 --- /dev/null +++ b/litellm-rust/crates/python-bridge/src/model_registry.rs @@ -0,0 +1,30 @@ +use pyo3::prelude::*; +use pyo3::sync::PyOnceLock; +use pyo3::types::PyBytes; + +#[pyfunction] +pub fn bundled_model_cost_map(py: Python<'_>) -> Bound<'_, PyBytes> { + static BYTES: PyOnceLock> = PyOnceLock::new(); + BYTES + .get_or_init(py, || { + PyBytes::new(py, litellm_model_catalog::BUNDLED_REGISTRY).unbind() + }) + .bind(py) + .clone() +} + +#[cfg(test)] +mod tests { + use super::*; + + #[rstest::rstest] + fn repeated_calls_reuse_the_bundled_registry_bytes() { + Python::initialize(); + let first = Python::attach(|py| bundled_model_cost_map(py).unbind()); + Python::attach(|py| { + let second = bundled_model_cost_map(py); + assert!(first.bind(py).is(&second)); + assert_eq!(second.as_bytes(), litellm_model_catalog::BUNDLED_REGISTRY); + }); + } +} diff --git a/litellm/rust_bridge/_native.pyi b/litellm/rust_bridge/_native.pyi index 7364a1d72f2..62b48621c05 100644 --- a/litellm/rust_bridge/_native.pyi +++ b/litellm/rust_bridge/_native.pyi @@ -14,6 +14,7 @@ class ForkedAfterNativeRuntimeStarted(RuntimeError): ... class ProcessReservedForForking(RuntimeError): ... def trace_encode_error(message: str) -> bytes: ... +def bundled_model_cost_map() -> bytes: ... @final class NativeClickHouseSpendConfig: @@ -220,6 +221,7 @@ __all__ = [ "aocr", "aresponses", "atranscription", + "bundled_model_cost_map", "completion", "embedding", "gil_stats", diff --git a/migrations/Dockerfile b/migrations/Dockerfile index 2b1fd029143..3ffa9140543 100644 --- a/migrations/Dockerfile +++ b/migrations/Dockerfile @@ -43,6 +43,7 @@ ENV UV_PYTHON_DOWNLOADS=0 COPY pyproject.toml README.md rust-toolchain.toml ./ COPY .cargo ./.cargo COPY litellm-rust ./litellm-rust +COPY model_prices_and_context_window.json ./ RUN --mount=type=cache,target=/root/.cache/uv \ mkdir -p litellm/rust_bridge /wheels && \ diff --git a/packaging/litellm-core/pyproject.toml b/packaging/litellm-core/pyproject.toml index 27681d9116c..db50589a675 100644 --- a/packaging/litellm-core/pyproject.toml +++ b/packaging/litellm-core/pyproject.toml @@ -56,6 +56,7 @@ editable-profile = "dev" include = [ { path = "rust-toolchain.toml", format = "sdist" }, { path = ".cargo/config.toml", format = "sdist" }, + { path = "model_prices_and_context_window.json", format = "sdist" }, "litellm/router_strategy/complexity_router/artifacts/*.json", "litellm/router_strategy/complexity_router/fuse_presets.json", "litellm/proxy/model_insights_tasks.json", diff --git a/pyproject.toml b/pyproject.toml index 200bb463313..e6d34a3d90c 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -335,6 +335,7 @@ include = [ "litellm/proxy/public_endpoints/whats_new.json", "litellm/proxy/common_utils/codex_base_instructions.md", "litellm/proxy/common_utils/codex_bundled_models_0.159.3.json", + { path = "model_prices_and_context_window.json", format = "sdist" }, ] exclude = [ "litellm/proxy/enterprise", @@ -357,6 +358,7 @@ cache-keys = [ { file = "litellm-rust/Cargo.lock" }, { file = "litellm-rust/Cargo.toml" }, { file = "litellm-rust/crates/**/*" }, + { file = "model_prices_and_context_window.json" }, ] constraint-dependencies = [ "tornado>=6.5.8", diff --git a/scripts/build_core_distribution.py b/scripts/build_core_distribution.py index 36ec20a9031..41d7208d2c3 100644 --- a/scripts/build_core_distribution.py +++ b/scripts/build_core_distribution.py @@ -9,7 +9,15 @@ from pathlib import Path from typing import Final ROOT: Final = Path(__file__).resolve().parents[1] -SOURCES: Final = ("litellm", "litellm-rust", ".cargo", "rust-toolchain.toml", "README.md", "LICENSE") +SOURCES: Final = ( + "litellm", + "litellm-rust", + ".cargo", + "rust-toolchain.toml", + "model_prices_and_context_window.json", + "README.md", + "LICENSE", +) def stage_core_distribution(source: Path, destination: Path) -> None: diff --git a/tests/base_sdk_tests/test_core_distribution.py b/tests/base_sdk_tests/test_core_distribution.py index ba68d08e4d1..c22fa4e54fc 100644 --- a/tests/base_sdk_tests/test_core_distribution.py +++ b/tests/base_sdk_tests/test_core_distribution.py @@ -132,7 +132,9 @@ def test_core_artifacts_exclude_local_configuration(distributions: tuple[Path, P assert not any(member.name.endswith(f"/{name}") for member in archive.getmembers()) -@pytest.mark.parametrize("relative_path", ["rust-toolchain.toml", ".cargo/config.toml"]) +@pytest.mark.parametrize( + "relative_path", ["rust-toolchain.toml", ".cargo/config.toml", "model_prices_and_context_window.json"] +) def test_core_sdist_preserves_native_build_configuration(distribution_directory: Path, relative_path: str) -> None: sdists: Final = tuple(distribution_directory.glob("litellm_core-*.tar.gz")) assert len(sdists) == 1 From 2a61f12f78d09c0cf31f7d57156d51760879d00e Mon Sep 17 00:00:00 2001 From: yujonglee Date: Thu, 29 Oct 2026 07:55:02 -0700 Subject: [PATCH 3/8] ci(model-catalog): check the cost map by loading it through model-catalog (#48908) Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .github/workflows/cost-map-guard.yml | 21 ++++ litellm-rust/crates/model-catalog/AGENTS.md | 2 + litellm-rust/crates/model-catalog/Cargo.toml | 2 +- .../model-catalog/src/bin/check-registry.rs | 26 ++++ .../crates/model-catalog/src/catalog.rs | 114 +++++++++++++++++- .../crates/model-catalog/src/error.rs | 10 +- .../crates/model-catalog/tests/catalog.rs | 18 ++- 7 files changed, 179 insertions(+), 14 deletions(-) create mode 100644 litellm-rust/crates/model-catalog/src/bin/check-registry.rs diff --git a/.github/workflows/cost-map-guard.yml b/.github/workflows/cost-map-guard.yml index 6519b65e71b..fb86237e504 100644 --- a/.github/workflows/cost-map-guard.yml +++ b/.github/workflows/cost-map-guard.yml @@ -46,3 +46,24 @@ jobs: HEAD_REF: ${{ github.event.pull_request.head.ref }} run: | uv run --no-project python ci_cd/cost_map_guard.py --base "$MERGE_BASE" --head "$HEAD_SHA" --head-ref "$HEAD_REF" + - name: Check whether the pull request changes the cost map + id: changed + env: + MERGE_BASE: ${{ steps.revisions.outputs.merge_base }} + HEAD_SHA: ${{ github.event.pull_request.head.sha }} + run: | + if git diff --quiet "$MERGE_BASE" "$HEAD_SHA" -- model_prices_and_context_window.json; then + echo "cost_map=false" >> "$GITHUB_OUTPUT" + else + echo "cost_map=true" >> "$GITHUB_OUTPUT" + fi + - if: steps.changed.outputs.cost_map == 'true' + run: rustup toolchain install --no-self-update + - name: Load the cost map through model-catalog + if: steps.changed.outputs.cost_map == 'true' + env: + HEAD_SHA: ${{ github.event.pull_request.head.sha }} + run: | + git show "${HEAD_SHA}:model_prices_and_context_window.json" > "$RUNNER_TEMP/model_prices_and_context_window.json" + cargo run --locked --quiet --manifest-path litellm-rust/Cargo.toml -p litellm-model-catalog --bin check-registry \ + -- "$RUNNER_TEMP/model_prices_and_context_window.json" diff --git a/litellm-rust/crates/model-catalog/AGENTS.md b/litellm-rust/crates/model-catalog/AGENTS.md index fc4b28d4ef0..ff24fb81143 100644 --- a/litellm-rust/crates/model-catalog/AGENTS.md +++ b/litellm-rust/crates/model-catalog/AGENTS.md @@ -25,3 +25,5 @@ A new `N`, image size or resolution in an existing key shape parses with no Rust ## Tests `BUNDLED_REGISTRY` embeds the root `model_prices_and_context_window.json`, so the tests over it fail a JSON change with a bad value or a key no type reads + +`Catalog::problems` is the one definition of a mergeable cost map. The bundled test asserts it is empty, and the cost map guard workflow runs it on a pull request's JSON through `cargo run --bin check-registry ` diff --git a/litellm-rust/crates/model-catalog/Cargo.toml b/litellm-rust/crates/model-catalog/Cargo.toml index 45d16515dcd..63b9ad14768 100644 --- a/litellm-rust/crates/model-catalog/Cargo.toml +++ b/litellm-rust/crates/model-catalog/Cargo.toml @@ -10,6 +10,7 @@ schema = ["dep:schemars", "litellm-cost/schema"] [dependencies] litellm-cost = { workspace = true, features = ["serde"] } +litellm-error.workspace = true indexmap = { version = "2.14.0", features = ["serde"] } jiff = { version = "0.2.37", default-features = false, features = ["std", "tzdb-bundle-always"] } @@ -24,5 +25,4 @@ time.workspace = true [dev-dependencies] jsonschema = { version = "0.55.1", default-features = false } -litellm-error.workspace = true rstest.workspace = true diff --git a/litellm-rust/crates/model-catalog/src/bin/check-registry.rs b/litellm-rust/crates/model-catalog/src/bin/check-registry.rs new file mode 100644 index 00000000000..6fdc7621a0a --- /dev/null +++ b/litellm-rust/crates/model-catalog/src/bin/check-registry.rs @@ -0,0 +1,26 @@ +use std::{env, fs, path::PathBuf, process::ExitCode}; + +use litellm_model_catalog::Catalog; + +fn main() -> ExitCode { + let Some(path) = env::args_os().nth(1).map(PathBuf::from) else { + eprintln!("usage: check-registry "); + return ExitCode::from(2); + }; + let body = match fs::read(&path) { + Ok(body) => body, + Err(error) => { + eprintln!("{}: {error}", path.display()); + return ExitCode::FAILURE; + } + }; + let problems = Catalog::problems(&body); + for problem in &problems { + eprintln!("{}", litellm_error::Report::new(problem)); + } + if problems.is_empty() { + ExitCode::SUCCESS + } else { + ExitCode::FAILURE + } +} diff --git a/litellm-rust/crates/model-catalog/src/catalog.rs b/litellm-rust/crates/model-catalog/src/catalog.rs index ac5f4ffcf25..98af785b17e 100644 --- a/litellm-rust/crates/model-catalog/src/catalog.rs +++ b/litellm-rust/crates/model-catalog/src/catalog.rs @@ -5,7 +5,7 @@ use indexmap::IndexMap; use serde::Deserialize; use serde_json::Value; -use crate::error::{Error, Loaded, Rejects, Repr}; +use crate::error::{Error, Loaded, Rejects, Repr, pointer_child, pointer_for_key}; use crate::fallback::{FallbackGeneralizations, FallbackRule}; use crate::model::Model; use crate::store::{Rows, Store}; @@ -211,6 +211,36 @@ impl Catalog { }) } + pub fn problems(body: &[u8]) -> Vec { + let Loaded { + value: catalog, + errors, + } = match Self::parse(body) { + Ok(loaded) => loaded, + Err(fatal) => return vec![fatal], + }; + let unread = catalog.rows().entries.iter().flat_map(|(model, entry)| { + EntryWire::read(entry.fields()) + .unknown + .into_iter() + .map(move |(key, _)| { + Error::new( + pointer_child(&pointer_for_key(model), &key), + Repr::UnreadField, + ) + }) + }); + let no_fallback_rules = catalog + .fallback_rules() + .is_none_or(<[_]>::is_empty) + .then(|| Error::from_key(FALLBACK_GENERALIZATIONS, Repr::NoFallbackRules)); + errors + .into_iter() + .chain(unread) + .chain(no_fallback_rules) + .collect() + } + pub fn row_errors(model: &str, fields: &Fields) -> Vec { let (_, _, errors) = read_row(fields); errors @@ -404,3 +434,85 @@ fn unknown_timezone(wire: &EntryWire) -> Option { ) }) } + +#[cfg(test)] +mod tests { + use rstest::rstest; + + use super::*; + + const RULES: &str = r#""fallback_generalizations":{"rules":[{"name":"family","pattern":"^new-","model_info":{"mode":"chat"}}]}"#; + + #[rstest] + #[case::clean(format!(r#"{{{RULES},"m":{{"litellm_provider":"test"}}}}"#), vec![])] + #[case::not_json("{".to_owned(), vec![("", "json")])] + #[case::row_error(format!(r#"{{{RULES},"m":{{}}}}"#), vec![("/m/litellm_provider", "missing_provider")])] + #[case::unread_key( + format!(r#"{{{RULES},"m":{{"litellm_provider":"test","typo_cost":1}}}}"#), + vec![("/m/typo_cost", "unread_field")] + )] + #[case::empty_model_key( + format!(r#"{{{RULES},"":{{"litellm_provider":"test","future_field":true}}}}"#), + vec![("//future_field", "unread_field")] + )] + #[case::unread_key_on_invalid_row( + format!(r#"{{{RULES},"m":{{"typo_cost":1}}}}"#), + vec![("/m/litellm_provider", "missing_provider"), ("/m/typo_cost", "unread_field")] + )] + #[case::escaped_unread_key_on_invalid_row( + format!(r#"{{{RULES},"open/router~v1":{{"typo/cost~":1}}}}"#), + vec![("/open~1router~0v1/litellm_provider", "missing_provider"), ("/open~1router~0v1/typo~1cost~0", "unread_field")] + )] + #[case::no_rules( + r#"{"m":{"litellm_provider":"test"}}"#.to_owned(), + vec![("/fallback_generalizations", "no_fallback_rules")] + )] + #[case::empty_rules( + r#"{"fallback_generalizations":{"rules":[]},"m":{"litellm_provider":"test"}}"#.to_owned(), + vec![("/fallback_generalizations", "no_fallback_rules")] + )] + fn problems_list_what_a_cost_map_change_must_fix( + #[case] body: String, + #[case] expected: Vec<(&str, &str)>, + ) { + let problems: Vec<_> = Catalog::problems(body.as_bytes()) + .iter() + .map(|problem| (problem.path().to_owned(), <&str>::from(problem.kind()))) + .collect(); + let expected: Vec<_> = expected + .into_iter() + .map(|(path, kind)| (path.to_owned(), kind)) + .collect(); + assert_eq!(problems, expected); + } + + #[rstest] + fn runtime_parsing_preserves_custom_fields_and_recovers_valid_rows() { + let body = br#"{ + "good":{"litellm_provider":"test","custom_metadata":true}, + "bad":{"litellm_provider":"test","input_cost_per_token":"wrong"} + }"#; + let loaded = Catalog::parse(body).unwrap(); + assert_eq!(loaded.errors.len(), 1); + assert_eq!(loaded.errors[0].path(), "/bad/input_cost_per_token"); + let (catalog, rejected) = loaded.value.without_rejected(&loaded.errors); + assert_eq!(rejected.len(), 1); + assert!(catalog.lookup("bad").is_none()); + let good = catalog.lookup("good").unwrap(); + assert!(good.entry.model().is_some()); + assert_eq!( + good.entry.field("custom_metadata"), + Some(&Value::Bool(true)) + ); + assert!( + Catalog::problems(body) + .iter() + .any(|problem| problem.kind() == crate::ErrorKind::UnreadField) + ); + assert!( + Catalog::problems(body) + .iter() + .any(|problem| problem.kind() == crate::ErrorKind::NoFallbackRules) + ); + } +} diff --git a/litellm-rust/crates/model-catalog/src/error.rs b/litellm-rust/crates/model-catalog/src/error.rs index 425ae47687f..89e27c05f46 100644 --- a/litellm-rust/crates/model-catalog/src/error.rs +++ b/litellm-rust/crates/model-catalog/src/error.rs @@ -109,6 +109,12 @@ pub(crate) enum Repr { #[source] cause: Option, }, + #[error("not a field litellm reads")] + #[strum_discriminants(strum(serialize = "unread_field"))] + UnreadField, + #[error("has no rules")] + #[strum_discriminants(strum(serialize = "no_fallback_rules"))] + NoFallbackRules, } /// What a problem takes out of the served map @@ -148,7 +154,7 @@ impl Error { | Repr::NoOffPeakHours | Repr::EmptyList | Repr::InvalidDate { .. } => Rejects::Row, - Repr::MissingProvider => Rejects::Nothing, + Repr::MissingProvider | Repr::UnreadField | Repr::NoFallbackRules => Rejects::Nothing, } } } @@ -334,6 +340,8 @@ mod tests { #[case::no_off_peak_hours(Repr::NoOffPeakHours, Rejects::Row)] #[case::empty_list(Repr::EmptyList, Rejects::Row)] #[case::invalid_date(Repr::InvalidDate { cause: None }, Rejects::Row)] + #[case::unread_field(Repr::UnreadField, Rejects::Nothing)] + #[case::no_fallback_rules(Repr::NoFallbackRules, Rejects::Nothing)] fn policy_classifies_each_error(#[case] error: Repr, #[case] expected: Rejects) { assert_eq!(Error::from(error).rejects(), expected); } diff --git a/litellm-rust/crates/model-catalog/tests/catalog.rs b/litellm-rust/crates/model-catalog/tests/catalog.rs index 30c070cfb52..31f907689d3 100644 --- a/litellm-rust/crates/model-catalog/tests/catalog.rs +++ b/litellm-rust/crates/model-catalog/tests/catalog.rs @@ -366,18 +366,14 @@ fn written_is_the_document_in_order_including_sections_and_bad_entries() { } #[rstest] -fn the_bundled_registry_parses_with_no_problems_and_every_key_understood() { - let loaded = Catalog::parse(BUNDLED_REGISTRY).unwrap(); - let catalog = loaded.value; - assert!(loaded.errors.is_empty(), "{:#?}", loaded.errors); - let unrecognized: Vec<_> = catalog.unrecognized().collect(); - assert!(unrecognized.is_empty(), "{unrecognized:#?}"); - assert!( - catalog - .fallback_rules() - .is_some_and(|rules| !rules.is_empty()) - ); +fn the_bundled_registry_has_no_problems() { + let problems: Vec<_> = Catalog::problems(BUNDLED_REGISTRY) + .iter() + .map(report) + .collect(); + assert!(problems.is_empty(), "{problems:#?}"); const REFERENCE_MODEL_COUNT: usize = 4303; + let catalog = Catalog::parse(BUNDLED_REGISTRY).unwrap().value; assert!(catalog.model_count() as f64 >= REFERENCE_MODEL_COUNT as f64 * 0.9); } From cd58014315ef584e505a05cb2ab4d4eb07004a6d Mon Sep 17 00:00:00 2001 From: Dan Lemon Date: Thu, 29 Oct 2026 15:58:38 +0100 Subject: [PATCH 4/8] feat(soniox): add realtime speech-to-text on /v1/realtime (#44822) * feat(soniox): add realtime speech-to-text on /v1/realtime Maps OpenAI realtime transcription sessions onto the Soniox websocket API (stt-rt-v5): server VAD uses Soniox endpoint detection, commit sends a manual finalize, deployment litellm_params pass Soniox options through, translation sessions transcribe the translated tokens, idle streams get keepalives, and cost is tracked from processed audio duration. * fix(soniox): bill realtime usage from integer milliseconds * fix(soniox): bound session-start queue and cover realtime backend edge paths * fix(soniox): bill realtime stream time, start sessions on connect and surface client protocol errors Soniox charges for the whole open stream, so session close now bills the connection's open time. The backend announces the session on connect, which flushes audio sent before any session.update, and a single upstream pump makes recv safe to cancel. Invalid client frames now get an OpenAI error event instead of a silent disconnect * fix(soniox): skip deferred setup so client frames reach Soniox in order Drops the connect marker frame, which counted as an upstream frame and hid refusals, and could start a buffered stream at the default rate before the client's session.update. Soniox now opts out of deferred setup, the handler announces a synthetic session.created whenever no setup went upstream, a session.update that would change a started stream gets an error event, and stream time is billed only once a stream started * fix(soniox): judge a follow-up session.update only on the settings it sends A partial follow-up session.update was rebuilt with defaults for the settings it left out, so restating part of a 16 kHz or VAD-off stream was rejected as a change * fix(soniox): treat an omitted rate in a follow-up session.update as unchanged * chore(soniox): drop comments that restate the code * fix(soniox): return speaker labels, log refused sessions as failures, and probe realtime health Deltas carry speaker and language when diarization or language identification is on, and the completed event lists segments per speaker and language. A session Soniox refuses before any audio is processed is logged as a failure with no spend. Billing starts at the start request. Deferred-setup providers (Meta, Vertex Chirp) send the error event and still end the session. Soniox realtime models get a websocket /health probe and cost rows for the stt-rt-v5 aliases. * fix(soniox): bill failed streams, probe health with deployment options, and keep token accumulation linear Only request rejections (400, 401, 402, 403, 429) before any processed audio count as a refused session, so a 408 or 5xx still bills the open stream. The /health probe starts Soniox with the deployment's options, sends 100 ms of silence, and passes only on Soniox's finished frame. A turn's final tokens append in constant time instead of copying the tuple per token. * fix(soniox): log every Soniox error before any processed audio as a failed session * fix(soniox): keep untranslated speech, support input_audio_buffer.clear, ignore an empty language * fix(soniox): reject frames without a type, keep strict language hints, queue finalizes in a deque * fix(soniox): answer a commit with no speech with an empty completed turn * fix(soniox): let a server VAD turn answer a commit that is still in flight * fix(soniox): let a server VAD turn answer only the oldest pending commit --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- .../health_check_helpers.py | 8 +- .../litellm_core_utils/realtime_streaming.py | 26 +- .../llms/base_llm/realtime/transformation.py | 9 +- litellm/llms/custom_httpx/llm_http_handler.py | 6 +- litellm/llms/soniox/common_utils.py | 8 +- litellm/llms/soniox/realtime/__init__.py | 0 .../llms/soniox/realtime/transformation.py | 609 +++++++++++++ ...odel_prices_and_context_window_backup.json | 120 +++ .../provider_endpoints_support_backup.json | 2 +- .../provider_create_fields.json | 2 +- litellm/realtime_api/README.md | 2 +- litellm/realtime_api/main.py | 14 +- litellm/utils.py | 5 + model_prices_and_context_window.json | 120 +++ provider_endpoints_support.json | 2 +- .../test_health_check_helpers.py | 5 +- .../test_realtime_streaming.py | 125 ++- .../custom_httpx/test_llm_http_handler.py | 154 ++++ tests/unit/llms/soniox/realtime/__init__.py | 1 + .../test_soniox_realtime_transformation.py | 807 ++++++++++++++++++ .../test_get_complete_url_trailing_slash.py | 1 + tests/unit/realtime_api/test_main.py | 101 ++- 22 files changed, 2093 insertions(+), 34 deletions(-) create mode 100644 litellm/llms/soniox/realtime/__init__.py create mode 100644 litellm/llms/soniox/realtime/transformation.py create mode 100644 tests/unit/llms/soniox/realtime/__init__.py create mode 100644 tests/unit/llms/soniox/realtime/test_soniox_realtime_transformation.py diff --git a/litellm/litellm_core_utils/health_check_helpers.py b/litellm/litellm_core_utils/health_check_helpers.py index b19bcd941fa..940730eeb9f 100644 --- a/litellm/litellm_core_utils/health_check_helpers.py +++ b/litellm/litellm_core_utils/health_check_helpers.py @@ -31,7 +31,13 @@ EXTRACTION_HEALTH_CHECK_FIELDS: Final[Mapping[str, Mapping[str, str]]] = Mapping ) -def native_health_check_mode(model: str, custom_llm_provider: str | None) -> Literal["anthropic_messages"] | None: +def native_health_check_mode( + model: str, custom_llm_provider: str | None +) -> Literal["anthropic_messages", "realtime"] | None: + if custom_llm_provider == "soniox": + from litellm.llms.soniox.common_utils import soniox_health_check_mode + + return soniox_health_check_mode(model) if custom_llm_provider != "bedrock_mantle": return None from litellm.llms.bedrock_mantle.common_utils import mantle_health_check_mode diff --git a/litellm/litellm_core_utils/realtime_streaming.py b/litellm/litellm_core_utils/realtime_streaming.py index 927caec3350..484bacdc72c 100644 --- a/litellm/litellm_core_utils/realtime_streaming.py +++ b/litellm/litellm_core_utils/realtime_streaming.py @@ -20,6 +20,7 @@ from litellm.constants import ( REALTIME_SESSION_SUCCESS_LOGGED_KEY, ) from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER +from litellm.llms.base_llm.realtime.transcription_protocol import RealtimeTranscriptionProtocolError from litellm.llms.base_llm.realtime.transformation import BaseRealtimeConfig, RealtimeBackend from litellm.types.llms.openai import ( OpenAIRealtimeEvents, @@ -1281,24 +1282,29 @@ class RealTimeStreaming: verbose_logger.exception("Connection closed in backend to client send messages - %s", e) close: Final = backend_close_from(e) self._flush_unbilled_transcription_usage() - if self._backend_refused_session(close): - await self.log_backend_refusal(e) - else: - await self.log_messages() + await self._log_session_end(e if self._backend_refused_session(close) else None) return close except asyncio.CancelledError: self._flush_unbilled_transcription_usage() - await self.log_messages() + await self._log_session_end(None) raise except Exception as e: verbose_logger.exception("Error in backend to client send messages: %s", e) self._flush_unbilled_transcription_usage() - await self.log_messages() + await self._log_session_end(None) return BackendClose(code=1011, reason="proxy failed while relaying the upstream websocket") def _backend_refused_session(self, close: BackendClose) -> bool: return close.code != 1000 and not self._backend_sent_frames + async def _log_session_end(self, transport_refusal: Exception | None) -> None: + provider_refusal: Final = None if self.provider_config is None else self.provider_config.session_refusal() + refusal: Final = provider_refusal or transport_refusal + if refusal is None: + await self.log_messages() + return + await self.log_backend_refusal(refusal) + async def log_backend_refusal(self, error: Exception) -> None: if not self.logging_obj: return @@ -1691,7 +1697,13 @@ class RealTimeStreaming: # ``_maybe_send_guardrail_turn_detection_update`` would retry. if msg_type == "session.update" and self.provider_config is None: await self._send_session_gate_once() - sent = await self._send_to_backend(message) + try: + sent = await self._send_to_backend(message) + except RealtimeTranscriptionProtocolError as e: + await self.websocket.send_text(realtime_error_event(str(e), error_type="invalid_request_error")) + if self._uses_deferred_backend_setup(): + raise + continue if guardrail_turn_detection_injected and sent: self._guardrail_turn_detection_update_sent = True diff --git a/litellm/llms/base_llm/realtime/transformation.py b/litellm/llms/base_llm/realtime/transformation.py index 1f4ad29fa74..f1539d16b30 100644 --- a/litellm/llms/base_llm/realtime/transformation.py +++ b/litellm/llms/base_llm/realtime/transformation.py @@ -4,7 +4,6 @@ from types import TracebackType from typing import TYPE_CHECKING, Any, Protocol import httpx -from typing_extensions import Self from litellm.types.llms.openai import OpenAIRealtimeStreamSessionEvents from litellm.types.realtime import ( @@ -24,7 +23,7 @@ else: class RealtimeBackend(Protocol): - async def __aenter__(self) -> Self: ... + async def __aenter__(self) -> object: ... async def __aexit__( self, @@ -97,9 +96,15 @@ class BaseRealtimeConfig(ABC): def unbilled_usage_on_session_close(self, model: str) -> RealtimeInputAudioTranscriptionUsage | None: return None + def session_refusal(self) -> Exception | None: + return None + async def open_backend(self, url: str, headers: Mapping[str, str]) -> RealtimeBackend | None: return None + def wrap_backend(self, backend: RealtimeBackend) -> RealtimeBackend: + return backend + def transform_session_created_event( self, model: str, diff --git a/litellm/llms/custom_httpx/llm_http_handler.py b/litellm/llms/custom_httpx/llm_http_handler.py index 4c0f84a003d..8a9ff62cfae 100644 --- a/litellm/llms/custom_httpx/llm_http_handler.py +++ b/litellm/llms/custom_httpx/llm_http_handler.py @@ -6923,7 +6923,7 @@ class BaseLLMHTTPHandler: try: ssl_context: Final = realtime_ssl_for_url(url) provider_backend: Final = await provider_config.open_backend(url, headers) - backend_ws: Final = ( + backend_ws: Final = provider_config.wrap_backend( provider_backend if provider_backend is not None else await self._open_realtime_backend_ws(websockets, url, headers, ssl_context) @@ -6961,9 +6961,7 @@ class BaseLLMHTTPHandler: await backend_ws.send(_session_config) realtime_streaming.session_configuration_request = _session_config - # For providers that defer setup until client session.update, optionally - # send synthetic session.created to unblock clients waiting on connect. - if not provider_config.requires_session_configuration(): + if not _session_config: synthetic_session: Final = provider_config.transform_session_created_event( model=model, logging_session_id=logging_obj.litellm_trace_id, diff --git a/litellm/llms/soniox/common_utils.py b/litellm/llms/soniox/common_utils.py index 8d83b4c9218..796f6ca12e4 100644 --- a/litellm/llms/soniox/common_utils.py +++ b/litellm/llms/soniox/common_utils.py @@ -3,7 +3,7 @@ Shared utilities for the Soniox provider (https://soniox.com). """ from collections.abc import Mapping, Sequence -from typing import Final, TypeAlias +from typing import Final, Literal, TypeAlias from litellm.litellm_core_utils.audio_utils.subtitle_utils import ( SubtitleToken, @@ -17,6 +17,8 @@ SonioxToken: TypeAlias = Mapping[str, object] # Soniox API base URL. SONIOX_API_BASE: Final[str] = "https://api.soniox.com" +SONIOX_REALTIME_MODEL_PREFIX: Final = "stt-rt-" + # Default polling interval in seconds when waiting for an async transcription # to finish. Mirrors the Soniox SDK default. SONIOX_DEFAULT_POLL_INTERVAL: Final[float] = 1.0 @@ -54,6 +56,10 @@ class SonioxException(BaseLLMException): """Provider-specific exception class for Soniox.""" +def soniox_health_check_mode(model: str) -> Literal["realtime"] | None: + return "realtime" if model.startswith(SONIOX_REALTIME_MODEL_PREFIX) else None + + def get_soniox_api_key(api_key: str | None = None) -> str | None: """Resolve the Soniox API key from arg or env var.""" # Local import to avoid a circular import: litellm.secret_managers.main diff --git a/litellm/llms/soniox/realtime/__init__.py b/litellm/llms/soniox/realtime/__init__.py new file mode 100644 index 00000000000..e69de29bb2d diff --git a/litellm/llms/soniox/realtime/transformation.py b/litellm/llms/soniox/realtime/transformation.py new file mode 100644 index 00000000000..ed7590b6f39 --- /dev/null +++ b/litellm/llms/soniox/realtime/transformation.py @@ -0,0 +1,609 @@ +import asyncio +import time +from collections import deque +from collections.abc import Awaitable, Callable, Mapping +from itertools import chain, groupby +from types import MappingProxyType, TracebackType +from typing import Final +from urllib.parse import urlparse, urlunparse + +from pydantic import BaseModel, ConfigDict, JsonValue, ValidationError +from typing_extensions import NotRequired, ReadOnly, Required, Self, TypedDict + +from litellm import verbose_logger +from litellm._uuid import uuid +from litellm.constants import REALTIME_WEBSOCKET_MAX_MESSAGE_SIZE_BYTES +from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj +from litellm.llms.base_llm.realtime.transcription_protocol import ( + SESSION_UPDATE_EVENT_TYPES, + RealtimeTranscriptionProtocolError, + TranscriptionSessionUpdate, + completed_event, + decode_pcm16_append, + delta_event, + duration_usage, + error_event, + json_object, + parse_transcription_session_update, + speech_event, + transcription_session, + transcription_session_created_event, + uses_server_vad, +) +from litellm.llms.base_llm.realtime.transformation import BaseRealtimeConfig, RealtimeBackend +from litellm.llms.custom_httpx.http_handler import realtime_ssl_for_url +from litellm.llms.soniox.common_utils import SonioxException, get_soniox_api_key +from litellm.types.llms.openai import ( + OpenAIRealtimeErrorEvent, + OpenAIRealtimeEvents, + OpenAIRealtimeInputAudioTranscriptionCompleted, + OpenAIRealtimeInputAudioTranscriptionDelta, + OpenAIRealtimeTranscriptionSessionCreated, +) +from litellm.types.realtime import ( + RealtimeInputAudioTranscriptionUsage, + RealtimeResponseTransformInput, + RealtimeResponseTypedDict, +) + +DEFAULT_SONIOX_REALTIME_URL: Final = "wss://stt-rt.soniox.com/transcribe-websocket" +WEBSOCKET_SCHEMES: Final = MappingProxyType({"wss": "wss", "ws": "ws", "https": "wss", "http": "ws"}) +DEFAULT_SAMPLE_RATE: Final = 24_000 +KEEPALIVE_INTERVAL_SECONDS: Final = 5.0 +_MAX_PENDING_FRAMES: Final = 64 +_FINALIZE: Final = '{"type":"finalize"}' +_KEEPALIVE: Final = '{"type":"keepalive"}' +_END_STREAM: Final = "" +_HEALTH_CHECK_SILENCE: Final = bytes(DEFAULT_SAMPLE_RATE // 5) +_CONTROL_FRAMES: Final = frozenset((_FINALIZE, _KEEPALIVE, _END_STREAM)) +SESSION_STARTED_FRAME: Final = '{"type":"litellm.soniox.session_started"}' +_ENDPOINT_TOKEN: Final = "" +_FINALIZED_TOKEN: Final = "" + + +class SonioxProtocolError(RealtimeTranscriptionProtocolError): + pass + + +class SonioxRealtimeOptions(BaseModel): + model_config = ConfigDict(frozen=True, extra="ignore") + + language_hints: tuple[str, ...] | None = None + language_hints_strict: bool | None = None + enable_language_identification: bool | None = None + enable_speaker_diarization: bool | None = None + context: JsonValue | None = None + translation: JsonValue | None = None + client_reference_id: str | None = None + max_endpoint_delay_ms: int | None = None + endpoint_sensitivity: float | None = None + endpoint_latency_adjustment_level: int | None = None + + +class SonioxStartRequest(SonioxRealtimeOptions): + model: str + audio_format: str + sample_rate: int + num_channels: int + enable_endpoint_detection: bool + + +class SonioxSpeechLabels(TypedDict, total=False): + speaker: ReadOnly[str] + language: ReadOnly[str] + + +class SonioxTranscriptSegment(SonioxSpeechLabels): + text: Required[ReadOnly[str]] + + +class SonioxTranscriptionDelta(OpenAIRealtimeInputAudioTranscriptionDelta, SonioxSpeechLabels): + pass + + +class SonioxTranscriptionCompleted(OpenAIRealtimeInputAudioTranscriptionCompleted): + segments: NotRequired[ReadOnly[tuple[SonioxTranscriptSegment, ...]]] + + +class _SonioxToken(BaseModel): + model_config = ConfigDict(frozen=True) + + text: str + is_final: bool = False + translation_status: str | None = None + speaker: str | None = None + language: str | None = None + + @property + def labelled(self) -> bool: + return self.speaker is not None or self.language is not None + + +class _SonioxResponse(BaseModel): + model_config = ConfigDict(frozen=True) + + tokens: tuple[_SonioxToken, ...] = () + total_audio_proc_ms: int | None = None + error_code: int | None = None + error_message: str | None = None + finished: bool = False + + +def _sample_rate(update: TranscriptionSessionUpdate | None) -> int: + audio_format: Final = None if update is None else update.audio_format + if audio_format is None: + return DEFAULT_SAMPLE_RATE + if not audio_format.is_pcm16: + raise SonioxProtocolError("Soniox realtime requires pcm16 input audio") + if audio_format.channels not in (None, 1): + raise SonioxProtocolError("Soniox realtime requires mono input audio") + if audio_format.rate is None: + return DEFAULT_SAMPLE_RATE + if audio_format.rate <= 0: + raise SonioxProtocolError("sample rate must be positive") + return audio_format.rate + + +def _language_hints( + language: str | None, configured: tuple[str, ...] | None, strict: bool | None +) -> tuple[str, ...] | None: + if not language or (strict and configured): + return configured + return (language, *(hint for hint in configured or () if hint != language)) + + +def _uses_endpoint_detection(update: TranscriptionSessionUpdate, drop_params: bool) -> bool: + return uses_server_vad(update, "Soniox realtime", drop_params, SonioxProtocolError) + + +def _ignored_settings(update: TranscriptionSessionUpdate) -> tuple[str, ...]: + vad_tuning: Final = (f"turn_detection.{key}" for key in sorted(update.turn_detection or {}) if key != "type") + return (*update.unsupported_transcription_keys, *vad_tuning) + + +def build_start_request( + model: str, update: TranscriptionSessionUpdate | None, options: SonioxRealtimeOptions, drop_params: bool = False +) -> SonioxStartRequest: + return SonioxStartRequest( + **options.model_dump(exclude={"language_hints"}), + language_hints=_language_hints( + None if update is None else update.language, options.language_hints, options.language_hints_strict + ), + model=model, + audio_format="pcm_s16le", + sample_rate=_sample_rate(update), + num_channels=1, + enable_endpoint_detection=update is None or _uses_endpoint_detection(update, drop_params), + ) + + +def _session_created(session_id: str, start_request: SonioxStartRequest) -> OpenAIRealtimeTranscriptionSessionCreated: + session: Final = transcription_session( + session_id=session_id, + model=start_request.model, + sample_rate=start_request.sample_rate, + language=start_request.language_hints[0] if start_request.language_hints else None, + server_vad=start_request.enable_endpoint_detection, + ) + return transcription_session_created_event(session) + + +def _changes_started_rate(update: TranscriptionSessionUpdate, started: SonioxStartRequest) -> bool: + audio_format: Final = update.audio_format + if audio_format is None: + return False + requested_rate: Final = _sample_rate(update) + return audio_format.rate is not None and requested_rate != started.sample_rate + + +def _changes_started_stream(update: TranscriptionSessionUpdate, started: SonioxStartRequest, drop_params: bool) -> bool: + turn_detection_set: Final = update.turn_detection is not None or update.turn_detection_disabled + return ( + _changes_started_rate(update, started) + or ( + update.language is not None + and _language_hints(update.language, started.language_hints, started.language_hints_strict) + != started.language_hints + ) + or (turn_detection_set and _uses_endpoint_detection(update, drop_params) != started.enable_endpoint_detection) + ) + + +def build_soniox_realtime_url(api_base: str | None) -> str: + if api_base is None: + return DEFAULT_SONIOX_REALTIME_URL + parsed: Final = urlparse(api_base) + scheme: Final = WEBSOCKET_SCHEMES.get(parsed.scheme) + if scheme is None: + raise ValueError("Soniox realtime api_base must be a wss://, ws://, https://, or http:// URL") + return urlunparse(parsed._replace(scheme=scheme)) + + +def _labels(token: _SonioxToken) -> SonioxSpeechLabels: + speaker: Final[SonioxSpeechLabels] = {} if token.speaker is None else {"speaker": token.speaker} + language: Final[SonioxSpeechLabels] = {} if token.language is None else {"language": token.language} + return {**speaker, **language} + + +def _delta(item_id: str, token: _SonioxToken) -> SonioxTranscriptionDelta: + return {**delta_event(item_id, token.text), **_labels(token)} + + +def _segment(run: tuple[_SonioxToken, ...]) -> SonioxTranscriptSegment: + return {**_labels(run[0]), "text": "".join(token.text for token in run).strip()} + + +def _segments(tokens: tuple[_SonioxToken, ...]) -> tuple[SonioxTranscriptSegment, ...]: + return tuple(_segment(tuple(run)) for _, run in groupby(tokens, key=lambda token: (token.speaker, token.language))) + + +def _new_item_id() -> str: + return f"item_{uuid.uuid4().hex}" + + +def _completed( + item_id: str, tokens: tuple[_SonioxToken, ...], usage: RealtimeInputAudioTranscriptionUsage | None +) -> OpenAIRealtimeInputAudioTranscriptionCompleted: + completed: Final = completed_event(item_id, "".join(token.text for token in tokens).strip(), usage) + if not any(token.labelled for token in tokens): + return completed + labelled: Final[SonioxTranscriptionCompleted] = {**completed, "segments": _segments(tokens)} + return labelled + + +def _soniox_error(response: _SonioxResponse) -> SonioxException | None: + if response.error_code is None and response.error_message is None: + return None + return SonioxException( + status_code=response.error_code or 500, + message=f"Soniox realtime error {response.error_code}: {response.error_message}", + ) + + +class SonioxEventTransformer: + def __init__(self, *, drops_originals: bool) -> None: + self._drops_originals: Final = drops_originals + self._item_id: str | None = None + self._final_tokens: Final[list[_SonioxToken]] = [] # mutable-ok: a turn's tokens append in O(1) each + self._finalizes: Final[deque[bool]] = deque() # mutable-ok: pending finalizes queue and pop in O(1) each + self._oldest_commit_answered: bool = False + self._processed_ms: int = 0 + self._billed_ms: int = 0 + self._refusal: SonioxException | None = None + + @property + def refusal(self) -> SonioxException | None: + return self._refusal + + def transform(self, payload: str) -> tuple[OpenAIRealtimeEvents, ...]: + try: + response: Final = _SonioxResponse.model_validate_json(payload) + except ValidationError: + raise SonioxProtocolError("Soniox returned an invalid realtime response") from None + self._bill(response.total_audio_proc_ms) + error: Final = _soniox_error(response) + if error is not None: + return (self._report(error),) + return tuple(chain.from_iterable(self._token_events(token) for token in response.tokens)) + + def queue_finalize(self, *, discards_turn: bool) -> None: + if discards_turn and not self._finalizes: + self._item_id = None + self._final_tokens.clear() + self._finalizes.append(discards_turn) + + def take_unbilled_usage(self, stream_ms: int = 0) -> RealtimeInputAudioTranscriptionUsage | None: + billable_ms: Final = max(self._processed_ms, stream_ms) + unbilled_ms: Final = billable_ms - self._billed_ms + if unbilled_ms <= 0: + return None + self._billed_ms = billable_ms + return duration_usage(unbilled_ms / 1000) + + def _report(self, error: SonioxException) -> OpenAIRealtimeErrorEvent: + if self._processed_ms == 0: + self._refusal = error + return error_event(error.message) + + def _bill(self, total_audio_proc_ms: int | None) -> None: + if total_audio_proc_ms is not None and total_audio_proc_ms > self._processed_ms: + self._processed_ms = total_audio_proc_ms + + def _token_events(self, token: _SonioxToken) -> tuple[OpenAIRealtimeEvents, ...]: + if token.text == _FINALIZED_TOKEN: + return self._finalized() + if self._discarding: + return () + if token.text == _ENDPOINT_TOKEN: + return self._end_turn() + started: Final = self._start() + if not token.is_final or (self._drops_originals and token.translation_status == "original"): + return started + self._final_tokens.append(token) + return (*started, _delta(self._current_item_id(), token)) + + @property + def _discarding(self) -> bool: + return bool(self._finalizes) and self._finalizes[0] + + def _finalized(self) -> tuple[OpenAIRealtimeEvents, ...]: + discarded: Final = self._finalizes.popleft() if self._finalizes else None + answered: Final = self._oldest_commit_answered + self._oldest_commit_answered = False + if discarded: + return () + if self._item_id is not None: + return self._complete(speech_stopped=False) + if discarded is None or answered: + return () + return (_completed(_new_item_id(), (), self.take_unbilled_usage()),) + + def _end_turn(self) -> tuple[OpenAIRealtimeEvents, ...]: + events: Final = self._complete(speech_stopped=True) + if events and self._finalizes: + self._oldest_commit_answered = True + return events + + def _start(self) -> tuple[OpenAIRealtimeEvents, ...]: + if self._item_id is not None: + return () + item_id: Final = _new_item_id() + self._item_id = item_id + return (speech_event("input_audio_buffer.speech_started", item_id),) + + def _current_item_id(self) -> str: + assert self._item_id is not None + return self._item_id + + def _complete(self, *, speech_stopped: bool) -> tuple[OpenAIRealtimeEvents, ...]: + item_id: Final = self._item_id + if item_id is None: + return () + tokens: Final = tuple(self._final_tokens) + self._item_id = None + self._final_tokens.clear() + stopped: Final[tuple[OpenAIRealtimeEvents, ...]] = ( + (speech_event("input_audio_buffer.speech_stopped", item_id),) if speech_stopped else () + ) + return (*stopped, _completed(item_id, tokens, self.take_unbilled_usage())) + + +class SonioxRealtimeBackend: + def __init__( + self, + inner: RealtimeBackend, + *, + keepalive_interval: float = KEEPALIVE_INTERVAL_SECONDS, + clock: Callable[[], float] = time.monotonic, + sleep: Callable[[float], Awaitable[None]] = asyncio.sleep, + ) -> None: + self._inner: Final = inner + self._interval: Final = keepalive_interval + self._clock: Final = clock + self._sleep: Final = sleep + self._frames: Final[asyncio.Queue[str | bytes | Exception]] = asyncio.Queue(maxsize=_MAX_PENDING_FRAMES) + self._last_sent: float = clock() + self._started: bool = False + self._ended: bool = False + self._tasks: tuple[asyncio.Task[None], ...] = () + + async def __aenter__(self) -> Self: + await self._inner.__aenter__() + self._tasks = (asyncio.create_task(self._pump()), asyncio.create_task(self._keepalive())) + return self + + async def __aexit__( + self, + exc_type: type[BaseException] | None, + exc_value: BaseException | None, + traceback: TracebackType | None, + ) -> None: + self._stop() + await self._inner.__aexit__(exc_type, exc_value, traceback) + + async def send(self, message: str | bytes) -> None: + self._last_sent = self._clock() + self._ended = self._ended or message == _END_STREAM + await self._inner.send(message) + if not self._started and isinstance(message, str) and message not in _CONTROL_FRAMES: + self._started = True + await self._frames.put(SESSION_STARTED_FRAME) + + async def recv(self, decode: bool | None = None) -> str | bytes: + frame: Final = await self._frames.get() + if isinstance(frame, Exception): + raise frame + return frame + + async def close(self) -> None: + self._stop() + await self._inner.close() + + def _stop(self) -> None: + for task in self._tasks: + task.cancel() + + async def _pump(self) -> None: + while True: + try: + frame = await self._inner.recv() + except Exception as e: # noqa: BLE001 # recv() re-raises it, so an upstream failure never hangs the reader + await self._frames.put(e) + return + await self._frames.put(frame) + + async def _keepalive(self) -> None: + while not self._ended: + await self._sleep(self._interval) + if self._ended or self._clock() - self._last_sent < self._interval: + continue + try: + await self.send(_KEEPALIVE) + except Exception as e: # noqa: BLE001 # stops only keepalives; the pump reports the broken socket + verbose_logger.debug("Soniox realtime: stopping keepalive after send failed: %s", e) + return + + +class SonioxRealtimeConfig(BaseRealtimeConfig): + def __init__( + self, + *, + options: SonioxRealtimeOptions | None = None, + wrap: Callable[[RealtimeBackend], RealtimeBackend] = SonioxRealtimeBackend, + clock: Callable[[], float] = time.monotonic, + drop_params: bool = False, + ) -> None: + self._options: Final = options or SonioxRealtimeOptions() + self._drop_params: Final = drop_params + self._wrap: Final = wrap + self._clock: Final = clock + self._transformer: Final = SonioxEventTransformer(drops_originals=self._options.translation is not None) + self._start_request: SonioxStartRequest | None = None + self._session_id: str = f"sess_{uuid.uuid4().hex}" + self._started_at: float | None = None + + @classmethod + def from_litellm_params(cls, litellm_params: Mapping[str, object], *, drop_params: bool = False) -> Self: + return cls(options=SonioxRealtimeOptions.model_validate(dict(litellm_params)), drop_params=drop_params) + + def requires_session_configuration(self) -> bool: + return True + + def validate_environment( + self, + headers: dict[str, str], # mutable-ok: BaseRealtimeConfig contract + model: str, + api_key: str | None = None, + ) -> dict[str, str]: # mutable-ok: BaseRealtimeConfig contract + resolved_key: Final = get_soniox_api_key(api_key) + if not resolved_key: + raise ValueError("Missing Soniox API key. Set SONIOX_API_KEY or api_key in the deployment litellm_params") + return {**headers, "Authorization": f"Bearer {resolved_key}"} + + def get_complete_url(self, api_base: str | None, model: str, api_key: str | None = None) -> str: + return build_soniox_realtime_url(api_base) + + def wrap_backend(self, backend: RealtimeBackend) -> RealtimeBackend: + return self._wrap(backend) + + def transform_session_created_event( + self, + model: str, + logging_session_id: str, + session_configuration_request: str | None = None, + ) -> OpenAIRealtimeTranscriptionSessionCreated: + self._session_id = logging_session_id + return _session_created(logging_session_id, build_start_request(model, None, self._options)) + + def transform_realtime_request( + self, + message: str, + model: str, + session_configuration_request: str | None = None, + ) -> tuple[str | bytes, ...]: + request: Final = json_object(message, SonioxProtocolError) + event_type: Final = request.get("type") + if event_type is None: + raise SonioxProtocolError("Soniox realtime takes OpenAI realtime events, and this message has no type") + if event_type in SESSION_UPDATE_EVENT_TYPES: + update: Final = parse_transcription_session_update(message, SonioxProtocolError) + ignored: Final = _ignored_settings(update) + if ignored: + verbose_logger.warning("Soniox realtime: ignoring unsupported session settings %s", ignored) + return self._start(model, update) + if event_type == "input_audio_buffer.append": + audio: Final = decode_pcm16_append(request.get("audio"), error=SonioxProtocolError) + return (*self._start(model, None), audio) + if event_type == "input_audio_buffer.commit": + return self._finalize(model, discards_turn=False) + if event_type == "input_audio_buffer.clear": + return () if self._start_request is None else self._finalize(model, discards_turn=True) + if event_type == "input_audio_buffer.end": + return (*self._start(model, None), _END_STREAM) + verbose_logger.debug("Soniox realtime: dropping unsupported client event %s", event_type) + return () + + def _finalize(self, model: str, *, discards_turn: bool) -> tuple[str, ...]: + start: Final = self._start(model, None) + self._transformer.queue_finalize(discards_turn=discards_turn) + return (*start, _FINALIZE) + + def session_refusal(self) -> Exception | None: + return self._transformer.refusal + + def unbilled_usage_on_session_close(self, model: str) -> RealtimeInputAudioTranscriptionUsage | None: + started_at: Final = self._started_at + if self._transformer.refusal is not None: + return None + if started_at is None: + return self._transformer.take_unbilled_usage() + return self._transformer.take_unbilled_usage(int((self._clock() - started_at) * 1000)) + + def transform_realtime_response( + self, + message: str | bytes, + model: str, + logging_obj: LiteLLMLoggingObj, + realtime_response_transform_input: RealtimeResponseTransformInput, + ) -> RealtimeResponseTypedDict: + payload: Final = message.decode("utf-8") if isinstance(message, bytes) else message + result: Final[RealtimeResponseTypedDict] = { + "response": list(self._backend_events(payload)), + "current_output_item_id": realtime_response_transform_input.get("current_output_item_id"), + "current_response_id": realtime_response_transform_input.get("current_response_id"), + "current_delta_chunks": realtime_response_transform_input.get("current_delta_chunks"), + "current_conversation_id": realtime_response_transform_input.get("current_conversation_id"), + "current_item_chunks": realtime_response_transform_input.get("current_item_chunks"), + "current_delta_type": realtime_response_transform_input.get("current_delta_type"), + "session_configuration_request": realtime_response_transform_input.get("session_configuration_request"), + } + return result + + def _backend_events(self, payload: str) -> tuple[OpenAIRealtimeEvents, ...]: + if payload != SESSION_STARTED_FRAME: + return self._transformer.transform(payload) + assert self._start_request is not None + return (_session_created(self._session_id, self._start_request),) + + def _start(self, model: str, update: TranscriptionSessionUpdate | None) -> tuple[str, ...]: + started: Final = self._start_request + if started is not None: + if update is not None and _changes_started_stream(update, started, self._drop_params): + raise SonioxProtocolError("Soniox realtime can't change the session once the stream has started") + return () + start_request: Final = build_start_request(model, update, self._options, self._drop_params) + self._start_request = start_request + self._started_at = self._clock() + return (start_request.model_dump_json(exclude_none=True),) + + +def _finishes_stream(frame: str | bytes) -> bool: + response: Final = _SonioxResponse.model_validate_json(frame) + error: Final = _soniox_error(response) + if error is not None: + raise error + return response.finished + + +async def soniox_realtime_health_check( + model: str, api_base: str | None, api_key: str | None, options: SonioxRealtimeOptions +) -> bool: + import websockets + + config: Final = SonioxRealtimeConfig() + url: Final = config.get_complete_url(api_base, model) + headers: Final = config.validate_environment({}, model, api_key) + start_request: Final = build_start_request(model, None, options) + async with websockets.connect( + url, + additional_headers=headers, + max_size=REALTIME_WEBSOCKET_MAX_MESSAGE_SIZE_BYTES, + ssl=realtime_ssl_for_url(url), + ) as backend: + await backend.send(start_request.model_dump_json(exclude_none=True)) + await backend.send(_HEALTH_CHECK_SILENCE) + await backend.send(_END_STREAM) + async for frame in backend: + if _finishes_stream(frame): + return True + raise SonioxException(status_code=502, message="Soniox closed the realtime stream without finishing it") diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 4fad4857c60..1256fef9181 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -61177,6 +61177,126 @@ ], "supports_audio_input": true }, + "soniox/stt-rt-v5": { + "input_cost_per_second": 3.333e-05, + "litellm_provider": "soniox", + "metadata": { + "calculation": "$0.12/3600 seconds = $0.00003333 per second", + "original_pricing_per_hour": 0.12 + }, + "mode": "audio_transcription", + "source": "https://soniox.com/pricing", + "supported_endpoints": [ + "/v1/realtime" + ], + "supported_modalities": [ + "audio" + ], + "supported_output_modalities": [ + "text" + ], + "supports_audio_input": true + }, + "soniox/stt-rt-preview": { + "input_cost_per_second": 3.333e-05, + "litellm_provider": "soniox", + "metadata": { + "calculation": "$0.12/3600 seconds = $0.00003333 per second", + "original_pricing_per_hour": 0.12 + }, + "mode": "audio_transcription", + "source": "https://soniox.com/pricing", + "supported_endpoints": [ + "/v1/realtime" + ], + "supported_modalities": [ + "audio" + ], + "supported_output_modalities": [ + "text" + ], + "supports_audio_input": true + }, + "soniox/stt-rt-preview-v2": { + "input_cost_per_second": 3.333e-05, + "litellm_provider": "soniox", + "metadata": { + "calculation": "$0.12/3600 seconds = $0.00003333 per second", + "original_pricing_per_hour": 0.12 + }, + "mode": "audio_transcription", + "source": "https://soniox.com/pricing", + "supported_endpoints": [ + "/v1/realtime" + ], + "supported_modalities": [ + "audio" + ], + "supported_output_modalities": [ + "text" + ], + "supports_audio_input": true + }, + "soniox/stt-rt-v3": { + "input_cost_per_second": 3.333e-05, + "litellm_provider": "soniox", + "metadata": { + "calculation": "$0.12/3600 seconds = $0.00003333 per second", + "original_pricing_per_hour": 0.12 + }, + "mode": "audio_transcription", + "source": "https://soniox.com/pricing", + "supported_endpoints": [ + "/v1/realtime" + ], + "supported_modalities": [ + "audio" + ], + "supported_output_modalities": [ + "text" + ], + "supports_audio_input": true + }, + "soniox/stt-rt-v3-preview": { + "input_cost_per_second": 3.333e-05, + "litellm_provider": "soniox", + "metadata": { + "calculation": "$0.12/3600 seconds = $0.00003333 per second", + "original_pricing_per_hour": 0.12 + }, + "mode": "audio_transcription", + "source": "https://soniox.com/pricing", + "supported_endpoints": [ + "/v1/realtime" + ], + "supported_modalities": [ + "audio" + ], + "supported_output_modalities": [ + "text" + ], + "supports_audio_input": true + }, + "soniox/stt-rt-v4": { + "input_cost_per_second": 3.333e-05, + "litellm_provider": "soniox", + "metadata": { + "calculation": "$0.12/3600 seconds = $0.00003333 per second", + "original_pricing_per_hour": 0.12 + }, + "mode": "audio_transcription", + "source": "https://soniox.com/pricing", + "supported_endpoints": [ + "/v1/realtime" + ], + "supported_modalities": [ + "audio" + ], + "supported_output_modalities": [ + "text" + ], + "supports_audio_input": true + }, "tensormesh/Qwen/Qwen3.5-397B-A17B-FP8": { "litellm_provider": "tensormesh", "mode": "chat", diff --git a/litellm/provider_endpoints_support_backup.json b/litellm/provider_endpoints_support_backup.json index 4188866de08..9f5989e765c 100644 --- a/litellm/provider_endpoints_support_backup.json +++ b/litellm/provider_endpoints_support_backup.json @@ -3724,7 +3724,7 @@ "ocr": false, "search": false, "moderations": false, - "realtime": false, + "realtime": true, "video_generations": false, "batches": false, "messages": false, diff --git a/litellm/proxy/public_endpoints/provider_create_fields.json b/litellm/proxy/public_endpoints/provider_create_fields.json index 4237517508b..04e6c5d8ba9 100644 --- a/litellm/proxy/public_endpoints/provider_create_fields.json +++ b/litellm/proxy/public_endpoints/provider_create_fields.json @@ -3612,7 +3612,7 @@ "key": "api_key", "label": "Soniox API Key", "placeholder": null, - "tooltip": "Currently only the async Speech-to-Text REST API (api.soniox.com) is supported. Realtime STT (stt-rt.soniox.com) and TTS (tts-rt.soniox.com) are not yet available.", + "tooltip": "Supports async Speech-to-Text (api.soniox.com) on /v1/audio/transcriptions and realtime Speech-to-Text (stt-rt.soniox.com) on /v1/realtime. TTS (tts-rt.soniox.com) is not yet available.", "required": true, "field_type": "password", "options": null, diff --git a/litellm/realtime_api/README.md b/litellm/realtime_api/README.md index d810de2f24f..987e8be1ed1 100644 --- a/litellm/realtime_api/README.md +++ b/litellm/realtime_api/README.md @@ -4,6 +4,6 @@ Supported endpoints: - WebSocket: `/v1/realtime` (with `intent=transcription` for transcription-only sessions) - HTTP: `/v1/realtime/client_secrets`, `/v1/realtime/transcription_sessions` -Supported providers: OpenAI, Azure OpenAI, Bedrock, Vertex AI, xAI. +Supported providers: OpenAI, Azure OpenAI, Bedrock, Vertex AI, xAI, Gemini, Meta, Soniox. For user-facing documentation and usage examples, see the litellm-docs repo. \ No newline at end of file diff --git a/litellm/realtime_api/main.py b/litellm/realtime_api/main.py index e9b534bdd89..919fd54e7d6 100644 --- a/litellm/realtime_api/main.py +++ b/litellm/realtime_api/main.py @@ -7,6 +7,7 @@ from dataclasses import dataclass from types import MappingProxyType from typing import TYPE_CHECKING, Any, Final, Literal, NoReturn, cast +from pydantic import TypeAdapter from typing_extensions import ReadOnly, TypedDict, assert_never import litellm @@ -55,6 +56,7 @@ from ..llms.openai.realtime.handler import ( confirm_session_started, ) from ..llms.openai.workload_identity import openai_static_api_key, resolve_openai_bearer_token_async +from ..llms.soniox.realtime.transformation import SonioxRealtimeOptions, soniox_realtime_health_check from ..llms.vertex_ai.audio_transcription.realtime_transformation import is_vertex_speech_to_text_model from ..llms.vertex_ai.realtime.transformation import VertexAIRealtimeConfig, vertex_realtime_config from ..llms.vertex_ai.vertex_llm_base import VertexBase @@ -75,6 +77,7 @@ vertex_llm_base: Final = VertexBase() base_llm_http_handler = BaseLLMHTTPHandler() _EMPTY_MODEL_PARAMS: Final[Mapping[str, object]] = MappingProxyType({}) _EMPTY_AUTH_HEADERS: Final[Mapping[str, str]] = MappingProxyType({}) +_DEPLOYMENT_PARAMS: Final = TypeAdapter(Mapping[str, object]) def _model_params_with_stored_credentials(model_params: Mapping[str, object]) -> Mapping[str, object]: @@ -527,6 +530,7 @@ async def _arealtime( custom_llm_provider=_custom_llm_provider, ) + deployment_params: Final = _DEPLOYMENT_PARAMS.validate_python(kwargs) if live_session_start is not None: await _alive_session( model=model, @@ -537,7 +541,7 @@ async def _arealtime( api_base=dynamic_api_base or litellm_params.api_base, api_key=dynamic_api_key, litellm_metadata=_build_litellm_metadata(kwargs), - litellm_params=kwargs, + litellm_params=deployment_params, ) return @@ -548,6 +552,7 @@ async def _arealtime( model=model, provider=LlmProviders(_custom_llm_provider), drop_params=drop_params, + litellm_params=deployment_params, ) if provider_config is not None: await base_llm_http_handler.async_realtime( @@ -882,6 +887,13 @@ async def realtime_health_check( ) elif custom_llm_provider == "openai": url = openai_realtime.construct_url(api_base=openai_api_base, query_params={"model": model}) + elif custom_llm_provider == "soniox": + return await soniox_realtime_health_check( + model=model, + api_base=resolved_api_base, + api_key=health_check_api_key, + options=SonioxRealtimeOptions.model_validate(dict(resolved_params)), + ) elif custom_llm_provider == "xai": url = xai_realtime.construct_url( api_base=resolved_api_base or "https://api.x.ai/v1", query_params={"model": model} diff --git a/litellm/utils.py b/litellm/utils.py index a656d8ebca3..342a2f12310 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -10759,6 +10759,7 @@ class ProviderConfigManager: model: str, provider: LlmProviders, drop_params: bool = False, + litellm_params: Mapping[str, object] = MappingProxyType({}), ) -> BaseRealtimeConfig | None: if LlmProviders.GEMINI == provider: from litellm.llms.gemini.realtime.transformation import GeminiRealtimeConfig @@ -10768,6 +10769,10 @@ class ProviderConfigManager: from litellm.llms.meta.realtime.transformation import MetaRealtimeConfig return MetaRealtimeConfig(drop_params=drop_params) + if LlmProviders.SONIOX == provider: + from litellm.llms.soniox.realtime.transformation import SonioxRealtimeConfig + + return SonioxRealtimeConfig.from_litellm_params(litellm_params, drop_params=drop_params) return None @staticmethod diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 4fad4857c60..1256fef9181 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -61177,6 +61177,126 @@ ], "supports_audio_input": true }, + "soniox/stt-rt-v5": { + "input_cost_per_second": 3.333e-05, + "litellm_provider": "soniox", + "metadata": { + "calculation": "$0.12/3600 seconds = $0.00003333 per second", + "original_pricing_per_hour": 0.12 + }, + "mode": "audio_transcription", + "source": "https://soniox.com/pricing", + "supported_endpoints": [ + "/v1/realtime" + ], + "supported_modalities": [ + "audio" + ], + "supported_output_modalities": [ + "text" + ], + "supports_audio_input": true + }, + "soniox/stt-rt-preview": { + "input_cost_per_second": 3.333e-05, + "litellm_provider": "soniox", + "metadata": { + "calculation": "$0.12/3600 seconds = $0.00003333 per second", + "original_pricing_per_hour": 0.12 + }, + "mode": "audio_transcription", + "source": "https://soniox.com/pricing", + "supported_endpoints": [ + "/v1/realtime" + ], + "supported_modalities": [ + "audio" + ], + "supported_output_modalities": [ + "text" + ], + "supports_audio_input": true + }, + "soniox/stt-rt-preview-v2": { + "input_cost_per_second": 3.333e-05, + "litellm_provider": "soniox", + "metadata": { + "calculation": "$0.12/3600 seconds = $0.00003333 per second", + "original_pricing_per_hour": 0.12 + }, + "mode": "audio_transcription", + "source": "https://soniox.com/pricing", + "supported_endpoints": [ + "/v1/realtime" + ], + "supported_modalities": [ + "audio" + ], + "supported_output_modalities": [ + "text" + ], + "supports_audio_input": true + }, + "soniox/stt-rt-v3": { + "input_cost_per_second": 3.333e-05, + "litellm_provider": "soniox", + "metadata": { + "calculation": "$0.12/3600 seconds = $0.00003333 per second", + "original_pricing_per_hour": 0.12 + }, + "mode": "audio_transcription", + "source": "https://soniox.com/pricing", + "supported_endpoints": [ + "/v1/realtime" + ], + "supported_modalities": [ + "audio" + ], + "supported_output_modalities": [ + "text" + ], + "supports_audio_input": true + }, + "soniox/stt-rt-v3-preview": { + "input_cost_per_second": 3.333e-05, + "litellm_provider": "soniox", + "metadata": { + "calculation": "$0.12/3600 seconds = $0.00003333 per second", + "original_pricing_per_hour": 0.12 + }, + "mode": "audio_transcription", + "source": "https://soniox.com/pricing", + "supported_endpoints": [ + "/v1/realtime" + ], + "supported_modalities": [ + "audio" + ], + "supported_output_modalities": [ + "text" + ], + "supports_audio_input": true + }, + "soniox/stt-rt-v4": { + "input_cost_per_second": 3.333e-05, + "litellm_provider": "soniox", + "metadata": { + "calculation": "$0.12/3600 seconds = $0.00003333 per second", + "original_pricing_per_hour": 0.12 + }, + "mode": "audio_transcription", + "source": "https://soniox.com/pricing", + "supported_endpoints": [ + "/v1/realtime" + ], + "supported_modalities": [ + "audio" + ], + "supported_output_modalities": [ + "text" + ], + "supports_audio_input": true + }, "tensormesh/Qwen/Qwen3.5-397B-A17B-FP8": { "litellm_provider": "tensormesh", "mode": "chat", diff --git a/provider_endpoints_support.json b/provider_endpoints_support.json index 4188866de08..9f5989e765c 100644 --- a/provider_endpoints_support.json +++ b/provider_endpoints_support.json @@ -3724,7 +3724,7 @@ "ocr": false, "search": false, "moderations": false, - "realtime": false, + "realtime": true, "video_generations": false, "batches": false, "messages": false, diff --git a/tests/unit/litellm_core_utils/test_health_check_helpers.py b/tests/unit/litellm_core_utils/test_health_check_helpers.py index 44b36a355c6..6979d72ff14 100644 --- a/tests/unit/litellm_core_utils/test_health_check_helpers.py +++ b/tests/unit/litellm_core_utils/test_health_check_helpers.py @@ -795,9 +795,12 @@ async def test_ahealth_check_keeps_the_noul_probe_for_other_decisions_providers( ("us.anthropic.claude-haiku-4-5-20251001-v1:0", "bedrock", None), ("claude-haiku-4-5", "anthropic", None), ("anthropic.claude-haiku-4-5", None, None), + ("stt-rt-v5", "soniox", "realtime"), + ("stt-rt-preview", "soniox", "realtime"), + ("stt-async-v5", "soniox", None), ), ) -def test_native_health_check_mode_is_messages_only_for_mantle_claude( +def test_native_health_check_mode_routes_mantle_claude_to_messages_and_soniox_realtime_models_to_realtime( model: str, custom_llm_provider: str | None, expected: str | None ) -> None: assert native_health_check_mode(model=model, custom_llm_provider=custom_llm_provider) == expected diff --git a/tests/unit/litellm_core_utils/test_realtime_streaming.py b/tests/unit/litellm_core_utils/test_realtime_streaming.py index 7d85da3f756..a8543ebcfc2 100644 --- a/tests/unit/litellm_core_utils/test_realtime_streaming.py +++ b/tests/unit/litellm_core_utils/test_realtime_streaming.py @@ -24,6 +24,7 @@ from litellm.litellm_core_utils.realtime_streaming import ( client_sent_openai_beta_realtime_header, ) from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj +from litellm.llms.base_llm.realtime.transcription_protocol import RealtimeTranscriptionProtocolError from litellm.llms.base_llm.realtime.transformation import BaseRealtimeConfig from litellm.llms.xai.realtime.transformation import XAIRealtimeNormalizer from litellm.llms.xai.realtime.usage_meter import XAIRealtimeUsageMeter @@ -827,9 +828,7 @@ async def test_usage_meter_snapshot_stored_in_model_call_details(): backend_ws.send = AsyncMock() backend_ws.recv = AsyncMock( side_effect=[ - json.dumps( - {"type": "session.created", "session": {"turn_detection": {"type": None}}} - ).encode(), + json.dumps({"type": "session.created", "session": {"turn_detection": {"type": None}}}).encode(), json.dumps( {"type": "response.output_audio.delta", "delta": base64.b64encode(b"x" * 24000).decode()} ).encode(), @@ -3187,6 +3186,7 @@ async def test_session_close_flushes_unbilled_transcription_usage(): provider_config: Final = MagicMock() provider_config.transform_realtime_response = MagicMock(return_value=transcript_frame) provider_config.unbilled_usage_on_session_close = MagicMock(return_value=usage) + provider_config.session_refusal = MagicMock(return_value=None) streaming: Final = RealTimeStreaming( client_ws, @@ -3320,10 +3320,14 @@ def _backend_ws_closing_with(*frames: bytes | Exception) -> MagicMock: return backend_ws -def _relay_session(client_ws: MagicMock, backend_ws: MagicMock) -> _RelaySession: +def _relay_session( + client_ws: MagicMock, backend_ws: MagicMock, provider_config: BaseRealtimeConfig | None = None +) -> _RelaySession: logging: Final = _RecordingLogging() worker: Final = _InlineLoggingWorker() - streaming: Final = RealTimeStreaming(client_ws, backend_ws, logging, model="gpt-realtime", logging_worker=worker) + streaming: Final = RealTimeStreaming( + client_ws, backend_ws, logging, provider_config=provider_config, model="gpt-realtime", logging_worker=worker + ) return _RelaySession(streaming=streaming, logging=logging, worker=worker) @@ -3408,6 +3412,103 @@ async def test_upstream_refusal_before_any_frame_logs_a_failure_not_a_success(): assert session.logging.logged_sessions == () +class _RefusingRealtimeConfig(_PassthroughRealtimeConfig): + def __init__(self, refusal: Exception | None) -> None: + self._refusal: Final = refusal + + def requires_session_configuration(self) -> bool: + return True + + def session_refusal(self) -> Exception | None: + return self._refusal + + +@pytest.mark.asyncio +async def test_a_session_the_provider_refused_in_band_logs_a_failure_even_after_a_normal_close(): + refusal: Final = ValueError("Soniox realtime error 401: Invalid API key.") + session: Final = _relay_session( + _client_ws_that_never_sends(), + _backend_ws_closing_with( + json.dumps({"type": "error", "error": {"message": str(refusal)}}).encode(), + ConnectionClosed(Close(1000, ""), None), + ), + _RefusingRealtimeConfig(refusal), + ) + + await session.run() + + assert session.logging.logged_failures == (refusal,) + assert session.logging.logged_sessions == () + + +@pytest.mark.asyncio +async def test_a_provider_session_without_a_refusal_logs_a_success_after_a_normal_close(): + session: Final = _relay_session( + _client_ws_that_never_sends(), + _backend_ws_closing_with( + json.dumps({"type": "session.created", "session": {"id": "sess_1"}}).encode(), + ConnectionClosed(Close(1000, ""), None), + ), + _RefusingRealtimeConfig(None), + ) + + await session.run() + + assert session.logging.logged_failures == () + assert len(session.logging.logged_sessions) == 1 + + +class _TranscriptionConfigRejectingSessionUpdates(_PassthroughRealtimeConfig): + def __init__(self, *, defers_setup: bool) -> None: + self._defers_setup: Final = defers_setup + + def requires_session_configuration(self) -> bool: + return not self._defers_setup + + def transform_realtime_request( + self, message: str, model: str, session_configuration_request: str | None = None + ) -> Sequence[str | bytes]: + if json.loads(message).get("type") == "session.update": + raise RealtimeTranscriptionProtocolError("requires pcm16 input audio") + return (message,) + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "defers_setup,client_frames_read,audio_reaches_backend", + ((True, 1, False), (False, 3, True)), + ids=["deferred_setup_session_ends", "eager_setup_session_continues"], +) +async def test_a_rejected_transcription_session_update_sends_an_error_and_ends_only_a_deferred_setup_session( + defers_setup: bool, client_frames_read: int, audio_reaches_backend: bool +): + append: Final = json.dumps({"type": "input_audio_buffer.append", "audio": "AAE="}) + client_ws: Final = MagicMock() + client_ws.headers = {} + client_ws.receive_text = AsyncMock( + side_effect=[json.dumps({"type": "session.update", "session": {"type": "transcription"}}), append, RuntimeError] + ) + client_ws.send_text = AsyncMock() + backend_ws: Final = MagicMock() + backend_ws.send = AsyncMock() + streaming: Final = RealTimeStreaming( + client_ws, + backend_ws, + _RecordingLogging(), + provider_config=_TranscriptionConfigRejectingSessionUpdates(defers_setup=defers_setup), + model="chirp_3", + logging_worker=_InlineLoggingWorker(), + ) + + await asyncio.wait_for(streaming.client_ack_messages(), timeout=2) + + assert _error_events_sent_to(client_ws) == [ + {"type": "error", "error": {"type": "invalid_request_error", "message": "requires pcm16 input audio"}} + ] + assert client_ws.receive_text.await_count == client_frames_read + assert (append in [call.args[0] for call in backend_ws.send.await_args_list]) is audio_reaches_backend + + @pytest.mark.asyncio async def test_upstream_refusal_after_a_synthetic_session_created_still_logs_a_failure(): """LIT-6973: deferred Gemini Live setup stores a synthetic ``session.created`` before @@ -4056,9 +4157,7 @@ _XAI_TYPED_VAD_SESSION: Final = { pytest.param(({"type": "realtime", "instructions": "hi"},), id="no-turn-detection"), pytest.param(({"type": "realtime", "audio": {"input": {"turn_detection": None}}},), id="push-to-talk"), pytest.param((_XAI_TYPED_VAD_SESSION,), id="typed-server-vad"), - pytest.param( - ({"type": "realtime", "instructions": "hi"}, _XAI_TYPED_VAD_SESSION), id="later-typed-server-vad" - ), + pytest.param(({"type": "realtime", "instructions": "hi"}, _XAI_TYPED_VAD_SESSION), id="later-typed-server-vad"), ], ) async def test_xai_session_update_gate_leaves_out_the_vad_type( @@ -4070,9 +4169,9 @@ async def test_xai_session_update_gate_leaves_out_the_vad_type( monkeypatch, *client_sessions, event_normalizer=XAIRealtimeNormalizer() ) - assert [session["audio"]["input"]["turn_detection"] for session in forwarded] == [ - {"create_response": False} - ] * (len(client_sessions) + 1) + assert [session["audio"]["input"]["turn_detection"] for session in forwarded] == [{"create_response": False}] * ( + len(client_sessions) + 1 + ) @pytest.mark.asyncio @@ -4206,7 +4305,9 @@ async def test_client_never_sees_the_error_caused_by_the_guardrail_cancel(monkey await RealTimeStreaming(client_ws, backend_ws, logging_obj).backend_to_client_send_messages() client_errors: Final = [ - event["error"] for call in client_ws.send_text.await_args_list if (event := json.loads(call.args[0]))["type"] == "error" + event["error"] + for call in client_ws.send_text.await_args_list + if (event := json.loads(call.args[0]))["type"] == "error" ] assert [(error["code"], error.get("event_id")) for error in client_errors] == [ ("content_policy_violation", None), diff --git a/tests/unit/llms/custom_httpx/test_llm_http_handler.py b/tests/unit/llms/custom_httpx/test_llm_http_handler.py index d8326536d03..dca21db5b6f 100644 --- a/tests/unit/llms/custom_httpx/test_llm_http_handler.py +++ b/tests/unit/llms/custom_httpx/test_llm_http_handler.py @@ -4911,6 +4911,160 @@ async def test_async_realtime_bridges_a_transcription_session_through_the_provid assert [bytes(request.audio) for request in speech_client.requests[1:]] == [b"\x00\x01" * 800, b"\x00\x01" * 800] +class _FakeSonioxSocket: + def __init__(self) -> None: + self.sent: Final[list[str | bytes]] = [] + self._responses: Final[asyncio.Queue[str]] = asyncio.Queue() + + async def __aenter__(self) -> "_FakeSonioxSocket": + return self + + async def __aexit__(self, exc_type, exc_value, traceback) -> None: + return None + + async def send(self, message: str | bytes) -> None: + self.sent.append(message) + if message == '{"type":"finalize"}': + tokens = [{"text": "Hello", "is_final": True}, {"text": "", "is_final": True}] + await self._responses.put(json.dumps({"tokens": tokens, "total_audio_proc_ms": 1200})) + + async def recv(self, decode: bool | None = None) -> str: + return await self._responses.get() + + async def close(self) -> None: + return None + + +@pytest.mark.asyncio +async def test_async_realtime_bridges_a_soniox_transcription_through_the_wrapped_websocket(): + import websockets.exceptions # noqa: F401 # binds the submodule so async_realtime's except clause resolves, as in the proxy process + + from litellm.llms.soniox.realtime.transformation import SonioxRealtimeBackend, SonioxRealtimeConfig + + upstream = _FakeSonioxSocket() + wrapped: list[object] = [] + provider_config = SonioxRealtimeConfig( + wrap=lambda backend: wrapped.append(backend) or SonioxRealtimeBackend(backend) + ) + audio = b"\x00\x01" * 800 + client_ws = _ScriptedClientWebSocket( + [ + json.dumps( + { + "type": "session.update", + "session": { + "type": "transcription", + "audio": { + "input": { + "format": {"type": "audio/pcm", "rate": 16000}, + "transcription": {"model": "stt-rt-v5", "language": "en"}, + "turn_detection": None, + } + }, + }, + } + ), + json.dumps({"type": "input_audio_buffer.append", "audio": base64.b64encode(audio).decode()}), + json.dumps({"type": "input_audio_buffer.commit"}), + ], + last_event_type="conversation.item.input_audio_transcription.completed", + ) + logging_obj = Mock() + logging_obj.litellm_trace_id = "trace_1" + logging_obj.model_call_details = {} + logging_obj.dispatch_success_handlers = AsyncMock() + logging_obj.dispatch_failure_handlers = AsyncMock() + handler = BaseLLMHTTPHandler() + + with patch.object(handler, "_open_realtime_backend_ws", AsyncMock(return_value=upstream)) as dial: + await handler.async_realtime( + model="stt-rt-v5", + websocket=client_ws, + logging_obj=logging_obj, + provider_config=provider_config, + headers={}, + api_key="soniox-key", + query_params={"model": "stt-rt-v5", "intent": "transcription"}, + ) + + assert dial.await_args.args[1:3] == ( + "wss://stt-rt.soniox.com/transcribe-websocket", + {"Authorization": "Bearer soniox-key"}, + ) + assert wrapped == [upstream] + assert json.loads(upstream.sent[0]) == { + "model": "stt-rt-v5", + "audio_format": "pcm_s16le", + "sample_rate": 16000, + "num_channels": 1, + "enable_endpoint_detection": False, + "language_hints": ["en"], + } + assert upstream.sent[1:] == [audio, '{"type":"finalize"}'] + events = client_ws.sent_events() + assert [event["type"] for event in events] == [ + "session.created", + "session.updated", + "input_audio_buffer.speech_started", + "conversation.item.input_audio_transcription.delta", + "conversation.item.input_audio_transcription.completed", + ] + assert events[-1]["transcript"] == "Hello" + assert events[-1]["usage"] == {"type": "duration", "seconds": 1.2} + + +@pytest.mark.asyncio +async def test_async_realtime_soniox_rejects_a_bad_session_update_and_still_transcribes_audio_sent_without_one(): + import websockets.exceptions # noqa: F401 # binds the submodule so async_realtime's except clause resolves, as in the proxy process + + from litellm.llms.soniox.realtime.transformation import SonioxRealtimeConfig + + upstream = _FakeSonioxSocket() + audio = b"\x00\x01" * 800 + client_ws = _ScriptedClientWebSocket( + [ + json.dumps( + { + "type": "session.update", + "session": {"type": "transcription", "audio": {"input": {"format": {"type": "audio/pcmu"}}}}, + } + ), + json.dumps({"type": "input_audio_buffer.append", "audio": base64.b64encode(audio).decode()}), + json.dumps({"type": "input_audio_buffer.commit"}), + ], + last_event_type="conversation.item.input_audio_transcription.completed", + ) + logging_obj = Mock() + logging_obj.litellm_trace_id = "trace_1" + logging_obj.model_call_details = {} + logging_obj.dispatch_success_handlers = AsyncMock() + logging_obj.dispatch_failure_handlers = AsyncMock() + handler = BaseLLMHTTPHandler() + + with patch.object(handler, "_open_realtime_backend_ws", AsyncMock(return_value=upstream)): + await handler.async_realtime( + model="stt-rt-v5", + websocket=client_ws, + logging_obj=logging_obj, + provider_config=SonioxRealtimeConfig(), + headers={}, + api_key="soniox-key", + query_params={"model": "stt-rt-v5", "intent": "transcription"}, + ) + + assert json.loads(upstream.sent[0])["sample_rate"] == 24_000 + assert upstream.sent[1:] == [audio, '{"type":"finalize"}'] + events = client_ws.sent_events() + errors = [event for event in events if event["type"] == "error"] + assert errors == [ + { + "type": "error", + "error": {"type": "invalid_request_error", "message": "Soniox realtime requires pcm16 input audio"}, + } + ] + assert events[-1]["transcript"] == "Hello" + + @pytest.mark.asyncio async def test_responses_agentic_followup_does_not_repeat_request_params_from_plan_kwargs(monkeypatch): """A plan whose kwargs repeat a request param must not crash the Responses follow-up with a duplicate keyword""" diff --git a/tests/unit/llms/soniox/realtime/__init__.py b/tests/unit/llms/soniox/realtime/__init__.py new file mode 100644 index 00000000000..b503dff241f --- /dev/null +++ b/tests/unit/llms/soniox/realtime/__init__.py @@ -0,0 +1 @@ +"""Soniox realtime tests.""" diff --git a/tests/unit/llms/soniox/realtime/test_soniox_realtime_transformation.py b/tests/unit/llms/soniox/realtime/test_soniox_realtime_transformation.py new file mode 100644 index 00000000000..653f73ce22c --- /dev/null +++ b/tests/unit/llms/soniox/realtime/test_soniox_realtime_transformation.py @@ -0,0 +1,807 @@ +import asyncio +import base64 +import json +from typing import Final +from unittest.mock import MagicMock + +import pytest + +from litellm.llms.soniox.common_utils import SonioxException +from litellm.llms.soniox.realtime.transformation import ( + DEFAULT_SONIOX_REALTIME_URL, + SESSION_STARTED_FRAME, + SonioxEventTransformer, + SonioxRealtimeBackend, + SonioxProtocolError, + SonioxRealtimeConfig, + SonioxRealtimeOptions, +) +from litellm.types.llms.openai import OpenAIRealtimeEvents +from litellm.types.realtime import RealtimeResponseTransformInput +from litellm.types.utils import LlmProviders +from litellm.utils import ProviderConfigManager + +MODEL: Final = "stt-rt-v5" +EMPTY_TRANSFORM_INPUT: Final[RealtimeResponseTransformInput] = { + "session_configuration_request": None, + "current_output_item_id": None, + "current_response_id": None, + "current_delta_chunks": None, + "current_item_chunks": None, + "current_conversation_id": None, + "current_delta_type": None, +} +PCM: Final = b"\x01\x02" * 160 + + +def _event(event_type: str, **fields: object) -> str: + return json.dumps({"type": event_type, **fields}) + + +def _session_update(audio_input: dict[str, object]) -> str: + return _event("session.update", session={"type": "transcription", "audio": {"input": audio_input}}) + + +def _append() -> str: + return _event("input_audio_buffer.append", audio=base64.b64encode(PCM).decode()) + + +def _start_request(frames: tuple[str | bytes, ...]) -> dict[str, object]: + assert isinstance(frames[0], str) + return json.loads(frames[0]) + + +def _response(tokens: list[dict[str, object]], total_audio_proc_ms: int | None = None) -> str: + return json.dumps({"tokens": tokens, "total_audio_proc_ms": total_audio_proc_ms}) + + +def _token(text: str, is_final: bool = True, **fields: object) -> dict[str, object]: + return {"text": text, "is_final": is_final, **fields} + + +def _types(events: tuple[OpenAIRealtimeEvents, ...]) -> list[object]: + return [event["type"] for event in events] + + +def test_session_update_starts_the_stream_with_client_audio_and_deployment_options(): + config = SonioxRealtimeConfig( + options=SonioxRealtimeOptions(language_hints=("en", "fr"), enable_speaker_diarization=True) + ) + + frames = config.transform_realtime_request( + _session_update( + { + "format": {"type": "audio/pcm", "rate": 16_000}, + "transcription": {"model": "client-chosen-model", "language": "fr"}, + "turn_detection": None, + } + ), + MODEL, + ) + + assert _start_request(frames) == { + "model": MODEL, + "audio_format": "pcm_s16le", + "sample_rate": 16_000, + "num_channels": 1, + "enable_endpoint_detection": False, + "language_hints": ["fr", "en"], + "enable_speaker_diarization": True, + } + + +def test_an_empty_transcription_language_keeps_the_deployment_language_hints(): + config = SonioxRealtimeConfig(options=SonioxRealtimeOptions(language_hints=("fa",))) + + frames = config.transform_realtime_request(_session_update({"transcription": {"language": ""}}), MODEL) + + assert _start_request(frames)["language_hints"] == ["fa"] + + +def test_a_client_language_cannot_widen_strict_deployment_language_hints(): + config = SonioxRealtimeConfig(options=SonioxRealtimeOptions(language_hints=("fa",), language_hints_strict=True)) + + frames = config.transform_realtime_request(_session_update({"transcription": {"language": "en"}}), MODEL) + + assert _start_request(frames)["language_hints"] == ["fa"] + assert config.transform_realtime_request(_session_update({"transcription": {"language": "en"}}), MODEL) == () + + +def test_a_soniox_native_config_frame_is_rejected_without_starting_the_stream(): + config = SonioxRealtimeConfig() + + with pytest.raises(SonioxProtocolError): + config.transform_realtime_request(json.dumps({"model": "stt-rt-v5", "audio_format": "auto"}), MODEL) + assert _start_request(config.transform_realtime_request(_append(), MODEL))["sample_rate"] == 24_000 + + +def test_audio_before_session_update_starts_a_default_stream_once_and_rejects_a_later_change(): + config = SonioxRealtimeConfig() + + first = config.transform_realtime_request(_append(), MODEL) + second = config.transform_realtime_request(_append(), MODEL) + with pytest.raises(SonioxProtocolError, match="once the stream has started"): + config.transform_realtime_request(_session_update({"turn_detection": None}), MODEL) + + assert _start_request(first) == { + "model": MODEL, + "audio_format": "pcm_s16le", + "sample_rate": 24_000, + "num_channels": 1, + "enable_endpoint_detection": True, + } + assert first[1:] == (PCM,) + assert second == (PCM,) + assert config.transform_realtime_request(_append(), MODEL) == (PCM,) + + +@pytest.mark.parametrize( + ("event_type", "control_frame"), + [("input_audio_buffer.commit", '{"type":"finalize"}'), ("input_audio_buffer.end", "")], +) +def test_buffer_control_events_map_to_soniox_control_frames(event_type: str, control_frame: str): + config = SonioxRealtimeConfig() + config.transform_realtime_request(_append(), MODEL) + + assert config.transform_realtime_request(_event(event_type), MODEL) == (control_frame,) + + +@pytest.mark.parametrize("event_type", ["response.create", "input_audio_buffer.clear"]) +def test_events_with_nothing_to_send_upstream_are_dropped_without_starting_the_stream(event_type: str): + config = SonioxRealtimeConfig() + + assert config.transform_realtime_request(_event(event_type), MODEL) == () + assert _start_request(config.transform_realtime_request(_append(), MODEL))["model"] == MODEL + + +def _events(config: SonioxRealtimeConfig, tokens: list[dict[str, object]]) -> tuple[OpenAIRealtimeEvents, ...]: + return tuple( + config.transform_realtime_response(_response(tokens), MODEL, MagicMock(), EMPTY_TRANSFORM_INPUT)["response"] + ) + + +def test_clear_discards_the_open_turn_and_everything_soniox_finalizes_for_it(): + config = SonioxRealtimeConfig() + config.transform_realtime_request(_append(), MODEL) + _events(config, [_token("Hel")]) + + assert config.transform_realtime_request(_event("input_audio_buffer.clear"), MODEL) == ('{"type":"finalize"}',) + cleared = _events(config, [_token("lo"), _token(""), _token("")]) + config.transform_realtime_request(_event("input_audio_buffer.commit"), MODEL) + kept = _events(config, [_token("Bye"), _token("")]) + + assert cleared == () + assert kept[-1]["transcript"] == "Bye" + + +def test_clear_after_a_commit_keeps_the_committed_turn_and_drops_only_audio_sent_after_it(): + config = SonioxRealtimeConfig() + config.transform_realtime_request(_append(), MODEL) + _events(config, [_token("One")]) + config.transform_realtime_request(_event("input_audio_buffer.commit"), MODEL) + config.transform_realtime_request(_event("input_audio_buffer.clear"), MODEL) + + committed = _events(config, [_token("")]) + cleared = _events(config, [_token(" two"), _token("")]) + later = _events(config, [_token("Three"), _token("")]) + + assert committed[-1]["transcript"] == "One" + assert cleared == () + assert later[-1]["transcript"] == "Three" + + +@pytest.mark.parametrize( + "audio_input", + [ + {"format": {"type": "audio/pcmu"}}, + {"format": {"type": "audio/pcm", "rate": 24_000, "channels": 2}}, + {"format": {"type": "audio/pcm", "rate": 0}}, + ], +) +def test_unsupported_audio_formats_are_rejected_without_starting_the_stream(audio_input: dict[str, object]): + config = SonioxRealtimeConfig() + + with pytest.raises(SonioxProtocolError): + config.transform_realtime_request(_session_update(audio_input), MODEL) + assert _start_request(config.transform_realtime_request(_append(), MODEL))["sample_rate"] == 24_000 + + +@pytest.mark.parametrize( + ("audio_input", "sample_rate"), + [({"format": {"type": "audio/pcm"}}, 24_000), ({"input_audio_format": "pcm16"}, 24_000)], +) +def test_pcm16_without_a_rate_streams_at_the_openai_default_rate(audio_input: dict[str, object], sample_rate: int): + config = SonioxRealtimeConfig() + + frames = config.transform_realtime_request(_session_update(audio_input), MODEL) + + assert _start_request(frames)["sample_rate"] == sample_rate + + +def test_validate_environment_sends_the_key_as_a_bearer_header(monkeypatch): + monkeypatch.setenv("SONIOX_API_KEY", "env-key") + config = SonioxRealtimeConfig() + + assert config.validate_environment({"x": "y"}, MODEL, "deployment-key") == { + "x": "y", + "Authorization": "Bearer deployment-key", + } + assert config.validate_environment({}, MODEL) == {"Authorization": "Bearer env-key"} + + +def test_validate_environment_rejects_a_missing_key(monkeypatch): + monkeypatch.delenv("SONIOX_API_KEY", raising=False) + + with pytest.raises(ValueError, match="SONIOX_API_KEY"): + SonioxRealtimeConfig().validate_environment({}, MODEL) + + +def test_get_complete_url_defaults_to_soniox_and_opens_a_websocket_for_http_bases(): + config = SonioxRealtimeConfig() + + assert config.get_complete_url(None, MODEL) == DEFAULT_SONIOX_REALTIME_URL + assert config.get_complete_url("wss://proxy.example/ws", MODEL) == "wss://proxy.example/ws" + assert config.get_complete_url("https://proxy.example/ws?region=eu", MODEL) == "wss://proxy.example/ws?region=eu" + assert config.get_complete_url("http://localhost:8080/ws", MODEL) == "ws://localhost:8080/ws" + with pytest.raises(ValueError, match="wss://"): + config.get_complete_url("ftp://proxy.example/ws", MODEL) + + +def test_endpoint_detection_emits_one_openai_turn_billed_by_processed_audio(): + transformer = SonioxEventTransformer(drops_originals=False) + + provisional = transformer.transform(_response([_token("Hel", is_final=False)], total_audio_proc_ms=500)) + finals = transformer.transform( + _response([_token("Hello"), _token(" world"), _token(" next", is_final=False)], total_audio_proc_ms=1500) + ) + endpoint = transformer.transform(_response([_token("")], total_audio_proc_ms=2000)) + + assert _types(provisional) == ["input_audio_buffer.speech_started"] + item_id = provisional[0]["item_id"] + assert [(event["type"], event.get("delta")) for event in finals] == [ + ("conversation.item.input_audio_transcription.delta", "Hello"), + ("conversation.item.input_audio_transcription.delta", " world"), + ] + assert _types(endpoint) == [ + "input_audio_buffer.speech_stopped", + "conversation.item.input_audio_transcription.completed", + ] + assert {event["item_id"] for event in (*finals, *endpoint)} == {item_id} + assert all("speaker" not in event and "language" not in event for event in finals) + assert endpoint[1]["transcript"] == "Hello world" + assert "segments" not in endpoint[1] + assert endpoint[1]["usage"] == {"type": "duration", "seconds": 2.0} + assert transformer.take_unbilled_usage() is None + + +def test_manual_finalization_completes_the_turn_without_speech_stopped_and_starts_a_new_item(): + transformer = SonioxEventTransformer(drops_originals=False) + + first = transformer.transform(_response([_token("One"), _token("")])) + second = transformer.transform(_response([_token("Two"), _token("")])) + + assert _types(first) == [ + "input_audio_buffer.speech_started", + "conversation.item.input_audio_transcription.delta", + "conversation.item.input_audio_transcription.completed", + ] + assert first[-1]["transcript"] == "One" + assert "usage" not in first[-1] + assert second[-1]["transcript"] == "Two" + assert second[0]["item_id"] != first[0]["item_id"] + + +def test_a_commit_without_speech_answers_with_an_empty_turn_billed_for_the_silence(): + config = SonioxRealtimeConfig() + config.transform_realtime_request(_append(), MODEL) + config.transform_realtime_request(_event("input_audio_buffer.commit"), MODEL) + + events = tuple( + config.transform_realtime_response( + _response([_token("")], total_audio_proc_ms=2000), MODEL, MagicMock(), EMPTY_TRANSFORM_INPUT + )["response"] + ) + + assert _types(events) == ["conversation.item.input_audio_transcription.completed"] + assert events[0]["transcript"] == "" + assert events[0]["usage"] == {"type": "duration", "seconds": 2.0} + assert _events(config, [_token("Next"), _token("")])[0]["item_id"] != events[0]["item_id"] + + +def test_a_commit_soniox_answers_after_its_endpoint_already_ended_the_turn_adds_no_second_turn(): + config = SonioxRealtimeConfig() + config.transform_realtime_request(_append(), MODEL) + _events(config, [_token("Hello")]) + config.transform_realtime_request(_event("input_audio_buffer.commit"), MODEL) + + events = _events(config, [_token(""), _token("")]) + + assert [event["transcript"] for event in events if "transcript" in event] == ["Hello"] + + +def test_a_server_vad_turn_answers_only_the_oldest_pending_commit(): + config = SonioxRealtimeConfig() + config.transform_realtime_request(_append(), MODEL) + _events(config, [_token("Hello")]) + config.transform_realtime_request(_event("input_audio_buffer.commit"), MODEL) + config.transform_realtime_request(_append(), MODEL) + config.transform_realtime_request(_event("input_audio_buffer.commit"), MODEL) + + events = _events(config, [_token(""), _token(""), _token("")]) + + assert [event["transcript"] for event in events if "transcript" in event] == ["Hello", ""] + + +def test_a_clear_without_speech_still_answers_nothing(): + config = SonioxRealtimeConfig() + config.transform_realtime_request(_append(), MODEL) + config.transform_realtime_request(_event("input_audio_buffer.clear"), MODEL) + + assert _events(config, [_token("")]) == () + + +def test_translation_sessions_transcribe_only_the_translated_tokens(): + transformer = SonioxEventTransformer(drops_originals=True) + + events = transformer.transform( + _response( + [ + _token("Hola", translation_status="original"), + _token("Hello", translation_status="translation"), + _token(""), + ] + ) + ) + + assert [event.get("delta") for event in events if "delta" in event] == ["Hello"] + assert events[-1]["transcript"] == "Hello" + + +def test_translation_sessions_keep_speech_soniox_leaves_untranslated(): + transformer = SonioxEventTransformer(drops_originals=True) + + events = transformer.transform( + _response( + [_token("Hola", translation_status="none"), _token(" amigos", translation_status="none"), _token("")] + ) + ) + + assert events[-1]["transcript"] == "Hola amigos" + + +def test_diarized_tokens_label_each_delta_and_split_the_completed_turn_into_speaker_segments(): + transformer = SonioxEventTransformer(drops_originals=False) + + events = transformer.transform( + _response( + [ + _token("Hi", speaker="1", language="en"), + _token(" doctor", speaker="1", language="en"), + _token(" Salam", speaker="2", language="fa"), + _token(" Hi", speaker="1", language="en"), + _token(""), + ] + ) + ) + + assert [(event["delta"], event["speaker"], event["language"]) for event in events if "delta" in event] == [ + ("Hi", "1", "en"), + (" doctor", "1", "en"), + (" Salam", "2", "fa"), + (" Hi", "1", "en"), + ] + assert events[-1]["transcript"] == "Hi doctor Salam Hi" + assert events[-1]["segments"] == ( + {"speaker": "1", "language": "en", "text": "Hi doctor"}, + {"speaker": "2", "language": "fa", "text": "Salam"}, + {"speaker": "1", "language": "en", "text": "Hi"}, + ) + + +def test_translation_sessions_announce_speech_on_the_first_spoken_token_before_any_translation(): + transformer = SonioxEventTransformer(drops_originals=True) + + events = transformer.transform(_response([_token("Hol", is_final=False, translation_status="original")])) + + assert _types(events) == ["input_audio_buffer.speech_started"] + + +def test_provider_errors_surface_as_openai_error_events(): + transformer = SonioxEventTransformer(drops_originals=False) + + events = transformer.transform(json.dumps({"error_code": 401, "error_message": "Invalid API key."})) + + assert events == ( + {"type": "error", "error": {"type": "server_error", "message": "Soniox realtime error 401: Invalid API key."}}, + ) + + +def test_malformed_provider_frames_are_rejected(): + with pytest.raises(SonioxProtocolError): + SonioxEventTransformer(drops_originals=False).transform('{"tokens": "nope"}') + + +def test_audio_processed_after_the_last_turn_is_billed_on_session_close(): + config = SonioxRealtimeConfig() + config.transform_realtime_response( + _response([_token("Hi"), _token("")], total_audio_proc_ms=1000), MODEL, MagicMock(), EMPTY_TRANSFORM_INPUT + ) + config.transform_realtime_response( + _response([], total_audio_proc_ms=3500), MODEL, MagicMock(), EMPTY_TRANSFORM_INPUT + ) + + assert config.unbilled_usage_on_session_close(MODEL) == {"type": "duration", "seconds": 2.5} + assert config.unbilled_usage_on_session_close(MODEL) is None + + +def test_translation_option_switches_the_config_to_translated_transcripts(): + config = SonioxRealtimeConfig( + options=SonioxRealtimeOptions(translation={"type": "one_way", "target_language": "en"}) + ) + + result = config.transform_realtime_response( + _response([_token("Hola", translation_status="original"), _token("Hello", translation_status="translation")]), + MODEL, + MagicMock(), + EMPTY_TRANSFORM_INPUT, + ) + + assert [event.get("delta") for event in result["response"] if "delta" in event] == ["Hello"] + + +def test_stream_time_since_the_start_request_is_billed_on_close_because_soniox_charges_the_whole_stream(): + now = [100.0] + config = SonioxRealtimeConfig(clock=lambda: now[0]) + config.wrap_backend(_RecordingBackend()) + now[0] = 110.0 + config.transform_realtime_request(_append(), MODEL) + config.transform_realtime_response( + _response([_token("Hi"), _token("")], total_audio_proc_ms=5000), MODEL, MagicMock(), EMPTY_TRANSFORM_INPUT + ) + now[0] = 130.5 + + assert config.unbilled_usage_on_session_close(MODEL) == {"type": "duration", "seconds": 15.5} + assert config.unbilled_usage_on_session_close(MODEL) is None + + +def _refusal_after(processed_ms: int, error_code: int = 401) -> SonioxRealtimeConfig: + now = [100.0] + config = SonioxRealtimeConfig(clock=lambda: now[0]) + config.transform_realtime_request(_append(), MODEL) + if processed_ms: + config.transform_realtime_response( + _response([_token("Hi")], total_audio_proc_ms=processed_ms), MODEL, MagicMock(), EMPTY_TRANSFORM_INPUT + ) + config.transform_realtime_response( + json.dumps({"error_code": error_code, "error_message": "Soniox error.", "total_audio_proc_ms": processed_ms}), + MODEL, + MagicMock(), + EMPTY_TRANSFORM_INPUT, + ) + now[0] = 103.0 + return config + + +@pytest.mark.parametrize("error_code", [400, 401, 402, 403, 408, 429, 500, 503]) +def test_an_error_before_any_audio_is_processed_is_a_refused_session_that_bills_nothing(error_code: int): + config = _refusal_after(processed_ms=0, error_code=error_code) + + refusal = config.session_refusal() + assert isinstance(refusal, SonioxException) + assert refusal.status_code == error_code + assert config.unbilled_usage_on_session_close(MODEL) is None + + +def test_an_error_after_audio_was_processed_keeps_the_session_a_billed_success(): + config = _refusal_after(processed_ms=1500) + + assert config.session_refusal() is None + assert config.unbilled_usage_on_session_close(MODEL) == {"type": "duration", "seconds": 3.0} + + +def test_a_connection_that_never_started_a_stream_bills_nothing(): + now = [100.0] + config = SonioxRealtimeConfig(clock=lambda: now[0]) + config.wrap_backend(_RecordingBackend()) + now[0] = 160.0 + + assert config.unbilled_usage_on_session_close(MODEL) is None + + +def test_soniox_opts_out_of_deferred_setup_so_client_frames_reach_it_in_order(): + assert SonioxRealtimeConfig().requires_session_configuration() + + +def test_a_session_update_matching_the_started_stream_is_a_no_op(): + config = SonioxRealtimeConfig() + config.transform_realtime_request(_append(), MODEL) + + assert ( + config.transform_realtime_request(_session_update({"format": {"type": "audio/pcm", "rate": 24_000}}), MODEL) + == () + ) + + +def _stream_started_at_16khz_in_french_without_vad() -> SonioxRealtimeConfig: + config = SonioxRealtimeConfig() + config.transform_realtime_request( + _session_update( + { + "format": {"type": "audio/pcm", "rate": 16_000}, + "transcription": {"language": "fr"}, + "turn_detection": None, + } + ), + MODEL, + ) + return config + + +@pytest.mark.parametrize( + "follow_up", + [ + {"transcription": {"language": "fr"}}, + {"format": {"type": "audio/pcm", "rate": 16_000}}, + {"format": {"type": "audio/pcm"}}, + {"turn_detection": None}, + ], +) +def test_a_partial_follow_up_update_that_restates_the_stream_is_a_no_op(follow_up: dict[str, object]): + config = _stream_started_at_16khz_in_french_without_vad() + + assert config.transform_realtime_request(_session_update(follow_up), MODEL) == () + + +def test_a_follow_up_update_with_an_unsupported_encoding_is_still_rejected(): + config = _stream_started_at_16khz_in_french_without_vad() + + with pytest.raises(SonioxProtocolError, match="requires pcm16"): + config.transform_realtime_request(_session_update({"format": {"type": "audio/pcmu"}}), MODEL) + + +@pytest.mark.parametrize( + "follow_up", + [ + {"transcription": {"language": "de"}}, + {"format": {"type": "audio/pcm", "rate": 24_000}}, + {"turn_detection": {"type": "server_vad"}}, + ], +) +def test_a_partial_follow_up_update_that_changes_a_sent_setting_is_rejected(follow_up: dict[str, object]): + config = _stream_started_at_16khz_in_french_without_vad() + + with pytest.raises(SonioxProtocolError, match="once the stream has started"): + config.transform_realtime_request(_session_update(follow_up), MODEL) + + +def test_session_created_event_advertises_the_deployment_model_for_cost_tracking(): + config = SonioxRealtimeConfig(options=SonioxRealtimeOptions(language_hints=("de",))) + + event = config.transform_session_created_event(MODEL, "session_1") + + assert event["session"]["id"] == "session_1" + assert event["session"]["audio"]["input"]["transcription"] == {"model": MODEL, "language": "de"} + + +def test_backend_session_start_is_announced_with_the_negotiated_session(): + config = SonioxRealtimeConfig() + config.transform_session_created_event(MODEL, "session_1") + config.transform_realtime_request( + _session_update( + { + "format": {"type": "audio/pcm", "rate": 16_000}, + "transcription": {"language": "it"}, + "turn_detection": None, + } + ), + MODEL, + ) + + result = config.transform_realtime_response(SESSION_STARTED_FRAME, MODEL, MagicMock(), EMPTY_TRANSFORM_INPUT) + + assert _types(tuple(result["response"])) == ["session.created"] + assert result["response"][0]["session"]["id"] == "session_1" + assert result["response"][0]["session"]["audio"]["input"] == { + "format": {"type": "audio/pcm", "rate": 16_000}, + "transcription": {"model": MODEL, "language": "it"}, + "turn_detection": None, + } + + +def test_provider_config_manager_passes_only_soniox_options_from_deployment_params(): + config = ProviderConfigManager.get_provider_realtime_config( + MODEL, + LlmProviders.SONIOX, + litellm_params={ + "api_key": "secret", + "model": "soniox/stt-rt-v5", + "audio_format": "mulaw", + "language_hints": ["pl"], + "context": {"terms": ["LiteLLM"]}, + "max_endpoint_delay_ms": 800, + }, + ) + + assert isinstance(config, SonioxRealtimeConfig) + assert _start_request(config.transform_realtime_request(_append(), MODEL)) == { + "model": MODEL, + "audio_format": "pcm_s16le", + "sample_rate": 24_000, + "num_channels": 1, + "enable_endpoint_detection": True, + "language_hints": ["pl"], + "context": {"terms": ["LiteLLM"]}, + "max_endpoint_delay_ms": 800, + } + + +SEMANTIC_VAD: Final = {"turn_detection": {"type": "semantic_vad", "eagerness": "high"}} + + +def _soniox_config(drop_params: bool) -> SonioxRealtimeConfig: + config = ProviderConfigManager.get_provider_realtime_config(MODEL, LlmProviders.SONIOX, drop_params=drop_params) + assert isinstance(config, SonioxRealtimeConfig) + return config + + +def test_unsupported_turn_detection_is_rejected_without_drop_params(): + config = _soniox_config(drop_params=False) + + with pytest.raises(SonioxProtocolError, match="Set drop_params to use server_vad instead"): + config.transform_realtime_request(_session_update(SEMANTIC_VAD), MODEL) + + +def test_unsupported_turn_detection_falls_back_to_endpoint_detection_with_drop_params(): + config = _soniox_config(drop_params=True) + + frames = config.transform_realtime_request(_session_update(SEMANTIC_VAD), MODEL) + + assert _start_request(frames)["enable_endpoint_detection"] is True + assert config.transform_realtime_request(_session_update(SEMANTIC_VAD), MODEL) == () + + +def test_a_follow_up_with_unsupported_turn_detection_is_rejected_without_drop_params(): + config = _soniox_config(drop_params=False) + config.transform_realtime_request(_append(), MODEL) + + with pytest.raises(SonioxProtocolError, match="Set drop_params to use server_vad instead"): + config.transform_realtime_request(_session_update(SEMANTIC_VAD), MODEL) + + +class _RecordingBackend: + def __init__(self) -> None: + self.sent: Final[list[str | bytes]] = [] + self.upstream: Final[asyncio.Queue[str | Exception]] = asyncio.Queue() + self.entered = False + self.exited = False + self.closed = False + + async def __aenter__(self) -> "_RecordingBackend": + self.entered = True + return self + + async def __aexit__(self, exc_type, exc_value, traceback) -> None: + self.exited = True + + async def send(self, message: str | bytes) -> None: + self.sent.append(message) + + async def recv(self, decode: bool | None = None) -> str | bytes: + frame = await self.upstream.get() + if isinstance(frame, Exception): + raise frame + return frame + + async def close(self) -> None: + self.closed = True + + +class _BrokenBackend(_RecordingBackend): + async def send(self, message: str | bytes) -> None: + raise ConnectionError("upstream gone") + + +@pytest.mark.asyncio +async def test_keepalive_is_sent_only_while_the_client_is_idle_and_stops_after_end_of_stream(): + now = [0.0] + ticks: asyncio.Queue[float] = asyncio.Queue() + + async def sleep(_: float) -> None: + now[0] = await ticks.get() + + async def tick(at: float) -> None: + await ticks.put(at) + for _ in range(5): + await asyncio.sleep(0) + + inner = _RecordingBackend() + backend = SonioxRealtimeBackend(inner, keepalive_interval=5.0, clock=lambda: now[0], sleep=sleep) + + async with backend: + await tick(5.0) + assert inner.sent == ['{"type":"keepalive"}'] + now[0] = 7.0 + await backend.send(PCM) + await tick(10.0) + assert inner.sent == ['{"type":"keepalive"}', PCM] + await backend.send("") + await tick(30.0) + assert inner.sent == ['{"type":"keepalive"}', PCM, ""] + + assert inner.entered and inner.exited + + +@pytest.mark.asyncio +async def test_backend_announces_the_session_once_the_start_request_is_sent_while_relaying_upstream_frames(): + inner = _RecordingBackend() + backend = SonioxRealtimeBackend(inner) + + async with backend: + pending = asyncio.create_task(backend.recv()) + await backend.send('{"type":"keepalive"}') + await asyncio.sleep(0) + assert not pending.done() + await backend.send('{"model":"stt-rt-v5"}') + assert await pending == SESSION_STARTED_FRAME + await backend.send('{"model":"ignored-second-config"}') + await inner.upstream.put('{"tokens":[]}') + assert await backend.recv() == '{"tokens":[]}' + + +@pytest.mark.asyncio +async def test_backend_relays_upstream_frames_that_arrive_before_the_session_starts(): + inner = _RecordingBackend() + + async with SonioxRealtimeBackend(inner) as backend: + await inner.upstream.put('{"error_code":401,"error_message":"Invalid API key."}') + + assert await backend.recv() == '{"error_code":401,"error_message":"Invalid API key."}' + + +@pytest.mark.asyncio +async def test_a_cancelled_read_loses_no_frames_and_upstream_errors_reach_the_reader(): + inner = _RecordingBackend() + + async with SonioxRealtimeBackend(inner) as backend: + cancelled = asyncio.create_task(backend.recv()) + await asyncio.sleep(0) + cancelled.cancel() + await asyncio.gather(cancelled, return_exceptions=True) + await inner.upstream.put('{"tokens":[]}') + await inner.upstream.put(ConnectionError("upstream closed")) + + assert await backend.recv() == '{"tokens":[]}' + with pytest.raises(ConnectionError, match="upstream closed"): + await backend.recv() + + +@pytest.mark.asyncio +async def test_closing_the_backend_cancels_pending_reads_and_closes_upstream(): + inner = _RecordingBackend() + backend = SonioxRealtimeBackend(inner) + await backend.__aenter__() + pending = asyncio.create_task(backend.recv()) + await backend.send('{"model":"stt-rt-v5"}') + assert await pending == SESSION_STARTED_FRAME + + await backend.close() + await inner.upstream.put('{"tokens":[]}') + await asyncio.sleep(0) + + assert inner.closed + assert inner.upstream.qsize() == 1 + + +@pytest.mark.asyncio +async def test_keepalive_stops_after_the_upstream_rejects_a_send(): + sleeps: list[float] = [] + + async def sleep(seconds: float) -> None: + sleeps.append(seconds) + await asyncio.sleep(0) + + async with SonioxRealtimeBackend(_BrokenBackend(), keepalive_interval=0.0, sleep=sleep): + for _ in range(5): + await asyncio.sleep(0) + + assert sleeps == [0.0] diff --git a/tests/unit/llms/test_get_complete_url_trailing_slash.py b/tests/unit/llms/test_get_complete_url_trailing_slash.py index cd778c20d9b..43a74e56028 100644 --- a/tests/unit/llms/test_get_complete_url_trailing_slash.py +++ b/tests/unit/llms/test_get_complete_url_trailing_slash.py @@ -143,6 +143,7 @@ API_BASE_IS_THE_FULL_URL: Final[Mapping[str, str]] = MappingProxyType( "duckduckgo.search.transformation.DuckDuckGoSearchConfig", "google_pse.search.transformation.GooglePSESearchConfig", "searchapi.search.transformation.SearchAPIConfig", + "soniox.realtime.transformation.SonioxRealtimeConfig", "tinyfish.search.transformation.TinyfishSearchConfig", "vertex_ai.text_to_speech.transformation.VertexAITextToSpeechConfig", "vertex_ai.vertex_ai_partner_models.anthropic.messages.transformation." diff --git a/tests/unit/realtime_api/test_main.py b/tests/unit/realtime_api/test_main.py index 5478b807691..46777b32319 100644 --- a/tests/unit/realtime_api/test_main.py +++ b/tests/unit/realtime_api/test_main.py @@ -618,7 +618,11 @@ async def test_bearer_health_check_sends_the_secret_behind_an_oidc_credential( monkeypatch.setattr( litellm, "credential_list", - [CredentialItem(credential_name="voice-sa", credential_values={"api_key": f"oidc/file/{token}"}, credential_info={})], + [ + CredentialItem( + credential_name="voice-sa", credential_values={"api_key": f"oidc/file/{token}"}, credential_info={} + ) + ], ) connect: Final = _CapturingConnect() with patch("websockets.connect", connect): @@ -1447,3 +1451,98 @@ async def test_openai_realtime_health_check_sends_the_federated_bearer_for_a_key model_params={"model": "openai/gpt-realtime", **federated_deployment}, ) assert connect.kwargs["additional_headers"] == {"Authorization": "Bearer federated-bearer"} + + +@pytest.mark.asyncio +async def test_arealtime_routes_soniox_with_the_deployment_options(monkeypatch): + from litellm.llms.soniox.realtime.transformation import SonioxRealtimeConfig + + captured: dict[str, object] = {} + + async def mock_async_realtime(**kwargs): + captured.update(kwargs) + + monkeypatch.setattr(realtime_main.base_llm_http_handler, "async_realtime", mock_async_realtime) + await realtime_main._arealtime.__wrapped__( + model="soniox/stt-rt-v5", + websocket=MagicMock(), + litellm_logging_obj=FakeLogging(), + query_params={"model": "soniox/stt-rt-v5", "intent": "transcription"}, + api_key="deployment-key", + language_hints=["de"], + ) + + provider_config = captured["provider_config"] + assert isinstance(provider_config, SonioxRealtimeConfig) + assert captured["model"] == "stt-rt-v5" + assert captured["api_key"] == "deployment-key" + start_request = provider_config.transform_realtime_request( + json.dumps({"type": "input_audio_buffer.commit"}), "stt-rt-v5" + )[0] + assert json.loads(start_request)["language_hints"] == ["de"] + + +async def _serve_soniox(reply: str | None) -> tuple[object, list[object]]: + import websockets + + received: list[object] = [] + + async def handler(connection) -> None: + received.append(connection.request.headers["Authorization"]) + received.append(json.loads(await connection.recv())) + received.append(await connection.recv()) + received.append(await connection.recv()) + if reply is not None: + await connection.send(reply) + + return await websockets.serve(handler, "127.0.0.1", 0), received + + +@pytest.mark.asyncio +async def test_soniox_realtime_health_check_starts_and_ends_a_stream_on_the_deployment_websocket(): + server, received = await _serve_soniox(json.dumps({"tokens": [], "finished": True})) + port = next(iter(server.sockets)).getsockname()[1] + async with server: + healthy = await realtime_main.realtime_health_check( + model="stt-rt-v5", + custom_llm_provider="soniox", + api_key="soniox-key", + api_base=f"http://127.0.0.1:{port}", + model_params={"model": "soniox/stt-rt-v5", "language_hints": ["fa", "en"]}, + ) + + assert healthy is True + assert received[0] == "Bearer soniox-key" + assert received[1]["model"] == "stt-rt-v5" + assert received[1]["language_hints"] == ["fa", "en"] + assert received[2] == bytes(len(received[2])) + assert len(received[2]) > 0 + assert received[3] == "" + + +@pytest.mark.asyncio +async def test_soniox_realtime_health_check_fails_when_soniox_refuses_the_stream(): + from litellm.llms.soniox.common_utils import SonioxException + + server, _ = await _serve_soniox(json.dumps({"error_code": 401, "error_message": "Invalid API key."})) + port = next(iter(server.sockets)).getsockname()[1] + async with server: + with pytest.raises(SonioxException) as refused: + await realtime_main.realtime_health_check( + model="stt-rt-v5", custom_llm_provider="soniox", api_key="bad-key", api_base=f"ws://127.0.0.1:{port}" + ) + + assert refused.value.status_code == 401 + + +@pytest.mark.asyncio +async def test_soniox_realtime_health_check_fails_when_soniox_closes_without_finishing_the_stream(): + from litellm.llms.soniox.common_utils import SonioxException + + server, _ = await _serve_soniox(None) + port = next(iter(server.sockets)).getsockname()[1] + async with server: + with pytest.raises(SonioxException): + await realtime_main.realtime_health_check( + model="stt-rt-v5", custom_llm_provider="soniox", api_key="soniox-key", api_base=f"ws://127.0.0.1:{port}" + ) From c13ecb308522841ba7485f571c7172056f19e8ca Mon Sep 17 00:00:00 2001 From: Otis Cui Date: Thu, 29 Oct 2026 23:05:47 +0800 Subject: [PATCH 5/8] fix(exceptions)!: honor status code when mapping OpenAI invalid_request_error exceptions (#32537) * fix: openai exception mapper ignored status code for invalid_request_error _map_openai_exception coerced anything whose body carried type invalid_request_error into BadRequestError, guarded only by excluding the single phrase Incorrect API key provided, and this ran before the status-code checks. OpenAI and openai-compatible providers return 401 and 404 bodies with type invalid_request_error, so an auth failure like Invalid API key mapped to a 400 instead of AuthenticationError, and a 404 mapped to BadRequestError instead of NotFoundError. That changes the status returned to proxy clients and router retry, cooldown, and fallback handling. Mirror the Azure mapper by only coercing to BadRequestError when status_code is None or 400, so 401 and 404 fall through to their existing handlers. * fix(exceptions): map OpenAI-family invalid_request_error 404s to NotFoundError Move the contributor's test to the mapped unit test file, assert the raised class, status code and message, and cover the 400 and status-less cases that must still raise BadRequestError Add a regression test for the flow the bug was found on. Retrieving a missing OpenAI vector store, sync or async, now raises NotFoundError with status 404 The router test that pinned an OpenAI 404 typed invalid_request_error as BadRequestError now expects NotFoundError * fix(a2a): answer a provider NotFoundError as JSON-RPC invalid params * fix(exceptions)!: limit the invalid_request_error status guard to 401 and 404 Every other status keeps raising BadRequestError as on main, so 402-415, 429 and 5xx bodies typed invalid_request_error keep their class and router retry behavior. A2A goes back to main's mapping, which answers NotFoundError and AuthenticationError as JSON-RPC internal errors. * fix(exceptions): keep the provider error fields on mapped 401 and 404 invalid_request_errors Raise AuthenticationError and NotFoundError straight from the invalid_request_error branch with the provider body, so code, param and type survive and the proxy's error.type matches main. The Rust mapper copy skips 401 and 404 the same way. --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- .../src/exception_mapping_utils/openai.rs | 18 +++ litellm/exceptions.py | 6 +- .../exception_mapping_utils.py | 19 +++ .../test_exception_mapping_utils.py | 133 ++++++++++++++++++ tests/unit/test_router/test_router.py | 2 +- tests/unit/vector_stores/test_main.py | 45 +++++- 6 files changed, 219 insertions(+), 4 deletions(-) diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/openai.