From ca0ff24726e08444058cab12caef4eb2a12914e9 Mon Sep 17 00:00:00 2001 From: "devin-ai-integration[bot]" <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Sat, 10 Oct 2026 23:58:16 -0700 Subject: [PATCH] fix(databricks): call the ai_decide AI Function route for decisions and the auto-router (#45722) * fix(databricks): call the ai_decide AI Function route for decisions and the auto-router Databricks serves ai_decide at POST /api/2.0/ai-functions/ai-decide, not on a serving endpoint's /invocations route. The decisions provider and the auto-router classifier now accept only the ai_decide model, post to that route without a model field, unwrap the response envelope, and map a noul answer's probability to the System One noul field. * test(databricks): bound the megabyte model name test by scaling, not wall time * test(databricks): assert the whole OpenAI-format decision and floor the liveliness bound * test(decisions): move the Databricks OpenAI-format wire row to the ai_decide route * fix(databricks): keep serving endpoint decisions beside the ai_decide route * test(databricks): assert the whole classifier request body in the test routing case * test(databricks): correct System One spend call type Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Co-authored-by: mateo Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../databricks/decisions/transformation.py | 82 ++++++- ...odel_prices_and_context_window_backup.json | 12 + .../complexity_router/config.py | 11 +- .../complexity_router/jev_classifier.py | 8 +- model_prices_and_context_window.json | 12 + ...uto_router_databricks_classifier_writes.py | 157 ++++++++----- .../management/test_model_health_check.py | 42 +++- .../test_databricks_decisions_wire.py | 219 ++++++++++++++---- .../test_decisions_openai_format_wire.py | 36 ++- ...complexity_router_databricks_classifier.py | 62 +++-- tests/unit/decisions/test_main.py | 79 ++++++- .../decisions/test_transformation.py | 72 +++++- .../decisions_endpoints/test_endpoints.py | 47 ++++ .../test_auto_router_permissions.py | 12 +- .../complexity_router/test_jev_classifier.py | 73 ++++-- .../add_model/ClassifierTypeRadios.tsx | 4 +- .../JevClassifierConfig.integration.test.tsx | 2 +- .../add_model/JevClassifierConfig.tsx | 4 +- ...d_auto_router_routing_test_request.test.ts | 2 + .../build_complexity_router_config.test.ts | 1 + .../add_model/jev_classifier_config.test.ts | 6 +- .../add_model/jev_classifier_config.ts | 9 +- 22 files changed, 761 insertions(+), 191 deletions(-) diff --git a/litellm/llms/databricks/decisions/transformation.py b/litellm/llms/databricks/decisions/transformation.py index a454da83e50..c2ac1f220b8 100644 --- a/litellm/llms/databricks/decisions/transformation.py +++ b/litellm/llms/databricks/decisions/transformation.py @@ -4,18 +4,63 @@ from dataclasses import dataclass, field from types import MappingProxyType from typing import Final +from pydantic import TypeAdapter, ValidationError + from litellm.llms.base_llm.decisions.transformation import BaseDecisionsConfig +from litellm.types.decisions import DecisionsIRRequest, UnsupportedDecisionsRequest +DATABRICKS_AI_DECIDE_MODEL: Final = "ai_decide" +_AI_DECIDE_PATH: Final = "/api/2.0/ai-functions/ai-decide" _SERVING_ENDPOINT_NAME: Final = re.compile(r"[A-Za-z0-9_-][A-Za-z0-9._-]*") +_MAPPING_ADAPTER: Final[TypeAdapter[Mapping[str, object]]] = TypeAdapter(Mapping[str, object]) -def validate_serving_endpoint_name(name: str) -> str: - if _SERVING_ENDPOINT_NAME.fullmatch(name) is None: - raise ValueError( - f"Databricks serving endpoint name {name!r} must be the bare endpoint name (letters, digits, '-', '_' " - "and '.', with no '/', '?', '#', spaces, or a leading '.'), e.g. databricks-openjev-qwen35-4b" - ) - return name +def is_ai_decide_model(model: str) -> bool: + return model == DATABRICKS_AI_DECIDE_MODEL + + +def validate_databricks_decisions_model(model: str) -> str: + if is_ai_decide_model(model) or _SERVING_ENDPOINT_NAME.fullmatch(model) is not None: + return model + raise ValueError( + f"Databricks decisions model must be {DATABRICKS_AI_DECIDE_MODEL!r} (the ai_decide AI Function) or a bare " + "serving endpoint name (letters, digits, '-', '_' and '.', with no '/', '?', '#', spaces, or a leading '.'), " + "e.g. databricks-openjev-qwen35-4b" + ) + + +def databricks_workspace_host(api_base: str) -> str: + return api_base.rstrip("/").removesuffix("/serving-endpoints") + + +def _mapping(value: object) -> Mapping[str, object] | None: + try: + return _MAPPING_ADAPTER.validate_python(value) + except