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/openai.rs index d45078c415e..1273fd4d7f1 100644 --- a/litellm-rust/crates/core-utils/src/exception_mapping_utils/openai.rs +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/openai.rs @@ -53,6 +53,7 @@ pub(super) const RULES: &[Rule] = &[ |mapping| { mapping.error_str.contains("invalid_request_error") && !mapping.error_str.contains("Incorrect API key provided") + && !matches!(mapping.status, Some(401 | 404)) }, PublicError::BadRequest, ), @@ -132,6 +133,23 @@ mod tests { assert_eq!(classified(Some(400), text), Some(expected)); } + #[rstest::rstest] + #[case::unauthorized(Some(401), None)] + #[case::not_found(Some(404), None)] + #[case::forbidden(Some(403), Some(PublicError::BadRequest))] + #[case::rate_limited(Some(429), Some(PublicError::BadRequest))] + #[case::server_error(Some(500), Some(PublicError::BadRequest))] + #[case::no_status(None, Some(PublicError::BadRequest))] + fn invalid_request_errors_leave_401_and_404_to_the_status_mapping( + #[case] status: Option, + #[case] expected: Option, + ) { + assert_eq!( + classified(status, "invalid_request_error bad field"), + expected + ); + } + #[rstest::rstest] #[case::rate_limit_before_context_window( "rate limit and This model's maximum context length is 10", diff --git a/litellm/exceptions.py b/litellm/exceptions.py index cd5139f5f79..61e3a8fb5dc 100644 --- a/litellm/exceptions.py +++ b/litellm/exceptions.py @@ -139,6 +139,7 @@ class AuthenticationError(openai.AuthenticationError): litellm_debug_info: str | None = None, max_retries: int | None = None, num_retries: int | None = None, + body: Mapping[str, object] | None = None, ): self.status_code = 401 self.message = f"litellm.AuthenticationError: {message}" @@ -152,7 +153,7 @@ class AuthenticationError(openai.AuthenticationError): request=httpx.Request(method="GET", url="https://litellm.ai"), # mock request object ) super().__init__( - self.message, response=self.response, body=None + self.message, response=self.response, body=body ) # Call the base class constructor with the parameters it needs def __str__(self): @@ -189,6 +190,7 @@ class NotFoundError(openai.NotFoundError): litellm_debug_info: str | None = None, max_retries: int | None = None, num_retries: int | None = None, + body: Mapping[str, object] | None = None, ): self.status_code = 404 self.message = f"litellm.NotFoundError: {message}" @@ -202,7 +204,7 @@ class NotFoundError(openai.NotFoundError): request=httpx.Request(method="GET", url="https://litellm.ai"), # mock request object ) super().__init__( - self.message, response=self.response, body=None + self.message, response=self.response, body=body ) # Call the base class constructor with the parameters it needs def __str__(self): diff --git a/litellm/litellm_core_utils/exception_mapping_utils.py b/litellm/litellm_core_utils/exception_mapping_utils.py index 703550dd2b5..64a95eb5f28 100644 --- a/litellm/litellm_core_utils/exception_mapping_utils.py +++ b/litellm/litellm_core_utils/exception_mapping_utils.py @@ -413,6 +413,25 @@ def _map_openai_exception( body=getattr(original_exception, "body", None), ) elif "invalid_request_error" in error_str and "Incorrect API key provided" not in error_str: + invalid_request_status: Final = getattr(original_exception, "status_code", None) + if invalid_request_status == 401: + raise AuthenticationError( + message=f"{exception_provider} - {message}", + llm_provider=custom_llm_provider, + model=model, + response=response, + litellm_debug_info=extra_information, + body=getattr(original_exception, "body", None), + ) + if invalid_request_status == 404: + raise NotFoundError( + message=f"{exception_provider} - {message}", + model=model, + llm_provider=custom_llm_provider, + response=response, + litellm_debug_info=extra_information, + body=getattr(original_exception, "body", None), + ) raise BadRequestError( message=f"{exception_provider} - {message}", llm_provider=custom_llm_provider, diff --git a/tests/unit/litellm_core_utils/test_exception_mapping_utils.py b/tests/unit/litellm_core_utils/test_exception_mapping_utils.py index 55e7c47e2c9..19e5387da9a 100644 --- a/tests/unit/litellm_core_utils/test_exception_mapping_utils.py +++ b/tests/unit/litellm_core_utils/test_exception_mapping_utils.py @@ -24,6 +24,7 @@ from litellm.llms.base_llm.chat.transformation import BaseLLMException from litellm.llms.bedrock.common_utils import BedrockError from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler from litellm.llms.openai.common_utils import OpenAIError +from litellm.proxy.common_utils.openai_error_payload import openai_error_type from litellm.types.utils import LlmProviders @@ -1015,6 +1016,138 @@ def test_azure_404_with_invalid_request_error_type_maps_to_not_found(): assert "Response with id 'resp_abc' not found." in excinfo.value.message +@pytest.mark.parametrize("provider", ["openai", "groq", "deepseek"]) +@pytest.mark.parametrize( + "status_code, message, expected_exception", + [ + (400, "Invalid 'temperature'", litellm.BadRequestError), + (401, "Invalid API key", litellm.AuthenticationError), + (404, "The model does not exist", litellm.NotFoundError), + ], +) +def test_openai_invalid_request_error_respects_status_code( + provider: str, status_code: int, message: str, expected_exception: type[openai.APIStatusError] +) -> None: + error_message: Final = json.dumps({"error": {"message": message, "type": "invalid_request_error"}}) + original_exception: Final = OpenAIError(status_code=status_code, message=error_message, headers={}) + + with pytest.raises(openai.APIError) as excinfo: + exception_type( + model="some-model", + original_exception=original_exception, + custom_llm_provider=provider, + ) + + assert type(excinfo.value) is expected_exception + assert excinfo.value.status_code == status_code + assert message in excinfo.value.message + + +@pytest.mark.parametrize("provider", ["openai", "groq", "deepseek"]) +def test_openai_invalid_request_error_without_a_status_code_maps_to_bad_request(provider: str) -> None: + message: Final = "Missing required parameter: 'input'." + original_exception: Final = Exception(json.dumps({"error": {"message": message, "type": "invalid_request_error"}})) + + with pytest.raises(openai.APIError) as excinfo: + exception_type( + model="some-model", + original_exception=original_exception, + custom_llm_provider=provider, + ) + + assert type(excinfo.value) is litellm.BadRequestError + assert message in excinfo.value.message + + +@pytest.mark.parametrize("provider", ["openai", "groq", "deepseek"]) +@pytest.mark.parametrize("status_code", [403, 409, 422, 429, 500, 503]) +def test_openai_invalid_request_error_keeps_bad_request_for_statuses_other_than_401_and_404( + provider: str, status_code: int +) -> None: + message: Final = "Cannot cancel a batch that is already completed." + error_message: Final = json.dumps({"error": {"message": message, "type": "invalid_request_error"}}) + original_exception: Final = OpenAIError(status_code=status_code, message=error_message, headers={}) + + with pytest.raises(openai.APIError) as excinfo: + exception_type( + model="some-model", + original_exception=original_exception, + custom_llm_provider=provider, + ) + + assert type(excinfo.value) is litellm.BadRequestError + assert message in excinfo.value.message + + +def _openai_sdk_error( + sdk_error: type[openai.APIStatusError], status_code: int, provider_error: dict[str, str | None] +) -> openai.APIStatusError: + request: Final = httpx.Request("POST", "https://api.openai.com/v1/chat/completions") + return sdk_error( + f"Error code: {status_code} - {{'error': {provider_error}}}", + response=httpx.Response(status_code, request=request), + body=provider_error, + ) + + +@pytest.mark.parametrize( + "sdk_error, status_code, expected_exception, param, code", + [ + (openai.AuthenticationError, 401, litellm.AuthenticationError, "organization", "mismatched_organization"), + (openai.NotFoundError, 404, litellm.NotFoundError, "id", None), + ], +) +def test_openai_invalid_request_error_401_and_404_keep_the_provider_error_fields( + sdk_error: type[openai.APIStatusError], + status_code: int, + expected_exception: type[openai.APIStatusError], + param: str, + code: str | None, +) -> None: + provider_error: Final = { + "message": "The request was rejected.", + "type": "invalid_request_error", + "param": param, + "code": code, + } + + with pytest.raises(openai.APIError) as excinfo: + exception_type( + model="gpt-5-mini", + original_exception=_openai_sdk_error(sdk_error, status_code, provider_error), + custom_llm_provider="openai", + ) + + assert type(excinfo.value) is expected_exception + assert (excinfo.value.status_code, excinfo.value.type, excinfo.value.param, excinfo.value.code) == ( + status_code, + "invalid_request_error", + param, + code, + ) + assert excinfo.value.body == provider_error + assert openai_error_type(excinfo.value, status_code) == "invalid_request_error" + assert excinfo.value.message.count(expected_exception.__name__) == 1 + + +def test_openai_incorrect_api_key_keeps_the_proxy_authentication_error_type() -> None: + provider_error: Final = { + "message": "Incorrect API key provided: sk-litellmqa.", + "type": "invalid_request_error", + "param": None, + "code": "invalid_api_key", + } + + with pytest.raises(litellm.AuthenticationError) as excinfo: + exception_type( + model="gpt-5-mini", + original_exception=_openai_sdk_error(openai.AuthenticationError, 401, provider_error), + custom_llm_provider="openai", + ) + + assert openai_error_type(excinfo.value, 401) == "authentication_error" + + class _UpstreamHTTPError(Exception): def __init__(self, status_code: int) -> None: super().__init__("upstream failure") diff --git a/tests/unit/test_router/test_router.py b/tests/unit/test_router/test_router.py index b55b3c5efc2..ee3a70c7c1c 100644 --- a/tests/unit/test_router/test_router.py +++ b/tests/unit/test_router/test_router.py @@ -21632,7 +21632,7 @@ async def test_router_retry_policy_controls_upstream_attempt_count( "error_body,error_type", [ ({"message": "model is down", "type": "server_error"}, litellm.NotFoundError), - ({"message": "Response with id 'resp_x' not found.", "type": "invalid_request_error"}, litellm.BadRequestError), + ({"message": "Response with id 'resp_x' not found.", "type": "invalid_request_error"}, litellm.NotFoundError), ], ) @pytest.mark.parametrize( diff --git a/tests/unit/vector_stores/test_main.py b/tests/unit/vector_stores/test_main.py index 25a378270b1..21b1eee3a3e 100644 --- a/tests/unit/vector_stores/test_main.py +++ b/tests/unit/vector_stores/test_main.py @@ -15,12 +15,13 @@ from unittest.mock import MagicMock, patch import httpx import pytest +import litellm import litellm.vector_stores.main as vector_stores_main from litellm.llms.base_llm.vector_store.transformation import ( RouterVectorStoreEmbeddingExecutor, ) from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler -from litellm.vector_stores.main import acreate, aupdate, search +from litellm.vector_stores.main import acreate, aretrieve, aupdate, retrieve, search MOCK_SEARCH_RESPONSE = { "object": "vector_store.search_results.page", @@ -200,3 +201,45 @@ async def test_direct_sdk_vector_store_writes_send_the_callers_metadata_unchange sent_metadata: Final = await _metadata_sent_to_provider(operation, callers_metadata) assert sent_metadata == [callers_metadata] + + +def _vector_store_not_found(request: httpx.Request) -> httpx.Response: + return httpx.Response( + 404, + json={ + "error": { + "message": "No vector store found with id 'vs_missing'.", + "type": "invalid_request_error", + "param": "vector_store_id", + "code": None, + } + }, + request=request, + ) + + +async def _retrieve_missing_openai_vector_store(is_async: bool) -> object: + transport: Final = httpx.MockTransport(_vector_store_not_found) + if is_async: + return await aretrieve( + vector_store_id="vs_missing", + custom_llm_provider="openai", + api_key="sk-test", + client=AsyncHTTPHandler(transport=transport), + ) + return retrieve( + vector_store_id="vs_missing", + custom_llm_provider="openai", + api_key="sk-test", + client=HTTPHandler(client=httpx.Client(transport=transport)), + ) + + +@pytest.mark.asyncio +@pytest.mark.parametrize("is_async", [True, False]) +async def test_retrieving_a_missing_openai_vector_store_raises_not_found(is_async: bool): + with pytest.raises(litellm.NotFoundError) as caught: + await _retrieve_missing_openai_vector_store(is_async) + + assert caught.value.status_code == 404 + assert "No vector store found with id 'vs_missing'." in caught.value.message From be6d082bdcbc2d83a6522c0e015077e007a9bddc Mon Sep 17 00:00:00 2001 From: Mateo Wang <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 29 Oct 2026 08:06:12 -0700 Subject: [PATCH 6/8] fix(anthropic)!: send /v1/messages PDF documents to OpenAI-shaped providers as file parts (#47166) * fix(anthropic): send /v1/messages PDF documents to OpenAI-shaped providers as file parts * refactor(anthropic): hold the chat completions document change for a later commit * fix(anthropic): send /v1/messages PDFs as file parts and keep file parts on azure_ai * fix(snowflake): turn file parts into document blocks for Cortex Claude * fix(prompt_security): sanitize chat file parts that carry file_data * fix: answer 400 or drop PDFs for models that cannot read them Models whose model_info or cost map entry sets supports_pdf_input to false, and Ollama, now get the PDF dropped under drop_params and a 400 from litellm without it. GigaChat uploads file parts, SAP keeps sending them as image_url data URLs, azure_ai passes dict content through, and annotations other than url_citation move to provider_specific_fields instead of failing the response. * fix(router): scope supports_pdf_input to its deployment and keep provider annotation keys Mark the azure_ai chat models in the cost map that cannot read PDFs, so drop_params drops the PDF and litellm answers its own 400 without it. Keep url_citation annotations whole, move only typed provider annotations to provider_specific_fields, and split them the same way when a stream is rebuilt. * fix(vertex_ai): send PDF file parts to Vertex partner models as image_url Vertex AI's OpenAI-compatible partner endpoint rejects the chat `file` part, so PDFs from /v1/messages reach those models as the image_url data URL they took before. Mark native DeepSeek chat models as unable to read PDFs, and treat a file part as a PDF by its format or filename when it has no data URL. * fix(cost-map): leave native DeepSeek PDF support unset /v1/messages reaches native DeepSeek through its own Anthropic endpoint, so the marks never applied there and would only change chat completions, which this PR does not otherwise touch. * fix(main): drop a user turn that held only a PDF under drop_params Dropping the PDF from a PDF-only user turn left an empty content list, which Foundry and OpenAI-shaped providers reject with a 400, so drop_params still failed for that request. The emptied user turn is now removed. * fix: read OCI PDFs, gate deployment base models, and stop empty drops OCI generic chat sends file parts as image data URLs again, so OCI Gemini keeps reading PDFs from /v1/messages and chat file parts. The PDF gate now reads model_info.base_model, drops PDFs under drop_params for azure_ai deployments that never declared PDF support (the old flattening behavior), and answers 400 when dropping the PDF leaves no user message. The cost map marks the Vertex partner text models that a live probe showed answering without reading the PDF. * test(integration): cover the PDF input gate, messages PDF file parts and provider annotations Wire cells for the supports_pdf_input check on chat, chat streaming, /v1/messages and the Responses bridge, its model_info, base_model, cost map and Ollama sources, drop_params at deployment and request level, azure_ai file parts, sibling deployment isolation, Anthropic SDK PDFs reaching azure as titled file parts, Gemini and Bedrock document controls, Prompt Security sanitizing a /v1/messages PDF, and OpenRouter file annotations live, cached and streamed * test(integration): cover Snowflake Claude and SAP in the messages PDF wire cells * fix(azure_ai): flatten non-PDF file parts to text as before and cover the GigaChat PDF upload * fix(main): PDF check reads the deployment's top-level base_model * fix(main): keep roles alternating after drop_params removes a PDF-only turn When ensure_alternating_roles is set, re-run role normalization on the messages left after the PDF check drops a turn, so two assistant turns never end up next to each other * test: type the PDF and annotation tests the PR added * test: type the GigaChat, SAP and Snowflake PDF tests --- .../convert_dict_to_response.py | 33 +- .../prompt_templates/common_utils.py | 103 +++++ .../streaming_chunk_builder_utils.py | 17 +- .../messages/adapters/transformation.py | 31 +- .../responses_adapters/transformation.py | 26 +- litellm/llms/azure_ai/chat/transformation.py | 24 +- litellm/llms/gigachat/chat/transformation.py | 15 +- litellm/llms/oci/chat/transformation.py | 5 +- litellm/llms/sap/chat/transformation.py | 3 +- litellm/llms/snowflake/chat/transformation.py | 45 +- .../vertex_ai_partner_models/main.py | 3 +- litellm/main.py | 163 ++++++- ...odel_prices_and_context_window_backup.json | 121 ++++- .../prompt_security/prompt_security.py | 5 +- litellm/types/utils.py | 7 +- model_prices_and_context_window.json | 121 ++++- ...st_anthropic_messages_pdf_document_wire.py | 421 ++++++++++++++++++ .../test_chat_provider_annotations_wire.py | 190 ++++++++ .../providers/test_pdf_input_support_wire.py | 390 ++++++++++++++++ .../test_convert_dict_to_response.py | 51 +++ ...hropic_messages_adapters_transformation.py | 34 ++ .../chat/test_azure_ai_transformation.py | 52 +++ .../chat/test_gigachat_chat_transformation.py | 15 + .../oci/chat/test_oci_chat_transformation.py | 30 ++ .../llms/sap/chat/test_sap_transformation.py | 18 + .../test_snowflake_chat_transformation.py | 30 ++ .../test_vertex_ai_gemma_global_endpoint.py | 50 +++ .../test_prompt_security_guardrails.py | 53 +++ tests/unit/test_main.py | 305 ++++++++++++- .../unit/test_router_model_cost_isolation.py | 57 +++ .../test_stream_chunk_builder_annotations.py | 45 ++ 31 files changed, 2374 insertions(+), 89 deletions(-) create mode 100644 tests/integration/messages_endpoint/chat_bridge/test_anthropic_messages_pdf_document_wire.py create mode 100644 tests/integration/providers/test_chat_provider_annotations_wire.py create mode 100644 tests/integration/providers/test_pdf_input_support_wire.py diff --git a/litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py b/litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py index 8b94c34942c..de911beabdd 100644 --- a/litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py +++ b/litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py @@ -19,6 +19,7 @@ from litellm.litellm_core_utils.prompt_templates.common_utils import ( from litellm.llms.ollama.common_utils import OllamaResponseIdNormalizer from litellm.types.llms.databricks import DatabricksTool from litellm.types.llms.openai import ( + ChatCompletionAnnotation, ChatCompletionReasoningItem, ChatCompletionThinkingBlock, ImageURLListItem, @@ -59,6 +60,31 @@ _CHOICES_FIELDS: Final[frozenset] = frozenset(Choices.model_fields.keys()) _MODEL_RESPONSE_FIELDS: Final[frozenset] = frozenset(ModelResponse.model_fields.keys()) | {"usage"} +_ANNOTATION_ENTRIES: Final = TypeAdapter(tuple[object, ...]) + + +def _is_provider_specific_annotation(annotation: object) -> bool: + match annotation: + case {"type": "url_citation"}: + return False + case {"type": _}: + return True + case _: + return False + + +def _annotation_entries(annotations: object) -> tuple[object, ...]: + return _ANNOTATION_ENTRIES.validate_python(annotations) if isinstance(annotations, list) else () + + +def partition_annotations(annotations: object) -> tuple[object, tuple[object, ...]]: + entries: Final = _annotation_entries(annotations) + others: Final = tuple(annotation for annotation in entries if _is_provider_specific_annotation(annotation)) + if not others: + return annotations, () + return [annotation for annotation in entries if not _is_provider_specific_annotation(annotation)] or None, others + + def _normalize_images_for_message( images: list[dict] | None, ) -> list[ImageURLListItem] | None: @@ -787,6 +813,9 @@ def convert_to_model_response_object( provider_specific_fields = dict(choice["message"].get("provider_specific_fields", None) or {}) for f in choice["message"].keys() - _MESSAGE_FIELDS: provider_specific_fields[f] = choice["message"][f] + annotations, other_annotations = partition_annotations(choice["message"].get("annotations", None)) + if other_annotations: + provider_specific_fields["annotations"] = list(other_annotations) # Handle reasoning models that display `reasoning_content` within `content` reasoning_content, content = extract_reasoning_content(choice["message"]) @@ -809,7 +838,9 @@ def convert_to_model_response_object( reasoning_content=reasoning_content, thinking_blocks=thinking_blocks, reasoning_items=choice["message"].get("reasoning_items", None), - annotations=choice["message"].get("annotations", None), + annotations=cast( # cast-ok: provider annotation dicts pass through whole, keeping keys such as url_citation.content + list[ChatCompletionAnnotation] | None, annotations + ), images=_normalize_images_for_message(choice["message"].get("images", None)), ) finish_reason = choice.get("finish_reason", None) diff --git a/litellm/litellm_core_utils/prompt_templates/common_utils.py b/litellm/litellm_core_utils/prompt_templates/common_utils.py index 9bc53791b36..b3f89b33d1d 100644 --- a/litellm/litellm_core_utils/prompt_templates/common_utils.py +++ b/litellm/litellm_core_utils/prompt_templates/common_utils.py @@ -424,6 +424,109 @@ def anthropic_image_source_to_openai_url(image_source: Mapping[str, object]) -> return None +@dataclass(frozen=True, slots=True) +class InlineDocument: + filename: str + file_data: str + + +def _document_media_type(source: object) -> str: + match source: + case {"media_type": str() as media_type} if media_type: + return media_type + case _: + return "application/pdf" + + +def anthropic_inline_document(document: Mapping[str, object]) -> InlineDocument | None: + source: Final = document.get("source") + match source: + case {"type": "base64", "data": str() as data} if data: + title: Final = document.get("title") + return InlineDocument( + filename=title if isinstance(title, str) and title else "document.pdf", + file_data=f"data:{_document_media_type(source)};base64,{data}", + ) + case _: + return None + + +def is_pdf_file_part(part: object) -> bool: + match part: + case {"type": "file", "file": {"file_data": str() as file_data}} if file_data.startswith("data:"): + return file_data.startswith("data:application/pdf") + case {"type": "file", "file": {"format": str() as file_format}}: + return file_format == "application/pdf" + case {"type": "file", "file": {"filename": str() as filename}}: + return filename.lower().endswith(".pdf") + case _: + return False + + +_CONTENT_PARTS: Final = TypeAdapter(tuple[object, ...]) + + +def content_parts(message: Mapping[str, object]) -> Sequence[object]: + content: Final = message.get("content") + return _CONTENT_PARTS.validate_python(content) if isinstance(content, list) else () + + +def _has_pdf_file_part(message: Mapping[str, object]) -> bool: + return any(is_pdf_file_part(part) for part in content_parts(message)) + + +def messages_have_pdf_file_part(messages: Sequence[Mapping[str, object]]) -> bool: + return any(_has_pdf_file_part(message) for message in messages) + + +def _message_without_pdf_file_parts(message: AllMessageValues) -> AllMessageValues: + if not _has_pdf_file_part(message): + return message + kept: Final = [part for part in content_parts(message) if not is_pdf_file_part(part)] + return cast(AllMessageValues, {**message, "content": kept}) # cast-ok: same message with fewer content parts + + +def _is_pdf_only_user_turn(message: AllMessageValues) -> bool: + parts: Final = content_parts(message) + return message["role"] == "user" and bool(parts) and all(is_pdf_file_part(part) for part in parts) + + +def messages_without_pdf_file_parts(messages: Sequence[AllMessageValues]) -> tuple[AllMessageValues, ...]: + return tuple( + _message_without_pdf_file_parts(message) for message in messages if not _is_pdf_only_user_turn(message) + ) + + +_CONTENT_PART: Final = TypeAdapter(Mapping[str, object]) + + +def _file_part_as_image_url(part: object) -> object: + match part: + case {"type": "file", "file": {"file_data": str() as file_data}} if file_data: + fields: Final = _CONTENT_PART.validate_python(part) + return {key: value for key, value in fields.items() if key != "file"} | { + "type": "image_url", + "image_url": {"url": file_data}, + } + case _: + return part + + +def message_with_file_parts_as_image_urls(message: Mapping[str, object]) -> Mapping[str, object]: + parts: Final = content_parts(message) + if not parts: + return message + return {**message, "content": [_file_part_as_image_url(part) for part in parts]} + + +def _typed_message_with_file_parts_as_image_urls(message: AllMessageValues) -> AllMessageValues: + return cast(AllMessageValues, message_with_file_parts_as_image_urls(message)) # cast-ok: same message, image parts + + +def messages_with_file_parts_as_image_urls(messages: Sequence[AllMessageValues]) -> tuple[AllMessageValues, ...]: + return tuple(_typed_message_with_file_parts_as_image_urls(message) for message in messages) + + def _image_part_url(part: Mapping[str, object]) -> str | None: """The image URL carried by one content part, whichever of the three dialects wrote it.""" part_type: Final = part.get("type") diff --git a/litellm/litellm_core_utils/streaming_chunk_builder_utils.py b/litellm/litellm_core_utils/streaming_chunk_builder_utils.py index 7c91ffdff76..8e1791c3c92 100644 --- a/litellm/litellm_core_utils/streaming_chunk_builder_utils.py +++ b/litellm/litellm_core_utils/streaming_chunk_builder_utils.py @@ -19,6 +19,7 @@ from litellm.litellm_core_utils.core_helpers import ( ) from litellm.litellm_core_utils.hidden_params import HIDDEN_PARAMS_ATTR, set_hidden_params from litellm.litellm_core_utils.llm_cost_calc.guardrail_cost import BEDROCK_TRACE_FIELDS, object_items, object_mapping +from litellm.litellm_core_utils.llm_response_utils.convert_dict_to_response import partition_annotations from litellm.litellm_core_utils.prompt_templates.common_utils import ( get_content_from_model_response, # pyright: ignore[reportUnknownVariableType] # response param accepts an untyped dict ) @@ -1700,10 +1701,17 @@ def stream_chunk_builder( # Merge annotations from ALL chunks — providers may spread # them across multiple streaming chunks or send them only in # the final chunk. - all_annotations: Final[list] = [] + all_annotations: Final[list[object]] = [] for ac in annotation_chunks: all_annotations.extend(ac["choices"][0]["delta"]["annotations"]) - response["choices"][0]["message"]["annotations"] = all_annotations + message_annotations, provider_annotations = partition_annotations(all_annotations) + if message_annotations is not None: + response["choices"][0]["message"]["annotations"] = message_annotations + if provider_annotations: + response.choices[0].message.provider_specific_fields = { + **(response.choices[0].message.provider_specific_fields or {}), + "annotations": list(provider_annotations), + } audio_chunks: Final = [ chunk @@ -1766,7 +1774,10 @@ def stream_chunk_builder( if combined_provider_fields: _choice = response.choices[0] - _choice.message.provider_specific_fields = combined_provider_fields + _choice.message.provider_specific_fields = { + **(_choice.message.provider_specific_fields or {}), + **combined_provider_fields, + } completion_output = get_content_from_model_response(response) diff --git a/litellm/llms/anthropic/messages/adapters/transformation.py b/litellm/llms/anthropic/messages/adapters/transformation.py index 3524c9e73d8..ddda759384f 100644 --- a/litellm/llms/anthropic/messages/adapters/transformation.py +++ b/litellm/llms/anthropic/messages/adapters/transformation.py @@ -123,6 +123,7 @@ from openai.types.chat.chat_completion_chunk import Choice as OpenAIStreamingCho from litellm.litellm_core_utils.prompt_templates.common_utils import ( anthropic_image_source_to_openai_url, + anthropic_inline_document, parse_tool_call_arguments, reasoning_content_from_thinking_blocks, with_prompt_cache_breakpoint, @@ -185,6 +186,8 @@ from litellm.types.llms.openai import ( AllMessageValues, ChatCompletionAssistantMessage, ChatCompletionAssistantToolCall, + ChatCompletionFileObject, + ChatCompletionFileObjectFile, ChatCompletionImageObject, ChatCompletionImageUrlObject, ChatCompletionRedactedThinkingBlock, @@ -498,7 +501,9 @@ class LiteLLMAnthropicMessagesAdapter: for m in ordered_messages: user_message: ChatCompletionUserMessage | None = None tool_message_list: list[ChatCompletionToolMessage] = [] - new_user_content_list: list[ChatCompletionTextObject | ChatCompletionImageObject] = [] + new_user_content_list: list[ + ChatCompletionTextObject | ChatCompletionImageObject | ChatCompletionFileObject + ] = [] if m["role"] == "system": system_message = self._translate_midturn_system_message_to_openai(m, model) if system_message is not None: @@ -531,13 +536,8 @@ class LiteLLMAnthropicMessagesAdapter: self._add_prompt_cache_breakpoint_if_present(content, image_obj) ) elif content.get("type") == "document": - # Convert Anthropic document format (PDF, etc.) to OpenAI format - source = content.get("source", {}) - openai_image_url = self._translate_anthropic_image_to_openai(cast(dict, source)) - - if openai_image_url: - image_url_obj = ChatCompletionImageUrlObject(url=openai_image_url) - doc_obj = ChatCompletionImageObject(type="image_url", image_url=image_url_obj) + doc_obj = self._translate_anthropic_document_to_openai(content) + if doc_obj is not None: self._add_cache_control_if_applicable(content, doc_obj, model) new_user_content_list.append(doc_obj) elif content.get("type") == "tool_result": @@ -1357,7 +1357,7 @@ class LiteLLMAnthropicMessagesAdapter: text: Final = tool_result_text_part(block) return None if text is None else ChatCompletionTextObject(type="text", text=text) case "image" | "document": - return self._tool_result_image_part(block.get("source")) + return self._source_image_url_part(block.get("source")) case "tool_reference": return ChatCompletionToolReferenceObject( type="tool_reference", tool_name=str(block.get("tool_name") or "") @@ -1365,7 +1365,18 @@ class LiteLLMAnthropicMessagesAdapter: case _: return None - def _tool_result_image_part(self, image_source: object) -> ChatCompletionImageObject | None: + def _translate_anthropic_document_to_openai( + self, document: Mapping[str, object] + ) -> ChatCompletionFileObject | ChatCompletionImageObject | None: + inline_document: Final = anthropic_inline_document(document) + if inline_document is None: + return self._source_image_url_part(document.get("source")) + return ChatCompletionFileObject( + type="file", + file=ChatCompletionFileObjectFile(filename=inline_document.filename, file_data=inline_document.file_data), + ) + + def _source_image_url_part(self, image_source: object) -> ChatCompletionImageObject | None: if not isinstance(image_source, dict): return None openai_image_url = self._translate_anthropic_image_to_openai(image_source) diff --git a/litellm/llms/anthropic/messages/responses_adapters/transformation.py b/litellm/llms/anthropic/messages/responses_adapters/transformation.py index 1dd07c1739b..94c608f90e6 100644 --- a/litellm/llms/anthropic/messages/responses_adapters/transformation.py +++ b/litellm/llms/anthropic/messages/responses_adapters/transformation.py @@ -13,6 +13,7 @@ from typing import Any, Final, cast from litellm.litellm_core_utils.prompt_templates.common_utils import ( TOOL_RESULT_IMAGE_BOUNDARY, TOOL_RESULT_IMAGE_PLACEHOLDER, + anthropic_inline_document, encrypted_reasoning_signature, responses_reasoning_items_from_thinking_blocks, with_prompt_cache_breakpoint, @@ -106,27 +107,18 @@ class LiteLLMAnthropicToResponsesAPIAdapter: block: Mapping[str, object], ) -> dict[str, str] | None: # mutable-ok: API message payload """Convert an Anthropic document block to a Responses input_file part.""" + inline_document: Final = anthropic_inline_document(block) + if inline_document is not None: + return { + "type": "input_file", + "filename": inline_document.filename, + "file_data": inline_document.file_data, + } raw_source: Final = block.get("source") if not isinstance(raw_source, Mapping): return None source: Final = cast(Mapping[str, object], raw_source) # cast-ok: untrusted client payload - source_type: Final = source.get("type") - if source_type == "base64": - data: Final = source.get("data") - if not isinstance(data, str) or not data: - return None - raw_media_type: Final = source.get("media_type") - media_type: Final = ( - raw_media_type if isinstance(raw_media_type, str) and raw_media_type else "application/pdf" - ) - raw_title: Final = block.get("title") - filename: Final = raw_title if isinstance(raw_title, str) and raw_title else "document.pdf" - return { - "type": "input_file", - "filename": filename, - "file_data": f"data:{media_type};base64,{data}", - } - if source_type == "url": + if source.get("type") == "url": url: Final = source.get("url") if not isinstance(url, str) or not url: return None diff --git a/litellm/llms/azure_ai/chat/transformation.py b/litellm/llms/azure_ai/chat/transformation.py index 8bb6072d0a9..d382d653a53 100644 --- a/litellm/llms/azure_ai/chat/transformation.py +++ b/litellm/llms/azure_ai/chat/transformation.py @@ -1,7 +1,7 @@ import copy import enum import re -from collections.abc import AsyncIterator, Collection, Iterator +from collections.abc import AsyncIterator, Collection, Iterator, Mapping from typing import TYPE_CHECKING, Final, cast import httpx @@ -10,10 +10,11 @@ from httpx import Response import litellm from litellm._logging import verbose_logger from litellm.litellm_core_utils.prompt_templates.common_utils import ( - audio_or_image_in_message_content, + content_parts, convert_content_list_to_str, drop_tool_reference_parts_from_tool_messages, filter_value_from_dict, + is_pdf_file_part, ) from litellm.llms.azure.common_utils import BaseAzureLLM from litellm.llms.azure_ai.common_utils import ( @@ -46,6 +47,20 @@ NON_OPENAI_SPEC_MESSAGE_FIELDS: Final = ( "cache_control", ) +LIST_ONLY_CONTENT_PART_TYPES: Final = frozenset({"image_url", "input_audio"}) + + +def _is_list_only_content_part(part: object) -> bool: + match part: + case {"type": str() as part_type} if part_type in LIST_ONLY_CONTENT_PART_TYPES: + return True + case _: + return is_pdf_file_part(part) + + +def _has_list_only_content_part(message: Mapping[str, object]) -> bool: + return any(_is_list_only_content_part(part) for part in content_parts(message)) + class AzureAIChatCompletionResponseIterator(OpenAIChatCompletionResponseIterator): def chunk_parser(self, chunk: dict[str, object]) -> ModelResponseStream: # mutable-ok: mirrors override contract @@ -246,7 +261,7 @@ class AzureAIStudioConfig(OpenAIConfig): these with "Extra inputs are not permitted", which breaks multi-turn Anthropic-format clients that echo thinking blocks back as history. 2. Transforms list content to a string. - 3. If message contains an image or audio, send as is (user-intended) + 3. If message contains an image, audio, or PDF file part, send as is (user-intended) Operates on a deep copy so the caller's messages keep their thinking blocks and provider metadata, which a fallback to another provider still needs. @@ -257,8 +272,7 @@ class AzureAIStudioConfig(OpenAIConfig): for field in NON_OPENAI_SPEC_MESSAGE_FIELDS: filter_value_from_dict(message_dict, field) - # Do nothing if the message contains an image or audio - if audio_or_image_in_message_content(message): + if _has_list_only_content_part(message): continue texts = convert_content_list_to_str(message=message) diff --git a/litellm/llms/gigachat/chat/transformation.py b/litellm/llms/gigachat/chat/transformation.py index 65921b628fb..81a186eadf4 100644 --- a/litellm/llms/gigachat/chat/transformation.py +++ b/litellm/llms/gigachat/chat/transformation.py @@ -38,6 +38,15 @@ else: _EMPTY_FUNCTION: Final[Mapping[str, object]] = MappingProxyType({}) _JSON_VALUE: Final = TypeAdapter(object) +_CONTENT_PART: Final = TypeAdapter(Mapping[str, object]) + + +def _file_part_data(part: Mapping[str, object]) -> str | None: + match part: + case {"type": "file", "file": {"file_data": str() as file_data}} if file_data: + return file_data + case _: + return None def is_valid_json(value: str) -> bool: @@ -295,7 +304,7 @@ class GigaChatConfig(BaseConfig): def _transform_list_content(self, content: Sequence) -> tuple[str, Sequence[object]]: """ - Extract text and image attachments from a multimodal message content list. + Extract text, image and file attachments from a multimodal message content list. Args: content: List of content parts (OpenAI multimodal format) @@ -320,6 +329,10 @@ class GigaChatConfig(BaseConfig): file_id = self._upload_image(upload_url) if file_id: attachments.append(file_id) + elif (file_data := _file_part_data(_CONTENT_PART.validate_python(part))) and ( + uploaded_file_id := self._upload_image(file_data) + ): + attachments.append(uploaded_file_id) text: Final = "\n".join(texts) if texts else "" return text, attachments diff --git a/litellm/llms/oci/chat/transformation.py b/litellm/llms/oci/chat/transformation.py index e8d37f857e6..8257a75b4a5 100644 --- a/litellm/llms/oci/chat/transformation.py +++ b/litellm/llms/oci/chat/transformation.py @@ -20,6 +20,7 @@ import litellm from litellm._logging import verbose_logger from litellm.constants import DEFAULT_OCI_CHAT_MAX_TOKENS from litellm.litellm_core_utils.logging_utils import track_llm_api_timing +from litellm.litellm_core_utils.prompt_templates.common_utils import messages_with_file_parts_as_image_urls from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper from litellm.litellm_core_utils.url_utils import join_api_base from litellm.llms.base_llm.chat.transformation import BaseConfig, BaseLLMException @@ -681,7 +682,9 @@ class OCIChatConfig(BaseConfig): servingMode=serving_mode, chatRequest=OCIChatRequestPayload( apiFormat=vendor.value, - messages=adapt_messages_to_generic_oci_standard(messages), + messages=adapt_messages_to_generic_oci_standard( + list(messages_with_file_parts_as_image_urls(messages)) + ), **self._get_optional_params(vendor, optional_params, model), ), ) diff --git a/litellm/llms/sap/chat/transformation.py b/litellm/llms/sap/chat/transformation.py index 622c28f3c90..0270f814328 100755 --- a/litellm/llms/sap/chat/transformation.py +++ b/litellm/llms/sap/chat/transformation.py @@ -11,6 +11,7 @@ import httpx from pydantic import BaseModel import litellm +from litellm.litellm_core_utils.prompt_templates.common_utils import message_with_file_parts_as_image_urls from litellm.litellm_core_utils.url_utils import join_api_base from litellm.types.llms.openai import AllMessageValues from litellm.types.utils import ModelResponse @@ -104,7 +105,7 @@ def _message_role_mapping(role: str) -> type[ChatMessage]: def _messages_to_sap_template(messages: list[AllMessageValues]) -> list: template: Final = [] for message in messages: - folded_message = _fold_message_cache_control(message) + folded_message = _fold_message_cache_control(message_with_file_parts_as_image_urls(message)) message_role_class = _message_role_mapping(str(folded_message["role"])) validated_message = validate_dict(folded_message, message_role_class) template.append(validated_message) diff --git a/litellm/llms/snowflake/chat/transformation.py b/litellm/llms/snowflake/chat/transformation.py index 030413030eb..928dcc3bc52 100644 --- a/litellm/llms/snowflake/chat/transformation.py +++ b/litellm/llms/snowflake/chat/transformation.py @@ -112,7 +112,7 @@ def _convert_image_url_to_anthropic(block: Mapping[str, object]) -> object: pdf/text data URIs become document blocks rather than malformed image blocks. """ image_url: Final = block.get("image_url") - url: Final = image_url if isinstance(image_url, str) else _image_url_field(image_url, "url") + url: Final = image_url if isinstance(image_url, str) else _str_field(image_url, "url") if not url: return block @@ -121,18 +121,31 @@ def _convert_image_url_to_anthropic(block: Mapping[str, object]) -> object: if select_anthropic_content_block_type_for_file(_data_uri_media_type(url)) == "document" else create_anthropic_image_param( image_url if isinstance(image_url, dict) else url, - format=_image_url_field(image_url, "format"), + format=_str_field(image_url, "format"), is_bedrock_invoke=True, ) ) + return _with_cache_control(converted, block) + + +def _convert_file_to_anthropic(block: Mapping[str, object]) -> object: + file_data: Final = _str_field(block.get("file"), "file_data") + if not file_data: + return block + return _with_cache_control( + anthropic_process_openai_file_message({"type": "file", "file": {"file_data": file_data}}), block + ) + + +def _with_cache_control(converted: Mapping[str, object], block: Mapping[str, object]) -> Mapping[str, object]: cache_control: Final = block.get("cache_control") if cache_control is None: return converted return {**converted, "cache_control": cache_control} -def _image_url_field(image_url: object, key: str) -> str | None: - value: Final = image_url.get(key) if isinstance(image_url, dict) else None +def _str_field(container: object, key: str) -> str | None: + value: Final = container.get(key) if isinstance(container, dict) else None return value if isinstance(value, str) else None @@ -141,15 +154,23 @@ def _data_uri_media_type(url: str) -> str: return match.group(1) if match else "" -def _convert_image_url_blocks_to_anthropic(content: object) -> object: +def _convert_media_block_to_anthropic(block: object) -> object: + mapping: Final = as_str_mapping(block) + if mapping is None: + return block + match mapping.get("type"): + case "image_url": + return _convert_image_url_to_anthropic(mapping) + case "file": + return _convert_file_to_anthropic(mapping) + case _: + return block + + +def _convert_media_blocks_to_anthropic(content: object) -> object: if not isinstance(content, list): return content - return [ - _convert_image_url_to_anthropic(block) - if isinstance(block, Mapping) and block.get("type") == "image_url" - else block - for block in content - ] + return [_convert_media_block_to_anthropic(block) for block in content] def _convert_tool_result_to_anthropic( @@ -391,7 +412,7 @@ class SnowflakeConfig(SnowflakeBaseConfig, OpenAIGPTConfig): conversation.append( { "role": role, - "content": _convert_image_url_blocks_to_anthropic(content), + "content": _convert_media_blocks_to_anthropic(content), } ) diff --git a/litellm/llms/vertex_ai/vertex_ai_partner_models/main.py b/litellm/llms/vertex_ai/vertex_ai_partner_models/main.py index bb200fd542d..a3a3427670d 100644 --- a/litellm/llms/vertex_ai/vertex_ai_partner_models/main.py +++ b/litellm/llms/vertex_ai/vertex_ai_partner_models/main.py @@ -9,6 +9,7 @@ import httpx import litellm from litellm import LlmProviders +from litellm.litellm_core_utils.prompt_templates.common_utils import messages_with_file_parts_as_image_urls from litellm.types.llms.vertex_ai import VertexPartnerProvider from litellm.utils import ModelResponse @@ -217,7 +218,7 @@ class VertexAIPartnerModels(VertexBase): return base_llm_http_handler.completion( model=model, stream=stream, - messages=messages, + messages=list(messages_with_file_parts_as_image_urls(messages)), acompletion=acompletion, api_base=api_base, model_response=model_response, diff --git a/litellm/main.py b/litellm/main.py index d1af2de3a13..50ab274a060 100644 --- a/litellm/main.py +++ b/litellm/main.py @@ -108,6 +108,10 @@ from litellm.litellm_core_utils.mock_functions import ( from litellm.litellm_core_utils.prompt_templates.common_utils import ( get_content_from_model_response as get_content_from_model_response, # noqa: PLC0414 # public re-export ) +from litellm.litellm_core_utils.prompt_templates.common_utils import ( + messages_have_pdf_file_part, + messages_without_pdf_file_parts, +) from litellm.litellm_core_utils.provider_affinity import add_provider_affinity_header from litellm.litellm_core_utils.request_timeout_resolver import ( get_configured_request_timeout, @@ -191,11 +195,13 @@ from litellm.utils import ( get_requester_metadata, get_secret, get_standard_openai_params, + is_explicitly_disabled_factory, mock_completion_streaming_obj, pre_process_non_default_params, read_config_args, should_run_mock_completion, supports_httpx_timeout, + supports_pdf_input, validate_and_fix_openai_messages, validate_and_fix_openai_tools, validate_and_fix_thinking_param, @@ -393,6 +399,125 @@ def _resolve_control_options(kwargs: Mapping[str, object], model: str) -> Contro return assert_never(control) +PROVIDERS_WITHOUT_DOCUMENT_INPUT: Final = frozenset({"ollama", "ollama_chat"}) +PROVIDERS_DROPPING_UNDECLARED_PDF_INPUT: Final = frozenset({"azure_ai"}) + + +def _cost_map_pdf_input_support(model: str, custom_llm_provider: str) -> bool | None: + if custom_llm_provider in PROVIDERS_WITHOUT_DOCUMENT_INPUT or is_explicitly_disabled_factory( + model, custom_llm_provider, "supports_pdf_input" + ): + return False + if supports_pdf_input(model, custom_llm_provider): + return True + return None + + +def _deployment_pdf_input_support(model: str, base_model: str, custom_llm_provider: str) -> bool | None: + supports: Final = ( + _cost_map_pdf_input_support(model, custom_llm_provider), + _cost_map_pdf_input_support(base_model, custom_llm_provider), + ) + if False in supports: + return False + if True in supports: + return True + return None + + +def _declared_pdf_input_support( + model: str, custom_llm_provider: str, model_info: Mapping[str, object] | None, base_model: str | None +) -> bool | None: + match model_info: + case {"supports_pdf_input": bool() as declared}: + return declared + case _ if base_model: + return _deployment_pdf_input_support(model, base_model, custom_llm_provider) + case _: + return _cost_map_pdf_input_support(model, custom_llm_provider) + + +def _pdf_input_unreadable( + model: str, + custom_llm_provider: str, + model_info: Mapping[str, object] | None, + base_model: str | None, + drop_params: bool, +) -> bool: + match _declared_pdf_input_support(model, custom_llm_provider, model_info, base_model): + case False: + return True + case None: + return drop_params and custom_llm_provider in PROVIDERS_DROPPING_UNDECLARED_PDF_INPUT + case True: + return False + + +def _messages_without_unreadable_pdfs( + messages: Sequence[AllMessageValues], + model: str, + custom_llm_provider: str, + model_info: Mapping[str, object] | None, + base_model: str | None, + drop_params: bool, +) -> tuple[AllMessageValues, ...] | None: + if not messages_have_pdf_file_part(messages) or not _pdf_input_unreadable( + model, custom_llm_provider, model_info, base_model, drop_params + ): + return None + if not drop_params: + raise litellm.BadRequestError( + message=( + f"{custom_llm_provider}/{model} does not accept PDF input. " + "Remove the PDF, or set drop_params=True to send the request without it" + ), + model=model, + llm_provider=custom_llm_provider, + ) + readable_messages: Final = messages_without_pdf_file_parts(messages) + if not any(message["role"] == "user" for message in readable_messages): + raise litellm.BadRequestError( + message=( + f"{custom_llm_provider}/{model} does not accept PDF input, " + "and no user message is left once the PDF is dropped" + ), + model=model, + llm_provider=custom_llm_provider, + ) + return readable_messages + + +def _messages_readable_by_model( + messages: Sequence[AllMessageValues], + model: str, + custom_llm_provider: str, + model_info: Mapping[str, object] | None, + base_model: str | None, + drop_params: bool, + ensure_alternating_roles: bool | None, + user_continue_message: ChatCompletionUserMessage | None, + assistant_continue_message: ChatCompletionAssistantMessage | None, +) -> tuple[AllMessageValues, ...] | None: + pdf_free_messages: Final = _messages_without_unreadable_pdfs( + messages=messages, + model=model, + custom_llm_provider=custom_llm_provider, + model_info=model_info, + base_model=base_model, + drop_params=drop_params, + ) + if pdf_free_messages is None: + return None + return tuple( + get_completion_messages( + messages=list(pdf_free_messages), + ensure_alternating_roles=ensure_alternating_roles or False, + user_continue_message=user_continue_message, + assistant_continue_message=assistant_continue_message, + ) + ) + + ####### COMPLETION ENDPOINTS ################ @@ -4234,7 +4359,6 @@ def _complete_sagemaker(ctx: CompletionDispatchContext) -> _CompletionDispatchRe _ADDITIONAL_DROP_PARAMS_ADAPTER: Final = TypeAdapter(list[str]) _OPTIONAL_PARAMS_ADAPTER: Final = TypeAdapter(dict[str, object]) _OLLAMA_EMBEDDING_HEADERS_ADAPTER: Final = TypeAdapter(dict[str, str]) -_REQUEST_KWARGS_ADAPTER: Final = TypeAdapter(Mapping[str, object]) _MODEL_INFO_ADAPTER: Final = TypeAdapter(Mapping[str, object]) _PROVIDER_ROUTING_HINTS: Final = TypeAdapter(ProviderRoutingHints) _BASE_URL_HINT: Final = TypeAdapter(BaseUrlRoutingHint) @@ -5558,9 +5682,15 @@ def completion( ) headers = kwargs.get("headers", None) or extra_headers - ensure_alternating_roles: Final[bool | None] = kwargs.get("ensure_alternating_roles", None) - user_continue_message: Final[ChatCompletionUserMessage | None] = kwargs.get("user_continue_message", None) - assistant_continue_message: ChatCompletionAssistantMessage | None = kwargs.get("assistant_continue_message", None) + ensure_alternating_roles: Final = cast( # cast-ok: read from untyped kwargs + bool | None, kwargs.get("ensure_alternating_roles", None) + ) + user_continue_message: Final = cast( # cast-ok: read from untyped kwargs + ChatCompletionUserMessage | None, kwargs.get("user_continue_message", None) + ) + assistant_continue_message: Final = cast( # cast-ok: read from untyped kwargs + ChatCompletionAssistantMessage | None, kwargs.get("assistant_continue_message", None) + ) headers = {} if headers is None else dict(headers) if extra_headers is not None: headers.update(extra_headers) @@ -5928,15 +6058,17 @@ def completion( def optional_params_for_model_info(deployment_info: Mapping[str, object] | None) -> dict[str, object]: return get_optional_params(**mapping_args, model_info=deployment_info, **mapping_special_params) - optional_params = optional_params_for_model_info( + validated_model_info: Final = ( _MODEL_INFO_ADAPTER.validate_python(model_info) if isinstance(model_info, dict) else None ) + deployment_base_model: Final = base_model if isinstance(base_model, str) else None + optional_params = optional_params_for_model_info(validated_model_info) fakes_stream: Final = partial( _completion_fakes_stream, model, custom_llm_provider, provider_config, - base_model=base_model if isinstance(base_model, str) else None, + base_model=deployment_base_model, bridges_to_responses_api=bridges_to_responses_api is True, ) resolve_timeout: Final = partial( @@ -5974,16 +6106,29 @@ def completion( functions_unsupported_model: Final[list[object]] = TypeAdapter(list[object]).validate_python( optional_params.pop("functions_unsupported_model") ) - request_kwargs: Final = _REQUEST_KWARGS_ADAPTER.validate_python(kwargs) messages = function_call_prompt( messages=messages, functions=functions_unsupported_model, model=model, llm_provider=custom_llm_provider, - drop_params=litellm.drop_params is True - or normalize_drop_params(request_kwargs.get("drop_params")) is True, + drop_params=drop_unsupported_reasoning, ) + readable_messages: Final = _messages_readable_by_model( + messages=cast(list[AllMessageValues], messages), # cast-ok: completion types messages as a bare List + model=model, + custom_llm_provider=custom_llm_provider, + model_info=validated_model_info, + base_model=deployment_base_model, + drop_params=drop_unsupported_reasoning, + ensure_alternating_roles=ensure_alternating_roles, + user_continue_message=user_continue_message, + assistant_continue_message=assistant_continue_message, + ) + messages = ( # rebind-ok: completion() rewrites messages step by step + messages if readable_messages is None else list(readable_messages) + ) + # For logging - save the values of the litellm-specific params passed in requested_litellm_params: Final = get_litellm_params( acompletion=acompletion, diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 1256fef9181..c6bbb63b6fc 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -4049,6 +4049,7 @@ "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_pdf_input": false, "supports_response_schema": true, "supports_tool_choice": true }, @@ -12437,7 +12438,8 @@ "supported_output_modalities": [ "text" ], - "supports_native_streaming": true + "supports_native_streaming": true, + "supports_pdf_input": false }, "azure_ai/FLUX-1.1-pro": { "deprecation_date": "2027-01-05", @@ -12502,6 +12504,7 @@ "output_cost_per_token": 1.85e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -12516,6 +12519,7 @@ "output_cost_per_token": 3.828e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -12531,6 +12535,7 @@ "output_cost_per_token": 3.52e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -12546,6 +12551,7 @@ "output_cost_per_token": 4.84e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -12560,6 +12566,7 @@ "output_cost_per_token": 4.84e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -12574,6 +12581,7 @@ "output_cost_per_token": 7.26e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -12594,6 +12602,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -12616,6 +12625,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true, @@ -12638,6 +12648,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true, @@ -12660,6 +12671,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true, @@ -12687,6 +12699,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true, @@ -12703,6 +12716,7 @@ "output_cost_per_token": 1.32e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -12724,6 +12738,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true, @@ -12744,6 +12759,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true, @@ -12765,6 +12781,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -12833,6 +12850,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -12847,6 +12865,7 @@ "output_cost_per_token": 7.1e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/Llama-4-Maverick-17B-128E-Instruct-FP8": { @@ -12859,6 +12878,7 @@ "output_cost_per_token": 1e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": true }, @@ -12872,6 +12892,7 @@ "output_cost_per_token": 7.8e-07, "source": "https://azure.microsoft.com/en-us/blog/introducing-the-llama-4-herd-in-azure-ai-foundry-and-azure-databricks/", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": true }, @@ -12884,6 +12905,7 @@ "mode": "chat", "output_cost_per_token": 3.7e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/Meta-Llama-3.1-70B-Instruct": { @@ -12895,6 +12917,7 @@ "mode": "chat", "output_cost_per_token": 3.54e-06, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/Phi-3-medium-128k-instruct": { @@ -12906,6 +12929,7 @@ "mode": "chat", "output_cost_per_token": 6.8e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": false }, @@ -12918,6 +12942,7 @@ "mode": "chat", "output_cost_per_token": 6.8e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": false }, @@ -12930,6 +12955,7 @@ "mode": "chat", "output_cost_per_token": 5.2e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": false }, @@ -12942,6 +12968,7 @@ "mode": "chat", "output_cost_per_token": 5.2e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": false }, @@ -12954,6 +12981,7 @@ "mode": "chat", "output_cost_per_token": 6e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": false }, @@ -12966,6 +12994,7 @@ "mode": "chat", "output_cost_per_token": 6e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": false }, @@ -12978,6 +13007,7 @@ "mode": "chat", "output_cost_per_token": 6.4e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": false }, @@ -12990,6 +13020,7 @@ "mode": "chat", "output_cost_per_token": 5.2e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": false }, @@ -13002,6 +13033,7 @@ "mode": "chat", "output_cost_per_token": 5.2e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": true }, @@ -13015,6 +13047,7 @@ "output_cost_per_token": 5e-07, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": false }, @@ -13027,7 +13060,8 @@ "mode": "chat", "output_cost_per_token": 3e-07, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", - "supports_function_calling": true + "supports_function_calling": true, + "supports_pdf_input": false }, "azure_ai/Phi-4-multimodal-instruct": { "deprecation_date": "2027-01-17", @@ -13041,6 +13075,7 @@ "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_audio_input": true, "supports_function_calling": true, + "supports_pdf_input": false, "supports_vision": true }, "azure_ai/Phi-4-mini-reasoning": { @@ -13052,7 +13087,8 @@ "mode": "chat", "output_cost_per_token": 3e-07, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", - "supports_function_calling": true + "supports_function_calling": true, + "supports_pdf_input": false }, "azure_ai/Phi-4-reasoning": { "deprecation_date": "2027-01-17", @@ -13064,6 +13100,7 @@ "output_cost_per_token": 5e-07, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true, "supports_reasoning": true }, @@ -13106,6 +13143,7 @@ "output_cost_per_token": 1e-05, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/doc-intelligence/prebuilt-read": { @@ -13144,6 +13182,7 @@ "mode": "chat", "output_cost_per_token": 5.4e-06, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_reasoning": true, "supports_tool_choice": true }, @@ -13202,6 +13241,7 @@ "source": "https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure", "supports_assistant_prefill": true, "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -13217,6 +13257,7 @@ "source": "https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure", "supports_assistant_prefill": true, "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -13230,6 +13271,7 @@ "mode": "chat", "output_cost_per_token": 4.56e-06, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/deepseek-v4-pro": { @@ -13242,6 +13284,7 @@ "output_cost_per_token": 3.48e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_reasoning": true, "supports_tool_choice": true, "cache_read_input_token_cost": 1.45e-07, @@ -13257,6 +13300,7 @@ "output_cost_per_token": 5.1e-07, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_reasoning": true, "supports_tool_choice": true, "cache_read_input_token_cost": 2.8e-08, @@ -13273,6 +13317,7 @@ "output_cost_per_token": 1.32e-06, "source": "https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -13306,6 +13351,7 @@ "output_cost_per_token": 1.5e-05, "source": "https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure", "supports_function_calling": true, + "supports_pdf_input": false, "supports_response_schema": true, "supports_tool_choice": true, "supports_web_search": true @@ -13323,6 +13369,7 @@ "output_cost_per_token_above_200k_tokens": 5e-06, "source": "https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, @@ -13343,6 +13390,7 @@ "output_cost_per_token_above_200k_tokens": 1.2e-05, "source": "https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/grok/", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, @@ -13361,6 +13409,7 @@ "output_cost_per_token": 2.5e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true, @@ -13379,6 +13428,7 @@ "output_cost_per_token": 2.5e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true, @@ -13395,6 +13445,7 @@ "mode": "chat", "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_response_schema": true, "supports_tool_choice": true, "supports_web_search": true @@ -13409,6 +13460,7 @@ "mode": "chat", "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true, @@ -13424,6 +13476,7 @@ "output_cost_per_token": 1.5e-06, "source": "https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure", "supports_function_calling": true, + "supports_pdf_input": false, "supports_response_schema": true, "supports_tool_choice": true, "supports_web_search": true @@ -13436,7 +13489,8 @@ "max_output_tokens": 8192, "mode": "chat", "output_cost_per_token": 0.00971, - "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models" + "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false }, "azure_ai/jamba-instruct": { "deprecation_date": "2025-03-01", @@ -13447,6 +13501,7 @@ "mode": "chat", "output_cost_per_token": 7e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/kimi-k2.5": { @@ -13459,6 +13514,7 @@ "output_cost_per_token": 3e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true, "supports_video_input": true, "supports_vision": true, @@ -13482,6 +13538,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_reasoning": true, "supports_tool_choice": true, "supports_vision": true, @@ -13498,6 +13555,7 @@ "output_cost_per_token": 4e-08, "source": "https://marketplace.microsoft.com/en/marketplace/apps/000-000.ministral-3b-2410-offer?tab=Overview", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/mistral-large": { @@ -13510,6 +13568,7 @@ "output_cost_per_token": 1.2e-05, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/mistral-large-2407": { @@ -13522,6 +13581,7 @@ "output_cost_per_token": 6e-06, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/mistral-large-latest": { @@ -13533,6 +13593,7 @@ "output_cost_per_token": 6e-06, "source": "https://marketplace.microsoft.com/en/marketplace/apps/000-000.mistral-ai-large-2407-offer?tab=Overview", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/mistral-large-3": { @@ -13544,6 +13605,7 @@ "output_cost_per_token": 1.5e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": true }, @@ -13557,6 +13619,7 @@ "output_cost_per_token": 2e-06, "supports_assistant_prefill": true, "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/mistral-nemo": { @@ -13568,7 +13631,8 @@ "mode": "chat", "output_cost_per_token": 1.5e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", - "supports_function_calling": true + "supports_function_calling": true, + "supports_pdf_input": false }, "azure_ai/mistral-small": { "deprecation_date": "2025-07-31", @@ -13580,6 +13644,7 @@ "output_cost_per_token": 3e-06, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/mistral-small-2503": { @@ -13591,6 +13656,7 @@ "mode": "chat", "output_cost_per_token": 3e-07, "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": true }, @@ -52006,6 +52072,7 @@ ], "supports_assistant_prefill": false, "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -52026,6 +52093,7 @@ ], "supports_assistant_prefill": false, "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -52777,6 +52845,7 @@ "output_cost_per_token": 2.5e-06, "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true, "supports_web_search": true }, @@ -52794,6 +52863,7 @@ "global" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_reasoning": true, "supports_tool_choice": true }, @@ -52811,6 +52881,7 @@ "global" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -52989,6 +53060,7 @@ "output_cost_per_token": 3.6e-07, "output_cost_per_token_batches": 1.8e-07, "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", + "supports_pdf_input": false, "supports_reasoning": true }, "vertex_ai/openai/gpt-oss-20b-maas": { @@ -53000,6 +53072,7 @@ "mode": "chat", "output_cost_per_token": 2.5e-07, "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", + "supports_pdf_input": false, "supports_reasoning": true, "cache_read_input_token_cost": 7e-09, "input_cost_per_token_batches": 3.5e-08, @@ -53131,6 +53204,7 @@ "us-south1" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "vertex_ai/qwen/qwen3-coder-480b-a35b-instruct-maas": { @@ -53149,6 +53223,7 @@ "global" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "vertex_ai/qwen/qwen3-next-80b-a3b-instruct-maas": { @@ -53164,6 +53239,7 @@ "global" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "vertex_ai/qwen/qwen3-next-80b-a3b-thinking-maas": { @@ -53179,6 +53255,7 @@ "global" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "vertex_ai/veo-3.1-generate-preview": { @@ -67322,6 +67399,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true, @@ -67336,6 +67414,7 @@ "mode": "chat", "output_cost_per_token": 1.2e-06, "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true, @@ -67378,6 +67457,7 @@ "output_cost_per_token": 1.5e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -67392,6 +67472,7 @@ "output_cost_per_token": 3.1e-07, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -67406,6 +67487,7 @@ "output_cost_per_token": 5.5e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -67420,6 +67502,7 @@ "output_cost_per_token": 6.25e-07, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -67436,6 +67519,7 @@ "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, @@ -67451,6 +67535,7 @@ "output_cost_per_token": 3.2e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_reasoning": true, "supports_tool_choice": true }, @@ -67469,6 +67554,7 @@ "supported_output_modalities": [ "text" ], + "supports_pdf_input": false, "supports_response_schema": true, "supports_vision": true }, @@ -67478,7 +67564,8 @@ "output_cost_per_token": 5.4e-06, "litellm_provider": "azure_ai", "mode": "chat", - "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'" + "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", + "supports_pdf_input": false }, "azure_ai/deepseek-v3-0324": { "deprecation_date": "2026-07-13", @@ -67486,7 +67573,8 @@ "output_cost_per_token": 4.56e-06, "litellm_provider": "azure_ai", "mode": "chat", - "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'" + "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", + "supports_pdf_input": false }, "azure_ai/deepseek-v3.1": { "deprecation_date": "2026-07-13", @@ -67494,7 +67582,8 @@ "output_cost_per_token": 4.94e-06, "litellm_provider": "azure_ai", "mode": "chat", - "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'" + "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", + "supports_pdf_input": false }, "azure_ai/grok-3": { "deprecation_date": "2026-05-01", @@ -67502,7 +67591,8 @@ "output_cost_per_token": 1.5e-05, "litellm_provider": "azure_ai", "mode": "chat", - "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'" + "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", + "supports_pdf_input": false }, "azure_ai/grok-3-mini": { "deprecation_date": "2026-05-01", @@ -67510,7 +67600,8 @@ "output_cost_per_token": 1.27e-06, "litellm_provider": "azure_ai", "mode": "chat", - "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'" + "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", + "supports_pdf_input": false }, "azure_ai/grok-4-fast-non-reasoning": { "deprecation_date": "2026-05-01", @@ -67518,7 +67609,8 @@ "output_cost_per_token": 5e-07, "litellm_provider": "azure_ai", "mode": "chat", - "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'" + "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", + "supports_pdf_input": false }, "azure_ai/grok-4-fast-reasoning": { "deprecation_date": "2026-05-01", @@ -67526,7 +67618,8 @@ "output_cost_per_token": 5e-07, "litellm_provider": "azure_ai", "mode": "chat", - "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'" + "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", + "supports_pdf_input": false }, "bedrock/us-gov-west-1/nvidia.nemotron-nano-3-30b": { "input_cost_per_token": 7.2e-08, @@ -73196,6 +73289,7 @@ "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_regions": ["global"], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, @@ -73811,6 +73905,7 @@ "output_cost_per_token": 3.5e-06, "source": "https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/microsoft/", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -74865,6 +74960,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": false, "supports_reasoning": true, "supports_tool_choice": true, @@ -75113,6 +75209,7 @@ "output_cost_per_token_above_200k_tokens": 1.2e-05, "source": "https://docs.x.ai/developers/models/grok-4.7", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, diff --git a/litellm/proxy/guardrails/guardrail_hooks/prompt_security/prompt_security.py b/litellm/proxy/guardrails/guardrail_hooks/prompt_security/prompt_security.py index 0f66135d038..8e2fea415ef 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/prompt_security/prompt_security.py +++ b/litellm/proxy/guardrails/guardrail_hooks/prompt_security/prompt_security.py @@ -647,9 +647,10 @@ class PromptSecurityGuardrail(CustomGuardrail): async def _process_document_item(self, item: dict, user_api_key_alias: str | None) -> dict: """Process and sanitize document/file items.""" doc_data: Final = item.get("document") or item.get("file") or item + url_key: Final = "file_data" if isinstance(doc_data, dict) and not doc_data.get("url") else "url" if isinstance(doc_data, dict): - url = doc_data.get("url", "") + url = doc_data.get(url_key) or "" doc_content = doc_data.get("data", "") else: url = doc_data if isinstance(doc_data, str) else "" @@ -698,7 +699,7 @@ class PromptSecurityGuardrail(CustomGuardrail): if url.startswith("data:") and header: sanitized_url: Final = f"{header},{sanitized_encoded}" if isinstance(doc_data, dict): - doc_data["url"] = sanitized_url + doc_data[url_key] = sanitized_url elif isinstance(doc_data, dict): doc_data["data"] = sanitized_encoded diff --git a/litellm/types/utils.py b/litellm/types/utils.py index e9984a35ad1..448bd022989 100644 --- a/litellm/types/utils.py +++ b/litellm/types/utils.py @@ -4147,7 +4147,12 @@ class CustomPricingLiteLLMParams(MirroredPricingParams): DEPLOYMENT_SCOPED_PRICING_FIELDS: Final[frozenset[str]] = frozenset({"off_peak_pricing"}) DEPLOYMENT_SCOPED_CAPABILITY_FIELDS: Final[frozenset[str]] = frozenset( - {"supports_regex_lookaround", "supports_assistant_prefill", "supports_bedrock_native_structured_output"} + { + "supports_regex_lookaround", + "supports_assistant_prefill", + "supports_bedrock_native_structured_output", + "supports_pdf_input", + } ) DEPLOYMENT_SCOPED_TOKEN_COUNT_FIELDS: Final[frozenset[str]] = frozenset({"token_count_framing", "token_count_model"}) diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 1256fef9181..c6bbb63b6fc 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -4049,6 +4049,7 @@ "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_pdf_input": false, "supports_response_schema": true, "supports_tool_choice": true }, @@ -12437,7 +12438,8 @@ "supported_output_modalities": [ "text" ], - "supports_native_streaming": true + "supports_native_streaming": true, + "supports_pdf_input": false }, "azure_ai/FLUX-1.1-pro": { "deprecation_date": "2027-01-05", @@ -12502,6 +12504,7 @@ "output_cost_per_token": 1.85e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -12516,6 +12519,7 @@ "output_cost_per_token": 3.828e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -12531,6 +12535,7 @@ "output_cost_per_token": 3.52e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -12546,6 +12551,7 @@ "output_cost_per_token": 4.84e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -12560,6 +12566,7 @@ "output_cost_per_token": 4.84e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -12574,6 +12581,7 @@ "output_cost_per_token": 7.26e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -12594,6 +12602,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -12616,6 +12625,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true, @@ -12638,6 +12648,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true, @@ -12660,6 +12671,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true, @@ -12687,6 +12699,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true, @@ -12703,6 +12716,7 @@ "output_cost_per_token": 1.32e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -12724,6 +12738,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true, @@ -12744,6 +12759,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true, @@ -12765,6 +12781,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -12833,6 +12850,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -12847,6 +12865,7 @@ "output_cost_per_token": 7.1e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/Llama-4-Maverick-17B-128E-Instruct-FP8": { @@ -12859,6 +12878,7 @@ "output_cost_per_token": 1e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": true }, @@ -12872,6 +12892,7 @@ "output_cost_per_token": 7.8e-07, "source": "https://azure.microsoft.com/en-us/blog/introducing-the-llama-4-herd-in-azure-ai-foundry-and-azure-databricks/", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": true }, @@ -12884,6 +12905,7 @@ "mode": "chat", "output_cost_per_token": 3.7e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/Meta-Llama-3.1-70B-Instruct": { @@ -12895,6 +12917,7 @@ "mode": "chat", "output_cost_per_token": 3.54e-06, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/Phi-3-medium-128k-instruct": { @@ -12906,6 +12929,7 @@ "mode": "chat", "output_cost_per_token": 6.8e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": false }, @@ -12918,6 +12942,7 @@ "mode": "chat", "output_cost_per_token": 6.8e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": false }, @@ -12930,6 +12955,7 @@ "mode": "chat", "output_cost_per_token": 5.2e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": false }, @@ -12942,6 +12968,7 @@ "mode": "chat", "output_cost_per_token": 5.2e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": false }, @@ -12954,6 +12981,7 @@ "mode": "chat", "output_cost_per_token": 6e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": false }, @@ -12966,6 +12994,7 @@ "mode": "chat", "output_cost_per_token": 6e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": false }, @@ -12978,6 +13007,7 @@ "mode": "chat", "output_cost_per_token": 6.4e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": false }, @@ -12990,6 +13020,7 @@ "mode": "chat", "output_cost_per_token": 5.2e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": false }, @@ -13002,6 +13033,7 @@ "mode": "chat", "output_cost_per_token": 5.2e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": true }, @@ -13015,6 +13047,7 @@ "output_cost_per_token": 5e-07, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": false }, @@ -13027,7 +13060,8 @@ "mode": "chat", "output_cost_per_token": 3e-07, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", - "supports_function_calling": true + "supports_function_calling": true, + "supports_pdf_input": false }, "azure_ai/Phi-4-multimodal-instruct": { "deprecation_date": "2027-01-17", @@ -13041,6 +13075,7 @@ "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_audio_input": true, "supports_function_calling": true, + "supports_pdf_input": false, "supports_vision": true }, "azure_ai/Phi-4-mini-reasoning": { @@ -13052,7 +13087,8 @@ "mode": "chat", "output_cost_per_token": 3e-07, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", - "supports_function_calling": true + "supports_function_calling": true, + "supports_pdf_input": false }, "azure_ai/Phi-4-reasoning": { "deprecation_date": "2027-01-17", @@ -13064,6 +13100,7 @@ "output_cost_per_token": 5e-07, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true, "supports_reasoning": true }, @@ -13106,6 +13143,7 @@ "output_cost_per_token": 1e-05, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/doc-intelligence/prebuilt-read": { @@ -13144,6 +13182,7 @@ "mode": "chat", "output_cost_per_token": 5.4e-06, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_reasoning": true, "supports_tool_choice": true }, @@ -13202,6 +13241,7 @@ "source": "https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure", "supports_assistant_prefill": true, "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -13217,6 +13257,7 @@ "source": "https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure", "supports_assistant_prefill": true, "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -13230,6 +13271,7 @@ "mode": "chat", "output_cost_per_token": 4.56e-06, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/deepseek-v4-pro": { @@ -13242,6 +13284,7 @@ "output_cost_per_token": 3.48e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_reasoning": true, "supports_tool_choice": true, "cache_read_input_token_cost": 1.45e-07, @@ -13257,6 +13300,7 @@ "output_cost_per_token": 5.1e-07, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_reasoning": true, "supports_tool_choice": true, "cache_read_input_token_cost": 2.8e-08, @@ -13273,6 +13317,7 @@ "output_cost_per_token": 1.32e-06, "source": "https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -13306,6 +13351,7 @@ "output_cost_per_token": 1.5e-05, "source": "https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure", "supports_function_calling": true, + "supports_pdf_input": false, "supports_response_schema": true, "supports_tool_choice": true, "supports_web_search": true @@ -13323,6 +13369,7 @@ "output_cost_per_token_above_200k_tokens": 5e-06, "source": "https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, @@ -13343,6 +13390,7 @@ "output_cost_per_token_above_200k_tokens": 1.2e-05, "source": "https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/grok/", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, @@ -13361,6 +13409,7 @@ "output_cost_per_token": 2.5e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true, @@ -13379,6 +13428,7 @@ "output_cost_per_token": 2.5e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_response_schema": true, "supports_tool_choice": true, @@ -13395,6 +13445,7 @@ "mode": "chat", "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_response_schema": true, "supports_tool_choice": true, "supports_web_search": true @@ -13409,6 +13460,7 @@ "mode": "chat", "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true, @@ -13424,6 +13476,7 @@ "output_cost_per_token": 1.5e-06, "source": "https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure", "supports_function_calling": true, + "supports_pdf_input": false, "supports_response_schema": true, "supports_tool_choice": true, "supports_web_search": true @@ -13436,7 +13489,8 @@ "max_output_tokens": 8192, "mode": "chat", "output_cost_per_token": 0.00971, - "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models" + "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false }, "azure_ai/jamba-instruct": { "deprecation_date": "2025-03-01", @@ -13447,6 +13501,7 @@ "mode": "chat", "output_cost_per_token": 7e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/kimi-k2.5": { @@ -13459,6 +13514,7 @@ "output_cost_per_token": 3e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true, "supports_video_input": true, "supports_vision": true, @@ -13482,6 +13538,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_reasoning": true, "supports_tool_choice": true, "supports_vision": true, @@ -13498,6 +13555,7 @@ "output_cost_per_token": 4e-08, "source": "https://marketplace.microsoft.com/en/marketplace/apps/000-000.ministral-3b-2410-offer?tab=Overview", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/mistral-large": { @@ -13510,6 +13568,7 @@ "output_cost_per_token": 1.2e-05, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/mistral-large-2407": { @@ -13522,6 +13581,7 @@ "output_cost_per_token": 6e-06, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/mistral-large-latest": { @@ -13533,6 +13593,7 @@ "output_cost_per_token": 6e-06, "source": "https://marketplace.microsoft.com/en/marketplace/apps/000-000.mistral-ai-large-2407-offer?tab=Overview", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/mistral-large-3": { @@ -13544,6 +13605,7 @@ "output_cost_per_token": 1.5e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": true }, @@ -13557,6 +13619,7 @@ "output_cost_per_token": 2e-06, "supports_assistant_prefill": true, "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/mistral-nemo": { @@ -13568,7 +13631,8 @@ "mode": "chat", "output_cost_per_token": 1.5e-07, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", - "supports_function_calling": true + "supports_function_calling": true, + "supports_pdf_input": false }, "azure_ai/mistral-small": { "deprecation_date": "2025-07-31", @@ -13580,6 +13644,7 @@ "output_cost_per_token": 3e-06, "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/retired-models", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "azure_ai/mistral-small-2503": { @@ -13591,6 +13656,7 @@ "mode": "chat", "output_cost_per_token": 3e-07, "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true, "supports_vision": true }, @@ -52006,6 +52072,7 @@ ], "supports_assistant_prefill": false, "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -52026,6 +52093,7 @@ ], "supports_assistant_prefill": false, "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -52777,6 +52845,7 @@ "output_cost_per_token": 2.5e-06, "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true, "supports_web_search": true }, @@ -52794,6 +52863,7 @@ "global" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_reasoning": true, "supports_tool_choice": true }, @@ -52811,6 +52881,7 @@ "global" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -52989,6 +53060,7 @@ "output_cost_per_token": 3.6e-07, "output_cost_per_token_batches": 1.8e-07, "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", + "supports_pdf_input": false, "supports_reasoning": true }, "vertex_ai/openai/gpt-oss-20b-maas": { @@ -53000,6 +53072,7 @@ "mode": "chat", "output_cost_per_token": 2.5e-07, "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", + "supports_pdf_input": false, "supports_reasoning": true, "cache_read_input_token_cost": 7e-09, "input_cost_per_token_batches": 3.5e-08, @@ -53131,6 +53204,7 @@ "us-south1" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "vertex_ai/qwen/qwen3-coder-480b-a35b-instruct-maas": { @@ -53149,6 +53223,7 @@ "global" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "vertex_ai/qwen/qwen3-next-80b-a3b-instruct-maas": { @@ -53164,6 +53239,7 @@ "global" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "vertex_ai/qwen/qwen3-next-80b-a3b-thinking-maas": { @@ -53179,6 +53255,7 @@ "global" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_tool_choice": true }, "vertex_ai/veo-3.1-generate-preview": { @@ -67322,6 +67399,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true, @@ -67336,6 +67414,7 @@ "mode": "chat", "output_cost_per_token": 1.2e-06, "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true, @@ -67378,6 +67457,7 @@ "output_cost_per_token": 1.5e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -67392,6 +67472,7 @@ "output_cost_per_token": 3.1e-07, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -67406,6 +67487,7 @@ "output_cost_per_token": 5.5e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -67420,6 +67502,7 @@ "output_cost_per_token": 6.25e-07, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -67436,6 +67519,7 @@ "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, "supports_parallel_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, @@ -67451,6 +67535,7 @@ "output_cost_per_token": 3.2e-06, "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", "supports_function_calling": true, + "supports_pdf_input": false, "supports_reasoning": true, "supports_tool_choice": true }, @@ -67469,6 +67554,7 @@ "supported_output_modalities": [ "text" ], + "supports_pdf_input": false, "supports_response_schema": true, "supports_vision": true }, @@ -67478,7 +67564,8 @@ "output_cost_per_token": 5.4e-06, "litellm_provider": "azure_ai", "mode": "chat", - "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'" + "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", + "supports_pdf_input": false }, "azure_ai/deepseek-v3-0324": { "deprecation_date": "2026-07-13", @@ -67486,7 +67573,8 @@ "output_cost_per_token": 4.56e-06, "litellm_provider": "azure_ai", "mode": "chat", - "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'" + "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", + "supports_pdf_input": false }, "azure_ai/deepseek-v3.1": { "deprecation_date": "2026-07-13", @@ -67494,7 +67582,8 @@ "output_cost_per_token": 4.94e-06, "litellm_provider": "azure_ai", "mode": "chat", - "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'" + "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", + "supports_pdf_input": false }, "azure_ai/grok-3": { "deprecation_date": "2026-05-01", @@ -67502,7 +67591,8 @@ "output_cost_per_token": 1.5e-05, "litellm_provider": "azure_ai", "mode": "chat", - "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'" + "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", + "supports_pdf_input": false }, "azure_ai/grok-3-mini": { "deprecation_date": "2026-05-01", @@ -67510,7 +67600,8 @@ "output_cost_per_token": 1.27e-06, "litellm_provider": "azure_ai", "mode": "chat", - "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'" + "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", + "supports_pdf_input": false }, "azure_ai/grok-4-fast-non-reasoning": { "deprecation_date": "2026-05-01", @@ -67518,7 +67609,8 @@ "output_cost_per_token": 5e-07, "litellm_provider": "azure_ai", "mode": "chat", - "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'" + "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", + "supports_pdf_input": false }, "azure_ai/grok-4-fast-reasoning": { "deprecation_date": "2026-05-01", @@ -67526,7 +67618,8 @@ "output_cost_per_token": 5e-07, "litellm_provider": "azure_ai", "mode": "chat", - "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'" + "source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'", + "supports_pdf_input": false }, "bedrock/us-gov-west-1/nvidia.nemotron-nano-3-30b": { "input_cost_per_token": 7.2e-08, @@ -73196,6 +73289,7 @@ "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing", "supported_regions": ["global"], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, @@ -73811,6 +73905,7 @@ "output_cost_per_token": 3.5e-06, "source": "https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/microsoft/", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_tool_choice": true @@ -74865,6 +74960,7 @@ "text" ], "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": false, "supports_reasoning": true, "supports_tool_choice": true, @@ -75113,6 +75209,7 @@ "output_cost_per_token_above_200k_tokens": 1.2e-05, "source": "https://docs.x.ai/developers/models/grok-4.7", "supports_function_calling": true, + "supports_pdf_input": false, "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, diff --git a/tests/integration/messages_endpoint/chat_bridge/test_anthropic_messages_pdf_document_wire.py b/tests/integration/messages_endpoint/chat_bridge/test_anthropic_messages_pdf_document_wire.py new file mode 100644 index 00000000000..130bab24f4c --- /dev/null +++ b/tests/integration/messages_endpoint/chat_bridge/test_anthropic_messages_pdf_document_wire.py @@ -0,0 +1,421 @@ +import json +import uuid +from collections.abc import Callable, Sequence +from dataclasses import dataclass +from pathlib import Path +from typing import Final + +import anthropic +import pytest +from integration._support.client import Gateway, Scenario, string_value +from integration._support.pdf_document import LETTER, Page, document, encoded, pdf_bytes +from integration._support.process import owned_proxy_process +from integration._support.wire import Reply, Request, Wire, wire_server +from pydantic import JsonValue, TypeAdapter + +_API_KEY: Final = "synthetic-pdf-document-key" +_ANSWER: Final = "The codeword is PELICAN-42." +_TITLE: Final = "Quarterly report" +_URL: Final = "https://example.com/brief.pdf" +_SAP_TOKEN_TARGET: Final = "/oauth/token" +_JSON_OBJECT: Final = TypeAdapter(dict[str, JsonValue]) +_PARTS: Final = TypeAdapter(list[dict[str, JsonValue]]) + + +def _pdf(text: str) -> str: + return encoded(pdf_bytes((Page(text=text), LETTER))) + + +def _chat_completion(identity: str, model: str) -> Reply: + return Reply( + body=json.dumps( + { + "id": identity, + "object": "chat.completion", + "created": 1, + "model": model, + "choices": [ + {"index": 0, "message": {"role": "assistant", "content": _ANSWER}, "finish_reason": "stop"} + ], + "usage": {"prompt_tokens": 914, "completion_tokens": 9, "total_tokens": 923}, + } + ).encode() + ) + + +def _only_body(wire: Wire) -> dict[str, JsonValue]: + received: Final = tuple( + request + for request in wire.drain() + if request.method == "POST" and not request.target.startswith(_SAP_TOKEN_TARGET) + ) + assert len(received) == 1, received + return _JSON_OBJECT.validate_json(received[0].body) + + +def _user_parts(body: dict[str, JsonValue], key: str) -> list[dict[str, JsonValue]]: + turns: Final = _PARTS.validate_python(body[key]) + assert len(turns) == 1, turns + return _PARTS.validate_python(turns[0]["content" if key == "messages" else "parts"]) + + +def test_anthropic_sdk_pdf_documents_reach_azure_as_file_parts_named_by_title(gateway: Gateway) -> None: + marker: Final = uuid.uuid4().hex + backend: Final = f"gpt-pdf-{marker[:8]}" + identity: Final = f"chatcmpl-{marker}" + titled: Final = _pdf(f"PELICAN-42 {marker}") + untitled: Final = _pdf(f"ORCHID-77 {marker}") + question: Final = f"What codeword does each document carry? {marker}" + with wire_server(lambda _: _chat_completion(identity, backend)) as wire, gateway.scenario() as scenario: + model: Final = scenario.model( + model=f"azure/{backend}", api_base=wire.url, api_key=_API_KEY, api_version="2024-10-21" + ) + client: Final = anthropic.Anthropic(base_url=str(gateway.client.base_url), api_key=gateway.key, max_retries=0) + message: Final = client.messages.create( + model=model, + max_tokens=64, + messages=[ + { + "role": "user", + "content": [ + { + "type": "document", + "title": _TITLE, + "source": {"type": "base64", "media_type": "application/pdf", "data": titled}, + }, + { + "type": "document", + "source": {"type": "base64", "media_type": "application/pdf", "data": untitled}, + }, + {"type": "document", "source": {"type": "url", "url": _URL}}, + {"type": "text", "text": question}, + ], + } + ], + ) + assert message.id == identity + assert [(block.type, getattr(block, "text", None)) for block in message.content] == [("text", _ANSWER)] + assert _user_parts(_only_body(wire), "messages") == [ + {"type": "file", "file": {"filename": _TITLE, "file_data": f"data:application/pdf;base64,{titled}"}}, + { + "type": "file", + "file": {"filename": "document.pdf", "file_data": f"data:application/pdf;base64,{untitled}"}, + }, + {"type": "image_url", "image_url": {"url": _URL}}, + {"type": "text", "text": question}, + ] + + +@dataclass(frozen=True, slots=True) +class _Reader: + deploy: Callable[[Scenario, Wire], str] + reply: Callable[[Request], Reply] + pdf_parts: Callable[[dict[str, JsonValue]], Sequence[JsonValue]] + + +def _gemini_reply(_: Request) -> Reply: + return Reply( + body=json.dumps( + { + "candidates": [ + {"content": {"parts": [{"text": _ANSWER}], "role": "model"}, "finishReason": "STOP", "index": 0} + ], + "usageMetadata": {"promptTokenCount": 300, "candidatesTokenCount": 9, "totalTokenCount": 309}, + "modelVersion": "gemini-2.5-flash", + } + ).encode() + ) + + +def _bedrock_reply(_: Request) -> Reply: + return Reply( + body=json.dumps( + { + "output": {"message": {"role": "assistant", "content": [{"text": _ANSWER}]}}, + "stopReason": "end_turn", + "usage": {"inputTokens": 300, "outputTokens": 9, "totalTokens": 309}, + "metrics": {"latencyMs": 1}, + } + ).encode() + ) + + +def _anthropic_reply(_: Request) -> Reply: + return Reply( + body=json.dumps( + { + "id": "msg_pdf", + "type": "message", + "role": "assistant", + "model": "claude-sonnet-4-5", + "content": [{"type": "text", "text": _ANSWER}], + "stop_reason": "end_turn", + "usage": {"input_tokens": 300, "output_tokens": 9}, + } + ).encode() + ) + + +def _sap_reply(request: Request) -> Reply: + if request.target.startswith(_SAP_TOKEN_TARGET): + return Reply(body=json.dumps({"access_token": "scripted", "expires_in": 3600}).encode()) + match request.target: + case "/v2/lm/deployments": + deployment: Final = { + "scenarioId": "orchestration", + "configurationId": "orchestration-config", + "deploymentUrl": f"http://{request.headers['host']}/sap-deployment", + "createdAt": "2026-01-01T00:00:00Z", + } + return Reply(body=json.dumps({"resources": [deployment]}).encode()) + case "/v2/lm/configurations/orchestration-config": + return Reply(body=json.dumps({"executableId": "orchestration"}).encode()) + case _: + completion: Final = _JSON_OBJECT.validate_json(_chat_completion("chatcmpl-sap-pdf", "claude").body) + return Reply(body=json.dumps({"final_result": completion}).encode()) + + +def _snowflake(scenario: Scenario, wire: Wire) -> str: + return scenario.model(model="snowflake/claude-sonnet-4-5", api_base=wire.url, api_key=_API_KEY) + + +def _sap(scenario: Scenario, wire: Wire) -> str: + service_key: Final = { + "clientid": "scripted", + "clientsecret": _API_KEY, + "url": wire.url, + "serviceurls": {"AI_API_URL": wire.url}, + } + return scenario.model(model="sap/anthropic--claude-4.5-sonnet", api_base=wire.url, api_key=json.dumps(service_key)) + + +def _snowflake_pdf_parts(body: dict[str, JsonValue]) -> Sequence[JsonValue]: + return tuple(part.get("source") for part in _user_parts(body, "messages") if part.get("type") == "document") + + +def _sap_pdf_parts(body: dict[str, JsonValue]) -> Sequence[JsonValue]: + config: Final = _JSON_OBJECT.validate_python(body["config"]) + templating: Final = _JSON_OBJECT.validate_python( + _JSON_OBJECT.validate_python(config["modules"])["prompt_templating"] + ) + prompt: Final = _JSON_OBJECT.validate_python(templating["prompt"]) + turns: Final = _PARTS.validate_python(prompt["template"]) + assert len(turns) == 1, turns + parts: Final = _PARTS.validate_python(turns[0]["content"]) + return tuple( + _JSON_OBJECT.validate_python(part["image_url"]).get("url") for part in parts if part.get("type") == "image_url" + ) + + +def _gemini(scenario: Scenario, wire: Wire) -> str: + return scenario.model(model="gemini/gemini-2.5-flash", api_base=wire.url, api_key=_API_KEY) + + +def _bedrock(scenario: Scenario, wire: Wire) -> str: + return scenario.model( + model="bedrock/converse/us.amazon.nova-2-lite-v1:0", + api_key=None, + aws_access_key_id="AKIDSYNTHETIC", + aws_secret_access_key="synthetic-secret", + aws_region_name="us-east-1", + aws_bedrock_runtime_endpoint=wire.url, + ) + + +def _gemini_pdf_parts(body: dict[str, JsonValue]) -> Sequence[JsonValue]: + return tuple(part.get("inline_data") for part in _user_parts(body, "contents") if "inline_data" in part) + + +def _bedrock_pdf_parts(body: dict[str, JsonValue]) -> Sequence[JsonValue]: + documents: Final = tuple( + _JSON_OBJECT.validate_python(part["document"]) for part in _user_parts(body, "messages") if "document" in part + ) + return tuple({"format": item.get("format"), "source": item.get("source")} for item in documents) + + +@pytest.mark.parametrize( + ("reader", "expected"), + [ + pytest.param( + _Reader(_gemini, _gemini_reply, _gemini_pdf_parts), + lambda data: [{"mime_type": "application/pdf", "data": data}], + id="gemini", + ), + pytest.param( + _Reader(_bedrock, _bedrock_reply, _bedrock_pdf_parts), + lambda data: [{"format": "pdf", "source": {"bytes": data}}], + id="bedrock-converse", + ), + pytest.param( + _Reader(_snowflake, _anthropic_reply, _snowflake_pdf_parts), + lambda data: [{"type": "base64", "media_type": "application/pdf", "data": data}], + id="snowflake-claude", + ), + pytest.param( + _Reader(_sap, _sap_reply, _sap_pdf_parts), + lambda data: [f"data:application/pdf;base64,{data}"], + id="sap-orchestration", + ), + ], +) +def test_messages_pdf_still_reaches_providers_that_read_pdfs_in_their_own_document_shape( + gateway: Gateway, reader: _Reader, expected: Callable[[str], Sequence[JsonValue]] +) -> None: + marker: Final = uuid.uuid4().hex + data: Final = _pdf(f"PELICAN-42 {marker}") + with wire_server(reader.reply) as wire, gateway.scenario() as scenario: + model: Final = reader.deploy(scenario, wire) + response: Final = gateway.request( + "POST", + "/v1/messages", + { + "model": model, + "max_tokens": 64, + "messages": [ + { + "role": "user", + "content": [ + document({"type": "base64", "media_type": "application/pdf", "data": data}, title=_TITLE), + {"type": "text", "text": f"What is the codeword? {marker}"}, + ], + } + ], + }, + headers={"anthropic-version": "2023-06-01"}, + ) + assert response.status_code == 200, response.text + assert _JSON_OBJECT.validate_json(response.content)["content"] == [{"type": "text", "text": _ANSWER}] + assert list(reader.pdf_parts(_only_body(wire))) == list(expected(data)) + + +def _prompt_security(job: str) -> Callable[[Request], Reply]: + def respond(request: Request) -> Reply: + match (request.method, request.target.partition("?")[0]): + case ("POST", "/api/sanitizeFile"): + return Reply(body=json.dumps({"jobId": job}).encode()) + case ("GET", "/api/sanitizeFile"): + return Reply( + body=json.dumps( + { + "status": "done", + "content": "", + "metadata": {"action": "block", "violations": ["Malicious File"]}, + } + ).encode() + ) + case _: + return Reply(body=json.dumps({"result": {"prompt": {"action": "allow"}}}).encode()) + + return respond + + +def _delete_guardrail(gateway: Gateway, identity: str) -> None: + response: Final = gateway.request("DELETE", f"/guardrails/{identity}") + assert response.status_code == 200, response.text + + +def test_prompt_security_still_sanitizes_a_messages_pdf_and_blocks_it_before_the_provider(gateway: Gateway) -> None: + marker: Final = uuid.uuid4().hex + pdf: Final = pdf_bytes((Page(text=f"PELICAN-42 {marker}"), LETTER)) + with ( + wire_server(lambda _: _chat_completion(f"chatcmpl-{marker}", "gpt-pdf")) as provider, + wire_server(_prompt_security(f"job-{marker}")) as sink, + gateway.scenario() as scenario, + ): + model: Final = scenario.model( + model=f"azure/gpt-pdf-{marker[:8]}", api_base=provider.url, api_key=_API_KEY, api_version="2024-10-21" + ) + name: Final = f"pdf-prompt-security-{marker}" + created: Final = gateway.post( + "/guardrails", + { + "guardrail": { + "guardrail_name": name, + "litellm_params": { + "guardrail": "prompt_security", + "mode": "pre_call", + "api_key": "synthetic-prompt-security-key", + "api_base": sink.url, + }, + } + }, + ) + scenario.cleanups.callback(_delete_guardrail, gateway, string_value(created["guardrail_id"])) + response: Final = gateway.request( + "POST", + "/v1/messages", + { + "model": model, + "max_tokens": 64, + "guardrails": [name], + "messages": [ + { + "role": "user", + "content": [ + document({"type": "base64", "media_type": "application/pdf", "data": encoded(pdf)}), + {"type": "text", "text": f"Summarize this PDF {marker}"}, + ], + } + ], + }, + headers={"anthropic-version": "2023-06-01"}, + ) + assert response.status_code == 400, response.text + assert "Malicious File" in response.text + assert provider.drain() == () + uploads: Final = tuple( + request for request in sink.drain() if (request.method, request.target) == ("POST", "/api/sanitizeFile") + ) + assert len(uploads) == 1, uploads + assert pdf in uploads[0].body + + +def _gigachat_reply(file_id: str) -> Callable[[Request], Reply]: + def respond(request: Request) -> Reply: + if request.target.endswith("/files"): + return Reply(body=json.dumps({"id": file_id, "object": "file", "purpose": "general"}).encode()) + return _chat_completion(f"chatcmpl-{file_id}", "GigaChat-Pro") + + return respond + + +def test_anthropic_sdk_pdf_document_reaches_gigachat_as_an_uploaded_attachment( + gateway: Gateway, tmp_path: Path +) -> None: + marker: Final = uuid.uuid4().hex + file_id: Final = f"file-{marker}" + pdf: Final = pdf_bytes((Page(text=f"PELICAN-42 {marker}"), LETTER)) + with ( + wire_server(_gigachat_reply(file_id)) as wire, + owned_proxy_process(gateway, tmp_path, {"GIGACHAT_ACCESS_TOKEN": _API_KEY}) as owned, + owned.gateway.scenario() as scenario, + ): + model: Final = scenario.model(model="gigachat/GigaChat-Pro", api_base=f"{wire.url}/v1", api_key=_API_KEY) + client: Final = anthropic.Anthropic( + base_url=str(owned.gateway.client.base_url), api_key=owned.gateway.key, max_retries=0 + ) + message: Final = client.messages.create( + model=model, + max_tokens=64, + messages=[ + { + "role": "user", + "content": [ + { + "type": "document", + "source": {"type": "base64", "media_type": "application/pdf", "data": encoded(pdf)}, + }, + {"type": "text", "text": f"What is the codeword? {marker}"}, + ], + } + ], + ) + assert [(block.type, getattr(block, "text", None)) for block in message.content] == [("text", _ANSWER)] + upload, chat = (request for request in wire.drain() if request.method == "POST") + assert upload.target == "/v1/files", upload.target + assert upload.headers["authorization"] == f"Bearer {_API_KEY}", sorted(upload.headers) + assert pdf in upload.body + assert b"application/pdf" in upload.body + assert chat.target == "/v1/chat/completions", chat.target + turns: Final = _PARTS.validate_python(_JSON_OBJECT.validate_json(chat.body)["messages"]) + assert turns[-1]["attachments"] == [file_id], chat.body diff --git a/tests/integration/providers/test_chat_provider_annotations_wire.py b/tests/integration/providers/test_chat_provider_annotations_wire.py new file mode 100644 index 00000000000..0b18d0837eb --- /dev/null +++ b/tests/integration/providers/test_chat_provider_annotations_wire.py @@ -0,0 +1,190 @@ +import json +import uuid +from collections.abc import Callable, Iterator +from itertools import chain +from typing import Final + +import httpx +from integration._support.client import Gateway +from integration._support.wire import Reply, Request, Wire, wire_server +from integration.caching.response_cache_case import await_entry, redis_store +from pydantic import JsonValue, TypeAdapter +from redis import Redis + +_BACKEND: Final = "z-ai/glm-5.2" +_API_KEY: Final = "synthetic-openrouter-key" +_ANSWER: Final = "The report covers revenue." +_URL_CITATION: Final[dict[str, JsonValue]] = { + "type": "url_citation", + "url_citation": { + "url": "https://example.com/report", + "title": "Report", + "start_index": 0, + "end_index": 10, + "content": "cited passage", + }, +} +_FILE_ANNOTATION: Final[dict[str, JsonValue]] = { + "type": "file", + "file": {"hash": "sha-report", "name": "report.pdf", "content": [{"type": "text", "text": "parsed page one"}]}, +} +_USAGE: Final[dict[str, JsonValue]] = {"prompt_tokens": 40, "completion_tokens": 6, "total_tokens": 46} +_JSON_OBJECT: Final = TypeAdapter(dict[str, JsonValue]) +_CHOICES: Final = TypeAdapter(list[dict[str, JsonValue]]) + + +def _completion(identity: str) -> bytes: + return json.dumps( + { + "id": identity, + "object": "chat.completion", + "created": 1, + "model": _BACKEND, + "choices": [ + { + "index": 0, + "message": { + "role": "assistant", + "content": _ANSWER, + "annotations": [_URL_CITATION, _FILE_ANNOTATION], + }, + "finish_reason": "stop", + } + ], + "usage": _USAGE, + } + ).encode() + + +def _frame(frame: dict[str, JsonValue]) -> bytes: + return b"data: " + json.dumps(frame).encode() + b"\n\n" + + +def _stream(identity: str) -> Reply: + chunk: Final[dict[str, JsonValue]] = { + "id": identity, + "object": "chat.completion.chunk", + "created": 1, + "model": _BACKEND, + } + return Reply( + content_type="text/event-stream", + chunks=( + _frame({**chunk, "choices": [{"index": 0, "delta": {"role": "assistant", "content": _ANSWER}}]}), + _frame({**chunk, "choices": [{"index": 0, "delta": {"annotations": [_URL_CITATION, _FILE_ANNOTATION]}}]}), + _frame({**chunk, "choices": [{"index": 0, "delta": {}, "finish_reason": "stop"}], "usage": _USAGE}), + b"data: [DONE]\n\n", + ), + ) + + +def _responder(identity: str) -> Callable[[Request], Reply]: + def respond(request: Request) -> Reply: + if _JSON_OBJECT.validate_json(request.body).get("stream") is True: + return _stream(identity) + return Reply(body=_completion(identity)) + + return respond + + +def _chat_body(model: str, marker: str, stream: bool) -> dict[str, JsonValue]: + return { + "model": model, + "messages": [{"role": "user", "content": f"Summarize the report {marker}"}], + "stream": stream, + } + + +def _posted(wire: Wire) -> int: + return sum(1 for request in wire.drain() if request.method == "POST") + + +def _message(response: httpx.Response) -> dict[str, JsonValue]: + assert response.status_code == 200, response.text + choices: Final = _CHOICES.validate_python(_JSON_OBJECT.validate_json(response.content)["choices"]) + return _JSON_OBJECT.validate_python(choices[0]["message"]) + + +def _assert_split(message: dict[str, JsonValue]) -> None: + assert message["content"] == _ANSWER, message + assert message["annotations"] == [_URL_CITATION], message + assert _JSON_OBJECT.validate_python(message["provider_specific_fields"])["annotations"] == [_FILE_ANNOTATION] + + +def test_file_annotation_reaches_the_caller_in_provider_specific_fields_and_survives_a_cache_hit( + gateway: Gateway, +) -> None: + marker: Final = uuid.uuid4().hex + identity: Final = f"gen-{marker}" + with wire_server(_responder(identity)) as wire, gateway.scenario() as scenario: + model: Final = scenario.model(model=f"openrouter/{_BACKEND}", api_base=wire.url, api_key=_API_KEY) + first: Final = gateway.request("POST", "/v1/chat/completions", _chat_body(model, marker, stream=False)) + _assert_split(_message(first)) + assert _posted(wire) == 1 + await_entry(redis_store(), identity) + cached: Final = gateway.request("POST", "/v1/chat/completions", _chat_body(model, marker, stream=False)) + assert cached.headers.get("x-litellm-cache-key"), dict(cached.headers) + _assert_split(_message(cached)) + assert _posted(wire) == 0 + + +def _deltas(response: httpx.Response) -> Iterator[dict[str, JsonValue]]: + for line in response.iter_lines(): + if not line.startswith("data: ") or line == "data: [DONE]": + continue + choices = _CHOICES.validate_python(_JSON_OBJECT.validate_json(line.removeprefix("data: "))["choices"]) + yield from (_JSON_OBJECT.validate_python(choice["delta"]) for choice in choices) + + +def _streamed(gateway: Gateway, body: dict[str, JsonValue]) -> tuple[dict[str, JsonValue], ...]: + with gateway.client.stream( + "POST", "/v1/chat/completions", json=body, headers={"Authorization": f"Bearer {gateway.key}"} + ) as response: + assert response.status_code == 200, response.read() + return tuple(_deltas(response)) + + +def _listed(value: JsonValue) -> list[JsonValue]: + return value if isinstance(value, list) else [] + + +def _provider_fields(delta: dict[str, JsonValue]) -> dict[str, JsonValue]: + return _JSON_OBJECT.validate_python(delta.get("provider_specific_fields") or {}) + + +def _annotations(deltas: tuple[dict[str, JsonValue], ...]) -> list[JsonValue]: + return list(chain.from_iterable(_listed(delta.get("annotations")) for delta in deltas)) + + +def _provider_annotations(deltas: tuple[dict[str, JsonValue], ...]) -> list[JsonValue]: + return list(chain.from_iterable(_listed(_provider_fields(delta).get("annotations")) for delta in deltas)) + + +def _text(deltas: tuple[dict[str, JsonValue], ...]) -> str: + return "".join(str(delta.get("content") or "") for delta in deltas) + + +def _cached_message(store: Redis, key: str) -> dict[str, JsonValue]: + raw: Final = store.get(key) + assert isinstance(raw, bytes), key + response: Final = _JSON_OBJECT.validate_json(raw)["response"] + stored: Final = _JSON_OBJECT.validate_json(response) if isinstance(response, str) else response + choices: Final = _CHOICES.validate_python(_JSON_OBJECT.validate_python(stored)["choices"]) + return _JSON_OBJECT.validate_python(choices[0]["message"]) + + +def test_streamed_annotations_reach_the_caller_as_sent_and_the_cached_stream_splits_them(gateway: Gateway) -> None: + marker: Final = uuid.uuid4().hex + identity: Final = f"gen-{marker}" + with wire_server(_responder(identity)) as wire, gateway.scenario() as scenario: + model: Final = scenario.model(model=f"openrouter/{_BACKEND}", api_base=wire.url, api_key=_API_KEY) + live: Final = _streamed(gateway, _chat_body(model, marker, stream=True)) + assert _text(live) == _ANSWER, live + assert _annotations(live) == [_URL_CITATION, _FILE_ANNOTATION], live + assert _posted(wire) == 1 + store: Final = redis_store() + _assert_split(_cached_message(store, await_entry(store, identity))) + replayed: Final = _streamed(gateway, _chat_body(model, marker, stream=True)) + assert _text(replayed) == _ANSWER, replayed + assert _provider_annotations(replayed) == [_FILE_ANNOTATION], replayed + assert _posted(wire) == 0 diff --git a/tests/integration/providers/test_pdf_input_support_wire.py b/tests/integration/providers/test_pdf_input_support_wire.py new file mode 100644 index 00000000000..b35815438f7 --- /dev/null +++ b/tests/integration/providers/test_pdf_input_support_wire.py @@ -0,0 +1,390 @@ +import json +import uuid +from collections.abc import Callable, Mapping, Sequence +from concurrent.futures import ThreadPoolExecutor +from dataclasses import dataclass +from pathlib import Path +from typing import Final + +import httpx +import pytest +from integration._support.client import Gateway, Scenario +from integration._support.openai_wire import chat_reply +from integration._support.pdf_document import ( + LETTER, + chat_body, + chat_file, + messages_body, + pdf_data_url, + pdf_document, + responses_body, + responses_input_file, +) +from integration._support.wire import Reply, Request, Wire, wire_server +from pydantic import JsonValue, TypeAdapter + +_API_KEY: Final = "synthetic-pdf-gate-key" +_ANSWER: Final = "The report covers revenue." +_FILENAME: Final = "report.pdf" +_PDF: Final = pdf_data_url((LETTER,)) +_JSON_OBJECT: Final = TypeAdapter(dict[str, JsonValue]) +_MESSAGES: Final = TypeAdapter(list[JsonValue]) +_COST_MAP_PATH: Final = Path(__file__).parents[3] / "litellm" / "model_prices_and_context_window_backup.json" + + +@dataclass(frozen=True, slots=True) +class _Entrypoint: + path: str + body: Callable[[str, str], Mapping[str, JsonValue]] + stream: bool = False + provider: str = "hosted_vllm" + + +_ENTRYPOINTS: Final = ( + pytest.param( + _Entrypoint("/v1/chat/completions", lambda model, text: chat_body(model, [chat_file(_PDF, _FILENAME)], text)), + id="chat-completions", + ), + pytest.param( + _Entrypoint( + "/v1/chat/completions", + lambda model, text: chat_body(model, [chat_file(_PDF, _FILENAME)], text, stream=True), + stream=True, + ), + id="chat-completions-stream", + ), + pytest.param( + _Entrypoint("/v1/messages", lambda model, text: messages_body(model, [pdf_document((LETTER,))], text)), + id="messages", + ), + pytest.param( + _Entrypoint( + "/v1/responses", + lambda model, text: responses_body(model, [responses_input_file(_PDF, _FILENAME)], text), + provider="lm_studio", + ), + id="responses-bridge", + ), +) + + +def _azure_ai_chat_model_pdf_input(entry: JsonValue) -> bool | None: + match entry: + case {"litellm_provider": "azure_ai", "mode": "chat", "supports_pdf_input": bool() as supports}: + return supports + case _: + return None + + +def _azure_ai_chat_model(reads_pdfs: bool) -> str: + entries: Final = _JSON_OBJECT.validate_json(_COST_MAP_PATH.read_bytes()) + matching: Final = sorted( + name for name, entry in entries.items() if _azure_ai_chat_model_pdf_input(entry) is reads_pdfs + ) + assert matching, f"the cost map marks no azure_ai chat model with supports_pdf_input {reads_pdfs}" + return matching[0] + + +def _ollama_reply(backend: str) -> Reply: + return Reply( + body=json.dumps( + { + "model": backend, + "created_at": "2026-01-01T00:00:00Z", + "message": {"role": "assistant", "content": _ANSWER}, + "done": True, + "done_reason": "stop", + "prompt_eval_count": 5, + "eval_count": 3, + } + ).encode() + ) + + +def _responder(backend: str) -> Callable[[Request], Reply]: + def respond(request: Request) -> Reply: + if request.target.endswith("/api/chat"): + return _ollama_reply(backend) + body: Final = _JSON_OBJECT.validate_json(request.body) + return chat_reply(f"chatcmpl-{uuid.uuid4().hex}", backend, _ANSWER, stream=body.get("stream") is True) + + return respond + + +def _post(gateway: Gateway, path: str, body: Mapping[str, JsonValue]) -> httpx.Response: + return gateway.request("POST", path, body, headers={"anthropic-version": "2023-06-01"}) + + +def _post_streaming(gateway: Gateway, path: str, body: Mapping[str, JsonValue]) -> httpx.Response: + with gateway.client.stream( + "POST", path, json=dict(body), headers={"Authorization": f"Bearer {gateway.key}"} + ) as response: + response.read() + return response + + +def _send(gateway: Gateway, entrypoint: _Entrypoint, model: str, text: str) -> httpx.Response: + body: Final = entrypoint.body(model, text) + if entrypoint.stream: + return _post_streaming(gateway, entrypoint.path, body) + return _post(gateway, entrypoint.path, body) + + +def _error_message(response: httpx.Response) -> str: + match _JSON_OBJECT.validate_json(response.content): + case {"error": {"message": str() as message}}: + return message + case _: + raise AssertionError(response.text) + + +def _assert_refused_before_the_provider(response: httpx.Response, wire: Wire, backend: str) -> None: + assert response.status_code == 400, response.text + message: Final = _error_message(response) + assert "PDF" in message and backend in message, message + assert wire.drain() == () + + +def _bodies(wire: Wire) -> tuple[dict[str, JsonValue], ...]: + return tuple(_JSON_OBJECT.validate_json(request.body) for request in wire.drain() if request.method == "POST") + + +def _only_body(wire: Wire) -> dict[str, JsonValue]: + bodies: Final = _bodies(wire) + assert len(bodies) == 1, bodies + return bodies[0] + + +def _assert_reaches_the_provider(gateway: Gateway, wire: Wire, model: str) -> None: + response: Final = _post(gateway, "/v1/chat/completions", chat_body(model, [], f"text only {uuid.uuid4().hex}")) + assert response.status_code == 200, response.text + assert len(_bodies(wire)) == 1 + + +def _openai_compatible( + scenario: Scenario, + wire: Wire, + backend: str, + model_info: Mapping[str, JsonValue], + provider: str = "hosted_vllm", + **parameters: JsonValue, +) -> str: + return scenario.model( + model=f"{provider}/{backend}", + api_base=f"{wire.url}/v1", + api_key=_API_KEY, + model_info=model_info, + **parameters, + ) + + +def _backend() -> str: + return f"pdf-gate-{uuid.uuid4().hex[:12]}" + + +@pytest.mark.parametrize("entrypoint", _ENTRYPOINTS) +def test_pdf_to_a_deployment_marked_unable_to_read_it_is_refused_before_the_provider( + gateway: Gateway, entrypoint: _Entrypoint +) -> None: + backend: Final = _backend() + with wire_server(_responder(backend)) as wire, gateway.scenario() as scenario: + model: Final = _openai_compatible(scenario, wire, backend, {"supports_pdf_input": False}, entrypoint.provider) + response: Final = _send(gateway, entrypoint, model, f"Summarize the report {uuid.uuid4().hex}") + _assert_refused_before_the_provider(response, wire, backend) + _assert_reaches_the_provider(gateway, wire, model) + + +def _base_model_deployment(scenario: Scenario, wire: Wire) -> tuple[str, str]: + backend: Final = f"pdf-gate-{uuid.uuid4().hex[:12]}" + model: Final = scenario.model( + model=f"azure_ai/{backend}", + api_base=wire.url, + api_key=_API_KEY, + model_info={"base_model": _azure_ai_chat_model(reads_pdfs=False)}, + ) + return model, backend + + +def _litellm_params_base_model_deployment(scenario: Scenario, wire: Wire) -> tuple[str, str]: + backend: Final = _backend() + model: Final = scenario.model( + model=f"azure_ai/{backend}", + api_base=wire.url, + api_key=_API_KEY, + base_model=_azure_ai_chat_model(reads_pdfs=False), + ) + return model, backend + + +def _cost_map_deployment(scenario: Scenario, wire: Wire) -> tuple[str, str]: + marked: Final = _azure_ai_chat_model(reads_pdfs=False) + return scenario.model(model=marked, api_base=wire.url, api_key=_API_KEY), marked.removeprefix("azure_ai/") + + +def _ollama_deployment(scenario: Scenario, wire: Wire) -> tuple[str, str]: + backend: Final = _backend() + return scenario.model(model=f"ollama_chat/{backend}", api_base=wire.url, api_key=None), backend + + +@pytest.mark.parametrize( + "deploy", + [ + pytest.param(_base_model_deployment, id="model-info-base-model"), + pytest.param(_litellm_params_base_model_deployment, id="litellm-params-base-model"), + pytest.param(_cost_map_deployment, id="cost-map"), + pytest.param(_ollama_deployment, id="ollama"), + ], +) +def test_pdf_to_a_model_its_declaration_marks_unable_is_refused_before_the_provider( + gateway: Gateway, deploy: Callable[[Scenario, Wire], tuple[str, str]] +) -> None: + with wire_server(_responder("scripted")) as wire, gateway.scenario() as scenario: + model, backend = deploy(scenario, wire) + response: Final = _post( + gateway, + "/v1/chat/completions", + chat_body(model, [chat_file(_PDF, _FILENAME)], f"Summarize the report {uuid.uuid4().hex}"), + ) + _assert_refused_before_the_provider(response, wire, backend) + _assert_reaches_the_provider(gateway, wire, model) + + +def _pdf_turns(question: str) -> list[JsonValue]: + return [ + {"role": "user", "content": [chat_file(_PDF, _FILENAME), {"type": "text", "text": question}]}, + {"role": "assistant", "content": "It is a quarterly report."}, + {"role": "user", "content": [chat_file(_PDF, "appendix.pdf")]}, + {"role": "user", "content": "Who wrote it?"}, + ] + + +def _answer(response: httpx.Response) -> JsonValue: + match _JSON_OBJECT.validate_json(response.content): + case {"choices": [{"message": {"content": content}}]}: + return content + case _: + raise AssertionError(response.text) + + +def _turns_without_pdfs(question: str) -> list[JsonValue]: + return [ + {"role": "user", "content": [{"type": "text", "text": question}]}, + {"role": "assistant", "content": "It is a quarterly report."}, + {"role": "user", "content": "Who wrote it?"}, + ] + + +@pytest.mark.parametrize( + ("deployment_params", "request_params"), + [ + pytest.param({"drop_params": True}, {}, id="deployment-drop-params"), + pytest.param({}, {"drop_params": True}, id="request-drop-params"), + ], +) +def test_drop_params_sends_the_conversation_without_the_pdfs_a_model_cannot_read( + gateway: Gateway, deployment_params: Mapping[str, JsonValue], request_params: Mapping[str, JsonValue] +) -> None: + backend: Final = _backend() + question: Final = f"Summarize the report {uuid.uuid4().hex}" + with wire_server(_responder(backend)) as wire, gateway.scenario() as scenario: + model: Final = _openai_compatible(scenario, wire, backend, {"supports_pdf_input": False}, **deployment_params) + response: Final = _post( + gateway, + "/v1/chat/completions", + {"model": model, "max_tokens": 64, "messages": _pdf_turns(question), **request_params}, + ) + assert response.status_code == 200, response.text + assert _answer(response) == _ANSWER + assert _only_body(wire)["messages"] == _turns_without_pdfs(question) + + +def test_drop_params_refuses_a_request_left_with_no_user_turn_once_the_pdf_is_dropped(gateway: Gateway) -> None: + backend: Final = _backend() + with wire_server(_responder(backend)) as wire, gateway.scenario() as scenario: + model: Final = _openai_compatible(scenario, wire, backend, {"supports_pdf_input": False}, drop_params=True) + response: Final = _post( + gateway, + "/v1/chat/completions", + { + "model": model, + "max_tokens": 64, + "messages": [ + {"role": "system", "content": f"Answer from the attachment {uuid.uuid4().hex}"}, + {"role": "user", "content": [chat_file(_PDF, _FILENAME)]}, + ], + }, + ) + _assert_refused_before_the_provider(response, wire, backend) + _assert_reaches_the_provider(gateway, wire, model) + + +def _file_and_question(question: str) -> list[JsonValue]: + return [{"role": "user", "content": [chat_file(_PDF, _FILENAME), {"type": "text", "text": question}]}] + + +@pytest.mark.parametrize( + ("model_info", "deployment_params", "keeps_pdf"), + [ + pytest.param({}, {}, True, id="undeclared"), + pytest.param({}, {"drop_params": True}, False, id="undeclared-drop-params"), + pytest.param({"supports_pdf_input": True}, {"drop_params": True}, True, id="declared-drop-params"), + pytest.param( + {}, + {"drop_params": True, "base_model": _azure_ai_chat_model(reads_pdfs=True)}, + True, + id="base-model-reads-drop-params", + ), + ], +) +def test_azure_ai_model_outside_the_cost_map_gets_the_pdf_unless_drop_params_finds_no_declaration( + gateway: Gateway, model_info: Mapping[str, JsonValue], deployment_params: Mapping[str, JsonValue], keeps_pdf: bool +) -> None: + backend: Final = _backend() + question: Final = f"Summarize the report {uuid.uuid4().hex}" + with wire_server(_responder(backend)) as wire, gateway.scenario() as scenario: + model: Final = scenario.model( + model=f"azure_ai/{backend}", + api_base=wire.url, + api_key=_API_KEY, + model_info=model_info, + **deployment_params, + ) + response: Final = _post( + gateway, "/v1/chat/completions", chat_body(model, [chat_file(_PDF, _FILENAME)], question) + ) + assert response.status_code == 200, response.text + expected: Final[list[JsonValue]] = ( + _file_and_question(question) if keeps_pdf else [{"role": "user", "content": question}] + ) + assert _only_body(wire)["messages"] == expected + + +def _sibling_pdf_request(gateway: Gateway, model: str) -> httpx.Response: + return _post( + gateway, + "/v1/chat/completions", + chat_body(model, [chat_file(_PDF, _FILENAME)], f"Summarize the report {uuid.uuid4().hex}"), + ) + + +def _forwarded_file_parts(bodies: Sequence[dict[str, JsonValue]]) -> tuple[JsonValue, ...]: + return tuple(_MESSAGES.validate_python(body["messages"])[0] for body in bodies) + + +def test_pdf_opt_out_on_one_deployment_leaves_its_sibling_on_the_same_model_reading_pdfs(gateway: Gateway) -> None: + backend: Final = _backend() + burst: Final = 6 + with wire_server(_responder(backend)) as wire, gateway.scenario() as scenario: + sibling: Final = _openai_compatible(scenario, wire, backend, {}) + marked: Final = _openai_compatible(scenario, wire, backend, {"supports_pdf_input": False}) + with ThreadPoolExecutor(max_workers=burst) as pool: + responses: Final = tuple(pool.map(lambda _: _sibling_pdf_request(gateway, sibling), range(burst))) + assert [response.status_code for response in responses] == [200] * burst, [r.text for r in responses] + forwarded: Final = _forwarded_file_parts(_bodies(wire)) + assert len(forwarded) == burst, forwarded + assert all( + _MESSAGES.validate_python(_JSON_OBJECT.validate_python(message)["content"])[0] == chat_file(_PDF, _FILENAME) + for message in forwarded + ), forwarded + refused: Final = _sibling_pdf_request(gateway, marked) + _assert_refused_before_the_provider(refused, wire, backend) diff --git a/tests/unit/litellm_core_utils/llm_response_utils/test_convert_dict_to_response.py b/tests/unit/litellm_core_utils/llm_response_utils/test_convert_dict_to_response.py index 0c601a6b2c8..cea12230fbd 100644 --- a/tests/unit/litellm_core_utils/llm_response_utils/test_convert_dict_to_response.py +++ b/tests/unit/litellm_core_utils/llm_response_utils/test_convert_dict_to_response.py @@ -2968,3 +2968,54 @@ def test_convert_to_model_response_object_function_output(): assert result.usage.completion_tokens_details == CompletionTokensDetailsWrapper(reasoning_tokens=0) assert result._hidden_params is not None + + +_URL_CITATION: Final = { + "type": "url_citation", + "url_citation": { + "url": "https://example.com", + "title": "Example", + "content": "cited snippet", + "start_index": 0, + "end_index": 4, + }, +} +_FILE_ANNOTATION: Final = { + "type": "file", + "file": {"hash": "abc123", "name": "report.pdf", "content": [{"type": "text", "text": "parsed page"}]}, +} + + +def _message_with_annotations(annotations: list[dict[str, object]]) -> Message: + response: Final = { + "id": "gen-1", + "object": "chat.completion", + "created": 1712697600, + "model": "z-ai/glm-5.3", + "choices": [ + { + "index": 0, + "finish_reason": "stop", + "message": {"role": "assistant", "content": "done", "annotations": annotations}, + } + ], + "usage": {"completion_tokens": 1, "prompt_tokens": 1, "total_tokens": 2}, + } + return convert_to_model_response_object( + response_object=response, model_response_object=ModelResponse(), response_type="completion" + ).choices[0].message + + +def test_annotations_typed_other_than_url_citation_move_to_provider_specific_fields() -> None: + untyped_annotation: Final = {"start_index": 0, "end_index": 4} + message: Final = _message_with_annotations([_URL_CITATION, untyped_annotation, _FILE_ANNOTATION]) + + assert message.annotations == [_URL_CITATION, untyped_annotation] + assert message.provider_specific_fields["annotations"] == [_FILE_ANNOTATION] + + +def test_url_citation_annotations_keep_provider_keys() -> None: + message: Final = _message_with_annotations([_URL_CITATION]) + + assert message.annotations == [_URL_CITATION] + assert "annotations" not in (message.provider_specific_fields or {}) diff --git a/tests/unit/llms/anthropic/messages/adapters/test_anthropic_messages_adapters_transformation.py b/tests/unit/llms/anthropic/messages/adapters/test_anthropic_messages_adapters_transformation.py index 51ec71aaea0..c30644e7fdc 100644 --- a/tests/unit/llms/anthropic/messages/adapters/test_anthropic_messages_adapters_transformation.py +++ b/tests/unit/llms/anthropic/messages/adapters/test_anthropic_messages_adapters_transformation.py @@ -5286,6 +5286,40 @@ def test_eager_input_streaming_tool_reaches_bedrock_converse_as_beta(): assert data["toolConfig"]["tools"][0]["toolSpec"]["inputSchema"]["json"] == EAGER_INPUT_SCHEMA +@pytest.mark.parametrize( + ("title", "expected_filename"), + [(None, "document.pdf"), ("Quarterly report", "Quarterly report")], +) +def test_user_pdf_document_reaches_openai_compatible_providers_as_file_part( + title: str | None, expected_filename: str +) -> None: + document: Final = _base64_pdf_block() | ({"title": title} if title else {}) + request_messages: Final = _run_chat_completions_pipeline( + [ + AnthropicMessagesUserMessageParam( + role="user", + content=[document, {"type": "text", "text": "What is the codeword?"}], + ) + ] + ) + + assert request_messages == [ + { + "role": "user", + "content": [ + { + "type": "file", + "file": { + "filename": expected_filename, + "file_data": f"data:application/pdf;base64,{TOOL_RESULT_PDF_B64}", + }, + }, + {"type": "text", "text": "What is the codeword?"}, + ], + } + ] + + _CHAT_BRIDGE_MODEL: Final = "azure/gpt-5.6" _CHAT_BRIDGE_API_BASE: Final = "https://tool-result-text.openai.azure.com" _CHAT_BRIDGE_URL: Final = r".*tool-result-text\.openai\.azure\.com/openai/deployments/gpt-5\.6/chat/completions.*" diff --git a/tests/unit/llms/azure_ai/chat/test_azure_ai_transformation.py b/tests/unit/llms/azure_ai/chat/test_azure_ai_transformation.py index 42dc222e420..6269fd367fe 100644 --- a/tests/unit/llms/azure_ai/chat/test_azure_ai_transformation.py +++ b/tests/unit/llms/azure_ai/chat/test_azure_ai_transformation.py @@ -1101,6 +1101,58 @@ def test_map_azure_model_group(model_group_header, expected_model): assert config._map_azure_model_group(model_group_header) == expected_model +def test_azure_ai_sends_file_parts_instead_of_flattening_them_to_text() -> None: + file_part: Final = { + "type": "file", + "file": {"filename": "document.pdf", "file_data": "data:application/pdf;base64,JVBERi0xLjQK"}, + } + content: Final = [file_part, {"type": "text", "text": "What is the codeword?"}] + + request: Final = AzureAIStudioConfig().transform_request( + model="Kimi-K2.6", + messages=[{"role": "user", "content": content}], + optional_params={}, + litellm_params={}, + headers={}, + ) + + assert request["messages"] == [{"role": "user", "content": content}] + + +def test_azure_ai_passes_dict_message_content_through_for_the_provider_to_judge() -> None: + content: Final = {"type": "text", "text": "Say hi"} + + request: Final = AzureAIStudioConfig().transform_request( + model="gpt-5.6-sol", + messages=[{"role": "user", "content": content}], + optional_params={}, + litellm_params={}, + headers={}, + ) + + assert request["messages"] == [{"role": "user", "content": content}] + + +@pytest.mark.parametrize( + "file", + [ + {"filename": "notes.txt", "file_data": "data:text/plain;base64,aGVsbG8="}, + {"file_id": "file-abc"}, + ], + ids=["text_data_url", "file_id_only"], +) +def test_azure_ai_flattens_a_non_pdf_file_part_to_its_text_as_before(file: dict[str, str]) -> None: + request: Final = AzureAIStudioConfig().transform_request( + model="Llama-3.3-70B-Instruct", + messages=[{"role": "user", "content": [{"type": "text", "text": "summarize"}, {"type": "file", "file": file}]}], + optional_params={}, + litellm_params={}, + headers={}, + ) + + assert request["messages"] == [{"role": "user", "content": "summarize"}] + + def test_translate_developer_role_hoists_a_later_developer_message_for_a_model_that_rejects_it( declare_mid_conversation_system: Callable[[str, bool], None], ) -> None: diff --git a/tests/unit/llms/gigachat/chat/test_gigachat_chat_transformation.py b/tests/unit/llms/gigachat/chat/test_gigachat_chat_transformation.py index f2224e84ff3..115d81bfb90 100644 --- a/tests/unit/llms/gigachat/chat/test_gigachat_chat_transformation.py +++ b/tests/unit/llms/gigachat/chat/test_gigachat_chat_transformation.py @@ -8,6 +8,7 @@ get_model_response_iterator, and get_error_class. """ import json +from typing import Final from unittest.mock import MagicMock, patch import httpx @@ -512,6 +513,20 @@ class TestTransformMessages: assert result[0]["content"] == "look" assert "file-456" in result[0]["attachments"] + def test_list_content_with_pdf_file_part_uploads_it_as_an_attachment(self) -> None: + pdf_data_url: Final = "data:application/pdf;base64,JVBERi0xLjQK" + content: Final = [ + {"type": "text", "text": "summarize"}, + {"type": "file", "file": {"filename": "report.pdf", "file_data": pdf_data_url}}, + ] + with patch.object(self.config, "_upload_image", return_value="file-789") as upload: + result: Final = self.config._transform_messages([ + {"role": "user", "content": content} + ]) + upload.assert_called_once_with(pdf_data_url) + assert result[0]["content"] == "summarize" + assert result[0]["attachments"] == ["file-789"] + class TestTransformRequest: def setup_method(self): diff --git a/tests/unit/llms/oci/chat/test_oci_chat_transformation.py b/tests/unit/llms/oci/chat/test_oci_chat_transformation.py index 8c96bd91049..0535279d177 100644 --- a/tests/unit/llms/oci/chat/test_oci_chat_transformation.py +++ b/tests/unit/llms/oci/chat/test_oci_chat_transformation.py @@ -1,4 +1,5 @@ import datetime +from typing import Final import httpx import pytest import json @@ -2164,3 +2165,32 @@ class TestOpenAIPrefixMaxTokensKey: chat_request = _oci_chat_request("openai.gpt-5", {"max_tokens": 64}) assert chat_request["maxCompletionTokens"] == 64 assert "maxTokens" not in chat_request + + +def test_transform_request_sends_a_pdf_file_part_as_an_image_data_url() -> None: + pdf_data_url: Final = "data:application/pdf;base64,JVBERi0xLjQK" + transformed_request: Final = OCIChatConfig().transform_request( + model="google.gemini-2.5-flash", + messages=[ + { + "role": "user", + "content": [ + {"type": "text", "text": "Summarize this."}, + {"type": "file", "file": {"filename": "report.pdf", "file_data": pdf_data_url}}, + ], + } + ], + optional_params={"oci_compartment_id": TEST_COMPARTMENT_ID}, + litellm_params={}, + headers={}, + ) + + assert transformed_request["chatRequest"]["messages"] == [ + { + "role": "USER", + "content": [ + {"type": "TEXT", "text": "Summarize this."}, + {"type": "IMAGE", "imageUrl": {"url": pdf_data_url}}, + ], + } + ] diff --git a/tests/unit/llms/sap/chat/test_sap_transformation.py b/tests/unit/llms/sap/chat/test_sap_transformation.py index 41e75b7bfc7..f95bf9b7845 100644 --- a/tests/unit/llms/sap/chat/test_sap_transformation.py +++ b/tests/unit/llms/sap/chat/test_sap_transformation.py @@ -1,4 +1,6 @@ import warnings +from typing import Final + import pytest from pydantic import ValidationError @@ -908,6 +910,22 @@ class TestCacheControl: assert "cache_control" not in content[0] assert content[1].get("cache_control") == {"type": "ephemeral"} + def test_pdf_file_part_reaches_sap_as_an_image_url_data_url(self) -> None: + pdf_data_url: Final = "data:application/pdf;base64,JVBERi0xLjQK" + messages: Final = [ + { + "role": "user", + "content": [ + {"type": "text", "text": "Summarize this."}, + {"type": "file", "file": {"filename": "report.pdf", "file_data": pdf_data_url}}, + ], + } + ] + body: Final = self._transform("anthropic--claude-3-5-sonnet", messages) + content: Final = self._template(body)[0]["content"] + assert content[1]["type"] == "image_url" + assert content[1]["image_url"]["url"] == pdf_data_url + def test_message_level_cache_control_folded_onto_trailing_image_block(self): messages = [ { diff --git a/tests/unit/llms/snowflake/chat/test_snowflake_chat_transformation.py b/tests/unit/llms/snowflake/chat/test_snowflake_chat_transformation.py index 36c319d9524..a31f15551e9 100644 --- a/tests/unit/llms/snowflake/chat/test_snowflake_chat_transformation.py +++ b/tests/unit/llms/snowflake/chat/test_snowflake_chat_transformation.py @@ -18,6 +18,7 @@ import pytest import litellm from litellm import completion, acompletion from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler +from litellm.llms.anthropic.messages.adapters.transformation import LiteLLMAnthropicMessagesAdapter from litellm.llms.snowflake.chat.transformation import SnowflakeConfig, SnowflakeStreamingHandler from litellm.types.utils import ModelResponse @@ -479,6 +480,35 @@ class TestSnowflakeCortexClaudeFixes: } ] + def test_messages_api_pdf_reaches_cortex_claude_as_a_document_block(self) -> None: + translated: Final = LiteLLMAnthropicMessagesAdapter().translate_anthropic_messages_to_openai( + messages=[ + { + "role": "user", + "content": [ + { + "type": "document", + "source": {"type": "base64", "media_type": "application/pdf", "data": "ZmFrZQ=="}, + "cache_control": {"type": "ephemeral"}, + }, + {"type": "text", "text": "What is the codeword?"}, + ], + } + ], + model="snowflake/claude-sonnet-4-6", + ) + + body: Final = self._transform(translated) + + assert body["messages"][0]["content"] == [ + { + "type": "document", + "source": {"type": "base64", "media_type": "application/pdf", "data": "ZmFrZQ=="}, + "cache_control": {"type": "ephemeral"}, + }, + {"type": "text", "text": "What is the codeword?"}, + ] + def test_multipart_tool_result_preserves_text_and_converts_image(self): body = self._transform( [ diff --git a/tests/unit/llms/vertex_ai/vertex_ai_partner_models/gemma/test_vertex_ai_gemma_global_endpoint.py b/tests/unit/llms/vertex_ai/vertex_ai_partner_models/gemma/test_vertex_ai_gemma_global_endpoint.py index 85a2124ab02..98d0d314977 100644 --- a/tests/unit/llms/vertex_ai/vertex_ai_partner_models/gemma/test_vertex_ai_gemma_global_endpoint.py +++ b/tests/unit/llms/vertex_ai/vertex_ai_partner_models/gemma/test_vertex_ai_gemma_global_endpoint.py @@ -13,6 +13,7 @@ These tests verify that: import json import os +from typing import Final from unittest.mock import AsyncMock, MagicMock, patch import pytest @@ -422,3 +423,52 @@ async def test_vertex_ai_gemma_vision_passthrough(): assert isinstance(content, list), f"Expected list content, got: {content}" image_parts = [p for p in content if p.get("type") == "image_url"] assert image_parts, f"No image_url part in forwarded message content: {content}" + + +@pytest.mark.asyncio +async def test_vertex_ai_gemma_pdf_file_part_is_sent_as_image_url() -> None: + pdf_data_url: Final = "data:application/pdf;base64,JVBERi0xLjQKJcfsj6IKMSAwIG9iago8PD4+CmVuZG9iagp0cmFpbGVyCjw8Pj4KJSVFT0YK" + messages: Final = [ + { + "role": "user", + "content": [ + {"type": "text", "text": "What does the PDF say?"}, + {"type": "file", "file": {"file_data": pdf_data_url, "filename": "report.pdf"}}, + ], + } + ] + + mock_response: Final = MagicMock() + mock_response.status_code = 200 + mock_response.headers = {} + mock_response.json.return_value = _MOCK_RESPONSE_JSON + + mock_vertexai: Final = MagicMock() + mock_vertexai.preview = MagicMock() + + with ( + patch("litellm.llms.custom_httpx.http_handler.AsyncHTTPHandler") as mock_http_handler, + patch( + "litellm.llms.vertex_ai.vertex_ai_partner_models.main.VertexAIPartnerModels._ensure_access_token", + return_value=("fake-token", "test-project"), + ), + patch.dict( + "sys.modules", + {"vertexai": mock_vertexai, "vertexai.preview": mock_vertexai.preview}, + ), + patch.dict(litellm.model_cost, _GEMMA_MODEL_COST_ENTRY, clear=False), + ): + mock_http_handler.return_value.post = AsyncMock(return_value=mock_response) + + await litellm.acompletion( + model="vertex_ai/google/gemma-4-26b-a4b-it-maas", + messages=messages, + vertex_ai_project="test-project", + ) + + body: Final = json.loads(mock_http_handler.return_value.post.call_args.kwargs["data"]) + user_msg: Final = next(m for m in body["messages"] if m["role"] == "user") + assert user_msg["content"] == [ + {"type": "text", "text": "What does the PDF say?"}, + {"type": "image_url", "image_url": {"url": pdf_data_url}}, + ] diff --git a/tests/unit/proxy/guardrails/test_prompt_security_guardrails.py b/tests/unit/proxy/guardrails/test_prompt_security_guardrails.py index e66e19dd1b4..b08646f1611 100644 --- a/tests/unit/proxy/guardrails/test_prompt_security_guardrails.py +++ b/tests/unit/proxy/guardrails/test_prompt_security_guardrails.py @@ -1,6 +1,7 @@ import asyncio import base64 from collections.abc import Mapping, Sequence +from typing import Final from unittest.mock import AsyncMock, patch import pytest @@ -8,6 +9,7 @@ from fastapi.exceptions import HTTPException from httpx import ReadTimeout, Request, Response import litellm +from litellm.llms.anthropic.chat.guardrail_translation.handler import AnthropicMessagesHandler from litellm.proxy._types import UserAPIKeyAuth from litellm.proxy.guardrails.guardrail_hooks.prompt_security.prompt_security import ( PromptSecurityGuardrail, @@ -1247,3 +1249,54 @@ async def test_check_tool_results_enabled(monkeypatch: pytest.MonkeyPatch): assert "indirect_prompt_injection" in str(excinfo.value.detail) + +@pytest.mark.asyncio +async def test_messages_api_pdf_reaches_file_sanitization_and_is_blocked() -> None: + sanitize_url: Final = "https://test.prompt.security/api/sanitizeFile" + upload: Final = AsyncMock( + return_value=Response(200, json={"jobId": "pdf-job"}, request=Request("POST", sanitize_url)) + ) + poll: Final = AsyncMock( + return_value=Response( + 200, + json={"status": "done", "content": "", "metadata": {"action": "block", "violations": ["Malicious File"]}}, + request=Request("GET", sanitize_url), + ) + ) + guardrail: Final = PromptSecurityGuardrail( + api_key="test-key", + api_base="https://test.prompt.security", + guardrail_name="test-guard", + event_hook="pre_call", + default_on=True, + ) + guardrail.poll_interval = 0 + pdf: Final = b"%PDF-1.4 pelican" + request: Final = { + "model": "gpt-5.6", + "max_tokens": 64, + "messages": [ + { + "role": "user", + "content": [ + { + "type": "document", + "source": { + "type": "base64", + "media_type": "application/pdf", + "data": base64.b64encode(pdf).decode(), + }, + }, + {"type": "text", "text": "Summarize this PDF"}, + ], + } + ], + } + + with patch.object(guardrail.async_handler, "post", upload), patch.object(guardrail.async_handler, "get", poll): + with pytest.raises(HTTPException) as blocked: + await AnthropicMessagesHandler().process_input_messages(request, guardrail) + + assert blocked.value.status_code == 400 + assert "Malicious File" in str(blocked.value.detail) + assert upload.call_args.kwargs["files"]["file"][1] == pdf diff --git a/tests/unit/test_main.py b/tests/unit/test_main.py index e119fe133d5..e10352e0b3d 100644 --- a/tests/unit/test_main.py +++ b/tests/unit/test_main.py @@ -23,7 +23,7 @@ import respx from openai import APITimeoutError, AsyncOpenAI, OpenAI from openai._legacy_response import LegacyAPIResponse from openai.types.chat.chat_completion import ChatCompletion -from pydantic import TypeAdapter +from pydantic import BaseModel, ConfigDict, TypeAdapter import litellm from litellm import ( @@ -10200,6 +10200,309 @@ async def test_atranscription_ignores_a_non_string_base_url_when_api_base_is_set assert response.text == "four score" +_PDF_FILE_PART: Final = { + "type": "file", + "file": {"filename": "report.pdf", "file_data": "data:application/pdf;base64,JVBERi0xLjQK"}, +} +_SUMMARIZE_TEXT_PART: Final = {"type": "text", "text": "Summarize this."} + + +def _pdf_messages() -> list[dict[str, object]]: + return [{"role": "user", "content": [_SUMMARIZE_TEXT_PART, _PDF_FILE_PART]}] + + +class _SentChatRequest(BaseModel): + model_config = ConfigDict(frozen=True) + + messages: list[dict[str, object]] + + +def _sent_messages(respx_mock: respx.MockRouter) -> list[dict[str, object]]: + request: Final[httpx.Request] = respx_mock.calls[0].request + return _SentChatRequest.model_validate_json(request.content).messages + + +def _mock_pdf_chat_route(respx_mock: respx.MockRouter, api_base: str) -> respx.Route: + return respx_mock.post(url__regex=rf"{api_base}/chat/completions.*").mock( + return_value=httpx.Response( + status_code=200, + json={ + "id": "chatcmpl-pdf", + "object": "chat.completion", + "created": 1712697600, + "model": "qa-pdf-model", + "choices": [{"index": 0, "message": {"role": "assistant", "content": "ok"}, "finish_reason": "stop"}], + "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2}, + }, + ) + ) + + +def test_drop_params_drops_a_pdf_the_model_cannot_read(respx_mock: respx.MockRouter) -> None: + api_base: Final = "http://localhost:12347/v1" + _mock_pdf_chat_route(respx_mock, api_base) + + litellm.completion( + model="openai/qa-text-only-model", + messages=_pdf_messages(), + api_base=api_base, + api_key="fake_openai_api_key", + model_info={"supports_pdf_input": False}, + drop_params=True, + ) + + assert _sent_messages(respx_mock) == [{"role": "user", "content": [_SUMMARIZE_TEXT_PART]}] + + +@pytest.mark.parametrize( + "pdf_file", + [ + {"filename": "report.pdf", "file_data": "JVBERi0xLjQK"}, + {"file_id": "file-report", "format": "application/pdf"}, + {"file_id": "file-report", "filename": "Report.PDF"}, + ], + ids=["bare_base64_named_pdf", "file_id_with_pdf_format", "file_id_named_pdf"], +) +def test_drop_params_drops_a_pdf_file_part_without_a_data_url( + respx_mock: respx.MockRouter, pdf_file: dict[str, str] +) -> None: + api_base: Final = "http://localhost:12350/v1" + _mock_pdf_chat_route(respx_mock, api_base) + + litellm.completion( + model="openai/qa-text-only-model", + messages=[{"role": "user", "content": [_SUMMARIZE_TEXT_PART, {"type": "file", "file": pdf_file}]}], + api_base=api_base, + api_key="fake_openai_api_key", + model_info={"supports_pdf_input": False}, + drop_params=True, + ) + + assert _sent_messages(respx_mock) == [{"role": "user", "content": [_SUMMARIZE_TEXT_PART]}] + + +def test_drop_params_drops_a_user_turn_that_held_only_a_pdf(respx_mock: respx.MockRouter) -> None: + api_base: Final = "http://localhost:12352/v1" + _mock_pdf_chat_route(respx_mock, api_base) + + litellm.completion( + model="openai/qa-text-only-model", + messages=[ + {"role": "user", "content": [_PDF_FILE_PART]}, + {"role": "user", "content": [_SUMMARIZE_TEXT_PART]}, + ], + api_base=api_base, + api_key="fake_openai_api_key", + model_info={"supports_pdf_input": False}, + drop_params=True, + ) + + assert _sent_messages(respx_mock) == [{"role": "user", "content": [_SUMMARIZE_TEXT_PART]}] + + +def test_drop_params_keeps_roles_alternating_after_dropping_a_pdf_only_turn(respx_mock: respx.MockRouter) -> None: + api_base: Final = "http://localhost:12353/v1" + _mock_pdf_chat_route(respx_mock, api_base) + + litellm.completion( + model="openai/qa-text-only-model", + messages=[ + {"role": "user", "content": [_SUMMARIZE_TEXT_PART]}, + {"role": "assistant", "content": "Send the file."}, + {"role": "user", "content": [_PDF_FILE_PART]}, + {"role": "assistant", "content": "Got it."}, + {"role": "user", "content": [_SUMMARIZE_TEXT_PART]}, + ], + api_base=api_base, + api_key="fake_openai_api_key", + model_info={"supports_pdf_input": False}, + drop_params=True, + ensure_alternating_roles=True, + ) + + sent_roles: Final = [message["role"] for message in _sent_messages(respx_mock)] + assert sent_roles == ["user", "assistant", "user", "assistant", "user"] + + +def test_drop_params_keeps_an_image_data_url_whose_filename_says_pdf(respx_mock: respx.MockRouter) -> None: + api_base: Final = "http://localhost:12351/v1" + _mock_pdf_chat_route(respx_mock, api_base) + image_part: Final = { + "type": "file", + "file": {"filename": "scan.pdf", "file_data": "data:image/png;base64,iVBORw0KGgo="}, + } + + litellm.completion( + model="openai/qa-text-only-model", + messages=[{"role": "user", "content": [_SUMMARIZE_TEXT_PART, image_part]}], + api_base=api_base, + api_key="fake_openai_api_key", + model_info={"supports_pdf_input": False}, + drop_params=True, + ) + + assert _sent_messages(respx_mock) == [{"role": "user", "content": [_SUMMARIZE_TEXT_PART, image_part]}] + + +def test_pdf_to_a_model_that_cannot_read_it_is_rejected_without_drop_params(respx_mock: respx.MockRouter) -> None: + api_base: Final = "http://localhost:12348/v1" + route: Final = _mock_pdf_chat_route(respx_mock, api_base) + + with pytest.raises(litellm.BadRequestError, match="does not accept PDF input"): + litellm.completion( + model="openai/qa-text-only-model", + messages=_pdf_messages(), + api_base=api_base, + api_key="fake_openai_api_key", + model_info={"supports_pdf_input": False}, + ) + + assert not route.called + + +def test_pdf_to_a_model_with_unknown_pdf_support_is_sent_even_with_drop_params(respx_mock: respx.MockRouter) -> None: + api_base: Final = "http://localhost:12349/v1" + _mock_pdf_chat_route(respx_mock, api_base) + + litellm.completion( + model="openai/qa-unlisted-model", + messages=_pdf_messages(), + api_base=api_base, + api_key="fake_openai_api_key", + drop_params=True, + ) + + assert _sent_messages(respx_mock) == _pdf_messages() + + +def test_pdf_to_ollama_is_rejected_without_drop_params() -> None: + with pytest.raises(litellm.BadRequestError, match="does not accept PDF input"): + litellm.completion(model="ollama_chat/llama3.2", messages=_pdf_messages(), api_base="http://127.0.0.1:9") + + +def _azure_ai_model_without_pdf_input() -> str: + return next( + name + for name, entry in litellm.model_cost.items() + if entry.get("litellm_provider") == "azure_ai" and entry.get("supports_pdf_input") is False + ) + + +def test_pdf_to_a_deployment_whose_base_model_cannot_read_it_is_rejected() -> None: + with pytest.raises(litellm.BadRequestError, match="does not accept PDF input"): + litellm.completion( + model="azure_ai/qa-custom-deployment", + messages=_pdf_messages(), + api_base="http://127.0.0.1:9", + api_key="fake_azure_ai_api_key", + model_info={"base_model": _azure_ai_model_without_pdf_input()}, + ) + + +def test_drop_params_drops_a_pdf_a_deployments_base_model_cannot_read(respx_mock: respx.MockRouter) -> None: + api_base: Final = "http://localhost:12353" + _mock_pdf_chat_route(respx_mock, api_base) + + litellm.completion( + model="azure_ai/qa-custom-deployment", + messages=_pdf_messages(), + api_base=api_base, + api_key="fake_azure_ai_api_key", + model_info={"base_model": _azure_ai_model_without_pdf_input()}, + drop_params=True, + ) + + assert _sent_messages(respx_mock) == [{"role": "user", "content": _SUMMARIZE_TEXT_PART["text"]}] + + +def _azure_ai_model_with_pdf_input() -> str: + return next( + name + for name, entry in litellm.model_cost.items() + if entry.get("litellm_provider") == "azure_ai" and entry.get("supports_pdf_input") is True + ) + + +def test_pdf_to_a_deployment_whose_base_model_param_cannot_read_it_is_rejected() -> None: + with pytest.raises(litellm.BadRequestError, match="does not accept PDF input"): + litellm.completion( + model="azure_ai/qa-custom-deployment", + messages=_pdf_messages(), + api_base="http://127.0.0.1:9", + api_key="fake_azure_ai_api_key", + base_model=_azure_ai_model_without_pdf_input(), + ) + + +def test_drop_params_keeps_a_pdf_a_deployments_base_model_param_reads(respx_mock: respx.MockRouter) -> None: + api_base: Final = "http://localhost:12357" + _mock_pdf_chat_route(respx_mock, api_base) + + litellm.completion( + model="azure_ai/qa-custom-deployment", + messages=_pdf_messages(), + api_base=api_base, + api_key="fake_azure_ai_api_key", + base_model=_azure_ai_model_with_pdf_input(), + drop_params=True, + ) + + assert _sent_messages(respx_mock) == _pdf_messages() + + +def test_drop_params_rejects_a_request_whose_only_user_turn_was_a_pdf(respx_mock: respx.MockRouter) -> None: + api_base: Final = "http://localhost:12354/v1" + route: Final = _mock_pdf_chat_route(respx_mock, api_base) + + with pytest.raises(litellm.BadRequestError, match="no user message is left"): + litellm.completion( + model="openai/qa-text-only-model", + messages=[ + {"role": "system", "content": "You read documents."}, + {"role": "user", "content": [_PDF_FILE_PART]}, + ], + api_base=api_base, + api_key="fake_openai_api_key", + model_info={"supports_pdf_input": False}, + drop_params=True, + ) + + assert not route.called + + +def test_drop_params_drops_a_pdf_an_unlisted_azure_ai_model_never_declared_it_reads( + respx_mock: respx.MockRouter, +) -> None: + api_base: Final = "http://localhost:12355" + _mock_pdf_chat_route(respx_mock, api_base) + + litellm.completion( + model="azure_ai/qa-unlisted-deployment", + messages=_pdf_messages(), + api_base=api_base, + api_key="fake_azure_ai_api_key", + drop_params=True, + ) + + assert _sent_messages(respx_mock) == [{"role": "user", "content": _SUMMARIZE_TEXT_PART["text"]}] + + +def test_drop_params_keeps_a_pdf_an_azure_ai_deployment_declares_it_reads(respx_mock: respx.MockRouter) -> None: + api_base: Final = "http://localhost:12356" + _mock_pdf_chat_route(respx_mock, api_base) + + litellm.completion( + model="azure_ai/qa-unlisted-deployment", + messages=_pdf_messages(), + api_base=api_base, + api_key="fake_azure_ai_api_key", + model_info={"supports_pdf_input": True}, + drop_params=True, + ) + + assert _sent_messages(respx_mock) == _pdf_messages() + + @pytest.mark.parametrize("stream", (False, True)) @pytest.mark.parametrize("asynchronous", (False, True)) @pytest.mark.parametrize("profile,custom_price", ((None, False), ("us.openai.gpt-6-sol", False), (None, True))) diff --git a/tests/unit/test_router_model_cost_isolation.py b/tests/unit/test_router_model_cost_isolation.py index 20b0055f729..91f0b2339c7 100644 --- a/tests/unit/test_router_model_cost_isolation.py +++ b/tests/unit/test_router_model_cost_isolation.py @@ -614,6 +614,63 @@ def test_capability_flag_stays_on_the_deployment_that_set_it(flag: str) -> None: _restore_model_cost_entries(model_keys) +def test_pdf_opt_out_on_one_deployment_does_not_block_its_sibling(respx_mock: respx.MockRouter) -> None: + api_base: Final = "http://localhost:12361/v1" + backend_model: Final = "openai/qa-shared-pdf-backend" + pdf_messages: Final = [ + { + "role": "user", + "content": [ + {"type": "text", "text": "Summarize this."}, + { + "type": "file", + "file": {"filename": "report.pdf", "file_data": "data:application/pdf;base64,JVBERi0xLjQK"}, + }, + ], + } + ] + route: Final = respx_mock.post(url__regex=rf"{api_base}/chat/completions.*").mock( + return_value=httpx.Response( + status_code=200, + json={ + "id": "chatcmpl-pdf", + "object": "chat.completion", + "created": 1712697600, + "model": "qa-shared-pdf-backend", + "choices": [{"index": 0, "message": {"role": "assistant", "content": "ok"}, "finish_reason": "stop"}], + "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2}, + }, + ) + ) + model_keys: Final = { + key: litellm.model_cost.get(key) for key in ("qa-pdf-marked", "qa-pdf-unmarked", backend_model) + } + try: + router: Final = Router( + model_list=[ + { + "model_name": "pdf-blocked", + "litellm_params": {"model": backend_model, "api_base": api_base, "api_key": "fake-key"}, + "model_info": {"id": "qa-pdf-marked", "supports_pdf_input": False}, + }, + { + "model_name": "pdf-allowed", + "litellm_params": {"model": backend_model, "api_base": api_base, "api_key": "fake-key"}, + "model_info": {"id": "qa-pdf-unmarked"}, + }, + ], + ) + + with pytest.raises(litellm.BadRequestError, match="does not accept PDF input"): + router.completion(model="pdf-blocked", messages=pdf_messages) + router.completion(model="pdf-allowed", messages=pdf_messages) + + assert route.call_count == 1 + assert json.loads(route.calls[0].request.content)["messages"] == pdf_messages + finally: + _restore_model_cost_entries(model_keys) + + def test_token_count_settings_stay_on_the_deployment_that_set_them() -> None: """A deployment's ``token_count_model`` and ``token_count_framing`` must not land on the shared ``{provider}/{model}`` key, or a sibling deployment of that model would count with them.""" diff --git a/tests/unit/test_stream_chunk_builder_annotations.py b/tests/unit/test_stream_chunk_builder_annotations.py index 9c7ad4126b0..c710697b39c 100644 --- a/tests/unit/test_stream_chunk_builder_annotations.py +++ b/tests/unit/test_stream_chunk_builder_annotations.py @@ -6,6 +6,8 @@ annotation chunk, losing any annotations that arrived in later chunks. This fix merges annotations from ALL chunks. """ +from typing import Final + from litellm import stream_chunk_builder from litellm.types.utils import Delta, ModelResponseStream, StreamingChoices @@ -189,3 +191,46 @@ def test_stream_chunk_builder_no_annotations(): message = response["choices"][0]["message"] assert not hasattr(message, "annotations") or message.annotations is None + + +def test_stream_chunk_builder_moves_typed_provider_annotations_like_non_streaming() -> None: + url_citation: Final = { + "type": "url_citation", + "url_citation": {"url": "https://example.com", "title": "Example", "start_index": 0, "end_index": 5}, + } + file_annotation: Final = {"type": "file", "file": {"hash": "abc123", "name": "report.pdf"}} + chunks: Final = [ + ModelResponseStream( + id="chatcmpl-test", + created=1700000000, + model="test-model", + object="chat.completion.chunk", + choices=[ + StreamingChoices( + finish_reason=None, + index=0, + delta=Delta(content="Hello", role="assistant", provider_specific_fields={"region": "eu"}), + ) + ], + ), + ModelResponseStream( + id="chatcmpl-test", + created=1700000000, + model="test-model", + object="chat.completion.chunk", + choices=[ + StreamingChoices( + finish_reason="stop", + index=0, + delta=Delta(content=None, annotations=[url_citation, file_annotation]), + ) + ], + ), + ] + + response: Final = stream_chunk_builder(chunks=chunks) + assert response is not None + + message: Final = response["choices"][0]["message"] + assert message.annotations == [url_citation] + assert message.provider_specific_fields == {"region": "eu", "annotations": [file_annotation]} From 77064d6c7086919c28de177d32864e8ffba7c21a Mon Sep 17 00:00:00 2001 From: Mateo Wang <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 29 Oct 2026 08:15:05 -0700 Subject: [PATCH 7/8] feat(vector_stores): add OpenSearch k-NN search as a vector store provider (#47759) * feat(vector_stores): add Amazon OpenSearch as a vector store provider * feat(ui): add Amazon OpenSearch to the vector store provider dropdown * feat(vector_stores): add OpenSearch k-NN search as a vector store provider * test(vector_stores): hold the OpenSearch provider tests alone for the first review The provider source comes back in the next commits on this PR * ci(tests): run the OpenSearch vector store unit tests * fix(vector_stores): sign on.aws and FIPS OpenSearch hosts and reject missing AWS credentials before embedding * fix(vector_stores): regenerate provider endpoint support for opensearch * fix(vector_stores): keep opensearch retrieve, update, and delete off the index and refuse half-set AWS keys and per-account endpoints * fix(vector_stores): return only scalar opensearch attributes and keep credential source output out of 401s * test(vector_stores): inject a failing opensearch credential source instead of running a shell * fix(vector_stores): sign opensearch for the endpoint's region, refuse list, add opensearch_index_name, honor drop_params * fix(vector_stores): refuse a caller-chosen opensearch_index_name on search and rag query * fix(vector_stores): refuse opensearch_index_name only on OpenSearch stores, keep the old 400 text Other providers drop the key again as they did before OpenSearch support, and the four embedding keys keep their original error sentence. * fix(vector_stores): refuse a caller index on slash-form OpenSearch stores --- .github/workflows/test-unit.yml | 1 + .../base_llm/vector_store/transformation.py | 3 + litellm/llms/custom_httpx/llm_http_handler.py | 4 +- litellm/llms/opensearch/__init__.py | 0 .../llms/opensearch/vector_stores/__init__.py | 0 .../vector_stores/transformation.py | 415 +++++++++++++++ .../provider_endpoints_support_backup.json | 27 + litellm/proxy/rag_endpoints/endpoints.py | 6 + .../proxy/vector_store_endpoints/endpoints.py | 26 + litellm/types/utils.py | 1 + litellm/utils.py | 35 +- provider_endpoints_support.json | 27 + tests/unit/AGENTS.md | 3 + tests/unit/llms/opensearch/__init__.py | 0 .../llms/opensearch/vector_stores/__init__.py | 0 .../test_opensearch_transformation.py | 482 ++++++++++++++++++ .../public_endpoints/test_public_endpoints.py | 1 + .../proxy/rag_endpoints/test_rag_endpoints.py | 56 ++ .../test_vector_store_endpoints.py | 52 ++ .../_components/CreateVectorStore.test.tsx | 6 +- .../_components/CreateVectorStore.tsx | 2 +- .../VectorStoreForm.integration.test.tsx | 66 ++- .../_components/VectorStoreForm.tsx | 40 +- .../vector_store_providers.test.tsx | 26 + .../src/components/vector_store_providers.tsx | 62 +++ 25 files changed, 1323 insertions(+), 18 deletions(-) create mode 100644 litellm/llms/opensearch/__init__.py create mode 100644 litellm/llms/opensearch/vector_stores/__init__.py create mode 100644 litellm/llms/opensearch/vector_stores/transformation.py create mode 100644 tests/unit/llms/opensearch/__init__.py create mode 100644 tests/unit/llms/opensearch/vector_stores/__init__.py create mode 100644 tests/unit/llms/opensearch/vector_stores/test_opensearch_transformation.py diff --git a/.github/workflows/test-unit.yml b/.github/workflows/test-unit.yml index e7865817048..509271cbf54 100644 --- a/.github/workflows/test-unit.yml +++ b/.github/workflows/test-unit.yml @@ -322,6 +322,7 @@ jobs: tests/unit/llms/openai_like tests/unit/llms/opencode tests/unit/llms/opensandbox + tests/unit/llms/opensearch tests/unit/llms/openrouter tests/unit/llms/ovhcloud tests/unit/llms/parallel_ai diff --git a/litellm/llms/base_llm/vector_store/transformation.py b/litellm/llms/base_llm/vector_store/transformation.py index 6885183aeba..4924c5a5acc 100644 --- a/litellm/llms/base_llm/vector_store/transformation.py +++ b/litellm/llms/base_llm/vector_store/transformation.py @@ -226,6 +226,9 @@ class BaseVectorStoreConfig: ``get_complete_url`` adds a query string, like Azure, override this.""" return f"{api_base}/{encoded_vector_store_id}" + def get_vector_store_list_url(self, api_base: str) -> str: + return api_base + def get_error_class(self, error_message: str, status_code: int, headers: dict | httpx.Headers) -> BaseLLMException: from ..chat.transformation import BaseLLMException diff --git a/litellm/llms/custom_httpx/llm_http_handler.py b/litellm/llms/custom_httpx/llm_http_handler.py index 8a9ff62cfae..c970b507288 100644 --- a/litellm/llms/custom_httpx/llm_http_handler.py +++ b/litellm/llms/custom_httpx/llm_http_handler.py @@ -11150,7 +11150,7 @@ class BaseLLMHTTPHandler: litellm_params=dict(litellm_params), ) - url: Final = api_base + url: Final = vector_store_provider_config.get_vector_store_list_url(api_base) params: Final[dict[str, object]] = {} if after is not None: @@ -11229,7 +11229,7 @@ class BaseLLMHTTPHandler: litellm_params=dict(litellm_params), ) - url: Final = api_base + url: Final = vector_store_provider_config.get_vector_store_list_url(api_base) params: Final[dict[str, object]] = {} if after is not None: diff --git a/litellm/llms/opensearch/__init__.py b/litellm/llms/opensearch/__init__.py new file mode 100644 index 00000000000..e69de29bb2d diff --git a/litellm/llms/opensearch/vector_stores/__init__.py b/litellm/llms/opensearch/vector_stores/__init__.py new file mode 100644 index 00000000000..e69de29bb2d diff --git a/litellm/llms/opensearch/vector_stores/transformation.py b/litellm/llms/opensearch/vector_stores/transformation.py new file mode 100644 index 00000000000..2d862bc8ea6 --- /dev/null +++ b/litellm/llms/opensearch/vector_stores/transformation.py @@ -0,0 +1,415 @@ +import hashlib +import json +import re +from collections.abc import Mapping, Sequence +from functools import partial +from typing import TYPE_CHECKING, Final, Literal, NoReturn +from urllib.parse import quote, urlsplit + +import httpx +from pydantic import ConfigDict, Field, TypeAdapter, ValidationError + +from litellm._logging import verbose_logger +from litellm.exceptions import AuthenticationError, BadRequestError +from litellm.litellm_core_utils.core_helpers import drops_unsupported_params +from litellm.litellm_core_utils.optional_imports import ensure_optional_import +from litellm.llms.base_llm.chat.transformation import BaseLLMException +from litellm.llms.base_llm.vector_store.transformation import ( + BaseQueryEmbeddingVectorStoreConfig, + VectorStoreEmbeddingExecutor, +) +from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM, SignsRequestsWithAWS, sign_aws_json_post +from litellm.types.llms.base import LiteLLMBaseModel +from litellm.types.llms.bedrock import AwsAuthParams +from litellm.types.router import GenericLiteLLMParams +from litellm.types.utils import EmbeddingResponse +from litellm.types.vector_stores import ( + BaseVectorStoreAuthCredentials, + VectorStoreCreateOptionalRequestParams, + VectorStoreIndexEndpoints, + VectorStoreResultContent, + VectorStoreSearchOptionalRequestParams, + VectorStoreSearchResponse, + VectorStoreSearchResult, +) + +if TYPE_CHECKING: + from botocore.credentials import Credentials + + from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj + +DEFAULT_EMBEDDING_FIELD: Final = "embedding" +DEFAULT_TEXT_FIELD: Final = "text" +DEFAULT_MAX_NUM_RESULTS: Final = 10 +MAX_MAX_NUM_RESULTS: Final = 50 +_PARAM_PREFIX: Final = "opensearch_" +_AWS_HOSTS: Final = ( + re.compile(r"\.(?P[a-z0-9-]+)\.(?Paoss|aoss-fips|es)\.amazonaws\.com(?:\.cn)?\Z"), + re.compile(r"\.(?Paoss|aos)\.(?P[a-z0-9-]+)\.on\.aws\Z"), +) +_ACCOUNT_ID_LABEL: Final = re.compile(r"\A\d{12}\.") +_QUERY_VECTOR: Final = TypeAdapter(tuple[float, ...]) +_ATTRIBUTE_VALUE_TYPES: Final = (str, int, float, bool) +_SEARCH_ONLY_MESSAGE: Final = ( + "The OpenSearch vector store is search-only. Create the k-NN index in OpenSearch, " + "then register it here with its name as opensearch_index_name or vector_store_id." +) +_PER_ACCOUNT_MESSAGE: Final = ( + "api_base is an OpenSearch Serverless per-account endpoint, which needs a collection header the vector store " + "does not send. Use the collection endpoint instead, for example https://..aoss.amazonaws.com." +) +_SIGV4_ONLY_MESSAGE: Final = ( + "The OpenSearch vector store signs every request with AWS SigV4 and does not support api_key. " + "Remove api_key and give the vector store AWS credentials instead." +) +_CREDENTIALS_HINT: Final = ( + "Set aws_access_key_id and aws_secret_access_key, aws_role_name, or aws_profile_name on the vector store, " + "or run LiteLLM with an AWS identity." +) + + +def _bad_request(message: str) -> BadRequestError: + return BadRequestError(message=message, model=None, llm_provider="opensearch") + + +def _unauthenticated(reason: str) -> AuthenticationError: + return AuthenticationError( + message=f"No usable AWS credentials to sign the OpenSearch request ({reason}). {_CREDENTIALS_HINT}", + llm_provider="opensearch", + model=None, + ) + + +def _usable_credentials(credentials: object) -> "Credentials": + from botocore.credentials import Credentials + + if isinstance(credentials, Credentials): + return credentials + raise _unauthenticated("none were found") + + +class _OpenSearchParams(LiteLLMBaseModel): + model_config = ConfigDict(frozen=True, extra="ignore") + + litellm_embedding_model: str | None = None + opensearch_embedding_field: str | None = None + opensearch_index_name: str | None = None + opensearch_text_field: str | None = None + + @property + def embedding_field(self) -> str: + return self.opensearch_embedding_field or DEFAULT_EMBEDDING_FIELD + + @property + def text_field(self) -> str: + return self.opensearch_text_field or DEFAULT_TEXT_FIELD + + def index_name(self, vector_store_id: str) -> str: + return self.opensearch_index_name or vector_store_id + + def require_embedding_model(self) -> str: + if not self.litellm_embedding_model: + raise _bad_request( + "litellm_embedding_model is required for the OpenSearch vector store. It must be the model that " + f"produced the vectors in '{self.embedding_field}', for example openai/text-embedding-3-small." + ) + return self.litellm_embedding_model + + +class _SigningRegion(LiteLLMBaseModel): + model_config = ConfigDict(frozen=True, extra="ignore") + + aws_region_name: str | None = None + + +class _SearchOptions(LiteLLMBaseModel): + model_config = ConfigDict(frozen=True, extra="ignore") + + max_num_results: int | None = None + filters: Mapping[str, object] | None = None + ranking_options: Mapping[str, object] | None = None + rewrite_query: bool | None = None + + @property + def unsupported(self) -> tuple[str, ...]: + requested: Final = ( + ("filters", self.filters), + ("ranking_options", self.ranking_options), + ("rewrite_query", self.rewrite_query), + ) + return tuple(name for name, value in requested if value) + + @property + def k(self) -> int: + if self.max_num_results is None: + return DEFAULT_MAX_NUM_RESULTS + if not 1 <= self.max_num_results <= MAX_MAX_NUM_RESULTS: + raise _bad_request( + f"max_num_results must be between 1 and {MAX_MAX_NUM_RESULTS}, got {self.max_num_results}" + ) + return self.max_num_results + + +_KNOWN_PARAMS: Final = frozenset(name for name in _OpenSearchParams.model_fields if name.startswith(_PARAM_PREFIX)) + + +class _LoggedSearch(LiteLLMBaseModel): + model_config = ConfigDict(frozen=True, extra="ignore") + + query: str | tuple[str, ...] = "" + litellm_params: _OpenSearchParams = _OpenSearchParams() + + @property + def search_query(self) -> str: + return BaseQueryEmbeddingVectorStoreConfig.query_text(self.query) + + +class _Hit(LiteLLMBaseModel): + model_config = ConfigDict(frozen=True) + + id: str = Field(alias="_id") + score: float | None = Field(default=None, alias="_score") + source: Mapping[str, object] = Field(default_factory=dict, alias="_source") + + +class _Hits(LiteLLMBaseModel): + model_config = ConfigDict(frozen=True) + + hits: Sequence[_Hit] + + +class _SearchResponse(LiteLLMBaseModel): + model_config = ConfigDict(frozen=True) + + hits: _Hits + + +def aws_signing_target(url: str) -> tuple[Literal["aoss", "es"], str | None]: + hostname: Final = urlsplit(url).hostname or "" + match: Final = next(filter(None, (pattern.search(hostname) for pattern in _AWS_HOSTS)), None) + if match is None: + return "es", None + return ("aoss" if match["service"].startswith("aoss") else "es"), match["region"] + + +def _search_result(hit: _Hit, params: _OpenSearchParams) -> VectorStoreSearchResult: + text: Final = hit.source.get(params.text_field) + filename: Final = hit.source.get("filename") + return VectorStoreSearchResult( + score=hit.score, + content=[VectorStoreResultContent(type="text", text=text)] if isinstance(text, str) else [], + file_id=hit.id, + filename=filename if isinstance(filename, str) else None, + attributes={ + key: value + for key, value in hit.source.items() + if key not in (params.text_field, params.embedding_field) and isinstance(value, _ATTRIBUTE_VALUE_TYPES) + }, + ) + + +class OpenSearchVectorStoreConfig(BaseQueryEmbeddingVectorStoreConfig, SignsRequestsWithAWS): + def __init__(self, aws: BaseAWSLLM | None = None) -> None: + self._aws: Final = aws if aws is not None else BaseAWSLLM() + + def get_auth_credentials(self, litellm_params: Mapping[str, object]) -> BaseVectorStoreAuthCredentials: + return BaseVectorStoreAuthCredentials() + + def get_vector_store_endpoints_by_type(self) -> VectorStoreIndexEndpoints: + return VectorStoreIndexEndpoints(read=[], write=[]) + + def validate_environment( + self, headers: Mapping[str, object], litellm_params: GenericLiteLLMParams | None + ) -> dict[str, object]: # mutable-ok: the shared HTTP handler requires writable headers + params: Final = dict(litellm_params or {}) + self._signing_credentials(params, self._signing_region(params, str(params.get("api_base") or ""))) + return {**headers, "Content-Type": "application/json"} + + def _signing_region(self, params: Mapping[str, object], api_base: str) -> str: + host_region: Final = aws_signing_target(api_base)[1] + return self._aws.get_aws_region_name_for_non_llm_api_calls( + host_region or _SigningRegion.model_validate(params).aws_region_name + ) + + def _signing_credentials(self, params: Mapping[str, object], region: str) -> "Credentials": + if params.get("api_key"): + raise _bad_request(_SIGV4_ONLY_MESSAGE) + auth: Final = AwsAuthParams.model_validate(params) + if bool(auth.aws_access_key_id) != bool(auth.aws_secret_access_key): + raise _unauthenticated("aws_access_key_id and aws_secret_access_key must be set together") + ensure_optional_import("botocore") + from botocore.exceptions import ( + CredentialRetrievalError, + NoCredentialsError, + PartialCredentialsError, + ProfileNotFound, + ) + + try: + return _usable_credentials(self._aws.resolve_credentials(auth, region)) + except CredentialRetrievalError as error: + verbose_logger.warning("OpenSearch vector store AWS credential source failed: %s", error) + raise _unauthenticated("the AWS credential source failed, see the LiteLLM logs") from None + except (NoCredentialsError, PartialCredentialsError, ProfileNotFound) as error: + raise _unauthenticated(str(error)) from None + + def get_complete_url(self, api_base: str | None, litellm_params: Mapping[str, object]) -> str: + parsed: Final = urlsplit(api_base or "") + if parsed.scheme not in ("http", "https") or not parsed.hostname: + raise _bad_request( + "api_base must be the OpenSearch endpoint URL, for example " + "https://..aoss.amazonaws.com or https://." + ) + if _ACCOUNT_ID_LABEL.match(parsed.hostname) and aws_signing_target(api_base or "")[0] == "aoss": + raise _bad_request(_PER_ACCOUNT_MESSAGE) + return (api_base or "").rstrip("/") + + def get_vector_store_url(self, api_base: str, encoded_vector_store_id: str) -> NoReturn: + raise _bad_request(_SEARCH_ONLY_MESSAGE) + + def get_vector_store_list_url(self, api_base: str) -> NoReturn: + raise _bad_request(_SEARCH_ONLY_MESSAGE) + + @staticmethod + def _params( + litellm_params: Mapping[str, object], + optional_params: VectorStoreSearchOptionalRequestParams, + extra_body: Mapping[str, object] | None, + ) -> tuple[_OpenSearchParams, int]: + unknown: Final = sorted( + key for key in litellm_params if key.startswith(_PARAM_PREFIX) and key not in _KNOWN_PARAMS + ) + if unknown: + raise _bad_request( + f"Unknown OpenSearch vector store parameter(s): {', '.join(unknown)}. " + f"Supported: {', '.join(sorted(_KNOWN_PARAMS))}." + ) + if extra_body: + raise _bad_request("The OpenSearch vector store does not support extra_body.") + try: + options: Final = _SearchOptions.model_validate(optional_params) + except ValidationError as error: + raise _bad_request(f"Invalid vector store search parameters: {error}") from None + if options.unsupported and not drops_unsupported_params(litellm_params.get("drop_params")): + raise _bad_request( + f"The OpenSearch vector store does not support {', '.join(options.unsupported)} yet. " + "Remove it from the request, or set drop_params to drop it." + ) + try: + return _OpenSearchParams.model_validate(litellm_params), options.k + except ValidationError: + raise _bad_request( + "opensearch_embedding_field, opensearch_index_name, opensearch_text_field, and litellm_embedding_model " + "must be strings." + ) from None + + @staticmethod + def _search_request( + vector_store_id: str, + embedding: EmbeddingResponse, + k: int, + params: _OpenSearchParams, + api_base: str, + ) -> tuple[str, dict[str, object]]: # mutable-ok: the provider contract returns a writable JSON request body + vector: Final = _QUERY_VECTOR.validate_python(embedding.data[0]["embedding"]) # pyright: ignore[reportUnknownMemberType] # EmbeddingResponse.data is an untyped list, the vector is validated here + return ( + f"{api_base}/{quote(params.index_name(vector_store_id), safe='')}/_search", + { + "size": k, + "query": {"knn": {params.embedding_field: {"vector": vector, "k": k}}}, + "_source": {"excludes": [params.embedding_field]}, + }, + ) + + def transform_search_vector_store_request( + self, + vector_store_id: str, + query: str | Sequence[str], + vector_store_search_optional_params: VectorStoreSearchOptionalRequestParams, + api_base: str, + litellm_logging_obj: "LiteLLMLoggingObj", + litellm_params: Mapping[str, object], + extra_body: Mapping[str, object] | None = None, + embedding_executor: VectorStoreEmbeddingExecutor | None = None, + ) -> tuple[str, dict[str, object]]: # mutable-ok: the provider contract returns a writable JSON request body + params, k = self._params(litellm_params, vector_store_search_optional_params, extra_body) + embedding: Final = self.query_embedding_executor(embedding_executor, None).embed( + params.require_embedding_model(), self.query_text(query), self.query_embedding_configuration(litellm_params) + ) + return self._search_request(vector_store_id, embedding, k, params, api_base) + + async def atransform_search_vector_store_request( + self, + vector_store_id: str, + query: str | Sequence[str], + vector_store_search_optional_params: VectorStoreSearchOptionalRequestParams, + api_base: str, + litellm_logging_obj: "LiteLLMLoggingObj", + litellm_params: Mapping[str, object], + extra_body: Mapping[str, object] | None = None, + embedding_executor: VectorStoreEmbeddingExecutor | None = None, + ) -> tuple[str, dict[str, object]]: # mutable-ok: the provider contract returns a writable JSON request body + params, k = self._params(litellm_params, vector_store_search_optional_params, extra_body) + embedding: Final = await self.query_embedding_executor(embedding_executor, None).aembed( + params.require_embedding_model(), self.query_text(query), self.query_embedding_configuration(litellm_params) + ) + return self._search_request(vector_store_id, embedding, k, params, api_base) + + def sign_request( + self, + headers: Mapping[str, str], + optional_params: Mapping[str, object], + request_data: Mapping[str, object], + api_base: str, + api_key: str | None = None, + ) -> tuple[dict[str, str], bytes]: # mutable-ok: the shared HTTP handler requires writable headers + region: Final = self._signing_region(optional_params, api_base) + body: Final = json.dumps(request_data) + signed: Final = sign_aws_json_post( + get_credentials=partial(self._signing_credentials, optional_params, region), + service_name=aws_signing_target(api_base)[0], + aws_region_name=region, + url=api_base, + body=body, + headers={ + "Content-Type": "application/json", + "x-amz-content-sha256": hashlib.sha256(body.encode()).hexdigest(), + }, + ) + signed_headers: Final = dict(signed.headers.items()) + signed_names: Final = frozenset(name.lower() for name in signed_headers) + return { + **{name: value for name, value in headers.items() if name.lower() not in signed_names}, + **signed_headers, + }, body.encode() + + def transform_search_vector_store_response( + self, response: httpx.Response, litellm_logging_obj: "LiteLLMLoggingObj" + ) -> VectorStoreSearchResponse: + if not response.is_success: + raise BaseLLMException(status_code=response.status_code, message=response.text, headers=response.headers) + try: + hits: Final = _SearchResponse.model_validate_json(response.content).hits.hits + except ValidationError: + raise BaseLLMException( + status_code=502, + message="OpenSearch answered the search without a hits.hits list.", + headers=response.headers, + ) from None + logged: Final = _LoggedSearch.model_validate(litellm_logging_obj.model_call_details) + return VectorStoreSearchResponse( + object="vector_store.search_results.page", + search_query=logged.search_query, + data=[_search_result(hit, logged.litellm_params) for hit in hits], + ) + + def validate_create_vector_store(self) -> NoReturn: + raise _bad_request(_SEARCH_ONLY_MESSAGE) + + def transform_create_vector_store_request( + self, vector_store_create_optional_params: VectorStoreCreateOptionalRequestParams, api_base: str + ) -> NoReturn: + raise _bad_request(_SEARCH_ONLY_MESSAGE) + + def transform_create_vector_store_response(self, response: httpx.Response) -> NoReturn: + raise _bad_request(_SEARCH_ONLY_MESSAGE) diff --git a/litellm/provider_endpoints_support_backup.json b/litellm/provider_endpoints_support_backup.json index 9f5989e765c..1d487f33835 100644 --- a/litellm/provider_endpoints_support_backup.json +++ b/litellm/provider_endpoints_support_backup.json @@ -4910,6 +4910,33 @@ "vector_stores_search": true } }, + "opensearch": { + "display_name": "Amazon OpenSearch (`opensearch`)", + "url": "https://docs.litellm.ai/docs/providers/opensearch_vector_stores", + "endpoints": { + "chat_completions": false, + "responses": false, + "text_completion": false, + "embeddings": false, + "image_generations": false, + "image_edits": false, + "audio_transcriptions": false, + "audio_speech": false, + "rerank": false, + "ocr": false, + "search": false, + "moderations": false, + "realtime": false, + "video_generations": false, + "batches": false, + "messages": false, + "interactions": false, + "systemone": false, + "a2a": false, + "generateContent": false, + "vector_stores_search": true + } + }, "helicone": { "display_name": "Helicone (`helicone`)", "url": "https://docs.litellm.ai/docs/providers/helicone", diff --git a/litellm/proxy/rag_endpoints/endpoints.py b/litellm/proxy/rag_endpoints/endpoints.py index a7146f13035..2d81dc32181 100644 --- a/litellm/proxy/rag_endpoints/endpoints.py +++ b/litellm/proxy/rag_endpoints/endpoints.py @@ -54,6 +54,7 @@ from litellm.proxy.rag_endpoints.upload_security import ( from litellm.proxy.vector_store_endpoints.endpoints import ( build_request_data_from_managed_vector_store, reject_caller_embedding_selection_params, + reject_caller_opensearch_index_name, ) from litellm.proxy.vector_store_endpoints.utils import ( assert_user_can_access_vector_store_id, @@ -913,6 +914,11 @@ async def rag_query( **retrieval_config, **store_data, } + reject_caller_opensearch_index_name( + caller_body=retrieval_config, + request_fields=merged_retrieval_config, + source="retrieval_config", + ) effective_vector_store_config: Final = ( managed_store_params if managed_store is not None diff --git a/litellm/proxy/vector_store_endpoints/endpoints.py b/litellm/proxy/vector_store_endpoints/endpoints.py index 1b54dddcf4f..b6a0ebe3c58 100644 --- a/litellm/proxy/vector_store_endpoints/endpoints.py +++ b/litellm/proxy/vector_store_endpoints/endpoints.py @@ -17,6 +17,7 @@ from litellm.proxy._types import CommonProxyErrors, UserAPIKeyAuth from litellm.proxy.auth.user_api_key_auth import user_api_key_auth from litellm.proxy.common_request_processing import ProxyBaseLLMRequestProcessing from litellm.proxy.common_utils.http_parsing_utils import read_raw_json_body +from litellm.proxy.route_llm_request import JSON_OBJECT_ADAPTER from litellm.proxy.utils import jsonify_object from litellm.proxy.vector_store_endpoints.utils import ( assert_proxy_admin_for_vector_store_index_management, @@ -24,6 +25,7 @@ from litellm.proxy.vector_store_endpoints.utils import ( get_litellm_managed_vector_store, ) from litellm.repositories.table_repositories import ManagedVectorStoreIndexRepository +from litellm.types.utils import LlmProviders from litellm.types.vector_stores import IndexCreateRequest, IndexListResponse from litellm.vector_stores.utils import VectorStoreWriteBody, callers_metadata_snapshot from litellm.vector_stores.vector_store_registry import VectorStoreIndexRegistry @@ -52,6 +54,24 @@ def reject_caller_embedding_selection_params(payload: Mapping[str, object], sour ) +def reject_caller_opensearch_index_name( + caller_body: Mapping[str, object], request_fields: Mapping[str, object], source: str +) -> None: + provider: Final = request_fields.get("custom_llm_provider") + if ( + isinstance(provider, str) + and provider.split("/", 1)[0] == LlmProviders.OPENSEARCH.value + and "opensearch_index_name" in caller_body + ): + raise HTTPException( + status_code=400, + detail={ + "error": f"'opensearch_index_name' cannot be set in {source}. " + "The index comes from the vector store's server-side registration." + }, + ) + + ######################################################## # OpenAI Compatible Endpoints ######################################################## @@ -138,11 +158,17 @@ async def vector_store_search( data = await read_request_body(request=request) # rebind-ok: pre-existing rebinding on a rename-only line reject_caller_embedding_selection_params(payload=data, source="the search request body") data["vector_store_id"] = vector_store_id + caller_body: Final = JSON_OBJECT_ADAPTER.validate_python(data) # Check for legacy vector store registry (non-managed vector stores) data = await _update_request_data_with_litellm_managed_vector_store_registry( data=data, vector_store_id=vector_store_id, user_api_key_dict=user_api_key_dict ) + reject_caller_opensearch_index_name( + caller_body=caller_body, + request_fields=JSON_OBJECT_ADAPTER.validate_python(data), + source="the search request body", + ) # The managed_vector_stores pre-call hook will handle: # 1. Decoding managed vector store IDs diff --git a/litellm/types/utils.py b/litellm/types/utils.py index 448bd022989..3bdcba75778 100644 --- a/litellm/types/utils.py +++ b/litellm/types/utils.py @@ -4564,6 +4564,7 @@ class LlmProviders(str, Enum): S3_VECTORS = "s3_vectors" VALKEY = "valkey" MONGODB = "mongodb" + OPENSEARCH = "opensearch" HELICONE = "helicone" HYPERBOLIC = "hyperbolic" RECRAFT = "recraft" diff --git a/litellm/utils.py b/litellm/utils.py index 342a2f12310..5fe3e63f31a 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -10460,6 +10460,22 @@ class ProviderConfigManager: return BedrockVectorStore.get_initialized_custom_logger() return None + @staticmethod + def _get_vertex_ai_vector_stores_config(api_type: str | None) -> BaseVectorStoreConfig | None: + if api_type == "rag_api" or api_type is None: + from litellm.llms.vertex_ai.vector_stores.rag_api.transformation import ( + VertexVectorStoreConfig, + ) + + return VertexVectorStoreConfig() + if api_type == "search_api": + from litellm.llms.vertex_ai.vector_stores.search_api.transformation import ( + VertexSearchAPIVectorStoreConfig, + ) + + return VertexSearchAPIVectorStoreConfig() + return None + @staticmethod def get_provider_vector_stores_config( provider: LlmProviders, @@ -10481,18 +10497,7 @@ class ProviderConfigManager: return AzureOpenAIVectorStoreConfig() elif litellm.LlmProviders.VERTEX_AI == provider: - if api_type == "rag_api" or api_type is None: # default to rag_api - from litellm.llms.vertex_ai.vector_stores.rag_api.transformation import ( - VertexVectorStoreConfig, - ) - - return VertexVectorStoreConfig() - elif api_type == "search_api": - from litellm.llms.vertex_ai.vector_stores.search_api.transformation import ( - VertexSearchAPIVectorStoreConfig, - ) - - return VertexSearchAPIVectorStoreConfig() + return ProviderConfigManager._get_vertex_ai_vector_stores_config(api_type) elif litellm.LlmProviders.BEDROCK == provider: from litellm.llms.bedrock.vector_stores.transformation import ( BedrockVectorStoreConfig, @@ -10547,6 +10552,12 @@ class ProviderConfigManager: ) return MongoDBVectorStoreConfig() + elif litellm.LlmProviders.OPENSEARCH == provider: + from litellm.llms.opensearch.vector_stores.transformation import ( + OpenSearchVectorStoreConfig, + ) + + return OpenSearchVectorStoreConfig() return None @staticmethod diff --git a/provider_endpoints_support.json b/provider_endpoints_support.json index 9f5989e765c..1d487f33835 100644 --- a/provider_endpoints_support.json +++ b/provider_endpoints_support.json @@ -4910,6 +4910,33 @@ "vector_stores_search": true } }, + "opensearch": { + "display_name": "Amazon OpenSearch (`opensearch`)", + "url": "https://docs.litellm.ai/docs/providers/opensearch_vector_stores", + "endpoints": { + "chat_completions": false, + "responses": false, + "text_completion": false, + "embeddings": false, + "image_generations": false, + "image_edits": false, + "audio_transcriptions": false, + "audio_speech": false, + "rerank": false, + "ocr": false, + "search": false, + "moderations": false, + "realtime": false, + "video_generations": false, + "batches": false, + "messages": false, + "interactions": false, + "systemone": false, + "a2a": false, + "generateContent": false, + "vector_stores_search": true + } + }, "helicone": { "display_name": "Helicone (`helicone`)", "url": "https://docs.litellm.ai/docs/providers/helicone", diff --git a/tests/unit/AGENTS.md b/tests/unit/AGENTS.md index ee6648c0a82..75c9cc94981 100644 --- a/tests/unit/AGENTS.md +++ b/tests/unit/AGENTS.md @@ -42,6 +42,9 @@ Green if `send_batched` drops every row. pydantic doubles in 12 of 203 files, fa `tests/unit/` mirrors `litellm/`, so a changed file selects its tests by path, not a mapping file. New unit tests go here +A new directory under `tests/unit` runs in CI only once a shard in `.github/workflows/test-unit.yml` names it. +`pytest tests/unit/test_assert_ci_coverage.py` checks that, so run it before you push + ## Writing it so a human can read it A class only when tests share an arrange diff --git a/tests/unit/llms/opensearch/__init__.py b/tests/unit/llms/opensearch/__init__.py new file mode 100644 index 00000000000..e69de29bb2d diff --git a/tests/unit/llms/opensearch/vector_stores/__init__.py b/tests/unit/llms/opensearch/vector_stores/__init__.py new file mode 100644 index 00000000000..e69de29bb2d diff --git a/tests/unit/llms/opensearch/vector_stores/test_opensearch_transformation.py b/tests/unit/llms/opensearch/vector_stores/test_opensearch_transformation.py new file mode 100644 index 00000000000..07836d57af0 --- /dev/null +++ b/tests/unit/llms/opensearch/vector_stores/test_opensearch_transformation.py @@ -0,0 +1,482 @@ +import hashlib +import json +import re +from collections.abc import Callable, Mapping +from datetime import datetime, timezone +from pathlib import Path +from typing import Final +from unittest.mock import MagicMock + +import httpx +import openai +import pytest +from botocore.auth import SigV4Auth +from botocore.awsrequest import AWSRequest +from botocore.credentials import Credentials +from botocore.exceptions import CredentialRetrievalError + +import litellm +from litellm.litellm_core_utils.litellm_logging import Logging +from litellm.llms.base_llm.chat.transformation import BaseLLMException +from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM +from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler +from litellm.llms.opensearch.vector_stores.transformation import OpenSearchVectorStoreConfig +from litellm.types.llms.bedrock import AwsAuthParams +from litellm.types.router import GenericLiteLLMParams +from litellm.types.utils import EmbeddingResponse +from litellm.types.vector_stores import VectorStoreSearchResponse +from litellm.vector_stores import main as vector_stores + +ACCESS_KEY: Final = "AKIDOPENSEARCHTEST" +SECRET_KEY: Final = "opensearch-test-secret" +QUERY: Final = "How many days a year can I work abroad?" +QUERY_VECTOR: Final = [0.1, 0.2, 0.3] +CREDENTIAL_SOURCE_OUTPUT: Final = "credential-process-stderr-detail" +SERVERLESS: Final = "https://abc123.us-east-1.aoss.amazonaws.com" +BASE_PARAMS: Final = { + "api_base": SERVERLESS, + "aws_access_key_id": ACCESS_KEY, + "aws_secret_access_key": SECRET_KEY, + "litellm_embedding_model": "embedding-alias", +} +_AUTHORIZATION: Final = re.compile( + r"\AAWS4-HMAC-SHA256 Credential=(?P[^/]+)/(?P\d{8})/(?P[^/]+)/(?P[^/]+)/aws4_request, " + r"SignedHeaders=(?P[^,]+), Signature=(?P[0-9a-f]{64})\Z" +) + + +class RecordingEmbeddingExecutor: + def __init__(self) -> None: + self.call: Final = MagicMock(return_value=EmbeddingResponse(data=[{"embedding": QUERY_VECTOR}])) + + def embed(self, model: str, query: str, configuration: Mapping[str, object]) -> EmbeddingResponse: + return self.call(model, query, configuration) + + async def aembed(self, model: str, query: str, configuration: Mapping[str, object]) -> EmbeddingResponse: + return self.call(model, query, configuration) + + +def _signature_scope(request: httpx.Request) -> tuple[str, str]: + authorization: Final = _AUTHORIZATION.match(request.headers["authorization"]) + assert authorization is not None, request.headers["authorization"] + signed_names: Final = authorization["signed"].split(";") + assert {"host", "x-amz-content-sha256", "x-amz-date"} <= set(signed_names), signed_names + assert request.headers["x-amz-content-sha256"] == hashlib.sha256(request.content).hexdigest() + signed_request: Final = AWSRequest( + method=request.method, + url=str(request.url), + data=request.content, + headers={name: request.headers[name] for name in signed_names}, + ) + signed_request.context["timestamp"] = request.headers["x-amz-date"] + signer: Final = SigV4Auth(Credentials(ACCESS_KEY, SECRET_KEY), authorization["service"], authorization["region"]) + expected: Final = signer.signature( + signer.string_to_sign(signed_request, signer.canonical_request(signed_request)), signed_request + ) + assert authorization["key"] == ACCESS_KEY + assert authorization["signature"] == expected, "SigV4 signature does not cover the request that was sent" + return authorization["service"], authorization["region"] + + +def _sent(respond: MagicMock) -> tuple[httpx.Request, ...]: + return tuple(call.args[0] for call in respond.call_args_list) + + +def _opensearch_hits(text_field: str, embedding_field: str) -> Mapping[str, object]: + return { + "took": 4, + "timed_out": False, + "hits": { + "total": {"value": 2, "relation": "eq"}, + "max_score": 0.581, + "hits": [ + { + "_index": "litellm-qa-docs", + "_id": "1%3A0%3Aw6ftW6EB", + "_score": 0.581, + "_source": { + text_field: "Employees may work abroad for at most 37 calendar days per year.", + embedding_field: [0.5, 0.5, 0.5], + "embedding_ada002": [0.25, 0.75], + "filename": "hr-remote-work-policy.md", + "category": "hr", + "revision": 7, + "published": True, + "tags": ["hr", "travel"], + "owner": {"team": "people-ops"}, + "reviewed_by": None, + }, + }, + {"_index": "litellm-qa-docs", "_id": "doc-2", "_score": 0.42, "_source": {"category": "facilities"}}, + ], + }, + } + + +EXPECTED_RESULTS: Final = [ + { + "score": 0.581, + "content": [{"type": "text", "text": "Employees may work abroad for at most 37 calendar days per year."}], + "file_id": "1%3A0%3Aw6ftW6EB", + "filename": "hr-remote-work-policy.md", + "attributes": {"filename": "hr-remote-work-policy.md", "category": "hr", "revision": 7, "published": True}, + }, + {"score": 0.42, "content": [], "file_id": "doc-2", "filename": None, "attributes": {"category": "facilities"}}, +] + + +async def _search( + asynchronous: bool, + respond: Callable[[httpx.Request], httpx.Response], + executor: RecordingEmbeddingExecutor, + kwargs: Mapping[str, object], +) -> VectorStoreSearchResponse: + request_kwargs: Final = { + **BASE_PARAMS, + "vector_store_id": "litellm-qa-docs", + "query": QUERY, + "custom_llm_provider": "opensearch", + "_direct_vector_store_embedding_executor": executor, + **kwargs, + } + if asynchronous: + return await litellm.vector_stores.asearch( + **request_kwargs, client=AsyncHTTPHandler(transport=httpx.MockTransport(respond)) + ) + with httpx.Client(transport=httpx.MockTransport(respond)) as transport: + return litellm.vector_stores.search(**request_kwargs, client=HTTPHandler(client=transport)) + + +@pytest.mark.parametrize("asynchronous", [False, True]) +@pytest.mark.parametrize( + "fields", + [{}, {"opensearch_embedding_field": "passage_vector", "opensearch_text_field": "passage"}], +) +@pytest.mark.parametrize( + "api_base,region_params,service,region", + [ + (SERVERLESS, {}, "aoss", "us-east-1"), + (SERVERLESS, {"aws_region_name": "us-west-2"}, "aoss", "us-east-1"), + ("https://search-docs-xyz.eu-west-1.es.amazonaws.com/", {}, "es", "eu-west-1"), + ("https://vpc-docs-xyz.cn-north-1.es.amazonaws.com.cn", {}, "es", "cn-north-1"), + ("https://abc123.us-gov-west-1.aoss-fips.amazonaws.com", {}, "aoss", "us-gov-west-1"), + ("https://abc123.aoss.eu-central-1.on.aws", {}, "aoss", "eu-central-1"), + ("https://search-docs-xyz.aos.ap-southeast-2.on.aws", {}, "es", "ap-southeast-2"), + ("https://search.docs.example", {"aws_region_name": "ap-south-1"}, "es", "ap-south-1"), + ], +) +@pytest.mark.asyncio +async def test_search_sends_a_signed_knn_query_and_returns_an_openai_results_page( + asynchronous: bool, + fields: Mapping[str, str], + api_base: str, + region_params: Mapping[str, str], + service: str, + region: str, +) -> None: + executor: Final = RecordingEmbeddingExecutor() + embedding_field: Final = fields.get("opensearch_embedding_field", "embedding") + text_field: Final = fields.get("opensearch_text_field", "text") + respond: Final = MagicMock(return_value=httpx.Response(200, json=_opensearch_hits(text_field, embedding_field))) + + response: Final = await _search( + asynchronous, + respond, + executor, + {**fields, **region_params, "api_base": api_base, "max_num_results": 3}, + ) + + assert response == {"object": "vector_store.search_results.page", "search_query": QUERY, "data": EXPECTED_RESULTS} + sent: Final = _sent(respond) + assert len(sent) == 1 + assert str(sent[0].url) == f"{api_base.rstrip('/')}/litellm-qa-docs/_search" + assert json.loads(sent[0].content) == { + "size": 3, + "query": {"knn": {embedding_field: {"vector": QUERY_VECTOR, "k": 3}}}, + "_source": {"excludes": [embedding_field]}, + } + assert _signature_scope(sent[0]) == (service, region) + executor.call.assert_called_once_with("embedding-alias", QUERY, {}) + + +@pytest.mark.asyncio +async def test_search_defaults_to_ten_results_like_openai() -> None: + respond: Final = MagicMock(return_value=httpx.Response(200, json={"hits": {"hits": []}})) + + response: Final = await _search(True, respond, RecordingEmbeddingExecutor(), {}) + + assert response == {"object": "vector_store.search_results.page", "search_query": QUERY, "data": []} + body: Final = json.loads(_sent(respond)[0].content) + assert (body["size"], body["query"]["knn"]["embedding"]["k"]) == (10, 10) + + +@pytest.mark.parametrize("asynchronous", [False, True]) +@pytest.mark.asyncio +async def test_a_list_query_is_embedded_and_echoed_as_one_string(asynchronous: bool) -> None: + executor: Final = RecordingEmbeddingExecutor() + respond: Final = MagicMock(return_value=httpx.Response(200, json={"hits": {"hits": []}})) + + response: Final = await _search(asynchronous, respond, executor, {"query": ["remote work", "abroad"]}) + + assert response["search_query"] == "remote work abroad" + executor.call.assert_called_once_with("embedding-alias", "remote work abroad", {}) + + +@pytest.mark.parametrize("asynchronous", [False, True]) +@pytest.mark.parametrize("status", [403, 404, 429, 500]) +@pytest.mark.asyncio +async def test_opensearch_error_keeps_its_status(asynchronous: bool, status: int) -> None: + def respond(request: httpx.Request) -> httpx.Response: + return httpx.Response(status, json={"status": status, "error": {"reason": "upstream refused", "type": "x"}}) + + with pytest.raises(openai.APIStatusError) as error: + await _search(asynchronous, respond, RecordingEmbeddingExecutor(), {}) + + assert error.value.status_code == status + + +@pytest.mark.parametrize( + "overrides,message", + [ + ({"filters": {"type": "eq", "key": "category", "value": "hr"}}, "filters"), + ({"ranking_options": {"score_threshold": 0.5}}, "ranking_options"), + ({"rewrite_query": True}, "rewrite_query"), + ({"max_num_results": 0}, "max_num_results"), + ({"max_num_results": 51}, "max_num_results"), + ({"max_num_results": "ten"}, "Invalid vector store search parameters"), + ({"litellm_embedding_model": None}, "litellm_embedding_model"), + ({"opensearch_text_feild": "passage"}, "opensearch_text_feild"), + ({"opensearch_text_field": 7}, "opensearch_text_field"), + ({"opensearch_index_name": ["docs"]}, "opensearch_index_name"), + ({"extra_body": {"min_score": 0.5}}, "extra_body"), + ({"api_base": None}, "api_base"), + ({"api_base": "abc123.us-east-1.aoss.amazonaws.com"}, "api_base"), + ({"api_base": "https://123456789012.aoss.us-east-1.on.aws"}, "per-account endpoint"), + ({"api_key": "opensearch-basic-auth"}, "api_key"), + ], +) +@pytest.mark.asyncio +async def test_unsupported_search_is_rejected_before_embedding_or_any_request( + overrides: Mapping[str, object], message: str +) -> None: + executor: Final = RecordingEmbeddingExecutor() + respond: Final = MagicMock(return_value=httpx.Response(200, json={"hits": {"hits": []}})) + + with pytest.raises(litellm.BadRequestError, match=message) as error: + await _search(True, respond, executor, overrides) + + assert error.value.status_code == 400 + respond.assert_not_called() + executor.call.assert_not_called() + + +_AMBIENT_AWS_ENV: Final = ( + "AWS_ACCESS_KEY_ID", + "AWS_SECRET_ACCESS_KEY", + "AWS_SESSION_TOKEN", + "AWS_PROFILE", + "AWS_PROFILE_NAME", + "AWS_ROLE_NAME", + "AWS_ROLE_ARN", + "AWS_SESSION_NAME", + "AWS_WEB_IDENTITY_TOKEN", + "AWS_WEB_IDENTITY_TOKEN_FILE", + "AWS_CONTAINER_CREDENTIALS_RELATIVE_URI", + "AWS_CONTAINER_CREDENTIALS_FULL_URI", +) + + +@pytest.mark.parametrize("asynchronous", [False, True]) +@pytest.mark.parametrize( + "credentials,env", + [ + pytest.param({"aws_session_name": "opensearch-no-credentials"}, {}, id="none-anywhere"), + pytest.param({"aws_profile_name": "opensearch-missing-profile"}, {}, id="missing-profile"), + pytest.param({}, {"AWS_ACCESS_KEY_ID": "AKIDPARTIAL"}, id="key-without-secret"), + pytest.param( + {"aws_access_key_id": "AKIDSTOREONLY"}, + {"AWS_ACCESS_KEY_ID": "AKIDAMBIENT", "AWS_SECRET_ACCESS_KEY": "ambient-secret"}, + id="store-key-without-secret-beside-ambient-credentials", + ), + pytest.param( + {"aws_secret_access_key": "store-secret-only"}, + {"AWS_ACCESS_KEY_ID": "AKIDAMBIENT", "AWS_SECRET_ACCESS_KEY": "ambient-secret"}, + id="store-secret-without-key-beside-ambient-credentials", + ), + ], +) +@pytest.mark.asyncio +async def test_missing_aws_credentials_are_a_401_before_embedding_or_any_request( + asynchronous: bool, + credentials: Mapping[str, str], + env: Mapping[str, str], + monkeypatch: pytest.MonkeyPatch, + tmp_path: Path, +) -> None: + for name in _AMBIENT_AWS_ENV: + monkeypatch.delenv(name, raising=False) + for name in ("AWS_CONFIG_FILE", "AWS_SHARED_CREDENTIALS_FILE", "BOTO_CONFIG"): + config_file: Path = tmp_path / name + config_file.write_text("") + monkeypatch.setenv(name, str(config_file)) + monkeypatch.setenv("AWS_EC2_METADATA_DISABLED", "true") + for name, value in env.items(): + monkeypatch.setenv(name, value) + executor: Final = RecordingEmbeddingExecutor() + respond: Final = MagicMock(return_value=httpx.Response(200, json={"hits": {"hits": []}})) + + with pytest.raises(litellm.AuthenticationError, match="No usable AWS credentials") as error: + await _search( + asynchronous, + respond, + executor, + {"aws_access_key_id": None, "aws_secret_access_key": None, **credentials}, + ) + + assert error.value.status_code == 401 + respond.assert_not_called() + executor.call.assert_not_called() + + +class _FailingCredentialSource(BaseAWSLLM): + def resolve_credentials(self, auth_params: AwsAuthParams, aws_region_name: str | None) -> Credentials: + raise CredentialRetrievalError(provider="container-role", error_msg=CREDENTIAL_SOURCE_OUTPUT) + + +def test_failing_credential_source_is_a_401_that_keeps_its_output_out_of_the_reply() -> None: + config: Final = OpenSearchVectorStoreConfig(aws=_FailingCredentialSource()) + + with pytest.raises(litellm.AuthenticationError, match="the AWS credential source failed") as error: + config.validate_environment({}, GenericLiteLLMParams(api_base=SERVERLESS)) + + assert error.value.status_code == 401 + assert CREDENTIAL_SOURCE_OUTPUT not in str(error.value) + + +@pytest.mark.parametrize("asynchronous", [False, True]) +@pytest.mark.asyncio +async def test_opensearch_index_name_picks_the_index_so_the_store_id_can_differ(asynchronous: bool) -> None: + respond: Final = MagicMock(return_value=httpx.Response(200, json=_opensearch_hits("text", "embedding"))) + + response: Final = await _search( + asynchronous, + respond, + RecordingEmbeddingExecutor(), + {"vector_store_id": "east-docs", "opensearch_index_name": "docs/v2"}, + ) + + assert response["data"] == EXPECTED_RESULTS + assert str(_sent(respond)[0].url) == f"{SERVERLESS}/docs%2Fv2/_search" + + +@pytest.mark.parametrize( + "unsupported", + [ + {"filters": {"type": "eq", "key": "category", "value": "hr"}}, + {"ranking_options": {"score_threshold": 0.5}}, + {"rewrite_query": True}, + ], +) +@pytest.mark.asyncio +async def test_unsupported_search_params_are_dropped_under_drop_params(unsupported: Mapping[str, object]) -> None: + respond: Final = MagicMock(return_value=httpx.Response(200, json=_opensearch_hits("text", "embedding"))) + + response: Final = await _search( + True, respond, RecordingEmbeddingExecutor(), {**unsupported, "drop_params": True, "max_num_results": 3} + ) + + assert response["data"] == EXPECTED_RESULTS + assert json.loads(_sent(respond)[0].content) == { + "size": 3, + "query": {"knn": {"embedding": {"vector": QUERY_VECTOR, "k": 3}}}, + "_source": {"excludes": ["embedding"]}, + } + + +@pytest.mark.asyncio +async def test_no_op_values_of_unsupported_params_are_accepted() -> None: + def respond(request: httpx.Request) -> httpx.Response: + return httpx.Response(200, json=_opensearch_hits("text", "embedding")) + + response: Final = await _search( + True, respond, RecordingEmbeddingExecutor(), {"filters": {}, "ranking_options": {}, "rewrite_query": False} + ) + + assert response["data"] == EXPECTED_RESULTS + + +def test_create_is_rejected_because_the_store_is_search_only() -> None: + with pytest.raises(litellm.BadRequestError, match="search-only"): + litellm.vector_stores.create(custom_llm_provider="opensearch") + + +async def _store_operation( + asynchronous: bool, operation: str, respond: Callable[[httpx.Request], httpx.Response] +) -> object: + request_kwargs: Final = {**BASE_PARAMS, "vector_store_id": "litellm-qa-docs", "custom_llm_provider": "opensearch"} + if asynchronous: + return await getattr(vector_stores, f"a{operation}")( + **request_kwargs, client=AsyncHTTPHandler(transport=httpx.MockTransport(respond)) + ) + with httpx.Client(transport=httpx.MockTransport(respond)) as transport: + return getattr(vector_stores, operation)(**request_kwargs, client=HTTPHandler(client=transport)) + + +@pytest.mark.parametrize("asynchronous", [False, True]) +@pytest.mark.parametrize("operation", ["retrieve", "update", "delete"]) +@pytest.mark.asyncio +async def test_store_operations_besides_search_send_nothing_to_the_index(asynchronous: bool, operation: str) -> None: + respond: Final = MagicMock(return_value=httpx.Response(200, json={"acknowledged": True})) + + with pytest.raises(litellm.BadRequestError, match="search-only") as error: + await _store_operation(asynchronous, operation, respond) + + assert error.value.status_code == 400 + respond.assert_not_called() + + +async def _list_stores(asynchronous: bool, respond: Callable[[httpx.Request], httpx.Response]) -> object: + request_kwargs: Final = {**BASE_PARAMS, "custom_llm_provider": "opensearch"} + if asynchronous: + return await vector_stores.alist( + **request_kwargs, client=AsyncHTTPHandler(transport=httpx.MockTransport(respond)) + ) + with httpx.Client(transport=httpx.MockTransport(respond)) as transport: + return vector_stores.list(**request_kwargs, client=HTTPHandler(client=transport)) + + +@pytest.mark.parametrize("asynchronous", [False, True]) +@pytest.mark.asyncio +async def test_listing_stores_sends_nothing_to_the_endpoint(asynchronous: bool) -> None: + respond: Final = MagicMock(return_value=httpx.Response(200, json={"cluster_name": "docs", "tagline": "OpenSearch"})) + + with pytest.raises(litellm.BadRequestError, match="search-only") as error: + await _list_stores(asynchronous, respond) + + assert error.value.status_code == 400 + respond.assert_not_called() + + +@pytest.mark.parametrize( + "response,status", + [ + (httpx.Response(403, json={"status": 403, "error": {"type": "Forbidden"}}), 403), + (httpx.Response(200, json={"took": 1}), 502), + (httpx.Response(200, text="gateway"), 502), + ], +) +def test_response_without_search_hits_is_an_error_not_an_empty_page(response: httpx.Response, status: int) -> None: + logging: Final = Logging( + model="", + messages=[], + stream=False, + call_type="vector_store_search", + start_time=datetime(2026, 10, 20, tzinfo=timezone.utc), + litellm_call_id="opensearch-test", + function_id="opensearch-test", + ) + + with pytest.raises(BaseLLMException) as error: + OpenSearchVectorStoreConfig().transform_search_vector_store_response(response, logging) + + assert error.value.status_code == status diff --git a/tests/unit/proxy/public_endpoints/test_public_endpoints.py b/tests/unit/proxy/public_endpoints/test_public_endpoints.py index e694bca3c8d..da608dbca0e 100644 --- a/tests/unit/proxy/public_endpoints/test_public_endpoints.py +++ b/tests/unit/proxy/public_endpoints/test_public_endpoints.py @@ -566,6 +566,7 @@ ADD_MODEL_UNLISTED_PROVIDERS: Final = frozenset( "mongodb", "nano-gpt", "neosantara", + "opensearch", "parasail", "pinstripes", "poe", diff --git a/tests/unit/proxy/rag_endpoints/test_rag_endpoints.py b/tests/unit/proxy/rag_endpoints/test_rag_endpoints.py index d9f3ceda1ed..bf4b5d5d77a 100644 --- a/tests/unit/proxy/rag_endpoints/test_rag_endpoints.py +++ b/tests/unit/proxy/rag_endpoints/test_rag_endpoints.py @@ -1671,6 +1671,62 @@ def test_rag_query_rejects_caller_embedding_selection_params(client_internal_use assert blocked_key in str(response.json()) +@pytest.mark.parametrize( + ("provider", "expected_status", "expected_searches"), + [("opensearch", 400, 0), ("opensearch/x", 400, 0), ("openai", 200, 1)], +) +def test_rag_query_rejects_caller_opensearch_index_name_only_on_opensearch_stores( + client_internal_user, provider, expected_status, expected_searches +): + """ + Regression: retrieval_config must not pick the index an OpenSearch store searches, + while other providers keep ignoring the key as they did before OpenSearch support. + """ + import litellm + from litellm import Router + from litellm.types.vector_stores import VectorStoreSearchResponse + + store = { + "vector_store_id": "kb", + "custom_llm_provider": provider, + "litellm_params": {"vector_store_id": "kb", "custom_llm_provider": provider}, + } + registry = MagicMock() + registry.get_litellm_managed_vector_store_from_registry.return_value = store + fake_search = AsyncMock( + return_value=VectorStoreSearchResponse(object="vector_store.search_results.page", search_query="q", data=[]) + ) + router = Router( + model_list=[ + { + "model_name": "gpt-4o-mini", + "litellm_params": {"model": "openai/gpt-4o-mini", "api_key": "sk-test", "mock_response": "hi"}, + } + ] + ) + + with ( + patch("litellm.vector_stores.asearch", new=fake_search), # test-quality-ok: the search boundary under test + patch.object(litellm, "vector_store_registry", registry), # test-quality-ok: seeds the store under test + patch("litellm.proxy.proxy_server.llm_router", router), # test-quality-ok: mock-response router for completion + patch( # test-quality-ok: store access is not under test, so the request reaches the search boundary + "litellm.proxy.vector_store_endpoints.utils.can_user_access_vector_store", + new=AsyncMock(return_value=True), + ), + ): + response = client_internal_user.post( + "/v1/rag/query", + json={ + "model": "gpt-4o-mini", + "messages": [{"role": "user", "content": "hello"}], + "retrieval_config": {"vector_store_id": "kb", "opensearch_index_name": "another-teams-index"}, + }, + ) + + assert response.status_code == expected_status, response.json() + assert fake_search.await_count == expected_searches + + EICAR = r"X5O!P%@AP[4\PZX54(P^)7CC)7}$EICAR-STANDARD-ANTIVIRUS-TEST-FILE!$H+H*" INGEST_REQUEST = '{"ingest_options":{"vector_store":{"custom_llm_provider":"openai"}}}' diff --git a/tests/unit/proxy/vector_store_endpoints/test_vector_store_endpoints.py b/tests/unit/proxy/vector_store_endpoints/test_vector_store_endpoints.py index 48642a9b50d..e128b5eb1f1 100644 --- a/tests/unit/proxy/vector_store_endpoints/test_vector_store_endpoints.py +++ b/tests/unit/proxy/vector_store_endpoints/test_vector_store_endpoints.py @@ -3066,6 +3066,58 @@ def test_vector_store_search_rejects_caller_embedding_selection_params(blocked_k assert blocked_key in str(response.json()) +@pytest.mark.parametrize( + ("provider", "expected_status", "expected_searches"), + [("opensearch", 400, 0), ("opensearch/x", 400, 0), ("openai", 200, 1)], +) +def test_vector_store_search_rejects_caller_opensearch_index_name_only_on_opensearch_stores( + provider, expected_status, expected_searches +): + """ + Regression: a caller must not pick the index an OpenSearch store searches, while + other providers keep ignoring the key as they did before OpenSearch support. + """ + from fastapi.testclient import TestClient + + from litellm.proxy.auth.user_api_key_auth import user_api_key_auth + from litellm.proxy.proxy_server import app + from litellm.types.vector_stores import VectorStoreSearchResponse + + store = { + "vector_store_id": "kb", + "custom_llm_provider": provider, + "litellm_params": {"vector_store_id": "kb", "custom_llm_provider": provider}, + } + registry = MagicMock() + registry.get_litellm_managed_vector_store_from_registry.return_value = store + fake_search = AsyncMock( + return_value=VectorStoreSearchResponse(object="vector_store.search_results.page", search_query="q", data=[]) + ) + mock_auth = UserAPIKeyAuth(user_id="test_internal_user", user_role=LitellmUserRoles.INTERNAL_USER.value) + original_overrides = app.dependency_overrides.copy() + app.dependency_overrides[user_api_key_auth] = lambda: mock_auth + try: + with ( + patch("litellm.vector_stores.main.asearch", new=fake_search), # test-quality-ok: the search boundary under test + patch.object(litellm, "vector_store_registry", registry), # test-quality-ok: seeds the store under test + patch( # test-quality-ok: store access is not under test, so the request reaches the search boundary + "litellm.proxy.vector_store_endpoints.utils.can_user_access_vector_store", + new=AsyncMock(return_value=True), + ), + ): + router = litellm.Router(model_list=[]) + with patch("litellm.proxy.proxy_server.llm_router", router): # test-quality-ok: router built over the fake search + response = TestClient(app).post( + "/v1/vector_stores/kb/search", + json={"query": "q", "opensearch_index_name": "another-teams-index"}, + ) + finally: + app.dependency_overrides = original_overrides + + assert response.status_code == expected_status, response.json() + assert fake_search.await_count == expected_searches + + class TestConfigOwnedVectorStores: """Stores declared under ``vector_store_registry`` in config.yaml are owned by the config file""" diff --git a/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/CreateVectorStore.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/CreateVectorStore.test.tsx index ba3d1d0fb7d..f8e2feaa635 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/CreateVectorStore.test.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/CreateVectorStore.test.tsx @@ -18,6 +18,7 @@ vi.mock("@/components/vector_store_providers", () => ({ AZURE_OPENAI: "Azure OpenAI", S3Vectors: "AWS S3 Vectors", Valkey: "Valkey", + OpenSearch: "Amazon OpenSearch", }, vectorStoreProviderMap: { BEDROCK: "bedrock", @@ -25,6 +26,7 @@ vi.mock("@/components/vector_store_providers", () => ({ AZURE_OPENAI: "azure_openai", S3Vectors: "s3_vectors", Valkey: "valkey", + OpenSearch: "opensearch", }, vectorStoreProviderLogoMap: { "Amazon Bedrock": "https://example.com/bedrock.png", @@ -32,6 +34,7 @@ vi.mock("@/components/vector_store_providers", () => ({ "Azure OpenAI": "https://example.com/azure.png", "AWS S3 Vectors": "https://example.com/aws.png", Valkey: "https://example.com/valkey.svg", + "Amazon OpenSearch": "https://example.com/aws.svg", }, getProviderSpecificFields: vi.fn((provider: string) => { if (provider === "s3_vectors") { @@ -254,7 +257,7 @@ describe("CreateVectorStore", () => { expect(screen.getByText("Vector Store Created Successfully")).toBeInTheDocument(); }); - it("should exclude valkey from the provider dropdown since it has no RAG ingestion", async () => { + it("should exclude search-only valkey and opensearch from the provider dropdown since they have no RAG ingestion", async () => { render(); const providerSelect = screen.getByRole("combobox"); @@ -267,6 +270,7 @@ describe("CreateVectorStore", () => { expect(screen.getByText("AWS S3 Vectors")).toBeInTheDocument(); }); expect(screen.queryByText("Valkey")).not.toBeInTheDocument(); + expect(screen.queryByText("Amazon OpenSearch")).not.toBeInTheDocument(); }); it("should display S3 Vectors provider-specific fields when selected", async () => { diff --git a/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/CreateVectorStore.tsx b/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/CreateVectorStore.tsx index 2422e83355d..2d7375431cd 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/CreateVectorStore.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/CreateVectorStore.tsx @@ -36,7 +36,7 @@ const ACCEPTED_DOCUMENT_TYPES = [ const MAX_DOCUMENT_BYTES = 50 * 1024 * 1024; -const RAG_INGEST_UNSUPPORTED_PROVIDERS = new Set(["valkey"]); +const RAG_INGEST_UNSUPPORTED_PROVIDERS = new Set(["valkey", "opensearch"]); const providerItems = Object.entries(VectorStoreProviders) .filter(([providerEnum]) => !RAG_INGEST_UNSUPPORTED_PROVIDERS.has(vectorStoreProviderMap[providerEnum])) diff --git a/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/VectorStoreForm.integration.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/VectorStoreForm.integration.test.tsx index d7ec325653f..f571ec6b9a0 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/VectorStoreForm.integration.test.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/VectorStoreForm.integration.test.tsx @@ -1,4 +1,4 @@ -import { render, screen } from "@testing-library/react"; +import { fireEvent, render, screen } from "@testing-library/react"; import userEvent from "@testing-library/user-event"; import { beforeEach, describe, expect, it, vi } from "vitest"; @@ -134,6 +134,70 @@ describe("VectorStoreForm submit payload", () => { expect(createdPayload().custom_llm_provider).toBe("milvus"); }); + it("registers an existing opensearch index with the proxy embedding model and default field names", async () => { + const user = setupUser(); + renderForm(); + + await chooseProvider(user, "Amazon OpenSearch"); + fireEvent.change(screen.getByPlaceholderText("my-knn-index (OpenSearch index name)"), { + target: { value: "litellm-qa-docs" }, + }); + fireEvent.change(screen.getByPlaceholderText("https://abc123.us-east-1.aoss.amazonaws.com"), { + target: { value: "https://yzatq55tpsuvlwozcal4.us-east-1.aoss.amazonaws.com" }, + }); + await chooseFromSelect(user, 1, "text-embedding-3-small"); + await submit(user); + + const expectedLitellmParams = { + api_base: "https://yzatq55tpsuvlwozcal4.us-east-1.aoss.amazonaws.com", + aws_region_name: undefined, + litellm_embedding_model: "text-embedding-3-small", + opensearch_embedding_field: "embedding", + opensearch_text_field: "text", + opensearch_index_name: undefined, + }; + + await vi.waitFor(() => expect(mockCreate).toHaveBeenCalledTimes(1)); + expect(createdPayload().custom_llm_provider).toBe("opensearch"); + expect(createdPayload().vector_store_id).toBe("litellm-qa-docs"); + expect(createdPayload().litellm_params).toStrictEqual(expectedLitellmParams); + }); + + it("registers an opensearch index under its own name when the vector store id differs", async () => { + const user = setupUser(); + renderForm(); + + await chooseProvider(user, "Amazon OpenSearch"); + fireEvent.change(screen.getByPlaceholderText("my-knn-index (OpenSearch index name)"), { + target: { value: "east-docs" }, + }); + fireEvent.change(screen.getByPlaceholderText("https://abc123.us-east-1.aoss.amazonaws.com"), { + target: { value: "https://yzatq55tpsuvlwozcal4.us-east-1.aoss.amazonaws.com" }, + }); + fireEvent.change(screen.getByPlaceholderText("my-knn-index"), { target: { value: "docs" } }); + await chooseFromSelect(user, 1, "text-embedding-3-small"); + await submit(user); + + await vi.waitFor(() => expect(mockCreate).toHaveBeenCalledTimes(1)); + expect(createdPayload().vector_store_id).toBe("east-docs"); + expect(createdPayload().litellm_params).toMatchObject({ opensearch_index_name: "docs" }); + }); + + it("blocks an opensearch submit that has no endpoint or embedding model", async () => { + const user = setupUser(); + renderForm(); + + await chooseProvider(user, "Amazon OpenSearch"); + fireEvent.change(screen.getByPlaceholderText("my-knn-index (OpenSearch index name)"), { + target: { value: "litellm-qa-docs" }, + }); + await submit(user); + + expect(await screen.findByText("Please input the opensearch endpoint")).toBeInTheDocument(); + expect(screen.getByText("Please select the embedding model")).toBeInTheDocument(); + expect(mockCreate).not.toHaveBeenCalled(); + }); + it("sends a provider field's seeded default even when the user never touches it", async () => { const user = setupUser(); renderForm(); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/VectorStoreForm.tsx b/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/VectorStoreForm.tsx index d82f56f3870..bf5b925eb16 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/VectorStoreForm.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/VectorStoreForm.tsx @@ -36,7 +36,7 @@ import { useZodForm } from "@/lib/forms/useZodForm"; import { SearchSelect } from "@/components/shared/SearchSelect"; import { credentialOptions } from "@/components/shared/credentialOptions"; -const EMBEDDING_MODEL_RENAME_PROVIDERS = new Set(["milvus", "valkey", "mongodb"]); +const EMBEDDING_MODEL_RENAME_PROVIDERS = new Set(["milvus", "valkey", "mongodb", "opensearch"]); export const buildVectorStoreLitellmParams = ( provider: string, @@ -83,6 +83,9 @@ const PROVIDER_FIELD_NAMES = [ "valkey_ssl", "valkey_text_field", "valkey_embedding_field", + "opensearch_embedding_field", + "opensearch_text_field", + "opensearch_index_name", ] as const; type ProviderFieldName = (typeof PROVIDER_FIELD_NAMES)[number]; @@ -119,6 +122,9 @@ const vectorStoreShape = { valkey_ssl: optionalText, valkey_text_field: optionalText, valkey_embedding_field: optionalText, + opensearch_embedding_field: optionalText, + opensearch_text_field: optionalText, + opensearch_index_name: optionalText, }; const vectorStoreSchema = z.object(vectorStoreShape).superRefine((values, ctx) => { @@ -143,6 +149,7 @@ const VECTOR_STORE_ID_PLACEHOLDERS: Record = { "vertex_ai/search_api": 'my-datastore_1234567890 (data store ID from Vertex AI / "Agent Search" console)', valkey: "my-search-index (FT index name in Valkey)", mongodb: "my-vector-index (MongoDB Vector Search index name)", + opensearch: "my-knn-index (OpenSearch index name)", }; const VERTEX_SEARCH_API_WITH_ENGINE_PLACEHOLDER = "Any identifier you'll use to reference this in LiteLLM"; @@ -159,6 +166,8 @@ const EMPTY_VALUES: VectorStoreFormValues = { valkey_ssl: "false", valkey_text_field: "text", valkey_embedding_field: "embedding", + opensearch_embedding_field: "embedding", + opensearch_text_field: "text", }; const labelWithHint = (label: string, hint: string): React.ReactNode => ( @@ -376,6 +385,35 @@ const VectorStoreForm: React.FC = ({ )} + {selectedProvider === "opensearch" && ( + + + OpenSearch Setup Required + +

+ LiteLLM searches a k-NN index you already have in an OpenSearch Serverless collection or an + OpenSearch Service domain. It does not create the index or upload documents. Before creating this + vector store, make sure: +

+
    +
  1. + The index has a knn_vector field holding each document's embedding. Enter the index name as + the Vector Store ID, or under Index Name when another endpoint already uses that ID +
    2. +
  2. + You know which embedding model created those embeddings. That model must be added to this proxy + under Models so you can pick it below. Using a different model returns wrong results +
    3. +
  3. + The proxy has AWS credentials that can read the index, either its own (environment variables or + an IAM role) or a saved credential picked under Existing Credentials. A Serverless collection + needs a data access policy for them, and a domain needs its access policy to allow them +
    4. +
+ + + )} + {selectedProvider === "vertex_rag_engine" && ( diff --git a/ui/litellm-dashboard/src/components/vector_store_providers.test.tsx b/ui/litellm-dashboard/src/components/vector_store_providers.test.tsx index 32d0f5dccc0..b4d890475c7 100644 --- a/ui/litellm-dashboard/src/components/vector_store_providers.test.tsx +++ b/ui/litellm-dashboard/src/components/vector_store_providers.test.tsx @@ -68,6 +68,32 @@ describe("getVectorStoreProviderLogoAndName", () => { expect(byName("mongodb_num_candidates")).toMatchObject({ required: false }); }); + it("registers opensearch with its endpoint, optional region, proxy embedding model, and default field names", () => { + const fields = getProviderSpecificFields("opensearch"); + const byName = (name: string) => fields.find((field) => field.name === name); + + expect(getVectorStoreProviderLogoAndName("opensearch")).toEqual({ + logo: expect.stringContaining("aws"), + displayName: VectorStoreProviders.OpenSearch, + }); + expect(fields.map((field) => field.name)).toEqual([ + "api_base", + "aws_region_name", + "embedding_model", + "opensearch_embedding_field", + "opensearch_text_field", + "opensearch_index_name", + ]); + expect(byName("api_base")).toMatchObject({ required: true }); + expect(byName("aws_region_name")).toMatchObject({ required: false }); + expect(byName("embedding_model")).toMatchObject({ type: "select", required: true }); + expect(byName("embedding_model")).not.toHaveProperty("options"); + expect(byName("opensearch_embedding_field")).toMatchObject({ required: false, initialValue: "embedding" }); + expect(byName("opensearch_text_field")).toMatchObject({ required: false, initialValue: "text" }); + expect(byName("opensearch_index_name")).toMatchObject({ required: false }); + expect(byName("opensearch_index_name")).not.toHaveProperty("initialValue"); + }); + it("registers valkey in the provider, logo, and field maps", () => { expect(vectorStoreProviderMap.Valkey).toBe("valkey"); expect(vectorStoreProviderLogoMap[VectorStoreProviders.Valkey]).toContain("valkey"); diff --git a/ui/litellm-dashboard/src/components/vector_store_providers.tsx b/ui/litellm-dashboard/src/components/vector_store_providers.tsx index 79e711f6f98..e953d6e3db6 100644 --- a/ui/litellm-dashboard/src/components/vector_store_providers.tsx +++ b/ui/litellm-dashboard/src/components/vector_store_providers.tsx @@ -1,4 +1,5 @@ import { getProviderLogoAndName, Providers, providerLogoMap } from "@/components/provider_info_helpers"; +import awsLogo from "../../public/assets/logos/aws.svg"; import milvusLogo from "../../public/assets/logos/milvus.svg"; import mongodbLogo from "../../public/assets/logos/mongodb.svg"; import postgresqlLogo from "../../public/assets/logos/postgresql.svg"; @@ -8,6 +9,7 @@ import valkeyLogo from "../../public/assets/logos/valkey.svg"; export enum VectorStoreProviders { Bedrock = "Amazon Bedrock", S3Vectors = "Amazon S3 Vectors", + OpenSearch = "Amazon OpenSearch", PgVector = "PostgreSQL pgvector (LiteLLM Connector)", VertexRagEngine = "Vertex AI RAG Engine", VertexAiSearch = "Vertex AI Search", @@ -28,6 +30,7 @@ export const vectorStoreProviderMap: Record = { Milvus: "milvus", MongoDB: "mongodb", S3Vectors: "s3_vectors", + OpenSearch: "opensearch", Valkey: "valkey", }; @@ -41,6 +44,7 @@ export const vectorStoreProviderLogoMap: Record = { [VectorStoreProviders.Milvus]: milvusLogo.src, [VectorStoreProviders.MongoDB]: mongodbLogo.src, [VectorStoreProviders.S3Vectors]: s3VectorLogo.src, + [VectorStoreProviders.OpenSearch]: awsLogo.src, [VectorStoreProviders.Valkey]: valkeyLogo.src, }; @@ -313,6 +317,64 @@ export const vectorStoreProviderFields: Record initialValue: "embedding", }, ], + opensearch: [ + { + name: "api_base", + label: "OpenSearch Endpoint", + tooltip: + "The endpoint of your OpenSearch Serverless collection or OpenSearch Service domain, starting with https://", + placeholder: "https://abc123.us-east-1.aoss.amazonaws.com", + required: true, + type: "text", + }, + { + name: "aws_region_name", + label: "AWS Region (optional)", + tooltip: + "The AWS region of the collection or domain. Leave blank to use the region in the endpoint, and set it when the endpoint is a custom domain", + placeholder: "us-east-1", + required: false, + type: "text", + }, + { + name: "embedding_model", + label: "Embedding Model", + tooltip: + "The embedding model on this proxy that created the vectors already stored in your index. LiteLLM embeds every search query with it, so it must be the same model. Add it under Models first if it is not listed", + placeholder: "text-embedding-3-small", + required: true, + type: "select", + }, + { + name: "opensearch_embedding_field", + label: "Vector Field Name", + tooltip: + "The top-level knn_vector field in your index mapping that holds each document's embedding (default: embedding)", + placeholder: "embedding", + required: false, + type: "text", + initialValue: "embedding", + }, + { + name: "opensearch_text_field", + label: "Text Field", + tooltip: + "The top-level field in each document that holds its readable text. LiteLLM returns this text in search results (default: text)", + placeholder: "text", + required: false, + type: "text", + initialValue: "text", + }, + { + name: "opensearch_index_name", + label: "Index Name (optional)", + tooltip: + "The OpenSearch index to search. Leave blank to use the Vector Store ID, and set it to register same-named indexes from two endpoints under different IDs", + placeholder: "my-knn-index", + required: false, + type: "text", + }, + ], s3_vectors: [ { name: "vector_bucket_name", From 6cdb9959df49132e7402b723ea6aeb79fd01242f Mon Sep 17 00:00:00 2001 From: Mateo Wang <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 29 Oct 2026 08:33:36 -0700 Subject: [PATCH 8/8] feat(router): opt in to non-chat deployment filtering (#46257) * feat(proxy): reject chat routes on models declared non-chat via model_info.supported_endpoints A model whose every deployment sets a non-chat model_info.mode and a supported_endpoints list with no chat endpoint now answers /v1/chat/completions, /v1/messages and /v1/responses with a 400 before any provider call. Models without both fields keep today's behavior. * fix(proxy): let fallbacks answer and cover more chat routes in the non-chat model check The check now steps aside when a request, key, team or router fallback would answer, skips such models in the rate-limit fallback loop, ignores mode and supported_endpoints copied from the cost map by old Admin UI saves, treats a blank mode as unset, and also covers /v1/completions, the Responses WebSocket, Gemini generateContent and /v1beta/interactions * fix(proxy): only step aside for fallbacks that can answer chat, gate the Responses WebSocket after alias resolution, and compare declarations to the loaded catalog * fix(proxy): gate the Responses WebSocket with the router after aliases resolve, and ignore client-set internal fallback flags * fix(proxy): walk fallback groups breadth first and release reservations on a rejected WebSocket Each fallback group is looked up once however often the request repeats it, and a WebSocket closed by the endpoint gate runs the failure hook so rate-limit reservations are released * fix(proxy): walk fallback layers without recursion and gate the compact Responses route The fallback walk now folds one layer at a time, so the recursion check passes, and /v1/responses/compact on a model declared non-chat gets the same 400 as the other chat-shaped routes * fix(proxy): keep chat on rows whose cost map entry serves chat A row that carries the cost map key of an entry that serves chat counts as declaring nothing, so a catalog edit that adds a chat path never turns old Admin UI copies into 400s * fix(proxy): check chat sent to a deployment id or backend model name * fix(proxy): check a model the read-through loads from the database * fix(proxy): step aside for fallbacks the check cannot follow and wildcards that serve a backend name A fallback target with no deployments of its own, a client routing stamp, and a wildcard that does not match the requested backend name no longer turn a working request into a 400 * fix(proxy): ignore a client model_group that is not a string when building rate-limit headers A client metadata.model_group sent as an int or a list crashed the error handler with a 500 once the embedding-only check rejected the call before routing. The header helper now treats a non-string stamp as no model group. Drops the pattern clause in the backend-name lookup, since the group lookup already returns wildcard matches. Adds the third audit round's routing-stamp cells. * fix(proxy): keep a non-string client model_group out of spend rows * fix(proxy): validate standard fallbacks before the typed fallback-group lookup * feat(router): opt in to non-chat deployment filtering * fix(router): skip unresolved local rate-limit fallbacks * fix(router): ignore caller-provided endpoint selector values * fix(router): normalize declared chat endpoint prefixes * test(router): cover preflight error-value behavior * test(router): derive replay metadata defaults from model info * fix(router): preserve fallback aliases and check synchronous text calls * test(routing): rewrite the WebSocket model with a pre-call callback, since main no longer runs guardrails at socket admission * test(routing): keep the endpoint gate's owned proxies on the worker's database * fix(router): keep the endpoint gate on model_info patches that omit it and on same-group stream retries A partial PATCH of model_info reset a stored enforce_supported_endpoints true back to false, because the field defaulted to False and the update merges model_dump(exclude_none=True). The field now defaults to None so an omitted flag keeps the stored value The same-group retry of a stream dropped before content called the provider without the endpoint selection context, so it could land on an enforced embedding-only deployment of the same group. It now runs inside endpoint_selection like every other router attempt * fix(router): probe rate-limit fallbacks through get_routable_deployments so the router surface does not grow Main's router fence lets code outside the router use only listed Router members and router_utils modules. The proxy's fallback probe now calls the listed get_routable_deployments with the request's call type, and the gate's 400 moves to litellm.exceptions as EndpointNotSupportedByModelError. It stays separate from main's UnsupportedEndpointError, which the router retries on the group's other deployments, while the gate refuses the whole group * test(responses): cover the WebSocket error frame for a model that refuses the endpoint * fix(proxy): let the rate-limit fallback probe try a fallback this worker has not loaded yet The probe now skips a fallback only when the endpoint check removed every deployment it has, so a model added on another worker still reaches route_request and its database lookup * fix(proxy): keep key router settings off the Responses WebSocket, read unrooted or single-string endpoint lists, and add status to the refusal frame * fix(router): warn when enforce_supported_endpoints has no usable list and match chat endpoints regardless of case, trailing slash, or Gemini path * test(router): type the load-warning test's caplog and router * test(routing): match the endpoint gate's integration tests to base fallback behavior on unloaded fallbacks and the Responses WebSocket --------- Co-authored-by: Mateo --- litellm/exceptions.py | 11 + litellm/proxy/common_request_processing.py | 30 + litellm/proxy/common_utils/callback_utils.py | 5 +- .../proxy/response_api_endpoints/endpoints.py | 16 +- litellm/proxy/route_llm_request.py | 22 + .../spend_tracking/spend_tracking_utils.py | 8 +- litellm/router.py | 60 +- litellm/router_utils/endpoint_eligibility.py | 120 ++ litellm/types/router.py | 1 + .../llm_nonconversational.yaml | 1 + ..._embedding_only_supported_endpoints_e2e.py | 119 ++ tests/e2e/models.py | 3 +- .../test_non_chat_model_endpoint_gate.py | 1515 +++++++++++++++++ .../proxy/common_utils/test_callback_utils.py | 13 + .../test_model_management_endpoints.py | 27 + .../response_api_endpoints/test_endpoints.py | 16 + .../test_spend_tracking_utils.py | 22 + .../proxy/test_common_request_processing.py | 58 +- tests/unit/proxy/test_route_llm_request.py | 25 + .../router_utils/test_endpoint_eligibility.py | 139 ++ tests/unit/test_router/test_router.py | 113 ++ ui/litellm-dashboard/src/lib/http/schema.d.ts | 2 + 22 files changed, 2300 insertions(+), 26 deletions(-) create mode 100644 litellm/router_utils/endpoint_eligibility.py create mode 100644 tests/e2e/llm_translation/test_embedding_only_supported_endpoints_e2e.py create mode 100644 tests/integration/routing/test_non_chat_model_endpoint_gate.py create mode 100644 tests/unit/router_utils/test_endpoint_eligibility.py diff --git a/litellm/exceptions.py b/litellm/exceptions.py index 61e3a8fb5dc..28211b3dab2 100644 --- a/litellm/exceptions.py +++ b/litellm/exceptions.py @@ -323,6 +323,17 @@ class UnsupportedEndpointError(BadRequestError): pass +class EndpointNotSupportedByModelError(BadRequestError): + def __init__(self, message: str, model: str, llm_provider: str = "") -> None: + super().__init__( # pyright: ignore[reportUnknownMemberType] # inherited public exception constructor is untyped + message=message, model=model, llm_provider=llm_provider + ) + self.message = message + self.type = "invalid_request_error" + self.param = "model" + self.code = "400" + + class UnprocessableEntityError(openai.UnprocessableEntityError): def __init__( self, diff --git a/litellm/proxy/common_request_processing.py b/litellm/proxy/common_request_processing.py index fc29a6e3a08..4884e75af4e 100644 --- a/litellm/proxy/common_request_processing.py +++ b/litellm/proxy/common_request_processing.py @@ -129,6 +129,7 @@ from litellm.proxy.guardrails.auto_router_compression import arm_pre_call as _ar from litellm.proxy.hooks.parallel_request_limiter_v3 import get_or_create_request_stash from litellm.proxy.native_compaction import with_proxy_compaction_executor from litellm.proxy.route_llm_request import ( + endpoint_unsupported_by_model, route_request, ) from litellm.proxy.telemetry.request_context import bind_call_id as bind_telemetry_call_id @@ -871,6 +872,17 @@ _DECODED_JSON: Final = TypeAdapter(object) _ROUTER_SETTINGS_ADAPTER: Final[TypeAdapter[Mapping[str, object] | None]] = TypeAdapter(Mapping[str, object] | None) +def _fallback_probe_data( + pristine: object, fallback_model: str, key_aliases: object, group_alias: object +) -> Mapping[str, object]: + return { + **_REQUEST_FIELDS.validate_python(pristine), + "model": resolve_served_model( + fallback_model, _DECODED_JSON.validate_python(litellm.model_alias_map), key_aliases, group_alias + ), + } + + def _request_model(data: Mapping[str, object]) -> str | list[str] | None: try: return _REQUEST_MODEL.validate_python(data.get("model"), strict=True) @@ -2676,6 +2688,17 @@ class ProxyBaseLLMRequestProcessing: original_model, fallback_models, ) + raw_router_settings: Final = _DECODED_JSON.validate_python( + _REQUEST_FIELDS.validate_python(rate_limited_data).get("router_settings_override") + ) + group_alias: Final = _DECODED_JSON.validate_python( + _REQUEST_FIELDS.validate_python(raw_router_settings).get("model_group_alias") + if isinstance(raw_router_settings, Mapping) + else None + ) + key_aliases: Final[object] = cast( # cast-ok: auth aliases are dynamically typed at runtime + object, user_api_key_dict.aliases + ) try: for fallback_model in fallback_models: if fallback_model == original_model: @@ -2691,6 +2714,13 @@ class ProxyBaseLLMRequestProcessing: self.data = independent_snapshot(pristine) self.data["model"] = fallback_model try: + if endpoint_unsupported_by_model( + route_type, + _fallback_probe_data(pristine, fallback_model, key_aliases, group_alias), + llm_router, + user_api_key_dict.team_id, + ): + continue return await self.common_processing_pre_call_logic( request=request, general_settings=general_settings, diff --git a/litellm/proxy/common_utils/callback_utils.py b/litellm/proxy/common_utils/callback_utils.py index f889748c504..d1787f3ff8b 100644 --- a/litellm/proxy/common_utils/callback_utils.py +++ b/litellm/proxy/common_utils/callback_utils.py @@ -436,10 +436,7 @@ def get_model_group_from_litellm_kwargs(kwargs: dict) -> str | None: def get_model_group_from_request_data(data: dict) -> str | None: _metadata: Final = data.get("metadata", None) or {} _model_group: Final = _metadata.get("model_group", None) - if _model_group is not None: - return _model_group - - return None + return _model_group if isinstance(_model_group, str) else None def get_remaining_tokens_and_requests_from_request_data(data: dict) -> dict[str, str]: diff --git a/litellm/proxy/response_api_endpoints/endpoints.py b/litellm/proxy/response_api_endpoints/endpoints.py index 9e642a02d49..772bb3e53d2 100644 --- a/litellm/proxy/response_api_endpoints/endpoints.py +++ b/litellm/proxy/response_api_endpoints/endpoints.py @@ -34,6 +34,7 @@ from litellm._logging import verbose_proxy_logger from litellm.anthropic_interface.exceptions.exception_mapping_utils import AnthropicExceptionMapping from litellm.caching.dual_cache import DualCache from litellm.constants import EMPTY_MAPPING, MANAGED_BACKGROUND_RESPONSE_METADATA_KEY, TERMINAL_RESPONSE_STATUSES +from litellm.exceptions import EndpointNotSupportedByModelError from litellm.integrations.custom_guardrail import ModifyResponseException from litellm.litellm_core_utils.core_helpers import as_str_mapping, get_metadata_variable_name_from_kwargs from litellm.litellm_core_utils.internal_call_metadata import is_background_response, requests_background @@ -2086,6 +2087,14 @@ def _routing_hints_from_first_ws_frame(first_message: str) -> Mapping[str, objec def _responses_ws_failure_frame(failure: Exception) -> str: + if isinstance(failure, EndpointNotSupportedByModelError): + return json.dumps( + { + "type": "error", + "status": 400, + "error": {"type": "invalid_request_error", "message": failure.message.split("\n\n", 1)[0]}, + } + ) proxy_exception: Final = handle_exception_on_proxy(failure) status: Final = int(proxy_exception.code) if str(proxy_exception.code).isdigit() else 500 error_type: Final = ( @@ -2444,7 +2453,12 @@ async def _responses_websocket_session( with contextlib.suppress(Exception): await websocket.send_text(_responses_ws_failure_frame(e)) with contextlib.suppress(Exception): - await websocket.close(code=1011, reason="Internal server error") + await websocket.close( + code=1008 if isinstance(e, EndpointNotSupportedByModelError) else 1011, + reason="Pre-call error" + if isinstance(e, EndpointNotSupportedByModelError) + else "Internal server error", + ) return crash_row return bill_socket_failure(outcome) if isinstance(outcome, Exception) else None diff --git a/litellm/proxy/route_llm_request.py b/litellm/proxy/route_llm_request.py index db1951ec9e2..fea4f342b91 100644 --- a/litellm/proxy/route_llm_request.py +++ b/litellm/proxy/route_llm_request.py @@ -431,6 +431,28 @@ def raise_if_required_body_param_missing( ) +def endpoint_unsupported_by_model( + route_type: str, + data: Mapping[str, object], + llm_router: LitellmRouter | None, + team_id: str | None, +) -> bool: + model: Final = data.get("model") + if not isinstance(model, str) or llm_router is None: + return False + metadata: Final = data.get("metadata") + request_kwargs: Final = { + **data, + "metadata": { + **(metadata if isinstance(metadata, Mapping) else {}), + "user_api_key_team_id": team_id, + }, + } + return bool(llm_router.get_routable_deployments(model, request_kwargs)) and not llm_router.get_routable_deployments( + model, request_kwargs, call_type=route_type + ) + + class MockTestingParamsDisabledError(HTTPException): def __init__(self, params: tuple[str, ...]): super().__init__( diff --git a/litellm/proxy/spend_tracking/spend_tracking_utils.py b/litellm/proxy/spend_tracking/spend_tracking_utils.py index 833ee71ae9c..df9ed611a47 100644 --- a/litellm/proxy/spend_tracking/spend_tracking_utils.py +++ b/litellm/proxy/spend_tracking/spend_tracking_utils.py @@ -464,7 +464,11 @@ def _sl_attribution_fallback( ) -> str: if standard_logging_payload is None: return "" - return standard_logging_payload.get(field) or "" + return _string_or_empty(standard_logging_payload.get(field)) + + +def _string_or_empty(value: object) -> str: + return value if isinstance(value, str) else "" def _deployment_provider(deployment: DeploymentTypedDict) -> str | None: @@ -682,7 +686,7 @@ def get_logging_payload( _model_id: Final = metadata.get("model_info", {}).get("id", "") or _sl_attribution_fallback( standard_logging_payload, "model_id" ) - _model_group: Final = metadata.get("model_group", "") or _sl_attribution_fallback( + _model_group: Final = _string_or_empty(metadata.get("model_group")) or _sl_attribution_fallback( standard_logging_payload, "model_group" ) _api_base: Final = litellm_params.get("api_base", "") or _sl_attribution_fallback( diff --git a/litellm/router.py b/litellm/router.py index ff31f48c49e..31d679f616e 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -287,6 +287,11 @@ from litellm.router_utils.embedding_batch_size_check import ( embedding_deployment_selection, log_rejection_before_provider_call, ) +from litellm.router_utils.endpoint_eligibility import ( + eligible_deployments, + endpoint_selection, + warn_on_unenforced_endpoint_flag, +) from litellm.router_utils.fallback_event_handlers import ( MID_STREAM_FALLBACK_CONTROLS_KEY, AttemptedFallbackTargets, @@ -1449,7 +1454,7 @@ class Router: self.assistants_config = assistants_config self.search_tools = search_tools or [] self.guardrail_list = guardrail_list or [] - self.deployment_names: list = [] # names of models under litellm_params. ex. azure/chatgpt-v-2 + self.deployment_names: list[str] = [] # names of models under litellm_params. ex. azure/chatgpt-v-2 self.deployment_latency_map = {} ### CACHING ### cache_type: Literal["local", "redis", "redis-semantic", "s3", "disk"] = "local" # default to an in-memory cache @@ -6054,12 +6059,13 @@ class Router: kwargs.setdefault("metadata", {}).update({"model_group": model}) # pick the one that is available (lowest TPM/RPM) - deployment: Final = self.get_available_deployment( - model=model, - messages=messages, - specific_deployment=kwargs.pop("specific_deployment", None), - request_kwargs=kwargs, - ) + with endpoint_selection("text_completion", model=model): + deployment: Final = self.get_available_deployment( + model=model, + messages=messages, + specific_deployment=kwargs.pop("specific_deployment", None), + request_kwargs=kwargs, + ) data: Final = deployment["litellm_params"].copy() for k, v in self.default_litellm_params.items(): @@ -6915,6 +6921,7 @@ class Router: """ model_group: Final = cast(str, initial_kwargs.get("model")) # cast-ok: kwargs always carries the model group retry_kwargs: Final = mid_stream_retry_kwargs(initial_kwargs) + handler: Final = retry_kwargs.get("original_generic_function") healthy_deployments, all_deployments = await self._async_get_healthy_deployments( model=model_group, parent_otel_span=get_parent_otel_span_from_kwargs(retry_kwargs), @@ -6945,7 +6952,10 @@ class Router: ) carry_superseded_attempt_usage(retry_kwargs) try: - response = await attempt_function(**retry_kwargs) + with endpoint_selection( + getattr(handler, "__name__", None), getattr(handler, "__module__", None), model=model_group + ): + response = await attempt_function(**retry_kwargs) except Exception as retry_error: # noqa: BLE001 # the fallback chain judges the failed retry wrapped = _stream_fallback_error_for_raised(retry_error, model_group, False, e.llm_provider) if wrapped is None: @@ -9315,9 +9325,18 @@ class Router: Handler for making a call to the .completion()/.embeddings()/etc. functions. """ model_group: Final = kwargs.get("model") - response = original_function(*args, **kwargs) - if coroutine_checker.is_async_callable(response) or inspect.isawaitable(response): - response = await response + generic_function: Final = _OPAQUE_VALUE_ADAPTER.validate_python(kwargs.get("original_generic_function")) + endpoint_function: Final = ( + generic_function if callable(generic_function) else _OPAQUE_VALUE_ADAPTER.validate_python(original_function) + ) + with endpoint_selection( + getattr(endpoint_function, "__name__", None), + getattr(endpoint_function, "__module__", None), + model=_OPAQUE_VALUE_ADAPTER.validate_python(model_group), + ): + response = original_function(*args, **kwargs) + if coroutine_checker.is_async_callable(response) or inspect.isawaitable(response): + response = await response await self.increment_deployment_usage_for_response(response=response, request_kwargs=kwargs) ## PROCESS RESPONSE HEADERS response = await self.set_response_headers(response=response, model_group=model_group, request_kwargs=kwargs) @@ -10638,6 +10657,7 @@ class Router: declared_params: Final = _LITELLM_PARAMS_ADAPTER.validate_python(_litellm_params) warn_on_provider_credential_mismatch(model_name=_model_name, litellm_params=declared_params) warn_on_ignored_model_capability_flags(model_name=_model_name, litellm_params=declared_params) + warn_on_unenforced_endpoint_flag(model_name=_model_name, model_info=_model_info) deployment = without_model_capability_flags( Deployment( **deployment_info, @@ -11441,6 +11461,9 @@ class Router: declared_params: Final = deployment.litellm_params.model_dump(exclude_none=True) warn_on_provider_credential_mismatch(model_name=deployment.model_name, litellm_params=declared_params) warn_on_ignored_model_capability_flags(model_name=deployment.model_name, litellm_params=declared_params) + warn_on_unenforced_endpoint_flag( + model_name=deployment.model_name, model_info=deployment.model_info.model_dump() + ) served: Final = with_admin_api_key_secret_reference(without_model_capability_flags(deployment)) # add to model list @@ -14797,9 +14820,15 @@ class Router: """ return [m for m in self.model_list if m["litellm_params"]["model"] == model] - def get_routable_deployments(self, model: str) -> tuple[DeploymentTypedDict, ...]: + def get_routable_deployments( + self, + model: str, + request_kwargs: Mapping[str, object] = MappingProxyType({}), + call_type: str | None = None, + ) -> tuple[DeploymentTypedDict, ...]: try: - _, candidates = self._common_checks_available_deployment(model=model, request_kwargs={}) # pyright: ignore[reportUnknownMemberType, reportUnknownVariableType] # its signature leaves input and the result untyped + with endpoint_selection(call_type, model=model): + _, candidates = self._common_checks_available_deployment(model=model, request_kwargs=request_kwargs) # pyright: ignore[reportUnknownMemberType, reportUnknownVariableType] # its signature leaves input and the result untyped except (litellm.BadRequestError, ValueError): return () if isinstance(candidates, dict): @@ -15075,7 +15104,10 @@ class Router: model=model, llm_provider="", ) - return selectable + eligible: Final = eligible_deployments(model, selectable) + if isinstance(eligible, tuple): + return list(eligible) + raise eligible def _filter_reserved_deployments( self, diff --git a/litellm/router_utils/endpoint_eligibility.py b/litellm/router_utils/endpoint_eligibility.py new file mode 100644 index 00000000000..4a87e582c97 --- /dev/null +++ b/litellm/router_utils/endpoint_eligibility.py @@ -0,0 +1,120 @@ +from collections.abc import Generator, Mapping, Sequence +from contextlib import contextmanager +from contextvars import ContextVar +from types import MappingProxyType +from typing import Final, TypeVar + +from pydantic import BaseModel, ConfigDict, TypeAdapter, ValidationError, field_validator + +from litellm._logging import verbose_router_logger +from litellm.exceptions import EndpointNotSupportedByModelError + +CHAT_ENDPOINT_BY_CALL: Final[Mapping[str, str]] = MappingProxyType( + { + "completion": "/v1/chat/completions", + "acompletion": "/v1/chat/completions", + "_completion": "/v1/chat/completions", + "_acompletion": "/v1/chat/completions", + "text_completion": "/v1/completions", + "atext_completion": "/v1/completions", + "_atext_completion": "/v1/completions", + "anthropic_messages": "/v1/messages", + "aresponses": "/v1/responses", + "agenerate_content": "generateContent", + "agenerate_content_stream": "streamGenerateContent", + "acreate_interaction": "/v1beta/interactions", + "litellm.interactions.main.acreate": "/v1beta/interactions", + "litellm.interactions.main.create": "/v1beta/interactions", + "_aresponses_websocket": "/v1/responses", + "acompact_responses": "/v1/responses/compact", + } +) + + +def _endpoint_path(endpoint: str) -> str: + rooted: Final = "/" + endpoint.strip().strip("/").casefold() + path: Final = rooted.removeprefix("/openai").removeprefix("/v1beta").removeprefix("/v1") + return "/" + path.rsplit(":", 1)[-1] if ":" in path else path + + +CHAT_ENDPOINTS: Final = frozenset(_endpoint_path(endpoint) for endpoint in CHAT_ENDPOINT_BY_CALL.values()) +_Deployment: Final = TypeVar("_Deployment", bound=Mapping[str, object]) +_ENFORCE_FLAG: Final = TypeAdapter(bool | None) +_REQUEST_SELECTION: Final[ContextVar[tuple[str | None, str | None]]] = ContextVar( + "router_request_endpoint", default=(None, None) +) + + +class _EndpointDeclaration(BaseModel): + model_config = ConfigDict(frozen=True) + + enforce_supported_endpoints: bool | None = None + supported_endpoints: tuple[str, ...] | None = None + + @field_validator("supported_endpoints", mode="before") + @classmethod + def _one_endpoint_as_a_list(cls, value: object) -> object: + return (value,) if isinstance(value, str) else value + + +@contextmanager +def endpoint_selection(call_type: object, module: object = None, model: object = None) -> Generator[None]: + qualified_call: Final = f"{module}.{call_type}" if module == "litellm.interactions.main" else call_type + endpoint: Final = CHAT_ENDPOINT_BY_CALL.get(qualified_call) if isinstance(qualified_call, str) else None + token: Final = _REQUEST_SELECTION.set((endpoint, model if isinstance(model, str) else None)) + try: + yield + finally: + _REQUEST_SELECTION.reset(token) + + +def _enforced_endpoints(model_info: object) -> tuple[str, ...] | None: + try: + declaration: Final = _EndpointDeclaration.model_validate(model_info) + except ValidationError: + return None + return declaration.supported_endpoints if declaration.enforce_supported_endpoints else None + + +def warn_on_unenforced_endpoint_flag(model_name: str, model_info: Mapping[str, object]) -> str | None: + try: + flagged: Final = _ENFORCE_FLAG.validate_python(model_info.get("enforce_supported_endpoints")) + except ValidationError: + return None + if not flagged or _enforced_endpoints(model_info) is not None: + return None + warning: Final = ( + f"Deployment '{model_name}' sets enforce_supported_endpoints without a supported_endpoints list of " + 'strings, so litellm enforces nothing for it. Set supported_endpoints, e.g. ["/v1/embeddings"].' + ) + verbose_router_logger.warning(warning) + return warning + + +def non_chat_endpoints(deployment: Mapping[str, object]) -> frozenset[str] | None: + declared: Final = _enforced_endpoints(deployment.get("model_info")) + if declared is None: + return None + endpoints: Final = frozenset(declared) + normalized: Final = frozenset(_endpoint_path(endpoint) for endpoint in endpoints) + return endpoints if normalized.isdisjoint(CHAT_ENDPOINTS) else None + + +def eligible_deployments( + model: str, deployments: Sequence[_Deployment] +) -> tuple[_Deployment, ...] | EndpointNotSupportedByModelError: + endpoint, requested_model = _REQUEST_SELECTION.get() + if endpoint is None or not deployments: + return tuple(deployments) + eligible: Final = tuple(deployment for deployment in deployments if non_chat_endpoints(deployment) is None) + if eligible: + return eligible + requested: Final = requested_model or model + supported: Final = ", ".join( + sorted(frozenset[str]().union(*(non_chat_endpoints(deployment) or frozenset() for deployment in deployments))) + ) + return EndpointNotSupportedByModelError( + message=f"{endpoint} is not supported by model={requested}. Supported endpoints: {supported or 'none'}", + model=model, + llm_provider="", + ) diff --git a/litellm/types/router.py b/litellm/types/router.py index db8be9e91a3..56dcad4051a 100644 --- a/litellm/types/router.py +++ b/litellm/types/router.py @@ -270,6 +270,7 @@ def _whole_float_as_int(value: object) -> object: class ModelInfo(MirroredPricingParams): + enforce_supported_endpoints: bool | None = None id: str | None # Allow id to be optional on input, but it will always be present as a str in the model instance db_model: bool = False # used for proxy - to separate models which are stored in the db vs. config. updated_at: datetime.datetime | None = None diff --git a/tests/e2e/coverage_registry/llm_nonconversational.yaml b/tests/e2e/coverage_registry/llm_nonconversational.yaml index 330ddf88fc5..367135eae26 100644 --- a/tests/e2e/coverage_registry/llm_nonconversational.yaml +++ b/tests/e2e/coverage_registry/llm_nonconversational.yaml @@ -2,6 +2,7 @@ - {id: llm.completions.openai.basic.nonstream.works, module: llm, tier: P0, subject_endpoint: completions, route: openai, capability: basic, streaming: nonstream, assertions: [works], source: "test_completions_endpoint_e2e.py", rationale: "Legacy text /completions endpoint, second-highest production request volume"} - {id: llm.embeddings.openai.basic.nonstream.works, module: llm, tier: P0, subject_endpoint: embeddings, route: openai, capability: basic, streaming: nonstream, assertions: [works], source: "test_embeddings_endpoint_e2e.py:23", rationale: "Core endpoint, live vector response"} - {id: llm.embeddings.openai.input_validation.nonstream.works, module: llm, tier: P1, subject_endpoint: embeddings, route: openai, capability: input_validation, streaming: nonstream, assertions: [works], source: "vendor strategy §9.3 / LIT-4778", rationale: "Missing model/input on /embeddings return client errors"} +- {id: llm.embeddings.openai.supported_endpoints_gate.nonstream.works, module: llm, tier: P1, subject_endpoint: embeddings, route: openai, capability: input_validation, streaming: nonstream, assertions: [works], fail_before_fix: proven, source: "test_embedding_only_supported_endpoints_e2e.py", rationale: "Explicit endpoint restrictions preserve real embeddings, refuse chat with a gateway 400, and route chat through a live fallback; zero-provider-call contracts live in integration/routing/test_non_chat_model_endpoint_gate.py"} - {id: llm.embeddings.openai.basic.nonstream.cost_logged, module: llm, tier: P0, subject_endpoint: embeddings, route: openai, capability: basic, streaming: nonstream, assertions: [cost_logged], source: "SPEND_TRACKING_COVERAGE_MATRIX.md:34", rationale: "Cost tracking on embeddings"} - {id: llm.embeddings.openai.passthrough.nonstream.cost_logged, module: llm, tier: P0, subject_endpoint: embeddings, route: openai, capability: basic, streaming: nonstream, assertions: [cost_logged], source: "test_passthrough_e2e.py", rationale: "POST /openai_passthrough/v1/embeddings is costed; the route wrote no spend row at all, so budgets never saw traffic OpenAI was billing for (GitHub issue #36646)"} - {id: llm.embeddings.azure_openai.basic.nonstream.works, module: llm, tier: P0, subject_endpoint: embeddings, route: azure_openai, capability: basic, streaming: nonstream, assertions: [works], source: "llms/azure/azure.py", rationale: "Azure embeddings via translation"} diff --git a/tests/e2e/llm_translation/test_embedding_only_supported_endpoints_e2e.py b/tests/e2e/llm_translation/test_embedding_only_supported_endpoints_e2e.py new file mode 100644 index 00000000000..c45adfeeb84 --- /dev/null +++ b/tests/e2e/llm_translation/test_embedding_only_supported_endpoints_e2e.py @@ -0,0 +1,119 @@ +"""Live provider checks for embeddings, the chat refusal, and chat fallback from an explicitly restricted model.""" + +from __future__ import annotations + +import dataclasses +from typing import Final + +import pytest +from e2e_config import unique_marker +from e2e_metadata import Domain, Mode, Provider, Route, Subject, meta +from lifecycle import ResourceManager +from models import LiteLLMParamsBody, ModelInfoBody, ModelMode, ModelNewBody +from openai import BadRequestError, OpenAI +from proxy_client import ProxyClient +from sdk_clients import NO_PROXY_CACHE, SdkClients + +pytestmark = pytest.mark.e2e + +EMBEDDING_BACKEND: Final = "openai/text-embedding-3-small" +CHAT_FALLBACK_BACKEND: Final = "openai/gpt-5.4-nano" +EMBEDDINGS_ONLY: Final = ("/v1/embeddings",) +PROMPT: Final = "Say this is a test!" + + +def _subject(route: Route) -> Subject: + return Subject( + domain=Domain.LLM_TRANSLATION, + route=route, + providers=(Provider.OPENAI,), + models=(EMBEDDING_BACKEND,), + mode=Mode.NONSTREAM, + ) + + +def _register( + proxy: ProxyClient, + resources: ResourceManager, + prefix: str, + supported_endpoints: tuple[str, ...] | None, + backend: str = EMBEDDING_BACKEND, + mode: ModelMode = "embedding", +) -> str: + model: Final = f"{prefix}-{unique_marker()}" + model_id: Final = proxy.register_model( + ModelNewBody( + model_name=model, + litellm_params=LiteLLMParamsBody(model=backend, api_key="os.environ/OPENAI_API_KEY"), + model_info=ModelInfoBody( + enforce_supported_endpoints=supported_endpoints is not None, + mode=mode, + supported_endpoints=None if supported_endpoints is None else list(supported_endpoints), + ), + ) + ) + resources.defer(lambda: proxy.delete_model(model_id)) + return model + + +def _openai_at(sdk: SdkClients, api_prefix: str, key: str) -> OpenAI: + return dataclasses.replace(sdk, base_url=f"{sdk.base_url}{api_prefix}").openai(key) + + +class TestEmbeddingOnlySupportedEndpoints: + @pytest.mark.covers("llm.embeddings.openai.supported_endpoints_gate.nonstream.works") + @meta(_subject(Route.EMBEDDINGS)) + @pytest.mark.parametrize( + ("prefix", "supported_endpoints"), + ( + pytest.param("e2e-embed-only", EMBEDDINGS_ONLY, id="embed-only"), + pytest.param("e2e-embed-control", None, id="control"), + ), + ) + def test_embeddings_return_a_vector( + self, + proxy: ProxyClient, + resources: ResourceManager, + sdk: SdkClients, + prefix: str, + supported_endpoints: tuple[str, ...] | None, + ) -> None: + model: Final = _register(proxy, resources, prefix, supported_endpoints) + embeddings: Final = sdk.openai(resources.key()).embeddings.create( + model=model, input=PROMPT, extra_body=NO_PROXY_CACHE + ) + assert embeddings.data, f"/embeddings returned no data for {model}: {embeddings!r}" + assert any(component != 0.0 for component in embeddings.data[0].embedding), ( + f"/embeddings returned an empty or all-zero vector for {model}" + ) + + @pytest.mark.covers("llm.embeddings.openai.supported_endpoints_gate.nonstream.works") + @meta(dataclasses.replace(_subject(Route.CHAT_COMPLETIONS), models=(EMBEDDING_BACKEND, CHAT_FALLBACK_BACKEND))) + def test_chat_completions_with_a_fallback_are_answered_by_the_fallback( + self, proxy: ProxyClient, resources: ResourceManager, sdk: SdkClients + ) -> None: + model: Final = _register(proxy, resources, "e2e-embed-only", EMBEDDINGS_ONLY) + fallback: Final = _register(proxy, resources, "e2e-chat-fallback", None, CHAT_FALLBACK_BACKEND, "chat") + completion: Final = _openai_at(sdk, "/v1", resources.key()).chat.completions.create( + model=model, + messages=[{"role": "user", "content": PROMPT}], + extra_body={**NO_PROXY_CACHE, "fallbacks": [fallback]}, + ) + assert completion.choices and completion.choices[0].message.content, ( + f"the fallback {fallback} should answer a chat call to {model}, got {completion!r}" + ) + + @pytest.mark.covers("llm.embeddings.openai.supported_endpoints_gate.nonstream.works") + @meta(_subject(Route.CHAT_COMPLETIONS)) + def test_chat_completions_without_a_fallback_are_refused_by_the_gateway( + self, proxy: ProxyClient, resources: ResourceManager, sdk: SdkClients + ) -> None: + model: Final = _register(proxy, resources, "e2e-embed-only", EMBEDDINGS_ONLY) + with pytest.raises(BadRequestError) as refused: + _ = _openai_at(sdk, "/v1", resources.key()).chat.completions.create( + model=model, messages=[{"role": "user", "content": PROMPT}], extra_body=NO_PROXY_CACHE + ) + assert ( + f"/v1/chat/completions is not supported by model={model}. Supported endpoints: /v1/embeddings" + in refused.value.message + ), f"chat on {model} should be refused by the gateway before any provider call, got {refused.value!r}" diff --git a/tests/e2e/models.py b/tests/e2e/models.py index afddc7e25d6..504adec5401 100644 --- a/tests/e2e/models.py +++ b/tests/e2e/models.py @@ -1416,7 +1416,7 @@ class LiteLLMParamsBody(BaseModel): order: int | None = None -ModelMode = Literal["chat", "batch", "realtime", "image_generation"] +ModelMode = Literal["chat", "embedding", "batch", "realtime", "image_generation"] class TokenCountModelBody(BaseModel, frozen=True): @@ -1425,6 +1425,7 @@ class TokenCountModelBody(BaseModel, frozen=True): class ModelInfoBody(BaseModel): + enforce_supported_endpoints: bool | None = None # id is left unset so the proxy assigns a unique model_id per deployment. # Pinning it to the model_name made re-registrations of a fixed-name model # (e.g. the batch suite's openai-batch) collide on the model_id unique diff --git a/tests/integration/routing/test_non_chat_model_endpoint_gate.py b/tests/integration/routing/test_non_chat_model_endpoint_gate.py new file mode 100644 index 00000000000..465f5801692 --- /dev/null +++ b/tests/integration/routing/test_non_chat_model_endpoint_gate.py @@ -0,0 +1,1515 @@ +import asyncio +import contextlib +import json +import uuid +from collections.abc import Iterator, Mapping +from dataclasses import dataclass +from pathlib import Path +from types import MappingProxyType +from typing import Final + +import httpx +import openai +import pytest +import websockets +import yaml +from integration._support.client import ( + JSON_OBJECT, + Gateway, + Scenario, + eventually, + gateway_from_environment, + held, + list_value, + object_value, + string_value, +) +from integration._support import responses_ws_peer +from integration._support.database import read_rows +from integration._support.openai_wire import answering_model_discovery, chat_reply, responses_reply +from integration._support.process import OwnedState, graceful_stop_seconds, owned_proxy_process +from integration._support.tls import write_self_signed_cert +from integration._support.wire import Reply, Request, Wire, wire_server +from pydantic import JsonValue, TypeAdapter +from websockets.exceptions import ConnectionClosed + +_PROVIDER_KEY: Final = "integration-provider-key" +_EMBEDDING_MODEL: Final = "text-embedding-3-small" +_CHAT_MODEL: Final = "gpt-4o-mini" +_CHAT_BACKEND_PREFIX: Final = "gate-chat-backend-" +_SERVED_TEXT: Final = "served by the chat upstream" +_VECTOR: Final = (0.25, 0.5, 0.75) +_COUNTED_TOKENS: Final = 17 +_CLOSE: Final = MappingProxyType({"connection": "close"}) +_RELOAD_SECONDS: Final = 5 +_EMBED_ONLY: Final[Mapping[str, JsonValue]] = MappingProxyType( + {"mode": "embedding", "supported_endpoints": ["/v1/embeddings"], "enforce_supported_endpoints": True} +) + + +def _embedding_reply() -> Reply: + return Reply( + body=json.dumps( + { + "object": "list", + "data": [{"object": "embedding", "index": 0, "embedding": list(_VECTOR)}], + "model": _EMBEDDING_MODEL, + "usage": {"prompt_tokens": 3, "total_tokens": 3}, + } + ).encode() + ) + + +def _not_a_chat_model() -> Reply: + return Reply( + status=404, + body=json.dumps( + {"error": {"message": "scripted: not a chat model", "type": "invalid_request_error", "code": None}} + ).encode(), + ) + + +def _text_completion_reply() -> Reply: + return Reply( + body=json.dumps( + { + "id": f"cmpl-{uuid.uuid4().hex}", + "object": "text_completion", + "created": 1, + "model": _CHAT_MODEL, + "choices": [{"index": 0, "text": _SERVED_TEXT, "finish_reason": "stop", "logprobs": None}], + "usage": {"prompt_tokens": 3, "completion_tokens": 3, "total_tokens": 6}, + } + ).encode() + ) + + +def _serves_chat(model: JsonValue) -> bool: + return model == _CHAT_MODEL or (isinstance(model, str) and model.startswith(_CHAT_BACKEND_PREFIX)) + + +def _respond(request: Request) -> Reply: + path: Final = request.target.split("?")[0] + body: Final = JSON_OBJECT.validate_json(request.body or b"{}") + streamed: Final = body.get("stream") is True + if path == "/v1/embeddings": + return _embedding_reply() + if path == "/v1/responses/input_tokens": + return Reply(body=json.dumps({"object": "response.input_tokens", "input_tokens": _COUNTED_TOKENS}).encode()) + if not _serves_chat(body.get("model")): + return _not_a_chat_model() + if path == "/v1/chat/completions": + return chat_reply(f"chatcmpl-{uuid.uuid4().hex}", _CHAT_MODEL, _SERVED_TEXT, stream=streamed) + if path in ("/v1/responses", "/v1/responses/compact"): + return responses_reply(f"resp_{uuid.uuid4().hex}", _CHAT_MODEL, _SERVED_TEXT, stream=streamed) + if path == "/v1/completions": + return _text_completion_reply() + return _not_a_chat_model() + + +class _Upstream: + def __init__(self, wire: Wire) -> None: + self.wire = wire + self.seen: tuple[Request, ...] = () + + def calls(self, marker: str) -> tuple[Request, ...]: + self.seen = (*self.seen, *self.wire.drain()) + return tuple(request for request in self.seen if marker.encode() in request.body or marker in request.target) + + def targets(self, marker: str) -> tuple[str, ...]: + return tuple(request.target for request in self.calls(marker)) + + +@dataclass(frozen=True, slots=True) +class _Deployment: + litellm_model: str + declarations: tuple[Mapping[str, JsonValue], ...] + + +def _single(declaration: Mapping[str, JsonValue], litellm_model: str = f"openai/{_CHAT_MODEL}") -> _Deployment: + return _Deployment(litellm_model, (declaration,)) + + +_SHARED: Final[Mapping[str, _Deployment]] = MappingProxyType( + { + "embedding_only": _single(_EMBED_ONLY, f"openai/{_EMBEDDING_MODEL}"), + "embedding_only_second": _single(_EMBED_ONLY, f"openai/{_EMBEDDING_MODEL}"), + "image_generation": _single( + {"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/images/generations"]} + ), + "chat_target": _single({"mode": "chat"}), + "chat_rate_limited": _single({"mode": "chat"}), + "mode_not_a_string": _single( + {"mode": 7, "supported_endpoints": ["/v1/embeddings"]}, "openai/gate-mode-not-a-string" + ), + } +) + + +@dataclass(frozen=True, slots=True) +class _Rig: + gateway: Gateway + upstream: _Upstream + names: Mapping[str, str] + api_base: str + + +def _create(gateway: Gateway, name: str, litellm_params: Mapping[str, JsonValue], info: Mapping[str, JsonValue]) -> str: + created: Final = gateway.post( + "/model/new", + { + "model_name": name, + "litellm_params": {"api_key": _PROVIDER_KEY, **litellm_params}, + "model_info": dict(info), + }, + ) + return string_value(object_value(created["model_info"])["id"]) + + +def _deploy(scenario: Scenario, api_base: str, label: str, deployment: _Deployment) -> tuple[str, tuple[str, ...]]: + name: Final = f"gate-{label}-{uuid.uuid4().hex[:12]}" + identities: Final = tuple( + _create(scenario.gateway, name, {"model": deployment.litellm_model, "api_base": api_base}, declaration) + for declaration in deployment.declarations + ) + tuple(scenario.cleanups.callback(scenario.delete_model, identity) for identity in identities) + return name, identities + + +def _listed_ids(gateway: Gateway) -> frozenset[str]: + listed: Final = gateway.request("GET", "/model/info", headers=_CLOSE) + assert listed.status_code == 200, listed.text + return frozenset( + string_value(object_value(object_value(entry)["model_info"])["id"]) + for entry in list_value(JSON_OBJECT.validate_json(listed.content)["data"]) + ) + + +def _served_by_every_worker(gateway: Gateway, identities: frozenset[str]) -> None: + held( + lambda: identities <= _listed_ids(gateway), + lambda served: served, + holding=_RELOAD_SECONDS * 2, + seconds=_RELOAD_SECONDS * 12, + ) + + +@pytest.fixture(scope="module") +def rig() -> Iterator[_Rig]: + with ( + wire_server(answering_model_discovery(_respond)) as wire, + gateway_from_environment() as gateway, + gateway.scenario() as scenario, + ): + api_base: Final = wire.url + "/v1" + deployed: Final = {label: _deploy(scenario, api_base, label, spec) for label, spec in _SHARED.items()} + _served_by_every_worker( + gateway, frozenset(identity for _name, identities in deployed.values() for identity in identities) + ) + yield _Rig( + gateway, + _Upstream(wire), + MappingProxyType({label: name for label, (name, _ids) in deployed.items()}), + api_base, + ) + + +def _marker() -> str: + return f"gate-{uuid.uuid4().hex}" + + +def _openai_error(endpoint: str, model: str, supported: str) -> dict[str, JsonValue]: + return { + "error": { + "message": f"{endpoint} is not supported by model={model}. Supported endpoints: {supported}", + "type": "invalid_request_error", + "param": "model", + "code": "400", + } + } + + +def _anthropic_error(model: str, supported: str) -> dict[str, JsonValue]: + return { + "type": "error", + "error": { + "type": "invalid_request_error", + "message": f"/v1/messages is not supported by model={model}. Supported endpoints: {supported}", + }, + } + + +def _chat_body(model: str, marker: str, *, stream: bool = False) -> dict[str, JsonValue]: + return {"model": model, "messages": [{"role": "user", "content": marker}], **({"stream": True} if stream else {})} + + +def _failure_row(call_id: str) -> dict[str, JsonValue]: + rows: Final = eventually( + lambda: read_rows('SELECT request_id, status FROM "LiteLLM_SpendLogs" WHERE request_id = %s', (call_id,)), + lambda found: len(found) >= 1, + seconds=70, + ) + assert len(rows) == 1, rows + return rows[0] + + +def _assert_error(response: httpx.Response, expected: dict[str, JsonValue]) -> None: + assert response.status_code == 400, response.text + body: Final = JSON_OBJECT.validate_json(response.content) + error: Final = object_value(body["error"]) + message: Final = string_value(error["message"]).split("\n\n", 1)[0] + assert {**body, "error": {**error, "message": message}} == expected, response.text + + +def _assert_rejected_before( + upstream: _Upstream, response: httpx.Response, marker: str, expected: dict[str, JsonValue] +) -> None: + _assert_error(response, expected) + assert upstream.calls(marker) == (), upstream.targets(marker) + assert _failure_row(response.headers["x-litellm-call-id"])["status"] == "failure" + + +def _assert_rejected(rig: _Rig, response: httpx.Response, marker: str, expected: dict[str, JsonValue]) -> None: + _assert_rejected_before(rig.upstream, response, marker, expected) + + +@dataclass(frozen=True, slots=True) +class _Route: + path: str + endpoint: str + anthropic_shape: bool + stream: bool + model_in_path: bool + shape: str + + +_ROUTES: Final[Mapping[str, _Route]] = MappingProxyType( + { + "interactions": _Route("/v1beta/interactions", "/v1beta/interactions", False, False, False, "interactions"), + "chat_v1": _Route("/v1/chat/completions", "/v1/chat/completions", False, False, False, "chat"), + "chat_v1_stream": _Route("/v1/chat/completions", "/v1/chat/completions", False, True, False, "chat"), + "chat_bare": _Route("/chat/completions", "/v1/chat/completions", False, False, False, "chat"), + "chat_engines": _Route("/engines/{model}/chat/completions", "/v1/chat/completions", False, False, True, "chat"), + "chat_azure_deployments": _Route( + "/openai/deployments/{model}/chat/completions", "/v1/chat/completions", False, False, True, "chat" + ), + "messages": _Route("/v1/messages", "/v1/messages", True, False, False, "messages"), + "messages_stream": _Route("/v1/messages", "/v1/messages", True, True, False, "messages"), + "responses_v1": _Route("/v1/responses", "/v1/responses", False, False, False, "responses"), + "responses_v1_stream": _Route("/v1/responses", "/v1/responses", False, True, False, "responses"), + "responses_bare": _Route("/responses", "/v1/responses", False, False, False, "responses"), + "cursor_responses_body": _Route("/cursor/chat/completions", "/v1/responses", False, False, False, "cursor"), + "cursor_chat_body": _Route("/cursor/chat/completions", "/v1/chat/completions", False, False, False, "chat"), + } +) + + +def _route_body(route: _Route, model: str, marker: str) -> dict[str, JsonValue]: + shaped: Final[Mapping[str, dict[str, JsonValue]]] = { + "chat": {"messages": [{"role": "user", "content": marker}]}, + "messages": {"max_tokens": 16, "messages": [{"role": "user", "content": marker}]}, + "responses": {"input": marker}, + "interactions": {"input": marker}, + "cursor": {"input": [{"role": "user", "content": marker}]}, + } + return { + **({} if route.model_in_path else {"model": model}), + **shaped[route.shape], + **({"stream": True} if route.stream else {}), + } + + +def _send(gateway: Gateway, route: _Route, model: str, marker: str) -> httpx.Response: + path: Final = route.path.format(model=model) + body: Final = _route_body(route, model, marker) + if not route.stream: + return gateway.request("POST", path, body) + with gateway.client.stream("POST", path, json=body, headers={"Authorization": f"Bearer {gateway.key}"}) as streamed: + streamed.read() + return streamed + + +@pytest.mark.parametrize("route", tuple(_ROUTES)) +def test_chat_shaped_route_rejects_an_embedding_only_model_before_the_upstream(rig: _Rig, route: str) -> None: + spec: Final = _ROUTES[route] + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + response: Final = _send(rig.gateway, spec, model, marker) + expected: Final = ( + _anthropic_error(model, "/v1/embeddings") + if spec.anthropic_shape + else _openai_error(spec.endpoint, model, "/v1/embeddings") + ) + _assert_rejected(rig, response, marker, expected) + + +def _sdk_embedding(rig: _Rig, model: str, marker: str) -> tuple[tuple[float, ...], str]: + async def call() -> tuple[tuple[float, ...], str]: + async with openai.AsyncOpenAI( + base_url=f"{rig.gateway.client.base_url}/v1", api_key=rig.gateway.key, max_retries=0 + ) as sdk: + raw: Final = await sdk.embeddings.with_raw_response.create( + model=model, input=marker, encoding_format="float" + ) + parsed: Final = raw.parse() + return tuple(parsed.data[0].embedding), raw.headers["x-litellm-call-id"] + + return asyncio.run(call()) + + +def _httpx_embedding(rig: _Rig, model: str, marker: str) -> tuple[tuple[float, ...], str]: + response: Final = rig.gateway.request( + "POST", "/v1/embeddings", {"model": model, "input": marker, "encoding_format": "float"} + ) + assert response.status_code == 200, response.text + data: Final = list_value(JSON_OBJECT.validate_json(response.content)["data"]) + vector: Final = list_value(object_value(data[0])["embedding"]) + return tuple(float(str(value)) for value in vector), response.headers["x-litellm-call-id"] + + +@pytest.mark.parametrize("client", ("httpx_sync", "openai_async_sdk")) +def test_embeddings_still_reach_the_embedding_only_model(rig: _Rig, client: str) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + vector, call_id = ( + _httpx_embedding(rig, model, marker) if client == "httpx_sync" else _sdk_embedding(rig, model, marker) + ) + assert vector == _VECTOR + assert rig.upstream.targets(marker) == ("/v1/embeddings",) + row: Final = eventually( + lambda: read_rows( + 'SELECT status, call_type FROM "LiteLLM_SpendLogs" WHERE request_id = %s OR litellm_call_id = %s', + (call_id, call_id), + ), + lambda found: len(found) >= 1, + seconds=70, + ) + assert [(entry["status"], entry["call_type"]) for entry in row] == [("success", "aembedding")], row + + +def _assert_served(rig: _Rig, response: httpx.Response, marker: str, target: str) -> None: + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert rig.upstream.targets(marker) == (target,) + + +def test_key_alias_to_an_embedding_only_model_is_rejected_under_the_resolved_name(rig: _Rig) -> None: + target: Final = rig.names["embedding_only"] + alias: Final = f"gate-alias-{uuid.uuid4().hex[:12]}" + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + key: Final = scenario.key(aliases={alias: target}) + response: Final = rig.gateway.request("POST", "/v1/chat/completions", _chat_body(alias, marker), key=key) + _assert_rejected(rig, response, marker, _openai_error("/v1/chat/completions", target, "/v1/embeddings")) + + +async def _websocket_first_frame_and_close( + gateway: Gateway, key: str, model: str, marker: str +) -> tuple[dict[str, JsonValue], int | None, str | None]: + url: Final = str(gateway.client.base_url).rstrip("/").replace("http://", "ws://", 1) + "/v1/responses" + async with websockets.connect(url, additional_headers={"Authorization": f"Bearer {key}"}, open_timeout=10) as ws: + await ws.send(json.dumps({"type": "response.create", "model": model, "input": marker})) + first: Final = JSON_OBJECT.validate_json(await asyncio.wait_for(ws.recv(), 30)) + with contextlib.suppress(ConnectionClosed): + await asyncio.wait_for(ws.recv(), 30) + return first, ws.close_code, ws.close_reason + + +def _websocket_rejection(model: str) -> tuple[dict[str, JsonValue], int | None, str | None]: + return ( + { + "type": "error", + "status": 400, + "error": { + "type": "invalid_request_error", + "message": f"/v1/responses is not supported by model={model}. Supported endpoints: /v1/embeddings", + }, + }, + 1008, + "Pre-call error", + ) + + +@pytest.mark.parametrize("aliased", [False, True]) +def test_responses_websocket_closes_on_an_embedding_only_model_before_the_upstream(rig: _Rig, aliased: bool) -> None: + target: Final = rig.names["embedding_only"] + alias: Final = f"gate-ws-alias-{uuid.uuid4().hex[:12]}" + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + key: Final = scenario.key(aliases={alias: target}) + closed: Final = asyncio.run( + _websocket_first_frame_and_close(rig.gateway, key, alias if aliased else target, marker) + ) + assert closed == _websocket_rejection(target) + assert rig.upstream.calls(marker) == () + + +_RESERVING_KEYS: Final[Mapping[str, Mapping[str, JsonValue]]] = MappingProxyType( + { + "max_parallel_requests": {"max_parallel_requests": 1}, + "tpm_limit": {"tpm_limit": 100, "default_estimated_output_tokens": 80}, + } +) + + +@pytest.mark.parametrize("limit", tuple(_RESERVING_KEYS)) +def test_responses_websocket_rejections_leave_the_key_reservation_free(rig: _Rig, limit: str) -> None: + target: Final = rig.names["embedding_only"] + markers: Final = tuple(_marker() for _ in range(3)) + served_marker: Final = _marker() + with rig.gateway.scenario() as scenario: + key: Final = scenario.key(**_RESERVING_KEYS[limit]) + closed: Final = tuple( + asyncio.run(_websocket_first_frame_and_close(rig.gateway, key, target, marker)) for marker in markers + ) + assert closed == (_websocket_rejection(target),) * 3 + served: Final = rig.gateway.request( + "POST", "/v1/chat/completions", _chat_body(rig.names["chat_target"], served_marker), key=key + ) + _assert_served(rig, served, served_marker, "/v1/chat/completions") + assert [rig.upstream.calls(marker) for marker in markers] == [()] * 3 + + +def test_team_scoped_embedding_only_model_is_rejected_under_its_public_name(rig: _Rig) -> None: + public: Final = f"gate-team-{uuid.uuid4().hex[:12]}" + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + team: Final = scenario.team() + identity: Final = _create( + rig.gateway, + public, + {"model": f"openai/{_EMBEDDING_MODEL}", "api_base": rig.api_base}, + {**_EMBED_ONLY, "team_id": team}, + ) + scenario.cleanups.callback(scenario.delete_model, identity) + key: Final = scenario.key(team_id=team) + _served_by_every_worker(rig.gateway, frozenset({identity})) + embedded: Final = rig.gateway.request( + "POST", "/v1/embeddings", {"model": public, "input": marker, "encoding_format": "float"}, key=key + ) + assert embedded.status_code == 200, embedded.text + assert rig.upstream.targets(marker) == ("/v1/embeddings",) + rejected_marker: Final = _marker() + response: Final = rig.gateway.request( + "POST", "/v1/chat/completions", _chat_body(public, rejected_marker), key=key + ) + _assert_rejected( + rig, response, rejected_marker, _openai_error("/v1/chat/completions", public, "/v1/embeddings") + ) + + +def test_wildcard_embedding_only_deployment_rejects_chat_for_any_matched_model(rig: _Rig) -> None: + prefix: Final = f"gate-wild-{uuid.uuid4().hex[:12]}" + requested: Final = f"{prefix}/{_EMBEDDING_MODEL}" + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + identity: Final = _create( + rig.gateway, f"{prefix}/*", {"model": "openai/*", "api_base": rig.api_base}, _EMBED_ONLY + ) + scenario.cleanups.callback(scenario.delete_model, identity) + _served_by_every_worker(rig.gateway, frozenset({identity})) + response: Final = rig.gateway.request("POST", "/v1/chat/completions", _chat_body(requested, marker)) + _assert_rejected(rig, response, marker, _openai_error("/v1/chat/completions", requested, "/v1/embeddings")) + + +def test_key_without_access_gets_the_access_error_and_learns_nothing_about_endpoints(rig: _Rig) -> None: + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + key: Final = scenario.key(models=[rig.names["chat_target"]]) + response: Final = rig.gateway.request( + "POST", "/v1/chat/completions", _chat_body(rig.names["embedding_only"], marker), key=key + ) + assert response.status_code == 403, response.text + assert object_value(JSON_OBJECT.validate_json(response.content)["error"])["type"] == "key_model_access_denied", ( + response.text + ) + assert "Supported endpoints" not in response.text, response.text + assert rig.upstream.calls(marker) == () + + +def test_unknown_model_keeps_the_invalid_model_error(rig: _Rig) -> None: + marker: Final = _marker() + response: Final = rig.gateway.request( + "POST", "/v1/chat/completions", _chat_body(f"gate-missing-{uuid.uuid4().hex}", marker) + ) + assert response.status_code == 400, response.text + assert "Invalid model name" in response.text and "Supported endpoints" not in response.text, response.text + assert rig.upstream.calls(marker) == () + + +def test_request_fallbacks_answer_for_an_embedding_only_primary(rig: _Rig) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + response: Final = rig.gateway.request( + "POST", "/v1/chat/completions", {**_chat_body(model, marker), "fallbacks": [rig.names["chat_target"]]} + ) + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert rig.upstream.targets(marker) == ("/v1/chat/completions",) + + +def test_an_empty_request_fallback_list_keeps_the_rejection(rig: _Rig) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + response: Final = rig.gateway.request( + "POST", "/v1/chat/completions", {**_chat_body(model, marker), "fallbacks": []} + ) + _assert_rejected(rig, response, marker, _openai_error("/v1/chat/completions", model, "/v1/embeddings")) + + +def test_fallbacks_that_cannot_answer_chat_keep_the_rejection(rig: _Rig) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + fallbacks: Final = [rig.names["image_generation"], model] + response: Final = rig.gateway.request( + "POST", "/v1/chat/completions", {**_chat_body(model, marker), "fallbacks": fallbacks} + ) + _assert_rejected(rig, response, marker, _openai_error("/v1/chat/completions", model, "/v1/embeddings")) + + +def test_request_fallbacks_do_not_rescue_an_unknown_primary(rig: _Rig) -> None: + marker: Final = _marker() + response: Final = rig.gateway.request( + "POST", + "/v1/chat/completions", + {**_chat_body(f"gate-missing-{uuid.uuid4().hex}", marker), "fallbacks": [rig.names["chat_target"]]}, + ) + assert response.status_code == 400, response.text + assert "Invalid model name" in response.text, response.text + assert rig.upstream.calls(marker) == () + + +def _chat(rig: _Rig, model: str, marker: str, *, key: str | None = None, **extra: JsonValue) -> httpx.Response: + return rig.gateway.request("POST", "/v1/chat/completions", {**_chat_body(model, marker), **extra}, key=key) + + +def _assert_rejected_chat(rig: _Rig, response: httpx.Response, marker: str, model: str) -> None: + _assert_rejected(rig, response, marker, _openai_error("/v1/chat/completions", model, "/v1/embeddings")) + + +def test_a_fallback_chain_that_reaches_chat_two_hops_away_answers(rig: _Rig) -> None: + first, second, chat = rig.names["embedding_only"], rig.names["embedding_only_second"], rig.names["chat_target"] + marker: Final = _marker() + response: Final = _chat(rig, first, marker, fallbacks=[{first: [second]}, {second: [chat]}]) + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert rig.upstream.targets(marker) == ("/v1/chat/completions",) + + +def test_a_fallback_chain_that_loops_between_embedding_only_groups_keeps_the_rejection(rig: _Rig) -> None: + first, second = rig.names["embedding_only"], rig.names["embedding_only_second"] + marker: Final = _marker() + response: Final = _chat(rig, first, marker, fallbacks=[{first: [second]}, {second: [first]}]) + _assert_rejected_chat(rig, response, marker, first) + + +def test_a_specific_chain_that_loops_beats_a_catch_all_that_could_answer(rig: _Rig) -> None: + first, second = rig.names["embedding_only"], rig.names["embedding_only_second"] + marker: Final = _marker() + response: Final = _chat( + rig, first, marker, fallbacks=[{"*": [rig.names["chat_target"]]}, {first: [second]}, {second: [first]}] + ) + _assert_rejected_chat(rig, response, marker, first) + + +def test_disable_fallbacks_keeps_the_rejection_over_request_fallbacks(rig: _Rig) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + response: Final = _chat(rig, model, marker, fallbacks=[rig.names["chat_target"]], disable_fallbacks=True) + _assert_rejected_chat(rig, response, marker, model) + + +def _assert_fallback_answered( + upstream: _Upstream, response: httpx.Response, marker: str, target: str = "/v1/chat/completions" +) -> None: + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert upstream.targets(marker) == (target,), response.text + + +def _through_a_group_with_no_deployments(rig: _Rig) -> Mapping[str, JsonValue]: + ghost: Final = f"gate-ghost-{uuid.uuid4().hex}" + return {"fallbacks": [{rig.names["embedding_only"]: [ghost]}, {ghost: [rig.names["chat_target"]]}]} + + +def test_request_fallbacks_naming_a_chat_backend_name_answer_for_an_embedding_only_primary(rig: _Rig) -> None: + model: Final = rig.names["embedding_only"] + backend: Final = f"openai/{_CHAT_BACKEND_PREFIX}{uuid.uuid4().hex[:12]}" + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + identity: Final = _create( + rig.gateway, f"gate-backend-{uuid.uuid4().hex[:12]}", {"model": backend, "api_base": rig.api_base}, {} + ) + scenario.cleanups.callback(scenario.delete_model, identity) + _served_by_every_worker(rig.gateway, frozenset({identity})) + response: Final = _chat(rig, model, marker, fallbacks=[{model: [backend]}]) + _assert_fallback_answered(rig.upstream, response, marker) + + +@pytest.mark.parametrize("level", ("key", "team")) +def test_router_settings_fallbacks_through_a_group_with_no_deployments_answer(rig: _Rig, level: str) -> None: + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + settings: Final = {"router_settings": dict(_through_a_group_with_no_deployments(rig))} + key: Final = scenario.key(**settings) if level == "key" else scenario.key(team_id=scenario.team(**settings)) + response: Final = _chat(rig, rig.names["embedding_only"], marker, key=key) + _assert_fallback_answered(rig.upstream, response, marker) + + +def test_team_router_settings_fallbacks_answer_for_an_embedding_only_primary(rig: _Rig) -> None: + model, chat = rig.names["embedding_only"], rig.names["chat_target"] + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + team: Final = scenario.team(router_settings={"fallbacks": [{model: [chat]}]}) + response: Final = _chat(rig, model, marker, key=scenario.key(team_id=team)) + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert rig.upstream.targets(marker) == ("/v1/chat/completions",) + + +def test_key_router_settings_fallbacks_answer_for_an_embedding_only_primary(rig: _Rig) -> None: + model, chat = rig.names["embedding_only"], rig.names["chat_target"] + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + key: Final = scenario.key(router_settings={"fallbacks": [{model: [chat]}]}) + response: Final = _chat(rig, model, marker, key=key) + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert rig.upstream.targets(marker) == ("/v1/chat/completions",) + + +def test_key_router_settings_that_cannot_answer_override_team_fallbacks_that_could(rig: _Rig) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + team: Final = scenario.team(router_settings={"fallbacks": [{model: [rig.names["chat_target"]]}]}) + key: Final = scenario.key( + team_id=team, router_settings={"fallbacks": [{model: [rig.names["embedding_only_second"]]}]} + ) + response: Final = _chat(rig, model, marker, key=key) + _assert_rejected_chat(rig, response, marker, model) + + +def test_an_empty_request_fallback_list_overrides_key_router_settings_fallbacks(rig: _Rig) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + key: Final = scenario.key(router_settings={"fallbacks": [{model: [rig.names["chat_target"]]}]}) + response: Final = _chat(rig, model, marker, key=key, fallbacks=[]) + _assert_rejected_chat(rig, response, marker, model) + + +def test_local_rate_limit_fallback_skips_an_embedding_only_hop(rig: _Rig) -> None: + limited, chat, model = rig.names["chat_rate_limited"], rig.names["chat_target"], rig.names["embedding_only"] + first, second = _marker(), _marker() + with rig.gateway.scenario() as scenario: + key: Final = scenario.key( + model_rpm_limit={limited: 1}, router_settings={"fallbacks": [{limited: [model, chat]}]} + ) + served_first: Final = _chat(rig, limited, first, key=key) + served_second: Final = _chat(rig, limited, second, key=key) + assert (served_first.status_code, served_first.headers.get("x-litellm-model-group")) == (200, limited), ( + served_first.text + ) + assert (served_second.status_code, served_second.headers.get("x-litellm-model-group")) == (200, chat), ( + served_second.text + ) + assert (rig.upstream.targets(first), rig.upstream.targets(second)) == ( + ("/v1/chat/completions",), + ("/v1/chat/completions",), + ) + + +_COMPACT_PATHS: Final = ("/v1/responses/compact", "/responses/compact", "/openai/v1/responses/compact") + + +def _compact_body(model: str, marker: str) -> dict[str, JsonValue]: + return {"model": model, "input": [{"role": "user", "content": marker}]} + + +@pytest.mark.parametrize("path", _COMPACT_PATHS) +def test_responses_compact_rejects_an_embedding_only_model_before_the_upstream(rig: _Rig, path: str) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + response: Final = rig.gateway.request("POST", path, _compact_body(model, marker)) + _assert_rejected(rig, response, marker, _openai_error("/v1/responses/compact", model, "/v1/embeddings")) + + +def test_responses_compact_keeps_reaching_the_upstream_for_a_chat_model(rig: _Rig) -> None: + marker: Final = _marker() + response: Final = rig.gateway.request( + "POST", "/v1/responses/compact", _compact_body(rig.names["chat_target"], marker) + ) + _assert_served(rig, response, marker, "/v1/responses/compact") + + +def test_count_tokens_keeps_counting_for_an_embedding_only_model(rig: _Rig) -> None: + marker: Final = _marker() + response: Final = rig.gateway.request( + "POST", + "/v1/messages/count_tokens", + {"model": rig.names["embedding_only"], "messages": [{"role": "user", "content": marker}]}, + ) + assert response.status_code == 200, response.text + assert JSON_OBJECT.validate_json(response.content)["input_tokens"] == _COUNTED_TOKENS, response.text + assert rig.upstream.targets(marker) == ("/v1/responses/input_tokens",) + + +def test_text_completions_reject_an_embedding_only_model_before_the_upstream(rig: _Rig) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + response: Final = rig.gateway.request("POST", "/v1/completions", {"model": model, "prompt": marker}) + _assert_rejected(rig, response, marker, _openai_error("/v1/completions", model, "/v1/embeddings")) + + +@pytest.mark.parametrize("method", ("generateContent", "streamGenerateContent")) +def test_gemini_routes_reject_an_embedding_only_model_before_the_upstream(rig: _Rig, method: str) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + response: Final = rig.gateway.request( + "POST", f"/v1beta/models/{model}:{method}", {"contents": [{"role": "user", "parts": [{"text": marker}]}]} + ) + assert response.status_code == 400, response.text + assert f"{method} is not supported by model={model}. Supported endpoints: /v1/embeddings" in response.text + assert rig.upstream.calls(marker) == (), rig.upstream.targets(marker) + + +def test_repeated_rejections_log_one_failure_row_per_call(rig: _Rig) -> None: + model: Final = rig.names["embedding_only"] + marker: Final = _marker() + responses: Final = tuple( + rig.gateway.request("POST", "/v1/chat/completions", _chat_body(model, marker)) for _ in range(3) + ) + expected: Final = _openai_error("/v1/chat/completions", model, "/v1/embeddings") + for response in responses: + _assert_error(response, expected) + call_ids: Final = tuple(item.headers["x-litellm-call-id"] for item in responses) + assert len(set(call_ids)) == 3, call_ids + assert [_failure_row(call_id)["status"] for call_id in call_ids] == ["failure"] * 3 + assert rig.upstream.calls(marker) == () + + +def _closed_chat(rig: _Rig, model: str, marker: str) -> httpx.Response: + return rig.gateway.request("POST", "/v1/chat/completions", _chat_body(model, marker), headers=_CLOSE) + + +def test_live_declaration_update_turns_the_gate_on_and_off_on_every_worker(rig: _Rig) -> None: + chat_capable: Final[dict[str, JsonValue]] = { + "mode": "chat", + "supported_endpoints": ["/v1/chat/completions", "/v1/embeddings"], + } + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + name, identities = _deploy(scenario, rig.api_base, "flip", _single(chat_capable)) + _served_by_every_worker(rig.gateway, frozenset(identities)) + served: Final = _closed_chat(rig, name, marker) + _assert_served(rig, served, marker, "/v1/chat/completions") + patched: Final = rig.gateway.request( + "PATCH", f"/model/{identities[0]}/update", {"model_info": dict(_EMBED_ONLY)} + ) + assert patched.status_code == 200, patched.text + rejected: Final = held( + lambda: _closed_chat(rig, name, _marker()), + lambda response: response.status_code == 400, + holding=_RELOAD_SECONDS * 2, + seconds=_RELOAD_SECONDS * 12, + ) + _assert_error(rejected, _openai_error("/v1/chat/completions", name, "/v1/embeddings")) + restored: Final = rig.gateway.request("PATCH", f"/model/{identities[0]}/update", {"model_info": chat_capable}) + assert restored.status_code == 200, restored.text + held( + lambda: _closed_chat(rig, name, _marker()), + lambda response: response.status_code == 200, + holding=_RELOAD_SECONDS * 2, + seconds=_RELOAD_SECONDS * 12, + ) + + +_TESTS_ROOT: Final = Path(__file__).resolve().parents[2] +_CONFIG: Final = TypeAdapter(dict[str, JsonValue]) + + +def _owned_embedding(owned: Gateway, model: str, marker: str) -> httpx.Response: + return owned.request("POST", "/v1/embeddings", {"model": model, "input": marker, "encoding_format": "float"}) + + +@dataclass(frozen=True, slots=True) +class _Routed: + gateway: Gateway + upstream: _Upstream + names: Mapping[str, str] + + +def _routed_entry(name: str, api_base: str, litellm_model: str, info: Mapping[str, JsonValue]) -> dict[str, JsonValue]: + return { + "model_name": name, + "litellm_params": {"model": litellm_model, "api_base": api_base, "api_key": _PROVIDER_KEY}, + "model_info": dict(info), + } + + +def _routed_config(directory: Path, api_base: str, names: Mapping[str, str]) -> Path: + shared: Final = _CONFIG.validate_python(yaml.safe_load((_TESTS_ROOT / "integration/proxy_config.yaml").read_text())) + config: Final = { + **shared, + "model_list": [ + _routed_entry(names["embed"], api_base, f"openai/{_EMBEDDING_MODEL}", _EMBED_ONLY), + _routed_entry(names["chat"], api_base, f"openai/{_CHAT_MODEL}", {"mode": "chat"}), + ], + "router_settings": { + **object_value(shared["router_settings"]), + "num_retries": 0, + "default_fallbacks": [names["chat"]], + "model_group_alias": {names["alias_embed"]: names["embed"], names["alias_chat"]: names["chat"]}, + }, + } + path: Final = directory / "gate-routed-config.yaml" + path.write_text(yaml.safe_dump(config, sort_keys=False)) + return path + + +@pytest.fixture(scope="module") +def routed(tmp_path_factory: pytest.TempPathFactory) -> Iterator[_Routed]: + directory: Final = tmp_path_factory.mktemp("gate-routed") + suffix: Final = uuid.uuid4().hex[:12] + names: Final = MappingProxyType( + {label: f"gate-routed-{label}-{suffix}" for label in ("embed", "chat", "alias_embed", "alias_chat")} + ) + with ( + wire_server(answering_model_discovery(_respond)) as wire, + gateway_from_environment() as rig_gateway, + owned_proxy_process( + rig_gateway, + directory, + {}, + config=_routed_config(directory, wire.url + "/v1", names), + workers=2, + state=OwnedState.SHARED_WITH_WORKER, + ) as owned, + ): + eventually( + lambda: owned.gateway.request("POST", "/v1/chat/completions", _chat_body(names["chat"], "probe")), + lambda response: response.status_code == 200, + seconds=30, + ) + yield _Routed(owned.gateway, _Upstream(wire), names) + + +_ROUTED_TIMEOUT: Final = pytest.mark.timeout(2 * graceful_stop_seconds() + 120) +_DEFAULT_FALLBACK_ROUTES: Final[Mapping[str, str]] = MappingProxyType( + {"chat_v1": "/v1/chat/completions", "messages": "/v1/responses", "responses_v1": "/v1/responses"} +) + + +@_ROUTED_TIMEOUT +@pytest.mark.parametrize("route", tuple(_DEFAULT_FALLBACK_ROUTES)) +def test_router_default_fallbacks_answer_for_an_embedding_only_group(routed: _Routed, route: str) -> None: + marker: Final = _marker() + response: Final = _send(routed.gateway, _ROUTES[route], routed.names["embed"], marker) + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert routed.upstream.targets(marker) == (_DEFAULT_FALLBACK_ROUTES[route],) + + +_OPT_OUTS: Final[Mapping[str, Mapping[str, JsonValue]]] = MappingProxyType( + { + "empty_fallbacks": {"fallbacks": []}, + "null_fallbacks": {"fallbacks": None}, + "disable_fallbacks": {"disable_fallbacks": True}, + } +) + + +@_ROUTED_TIMEOUT +@pytest.mark.parametrize("opt_out", tuple(_OPT_OUTS)) +def test_request_opt_outs_keep_the_rejection_over_router_default_fallbacks(routed: _Routed, opt_out: str) -> None: + model: Final = routed.names["embed"] + marker: Final = _marker() + response: Final = routed.gateway.request( + "POST", "/v1/chat/completions", {**_chat_body(model, marker), **_OPT_OUTS[opt_out]} + ) + _assert_rejected_before( + routed.upstream, response, marker, _openai_error("/v1/chat/completions", model, "/v1/embeddings") + ) + + +@_ROUTED_TIMEOUT +def test_model_group_alias_to_an_embedding_only_group_falls_back_to_the_default(routed: _Routed) -> None: + marker: Final = _marker() + response: Final = routed.gateway.request( + "POST", "/v1/chat/completions", _chat_body(routed.names["alias_embed"], marker) + ) + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert routed.upstream.targets(marker) == ("/v1/chat/completions",) + + +@_ROUTED_TIMEOUT +def test_model_group_alias_to_an_embedding_only_group_is_rejected_without_fallbacks(routed: _Routed) -> None: + alias: Final = routed.names["alias_embed"] + marker: Final = _marker() + response: Final = routed.gateway.request( + "POST", "/v1/chat/completions", {**_chat_body(alias, marker), "disable_fallbacks": True} + ) + _assert_rejected_before( + routed.upstream, response, marker, _openai_error("/v1/chat/completions", alias, "/v1/embeddings") + ) + + +@_ROUTED_TIMEOUT +def test_model_group_alias_to_a_chat_group_keeps_reaching_the_upstream(routed: _Routed) -> None: + marker: Final = _marker() + response: Final = routed.gateway.request( + "POST", "/v1/chat/completions", _chat_body(routed.names["alias_chat"], marker) + ) + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert routed.upstream.targets(marker) == ("/v1/chat/completions",) + + +@_ROUTED_TIMEOUT +def test_model_group_alias_to_an_embedding_only_group_keeps_serving_embeddings(routed: _Routed) -> None: + marker: Final = _marker() + response: Final = _owned_embedding(routed.gateway, routed.names["alias_embed"], marker) + assert response.status_code == 200, response.text + assert list_value( + object_value(list_value(JSON_OBJECT.validate_json(response.content)["data"])[0])["embedding"] + ) == list(_VECTOR), response.text + assert routed.upstream.targets(marker) == ("/v1/embeddings",) + + +_REWRITER: Final = """from litellm.integrations.custom_logger import CustomLogger + + +class WebSocketModelRewriter(CustomLogger): + async def async_pre_call_hook(self, user_api_key_dict, cache, data, call_type): + if call_type == "_aresponses_websocket" and data.get("model") == "__SOURCE__": + data["model"] = "__TARGET__" + return data + + +rewriter = WebSocketModelRewriter() +""" + + +@dataclass(frozen=True, slots=True) +class _Rewritten: + gateway: Gateway + upstream: _Upstream + source: str + target: str + + +def _rewritten_config(directory: Path, api_base: str, source: str, target: str, module: str) -> Path: + shared: Final = _CONFIG.validate_python(yaml.safe_load((_TESTS_ROOT / "integration/proxy_config.yaml").read_text())) + config: Final = { + **shared, + "model_list": [ + _routed_entry(source, api_base, f"openai/{_CHAT_MODEL}", {"mode": "chat"}), + _routed_entry(target, api_base, f"openai/{_EMBEDDING_MODEL}", _EMBED_ONLY), + ], + "router_settings": {**object_value(shared["router_settings"]), "num_retries": 0}, + "litellm_settings": {**object_value(shared["litellm_settings"]), "callbacks": [f"{module}.rewriter"]}, + } + path: Final = directory / "gate-rewritten-config.yaml" + path.write_text(yaml.safe_dump(config, sort_keys=False)) + return path + + +@pytest.fixture(scope="module") +def rewritten(tmp_path_factory: pytest.TempPathFactory) -> Iterator[_Rewritten]: + directory: Final = tmp_path_factory.mktemp("gate-rewritten") + suffix: Final = uuid.uuid4().hex[:12] + source: Final = f"gate-rewrite-chat-{suffix}" + target: Final = f"gate-rewrite-embed-{suffix}" + module: Final = f"gate_ws_rewriter_{suffix}" + (directory / f"{module}.py").write_text(_REWRITER.replace("__SOURCE__", source).replace("__TARGET__", target)) + with ( + wire_server(answering_model_discovery(_respond)) as wire, + gateway_from_environment() as rig_gateway, + owned_proxy_process( + rig_gateway, + directory, + {}, + config=_rewritten_config(directory, wire.url + "/v1", source, target, module), + workers=2, + state=OwnedState.SHARED_WITH_WORKER, + ) as owned, + ): + eventually( + lambda: owned.gateway.request("POST", "/v1/chat/completions", _chat_body(source, "probe")), + lambda response: response.status_code == 200, + seconds=30, + ) + yield _Rewritten(owned.gateway, _Upstream(wire), source, target) + + +@_ROUTED_TIMEOUT +@pytest.mark.parametrize("limit", tuple(_RESERVING_KEYS)) +def test_a_websocket_model_a_hook_rewrites_to_an_embedding_only_group_releases_the_key_reservation( + rewritten: _Rewritten, limit: str +) -> None: + markers: Final = tuple(_marker() for _ in range(3)) + served_marker: Final = _marker() + with rewritten.gateway.scenario() as scenario: + key: Final = scenario.key(**_RESERVING_KEYS[limit]) + closed: Final = tuple( + asyncio.run(_websocket_first_frame_and_close(rewritten.gateway, key, rewritten.source, marker)) + for marker in markers + ) + assert closed == (_websocket_rejection(rewritten.target),) * 3 + served: Final = rewritten.gateway.request( + "POST", "/v1/chat/completions", _chat_body(rewritten.source, served_marker), key=key + ) + assert served.status_code == 200, served.text + assert _SERVED_TEXT in served.text, served.text + assert [rewritten.upstream.calls(marker) for marker in markers] == [()] * 3 + assert rewritten.upstream.targets(served_marker) == ("/v1/chat/completions",) + + +@dataclass(frozen=True, slots=True) +class _Addressed: + gateway: Gateway + rig: Gateway + upstream: _Upstream + names: Mapping[str, str] + addresses: Mapping[str, str] + api_base: str + + +def _wildcard_backend_entry(name: str, api_base: str, backend: str) -> dict[str, JsonValue]: + entry: Final = _routed_entry(name, api_base, backend, _EMBED_ONLY) + return {**entry, "litellm_params": {**object_value(entry["litellm_params"]), "custom_llm_provider": "openai"}} + + +def _addressed_config( + directory: Path, api_base: str, names: Mapping[str, str], suffix: str, with_wildcards: bool +) -> Path: + shared: Final = _CONFIG.validate_python(yaml.safe_load((_TESTS_ROOT / "integration/proxy_config.yaml").read_text())) + wildcard: Final = f"gate-addr-wild-{suffix}" + wildcard_entries: Final = ( + ( + _routed_entry(f"{wildcard}/*", api_base, "openai/*", {}), + _wildcard_backend_entry(names["wildcard_backend"], api_base, f"{wildcard}/{_CHAT_MODEL}"), + ) + if with_wildcards + else () + ) + aliases: Final = {names["wildcard_alias"]: f"{wildcard}/{_CHAT_BACKEND_PREFIX}{suffix}"} if with_wildcards else {} + config: Final = { + **shared, + "model_list": [ + _routed_entry( + names["embed"], api_base, f"openai/gate-addr-backend-{suffix}", {**_EMBED_ONLY, "id": names["embed_id"]} + ), + _routed_entry(names["chat"], api_base, f"openai/{_CHAT_MODEL}", {"mode": "chat", "id": names["chat_id"]}), + _routed_entry(names["shared_embed"], api_base, f"openai/{_CHAT_MODEL}", _EMBED_ONLY), + *wildcard_entries, + _routed_entry(names["ghost_primary"], api_base, f"openai/{_EMBEDDING_MODEL}", _EMBED_ONLY), + ], + "router_settings": { + **object_value(shared["router_settings"]), + "num_retries": 0, + "model_group_alias": aliases, + "fallbacks": [{names["ghost_primary"]: [names["ghost"]]}, {names["ghost"]: [names["chat"]]}], + }, + "general_settings": { + **object_value(shared["general_settings"]), + "proxy_config_reload_interval_seconds": 3600, + }, + } + path: Final = directory / "gate-addressed-config.yaml" + path.write_text(yaml.safe_dump(config, sort_keys=False)) + return path + + +@contextlib.contextmanager +def _addressed_proxy(tmp_path_factory: pytest.TempPathFactory, with_wildcards: bool) -> Iterator[_Addressed]: + directory: Final = tmp_path_factory.mktemp("gate-addressed") + suffix: Final = uuid.uuid4().hex[:12] + names: Final = MappingProxyType( + { + label: f"gate-addr-{label}-{suffix}" + for label in ( + "embed", + "embed_id", + "chat", + "chat_id", + "shared_embed", + "wildcard_backend", + "wildcard_alias", + "ghost_primary", + "ghost", + ) + } + ) + with ( + wire_server(answering_model_discovery(_respond)) as wire, + gateway_from_environment() as rig_gateway, + owned_proxy_process( + rig_gateway, + directory, + {}, + config=_addressed_config(directory, wire.url + "/v1", names, suffix, with_wildcards), + workers=2, + state=OwnedState.SHARED_WITH_WORKER, + ) as owned, + ): + eventually( + lambda: owned.gateway.request("POST", "/v1/chat/completions", _chat_body(names["chat"], "probe")), + lambda response: response.status_code == 200, + seconds=30, + ) + yield _Addressed( + owned.gateway, + rig_gateway, + _Upstream(wire), + names, + MappingProxyType( + { + "deployment_id": names["embed_id"], + "backend_name": f"openai/gate-addr-backend-{suffix}", + "wildcard_backend_name": f"gate-addr-wild-{suffix}/{_CHAT_MODEL}", + } + ), + wire.url + "/v1", + ) + + +@pytest.fixture(scope="module") +def addressed(tmp_path_factory: pytest.TempPathFactory) -> Iterator[_Addressed]: + with _addressed_proxy(tmp_path_factory, with_wildcards=True) as proxy: + yield proxy + + +@pytest.fixture(scope="module") +def addressed_without_wildcards(tmp_path_factory: pytest.TempPathFactory) -> Iterator[_Addressed]: + with _addressed_proxy(tmp_path_factory, with_wildcards=False) as proxy: + yield proxy + + +_ADDRESSED_ROUTES: Final = ("chat_v1", "messages", "responses_v1") + + +@_ROUTED_TIMEOUT +@pytest.mark.parametrize("route", _ADDRESSED_ROUTES) +@pytest.mark.parametrize("address", ("deployment_id", "backend_name")) +def test_chat_sent_to_an_embedding_only_deployment_id_or_backend_name_is_rejected_before_the_upstream( + addressed: _Addressed, address: str, route: str +) -> None: + spec: Final = _ROUTES[route] + model: Final = addressed.addresses[address] + marker: Final = _marker() + response: Final = _send(addressed.gateway, spec, model, marker) + expected: Final = ( + _anthropic_error(model, "/v1/embeddings") + if spec.anthropic_shape + else _openai_error(spec.endpoint, model, "/v1/embeddings") + ) + _assert_rejected_before(addressed.upstream, response, marker, expected) + + +@_ROUTED_TIMEOUT +@pytest.mark.parametrize("address", ("deployment_id", "backend_name")) +def test_embeddings_sent_to_an_embedding_only_deployment_id_or_backend_name_keep_working( + addressed: _Addressed, address: str +) -> None: + marker: Final = _marker() + response: Final = _owned_embedding(addressed.gateway, addressed.addresses[address], marker) + assert response.status_code == 200, response.text + assert list_value( + object_value(list_value(JSON_OBJECT.validate_json(response.content)["data"])[0])["embedding"] + ) == list(_VECTOR), response.text + assert addressed.upstream.targets(marker) == ("/v1/embeddings",) + + +@_ROUTED_TIMEOUT +@pytest.mark.parametrize("address", ("chat_deployment_id", "backend_name_a_chat_deployment_shares")) +def test_chat_sent_to_a_chat_capable_deployment_id_or_backend_name_keeps_reaching_the_upstream( + addressed: _Addressed, address: str +) -> None: + model: Final = addressed.names["chat_id"] if address == "chat_deployment_id" else f"openai/{_CHAT_MODEL}" + marker: Final = _marker() + response: Final = addressed.gateway.request("POST", "/v1/chat/completions", _chat_body(model, marker)) + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert addressed.upstream.targets(marker) == ("/v1/chat/completions",) + + +@_ROUTED_TIMEOUT +def test_request_fallbacks_naming_a_chat_deployment_id_answer_for_an_embedding_only_primary( + addressed: _Addressed, +) -> None: + marker: Final = _marker() + response: Final = addressed.gateway.request( + "POST", + "/v1/chat/completions", + {**_chat_body(addressed.names["embed"], marker), "fallbacks": [addressed.names["chat_id"]]}, + ) + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert addressed.upstream.targets(marker) == ("/v1/chat/completions",) + + +@_ROUTED_TIMEOUT +def test_chat_sent_to_an_embedding_only_backend_name_a_wildcard_also_matches_is_served_by_the_wildcard( + addressed: _Addressed, +) -> None: + marker: Final = _marker() + response: Final = addressed.gateway.request( + "POST", "/v1/chat/completions", _chat_body(addressed.addresses["wildcard_backend_name"], marker) + ) + assert response.status_code == 200, response.text + assert _SERVED_TEXT in response.text, response.text + assert [JSON_OBJECT.validate_json(call.body)["model"] for call in addressed.upstream.calls(marker)] == [_CHAT_MODEL] + + +@_ROUTED_TIMEOUT +def test_request_fallbacks_naming_an_alias_only_a_wildcard_serves_answer_for_an_embedding_only_primary( + addressed: _Addressed, +) -> None: + marker: Final = _marker() + response: Final = addressed.gateway.request( + "POST", + "/v1/chat/completions", + {**_chat_body(addressed.names["embed"], marker), "fallbacks": [addressed.names["wildcard_alias"]]}, + ) + _assert_fallback_answered(addressed.upstream, response, marker) + + +@_ROUTED_TIMEOUT +def test_yaml_fallbacks_through_a_group_with_no_deployments_answer_for_an_embedding_only_group( + addressed: _Addressed, +) -> None: + marker: Final = _marker() + response: Final = addressed.gateway.request( + "POST", "/v1/chat/completions", _chat_body(addressed.names["ghost_primary"], marker) + ) + _assert_fallback_answered(addressed.upstream, response, marker) + + +@_ROUTED_TIMEOUT +def test_an_embedding_only_model_the_read_through_loads_from_the_database_is_rejected_on_every_call( + addressed_without_wildcards: _Addressed, +) -> None: + addressed: Final = addressed_without_wildcards + name: Final = f"gate-addr-loaded-{uuid.uuid4().hex[:12]}" + markers: Final = tuple(_marker() for _ in range(6)) + with addressed.rig.scenario() as scenario: + identity: Final = _create( + addressed.rig, name, {"model": f"openai/{_EMBEDDING_MODEL}", "api_base": addressed.api_base}, _EMBED_ONLY + ) + scenario.cleanups.callback(scenario.delete_model, identity) + responses: Final = tuple( + addressed.gateway.request("POST", "/v1/chat/completions", _chat_body(name, marker), headers=_CLOSE) + for marker in markers + ) + expected: Final = _openai_error("/v1/chat/completions", name, "/v1/embeddings") + for response in responses: + _assert_error(response, expected) + assert [addressed.upstream.calls(marker) for marker in markers] == [()] * len(markers) + + +@pytest.mark.parametrize("metadata", ({"model_group": 7}, {"model_group": ["wrong-group"]})) +def test_rejection_with_invalid_metadata_is_logged_without_blocking_embeddings( + rig: _Rig, metadata: dict[str, JsonValue] +) -> None: + marker: Final = _marker() + model: Final = rig.names["embedding_only"] + rejected: Final = _chat(rig, model, marker, metadata=metadata) + _assert_rejected_chat(rig, rejected, marker, model) + embedded_marker: Final = _marker() + embedded: Final = rig.gateway.request( + "POST", "/v1/embeddings", {"model": model, "input": embedded_marker, "encoding_format": "float"} + ) + assert embedded.status_code == 200, embedded.text + assert rig.upstream.targets(embedded_marker) == ("/v1/embeddings",) + + +def test_existing_non_chat_metadata_on_a_chat_backend_remains_advisory(rig: _Rig) -> None: + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + name, identities = _deploy( + scenario, + rig.api_base, + "legacy-audio", + _single({"mode": "audio_speech", "supported_endpoints": ["/v1/audio/speech"]}), + ) + _served_by_every_worker(rig.gateway, frozenset(identities)) + response: Final = _chat(rig, name, marker) + _assert_served(rig, response, marker, "/v1/chat/completions") + + +def test_a_mixed_group_skips_its_restricted_embedding_deployment(rig: _Rig) -> None: + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + name: Final = f"gate-mixed-{uuid.uuid4().hex[:12]}" + embed: Final = _create( + scenario.gateway, + name, + {"model": f"openai/{_EMBEDDING_MODEL}", "api_base": rig.api_base, "order": 0}, + _EMBED_ONLY, + ) + scenario.cleanups.callback(scenario.delete_model, embed) + chat: Final = _create( + scenario.gateway, name, {"model": f"openai/{_CHAT_MODEL}", "api_base": rig.api_base, "order": 1}, {} + ) + scenario.cleanups.callback(scenario.delete_model, chat) + _served_by_every_worker(rig.gateway, frozenset({embed, chat})) + response: Final = _chat(rig, name, marker) + _assert_served(rig, response, marker, "/v1/chat/completions") + calls: Final = rig.upstream.calls(marker) + assert [JSON_OBJECT.validate_json(call.body)["model"] for call in calls] == [_CHAT_MODEL] + + +@pytest.mark.parametrize( + "fallbacks, expected_status", + [(("embedding_only",), 429), (("embedding_only", "chat_target"), 200), (("missing",), 400)], +) +def test_local_rate_limit_skips_restricted_fallbacks_and_still_tries_unloaded_ones( + rig: _Rig, fallbacks: tuple[str, ...], expected_status: int +) -> None: + limited: Final = rig.names["chat_rate_limited"] + targets: Final = tuple(rig.names[name] if name != "missing" else _marker() for name in fallbacks) + with rig.gateway.scenario() as scenario: + key: Final = scenario.key( + model_rpm_limit={limited: 1}, router_settings={"fallbacks": [{limited: list(targets)}]} + ) + first: Final = _chat(rig, limited, _marker(), key=key) + assert first.status_code == 200, first.text + marker: Final = _marker() + second: Final = _chat(rig, limited, marker, key=key) + assert second.status_code == expected_status, second.text + if expected_status == 200: + _assert_served(rig, second, marker, "/v1/chat/completions") + return + assert "Supported endpoints" not in second.text, second.text + assert rig.upstream.calls(marker) == () + + +def _websocket_fallback_config(directory: Path, primary: tuple[str, str], fallback: tuple[str, str]) -> Path: + shared: Final = _CONFIG.validate_python(yaml.safe_load((_TESTS_ROOT / "integration/proxy_config.yaml").read_text())) + config: Final = { + **shared, + "model_list": [ + _routed_entry(primary[0], primary[1], f"openai/{_EMBEDDING_MODEL}", _EMBED_ONLY), + _routed_entry(fallback[0], fallback[1], f"openai/{_CHAT_MODEL}", {"mode": "chat"}), + ], + "router_settings": { + **object_value(shared["router_settings"]), + "num_retries": 0, + "fallbacks": [{primary[0]: [fallback[0]]}], + }, + } + path: Final = directory / "gate-websocket-fallback.yaml" + path.write_text(yaml.safe_dump(config, sort_keys=False)) + return path + + +def test_responses_websocket_uses_a_configured_chat_fallback_without_calling_the_embedding_provider( + rig: _Rig, tmp_path: Path +) -> None: + marker: Final = _marker() + suffix: Final = uuid.uuid4().hex[:12] + primary: Final = f"gate-ws-embed-{suffix}" + fallback: Final = f"gate-ws-chat-{suffix}" + certificate: Final = write_self_signed_cert(tmp_path) + with ( + responses_ws_peer.responses_peer( + certificate, lambda identity: responses_ws_peer.events(identity, _SERVED_TEXT) + ) as peer, + owned_proxy_process( + rig.gateway, + tmp_path, + {"SSL_CERT_FILE": str(certificate[0])}, + config=_websocket_fallback_config(tmp_path, (primary, f"{rig.api_base}/{marker}"), (fallback, peer.url)), + workers=2, + state=OwnedState.SHARED_WITH_WORKER, + ) as owned, + owned.gateway.scenario() as scenario, + ): + key: Final = scenario.key() + (received,) = asyncio.run( + responses_ws_peer.session(str(owned.gateway.client.base_url), key, primary, (marker,)) + ) + assert received[-1]["type"] == "response.completed", received + assert responses_ws_peer.completed(received)["status"] == "completed" + assert any(event.get("delta") == _SERVED_TEXT for event in received), received + (forwarded,) = responses_ws_peer.drain(peer.frames) + assert marker in json.dumps(forwarded), forwarded + assert rig.upstream.calls(marker) == () + + +def test_client_routing_fields_cannot_disable_an_explicit_restriction(rig: _Rig) -> None: + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + model, identities = _deploy(scenario, rig.api_base, "restricted-chat", _single(_EMBED_ONLY)) + _served_by_every_worker(rig.gateway, frozenset(identities)) + response: Final = _chat( + rig, + model, + marker, + original_generic_function="caller-value", + additional_drop_params=["original_generic_function"], + ) + _assert_rejected_chat(rig, response, marker, model) + + +def test_a_compaction_endpoint_declaration_keeps_the_model_available_for_compaction(rig: _Rig) -> None: + marker: Final = _marker() + with rig.gateway.scenario() as scenario: + model, identities = _deploy( + scenario, + rig.api_base, + "compact-declaration", + _single({"supported_endpoints": ["/responses/compact"], "enforce_supported_endpoints": True}), + ) + _served_by_every_worker(rig.gateway, frozenset(identities)) + response: Final = rig.gateway.request("POST", "/responses/compact", {"model": model, "input": marker}) + _assert_served(rig, response, marker, "/v1/responses/compact") + + +@pytest.mark.parametrize("alias_source", ["key_alias", "key_router", "team_router"]) +def test_local_rate_limit_fallbacks_resolve_saved_aliases(rig: _Rig, tmp_path: Path, alias_source: str) -> None: + suffix: Final = uuid.uuid4().hex[:12] + names: Final = {label: f"gate-fallback-alias-{label}-{suffix}" for label in ("primary", "target", "alias")} + shared: Final = _CONFIG.validate_python(yaml.safe_load((_TESTS_ROOT / "integration/proxy_config.yaml").read_text())) + config: Final = { + **shared, + "model_list": [ + _routed_entry(names["primary"], rig.api_base, f"openai/{_CHAT_MODEL}", {"mode": "chat"}), + _routed_entry(names["target"], rig.api_base, f"openai/{_CHAT_MODEL}", {"mode": "chat"}), + ], + "router_settings": { + **object_value(shared["router_settings"]), + "num_retries": 0, + "fallbacks": [{names["primary"]: [names["alias"]]}], + }, + } + path: Final = tmp_path / "fallback-alias.yaml" + path.write_text(yaml.safe_dump(config, sort_keys=False)) + aliases: Final = {names["alias"]: names["target"]} + with ( + owned_proxy_process( + rig.gateway, tmp_path, {}, config=path, workers=2, state=OwnedState.SHARED_WITH_WORKER + ) as owned, + owned.gateway.scenario() as scenario, + ): + team: Final = ( + scenario.team(router_settings={"model_group_alias": aliases}) if alias_source == "team_router" else None + ) + key: Final = scenario.key( + team_id=team, + aliases=aliases if alias_source == "key_alias" else {}, + router_settings={"model_group_alias": aliases} if alias_source == "key_router" else None, + model_rpm_limit={names["primary"]: 1}, + ) + first: Final = owned.gateway.request( + "POST", "/v1/chat/completions", _chat_body(names["primary"], _marker()), key=key + ) + assert first.status_code == 200, first.text + marker: Final = _marker() + second: Final = owned.gateway.request( + "POST", "/v1/chat/completions", _chat_body(names["primary"], marker), key=key + ) + _assert_fallback_answered(rig.upstream, second, marker) diff --git a/tests/unit/proxy/common_utils/test_callback_utils.py b/tests/unit/proxy/common_utils/test_callback_utils.py index ecb2375d495..c8d909ac2eb 100644 --- a/tests/unit/proxy/common_utils/test_callback_utils.py +++ b/tests/unit/proxy/common_utils/test_callback_utils.py @@ -49,6 +49,19 @@ def test_get_remaining_tokens_and_requests_from_request_data(): } +@pytest.mark.parametrize("client_model_group", [7, ["embed-only"], {"name": "embed-only"}]) +def test_remaining_limit_headers_ignore_a_model_group_that_is_not_a_string(client_model_group: object): + casedata: Final = { + "metadata": { + "model_group": client_model_group, + f"litellm-key-remaining-requests-{client_model_group}": 100, + f"litellm-key-remaining-tokens-{client_model_group}": 200, + } + } + + assert get_remaining_tokens_and_requests_from_request_data(casedata) == {} + + @patch( "litellm.proxy.common_utils.callback_utils.CustomLogger.get_callback_env_vars", return_value=["API_KEY", "MISSING_VAR"], diff --git a/tests/unit/proxy/management_endpoints/test_model_management_endpoints.py b/tests/unit/proxy/management_endpoints/test_model_management_endpoints.py index b892dc97e5c..835e7605955 100644 --- a/tests/unit/proxy/management_endpoints/test_model_management_endpoints.py +++ b/tests/unit/proxy/management_endpoints/test_model_management_endpoints.py @@ -3617,6 +3617,33 @@ class TestDeleteTeamModels: assert any(e[0] == "delete_many" for e in prisma.events) +@pytest.mark.parametrize( + ("model_info_patch", "expected_gate"), + [ + pytest.param({"tpm": 5}, True, id="unrelated-field"), + pytest.param({"supported_endpoints": ["/v1/embeddings", "/v1/rerank"]}, True, id="endpoint-list-only"), + pytest.param({"enforce_supported_endpoints": False}, False, id="explicit-false"), + ], +) +def test_model_info_patch_changes_the_endpoint_gate_only_when_it_sends_the_flag( + model_info_patch: dict[str, object], expected_gate: bool +): + from litellm.proxy.management_endpoints.model_management_endpoints import update_db_model + + result: Final = update_db_model( + db_model=Deployment( + model_name="embedding-only", + litellm_params=LiteLLM_Params(model="openai/text-embedding-3-small"), + model_info=ModelInfo( + id="embedding-only", enforce_supported_endpoints=True, supported_endpoints=["/v1/embeddings"] + ), + ), + updated_patch=updateDeployment.model_validate({"model_info": model_info_patch}), + ) + + assert json.loads(result["model_info"]).get("enforce_supported_endpoints") is expected_gate + + def _build_db_model_for_blocked_test(): from litellm.types.router import Deployment, LiteLLM_Params, ModelInfo diff --git a/tests/unit/proxy/response_api_endpoints/test_endpoints.py b/tests/unit/proxy/response_api_endpoints/test_endpoints.py index e3d514b2ea9..b11dcc58ef6 100644 --- a/tests/unit/proxy/response_api_endpoints/test_endpoints.py +++ b/tests/unit/proxy/response_api_endpoints/test_endpoints.py @@ -2055,6 +2055,22 @@ def test_responses_ws_failure_frame_uses_the_http_status_for_the_same_failure( assert frame["error"]["message"] == handle_exception_on_proxy(failure).message +def test_responses_ws_failure_frame_shows_only_the_refusal_for_a_model_that_refuses_the_endpoint(): + from litellm.exceptions import EndpointNotSupportedByModelError + from litellm.proxy.response_api_endpoints.endpoints import _responses_ws_failure_frame + + refusal: Final = "/v1/responses is not supported by model=embed-only. Supported endpoints: /v1/embeddings" + failure: Final = EndpointNotSupportedByModelError( + message=f"{refusal}\n\nLiteLLM: model group 'embed-only' failed with the error above", model="embed-only" + ) + + assert json.loads(_responses_ws_failure_frame(failure)) == { + "type": "error", + "status": 400, + "error": {"type": "invalid_request_error", "message": refusal}, + } + + class TestWSSessionCostTracking: @pytest.mark.asyncio async def test_router_budget_limiter_skips_arealtime_call_type(self): diff --git a/tests/unit/proxy/spend_tracking/test_spend_tracking_utils.py b/tests/unit/proxy/spend_tracking/test_spend_tracking_utils.py index 23f9ac0e2d8..740334252d2 100644 --- a/tests/unit/proxy/spend_tracking/test_spend_tracking_utils.py +++ b/tests/unit/proxy/spend_tracking/test_spend_tracking_utils.py @@ -5384,6 +5384,28 @@ def test_get_logging_payload_router_rejected_request_takes_provider_from_model_g assert payload["custom_llm_provider"] == expected_provider +@pytest.mark.parametrize("client_model_group", [7, ["embed-only"], {"name": "embed-only"}]) +def test_get_logging_payload_drops_a_client_model_group_that_is_not_a_string(client_model_group: object): + payload: Final = get_logging_payload( + kwargs={ + "model": "embed-only", + "litellm_params": { + "metadata": {"user_api_key": "test-key", "model_group": client_model_group, "status": "failure"} + }, + "standard_logging_object": { + **_make_failed_request_standard_logging_payload(), + "model_group": client_model_group, + }, + }, + response_obj={}, + start_time=datetime.datetime.now(timezone.utc), + end_time=datetime.datetime.now(timezone.utc), + llm_router=None, + ) + + assert payload["model_group"] == "" + + def test_get_logging_payload_router_rejected_request_without_router_leaves_provider_empty(): assert _router_rejected_failure_payload("openai-group", None)["custom_llm_provider"] == "" diff --git a/tests/unit/proxy/test_common_request_processing.py b/tests/unit/proxy/test_common_request_processing.py index fdcbcbfcd05..e58a1de9eb5 100644 --- a/tests/unit/proxy/test_common_request_processing.py +++ b/tests/unit/proxy/test_common_request_processing.py @@ -456,7 +456,7 @@ class TestProxyBaseLLMRequestProcessing: mock_add_litellm_data_to_request, ) mock_general_settings = {} - mock_user_api_key_dict = MagicMock(spec=UserAPIKeyAuth) + mock_user_api_key_dict = ProxyUserAPIKeyAuth() mock_proxy_config = MagicMock(spec=ProxyConfig) route_type = "acompletion" @@ -740,7 +740,7 @@ class TestProxyBaseLLMRequestProcessing: returned_data, logging_obj = await processing_obj.common_processing_pre_call_logic( request=mock_request, general_settings={}, - user_api_key_dict=MagicMock(spec=UserAPIKeyAuth), + user_api_key_dict=ProxyUserAPIKeyAuth(), proxy_logging_obj=mock_proxy_logging_obj, proxy_config=MagicMock(spec=ProxyConfig), route_type=route_type, @@ -1186,7 +1186,7 @@ class TestProxyBaseLLMRequestProcessing: ) mock_general_settings = {} - mock_user_api_key_dict = MagicMock(spec=UserAPIKeyAuth) + mock_user_api_key_dict = ProxyUserAPIKeyAuth() mock_proxy_config = MagicMock(spec=ProxyConfig) mock_router_settings = { @@ -2208,7 +2208,7 @@ class TestProxyBaseLLMRequestProcessing: mock_add_litellm_data_to_request, ) mock_general_settings = {} - mock_user_api_key_dict = MagicMock(spec=UserAPIKeyAuth) + mock_user_api_key_dict = ProxyUserAPIKeyAuth() mock_proxy_config = MagicMock(spec=ProxyConfig) route_type = "acompletion" @@ -7706,6 +7706,56 @@ class TestPreCallWithFallbacksOnLocalRateLimit: assert processor.data["model"] == fallback_model assert call_count == 2 + @pytest.mark.asyncio + async def test_local_rate_limit_fallback_skips_an_aliased_embedding_only_model(self): + from litellm.proxy.common_request_processing import ProxyBaseLLMRequestProcessing + from litellm.proxy.common_utils.proxy_rate_limit_error import ProxyRateLimitError + + router: Final = Router( + model_list=[ + { + "model_name": "embed-only", + "litellm_params": {"model": "openai/text-embedding-3-small", "api_key": "sk-synthetic"}, + "model_info": {"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings"]}, + }, + {"model_name": "chat-model", "litellm_params": {"model": "openai/gpt-5.4-nano", "api_key": "sk-synthetic"}}, + ], + fallbacks=[{"rate-limited": ["embed-alias", "chat-model"]}], + ) + processor: Final = ProxyBaseLLMRequestProcessing( + data={ + "model": "rate-limited", + "router_settings_override": {"model_group_alias": {"embed-alias": "embed-only"}}, + } + ) + attempted: Final[list[object]] = [] + + async def pre_call_logic(**_: object): + attempted.append(processor.data.get("model")) + if processor.data.get("model") == "rate-limited": + raise ProxyRateLimitError(detail="key rpm limit exceeded", headers={"retry-after": "30"}) + return processor.data, MagicMock() + + with patch.object(processor, "common_processing_pre_call_logic", side_effect=pre_call_logic): + await processor._pre_call_with_fallbacks( + request=MagicMock(), + general_settings={}, + proxy_logging_obj=MagicMock(), + user_api_key_dict=ProxyUserAPIKeyAuth(), + version=None, + proxy_config=MagicMock(), + user_model=None, + user_temperature=None, + user_request_timeout=None, + user_max_tokens=None, + user_api_base=None, + model="rate-limited", + route_type="acompletion", + llm_router=router, + ) + + assert attempted == ["rate-limited", "chat-model"] + @pytest.mark.asyncio async def test_raises_when_no_fallbacks_configured(self): from litellm.proxy.common_utils.proxy_rate_limit_error import ProxyRateLimitError diff --git a/tests/unit/proxy/test_route_llm_request.py b/tests/unit/proxy/test_route_llm_request.py index 9e31d3774f9..8f364e51e2b 100644 --- a/tests/unit/proxy/test_route_llm_request.py +++ b/tests/unit/proxy/test_route_llm_request.py @@ -28,6 +28,7 @@ from litellm.proxy.route_llm_request import ( JSON_OBJECT_ADAPTER, ProxyModelNotFoundError, UserConfigRouters, + endpoint_unsupported_by_model, route_request, user_config_routers, ) @@ -3027,3 +3028,27 @@ async def test_user_config_routers_discard_the_least_recently_used_router_past_c assert first_handler() is not None assert second_handler() is None routers.clear() + + +def test_fallback_probe_skips_only_a_loaded_model_that_refuses_the_route() -> None: + router: Final = litellm.Router( + model_list=[ + { + "model_name": "chat-fallback", + "litellm_params": {"model": "openai/gpt-5.4-nano", "api_key": "unit-provider-key"}, + }, + { + "model_name": "embed-only", + "litellm_params": {"model": "openai/text-embedding-3-small", "api_key": "unit-provider-key"}, + "model_info": {"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings"]}, + }, + ], + num_retries=0, + ) + + def skipped(model: str) -> bool: + return endpoint_unsupported_by_model("acompletion", {"model": model}, router, None) + + assert skipped("embed-only") is True + assert skipped("chat-fallback") is False + assert skipped("added-on-another-worker") is False diff --git a/tests/unit/router_utils/test_endpoint_eligibility.py b/tests/unit/router_utils/test_endpoint_eligibility.py new file mode 100644 index 00000000000..5e5d77254d0 --- /dev/null +++ b/tests/unit/router_utils/test_endpoint_eligibility.py @@ -0,0 +1,139 @@ +from collections.abc import Mapping +from typing import Final + +import pytest +from pydantic import TypeAdapter + +from litellm.exceptions import EndpointNotSupportedByModelError +from litellm.router_utils.endpoint_eligibility import ( + eligible_deployments, + endpoint_selection, + warn_on_unenforced_endpoint_flag, +) + +_DEPLOYMENT: Final = TypeAdapter(list[dict[str, object]]) + + +def _deployment(info: Mapping[str, object], identity: str = "embedding") -> dict[str, object]: + return _DEPLOYMENT.validate_python( + [ + { + "model_name": "shared", + "litellm_params": {"model": "openai/test-model"}, + "model_info": {"id": identity, **info}, + } + ] + )[0] + + +@pytest.mark.parametrize("call_type", ["acompletion", "anthropic_messages", "aresponses", "acreate_interaction"]) +def test_explicit_restriction_excludes_an_embedding_deployment_from_a_mixed_group(call_type: str) -> None: + embedding: Final = _deployment({"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings"]}) + chat: Final = _deployment({"supported_endpoints": ["/v1/chat/completions"]}, "chat") + with endpoint_selection(call_type): + result: Final = eligible_deployments("shared", (embedding, chat)) + assert result == (chat,) + + +@pytest.mark.parametrize( + "info", + [ + {"mode": "audio_speech", "supported_endpoints": ["/v1/audio/speech"]}, + {"enforce_supported_endpoints": False, "supported_endpoints": ["/v1/embeddings"]}, + {"enforce_supported_endpoints": True}, + {"enforce_supported_endpoints": True, "supported_endpoints": None}, + {"enforce_supported_endpoints": True, "supported_endpoints": 7}, + {"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/messages"]}, + ], +) +def test_advisory_or_incomplete_metadata_preserves_routing(info: Mapping[str, object]) -> None: + deployment: Final = _deployment(info) + with endpoint_selection("acompletion"): + result: Final = eligible_deployments("shared", (deployment,)) + assert result == (deployment,) + + +def test_a_single_endpoint_string_is_enforced_like_a_one_item_list() -> None: + deployment: Final = _deployment({"enforce_supported_endpoints": True, "supported_endpoints": "/v1/embeddings"}) + with endpoint_selection("acompletion"): + result: Final = eligible_deployments("shared", (deployment,)) + assert isinstance(result, EndpointNotSupportedByModelError) + assert result.message.endswith("Supported endpoints: /v1/embeddings") + + +def test_unsupported_error_explains_the_requested_operation_and_available_endpoints() -> None: + deployment: Final = _deployment({"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings"]}) + with endpoint_selection("acreate_interaction"): + result: Final = eligible_deployments("shared", (deployment,)) + assert isinstance(result, EndpointNotSupportedByModelError) + assert result.status_code == 400 + assert "/v1beta/interactions" in result.message + assert "/v1/embeddings" in result.message + + +@pytest.mark.parametrize("call_type", ["aembedding", "aspeech", "_arealtime", "aget_responses"]) +def test_chat_restriction_does_not_block_other_operations(call_type: str) -> None: + deployment: Final = _deployment({"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings"]}) + with endpoint_selection(call_type): + result: Final = eligible_deployments("shared", (deployment,)) + assert result == (deployment,) + + +def test_nested_non_chat_calls_do_not_inherit_the_outer_chat_restriction() -> None: + deployment: Final = _deployment({"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings"]}) + with endpoint_selection("acompletion"): + with endpoint_selection("aembedding"): + embedding_result: Final = eligible_deployments("shared", (deployment,)) + chat_result: Final = eligible_deployments("shared", (deployment,)) + assert embedding_result == (deployment,) + assert isinstance(chat_result, EndpointNotSupportedByModelError) + + +@pytest.mark.parametrize( + "endpoint", + [ + "/messages", + "/interactions", + "/responses/compact", + "/openai/v1/responses/compact", + "/openai/v1/chat/completions", + "chat/completions", + "v1/responses", + " /v1/chat/completions ", + "/v1/chat/completions/", + "/V1/Chat/Completions", + "/v1beta/models/gemini-3-flash:generateContent", + ], +) +def test_chat_endpoint_declarations_keep_their_meaning_across_prefixes(endpoint: str) -> None: + deployment: Final = _deployment({"enforce_supported_endpoints": True, "supported_endpoints": [endpoint]}) + with endpoint_selection("acompletion"): + result: Final = eligible_deployments("shared", (deployment,)) + assert result == (deployment,) + + +@pytest.mark.parametrize( + "info", + [ + {"enforce_supported_endpoints": True}, + {"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings", None]}, + {"enforce_supported_endpoints": "true", "supported_endpoints": 7}, + ], +) +def test_a_flag_without_a_usable_endpoint_list_is_reported(info: Mapping[str, object]) -> None: + warning: Final = warn_on_unenforced_endpoint_flag("embed-only", info) + assert warning is not None + assert "'embed-only'" in warning + + +@pytest.mark.parametrize( + "info", + [ + {"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings"]}, + {"enforce_supported_endpoints": True, "supported_endpoints": "/v1/embeddings"}, + {"enforce_supported_endpoints": False}, + {"supported_endpoints": ["/v1/embeddings", None]}, + ], +) +def test_an_enforceable_or_unflagged_deployment_is_not_reported(info: Mapping[str, object]) -> None: + assert warn_on_unenforced_endpoint_flag("embed-only", info) is None diff --git a/tests/unit/test_router/test_router.py b/tests/unit/test_router/test_router.py index ee3a70c7c1c..62510dac26d 100644 --- a/tests/unit/test_router/test_router.py +++ b/tests/unit/test_router/test_router.py @@ -12270,6 +12270,67 @@ def test_is_deployment_blocked_static_helper_reflects_blocked_flag(): ) +def test_synchronous_text_completion_rejects_a_restricted_model_before_the_provider() -> None: + model: Final = "text-completion-openai/unit-text-model" + router: Final = Router( + model_list=[ + { + "model_name": model, + "litellm_params": { + "model": model, + "api_base": "https://sync-router.invalid/v1", + "api_key": "unit-provider-key", + }, + "model_info": {"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings"]}, + } + ], + num_retries=0, + ) + with respx.mock(assert_all_called=False) as upstream: + provider: Final = upstream.post("https://sync-router.invalid/v1/completions").respond( + 200, + json={ + "id": "cmpl-unit", + "object": "text_completion", + "created": 1, + "model": "unit-text-model", + "choices": [{"text": "provider called", "index": 0, "finish_reason": "stop", "logprobs": None}], + "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2}, + }, + ) + with pytest.raises(litellm.BadRequestError, match="/v1/completions is not supported"): + router.text_completion(model=model, prompt="test") + assert provider.call_count == 0 + + +def test_routable_deployments_leave_out_models_that_refuse_the_call_type() -> None: + router: Final = Router( + model_list=[ + { + "model_name": "available", + "litellm_params": {"model": "openai/gpt-5.4-nano", "api_key": "unit-provider-key"}, + }, + { + "model_name": "embed-only", + "litellm_params": {"model": "openai/text-embedding-3-small", "api_key": "unit-provider-key"}, + "model_info": {"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings"]}, + }, + ], + num_retries=0, + ) + + def routable(model: str, call_type: str | None) -> tuple[str, ...]: + return tuple( + deployment["model_name"] for deployment in router.get_routable_deployments(model, {}, call_type=call_type) + ) + + assert routable("unregistered-fallback", "acompletion") == () + assert routable("available", "acompletion") == ("available",) + assert routable("embed-only", "acompletion") == () + assert routable("embed-only", "aembedding") == ("embed-only",) + assert routable("embed-only", None) == ("embed-only",) + + def test_routing_groups_legacy_property_forwards_to_public_attribute(): router = _router_with_two_deployments([False, False]) @@ -18803,6 +18864,35 @@ async def test_anthropic_messages_stream_dropped_before_content_is_retried_withi assert [(attempted, budget) for _, attempted, budget in provider.calls] == [(0, 2), (1, 2)] +@pytest.mark.asyncio +async def test_anthropic_messages_same_group_retry_skips_an_embedding_only_deployment(): + """The retry of a /v1/messages stream dropped before content must not land on an enforced embedding-only + deployment of the same group, even when that deployment is the group's preferred order.""" + router: Final = Router( + model_list=[ + {"model_name": "glm", "litellm_params": {"model": "anthropic/glm-chat", "api_key": "sk-test", "order": 2}}, + { + "model_name": "glm", + "litellm_params": {"model": "anthropic/glm-embed", "api_key": "sk-test", "order": 1}, + "model_info": {"enforce_supported_endpoints": True, "supported_endpoints": ["/v1/embeddings"]}, + }, + ], + num_retries=1, + ) + provider: Final = _AnthropicMessagesScriptedProvider( + _anthropic_messages_dropped_before_content, _anthropic_messages_retried_stream + ) + + async def anthropic_messages(**kwargs: object) -> object: + return await provider(**kwargs) + + stream: Final = await _anthropic_messages_stream_through_router(router, anthropic_messages) + body: Final = [chunk async for chunk in stream] + + assert body == [_anthropic_messages_message_start_chunk(), _anthropic_messages_content_chunk("pong")] + assert [model for model, _, _ in provider.calls] == ["anthropic/glm-chat", "anthropic/glm-chat"] + + class _AnthropicMessagesRetryStashRecorder: def __init__(self) -> None: self.observed: tuple[Usage | None, ...] = () @@ -31431,3 +31521,26 @@ def test_router_keeps_supported_fallback_shapes( router: Final = Router(model_list=list(_FALLBACK_VALIDATION_MODEL_LIST), **router_kwargs) assert router.fallbacks == expected_fallbacks + + +def test_router_warns_when_a_deployment_enforces_endpoints_without_a_list(caplog: pytest.LogCaptureFixture) -> None: + with caplog.at_level(logging.WARNING, logger="LiteLLM Router"): + router: Final = Router( + model_list=[ + { + "model_name": "embed-from-config", + "litellm_params": {"model": "openai/text-embedding-3-small", "api_key": "sk-fake"}, + "model_info": {"enforce_supported_endpoints": True}, + } + ] + ) + router.add_deployment( + Deployment( + model_name="embed-from-api", + litellm_params=LiteLLM_Params(model="openai/text-embedding-3-small", api_key="sk-fake"), + model_info=ModelInfo(enforce_supported_endpoints=True, supported_endpoints=["/v1/embeddings", 7]), + ) + ) + + assert "Deployment 'embed-from-config' sets enforce_supported_endpoints without" in caplog.text + assert "Deployment 'embed-from-api' sets enforce_supported_endpoints without" in caplog.text diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index f114c1c73d2..6faffd7a050 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -50433,6 +50433,8 @@ export interface components { discoverable?: boolean | null; /** Enable Tag Filtering */ enable_tag_filtering?: boolean | null; + /** Enforce Supported Endpoints */ + enforce_supported_endpoints?: boolean | null; /** Explicit Cache Read Input Token Cost */ explicit_cache_read_input_token_cost?: number | null; /** Id */