ValidationError: + return None + + +def _systemone_answer(answer: object) -> object: + fields: Final = _mapping(answer) + if fields is None or fields.get("type") != "noul" or "probability" not in fields: + return answer + return MappingProxyType( + {**{key: value for key, value in fields.items() if key != "probability"}, "noul": fields["probability"]} + ) + + +def ai_decide_response(payload: Mapping[str, object]) -> Mapping[str, object]: + response: Final = _mapping(payload.get("response")) + answers: Final = None if response is None else _mapping(response.get("answers")) + if response is None or answers is None: + return payload + return MappingProxyType( + { + **{key: value for key, value in payload.items() if key != "response"}, + **response, + "answers": MappingProxyType({key: _systemone_answer(answer) for key, answer in answers.items()}), + } + ) @dataclass(frozen=True, slots=True) @@ -35,13 +80,30 @@ class DatabricksDecisionsConfig(BaseDecisionsConfig): ) def canonical_model(self, model: str) -> str: - return validate_serving_endpoint_name(model) + return validate_databricks_decisions_model(model) def get_complete_url(self, api_base: str, model: str) -> str: - return f"{api_base.rstrip('/')}/{validate_serving_endpoint_name(model)}/invocations" + if is_ai_decide_model(model): + return f"{databricks_workspace_host(api_base)}{_AI_DECIDE_PATH}" + return f"{api_base.rstrip('/')}/{validate_databricks_decisions_model(model)}/invocations" + + def transform_decisions_request( + self, + model: str, + request: DecisionsIRRequest, + custom_llm_provider: str, + ) -> Mapping[str, object] | UnsupportedDecisionsRequest: + body: Final = super().transform_decisions_request(model, request, custom_llm_provider) + if isinstance(body, UnsupportedDecisionsRequest) or not is_ai_decide_model(model): + return body + return MappingProxyType({key: value for key, value in body.items() if key != "model"}) + + def unwrap_response(self, payload: object) -> object: + fields: Final = _mapping(payload) + return payload if fields is None else ai_decide_response(fields) def classifier_response(self, body: Mapping[str, object], requested_model: str) -> Mapping[str, object]: - return MappingProxyType({**body, "model": requested_model}) + return MappingProxyType({**ai_decide_response(body), "model": requested_model}) def connection(self, api_base: str | None, api_key: str | None) -> DatabricksDecisionsConnection: base: Final = self.resolve_api_base(api_base) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index bf7f1db7a31..782ef2f2ff2 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -21177,6 +21177,18 @@ "output_cost_per_token": 0.0, "source": "https://docs.databricks.com/aws/en/machine-learning/foundation-model-apis/supported-models" }, + "databricks/ai_decide": { + "cache_creation_input_token_cost": 0.0, + "cache_read_input_token_cost": 0.0, + "input_cost_per_token": 0.0, + "litellm_provider": "databricks", + "metadata": { + "notes": "The ai_decide AI Function (Beta), called through POST /api/2.0/ai-functions/ai-decide. Its response carries no token usage, so LiteLLM records no spend for it." + }, + "mode": "evaluation", + "output_cost_per_token": 0.0, + "source": "https://docs.databricks.com/api/ai-functions/v1/ai-decide" + }, "dataforseo/search": { "input_cost_per_query": 0.003, "litellm_provider": "dataforseo", diff --git a/litellm/router_strategy/complexity_router/config.py b/litellm/router_strategy/complexity_router/config.py index 992c098bc80..4a2c56daf00 100644 --- a/litellm/router_strategy/complexity_router/config.py +++ b/litellm/router_strategy/complexity_router/config.py @@ -769,14 +769,17 @@ class OpenSourceClassifierConfig(LiteLLMBaseModel): _ = validate_oss_api_base(self.provider, self.api_base) return self if self.provider == "databricks": - from litellm.llms.databricks.decisions.transformation import validate_serving_endpoint_name + from litellm.llms.databricks.decisions.transformation import ( + DATABRICKS_AI_DECIDE_MODEL, + validate_databricks_decisions_model, + ) if "model" not in self.model_fields_set: raise ValueError( - "opensource_classifier_config.model is required for provider 'databricks': the serving endpoint " - "name, e.g. databricks-openjev-qwen35-4b" + "opensource_classifier_config.model is required for provider 'databricks': " + f"{DATABRICKS_AI_DECIDE_MODEL!r} or a serving endpoint name, e.g. databricks-openjev-qwen35-4b" ) - _ = validate_serving_endpoint_name(self.model) + _ = validate_databricks_decisions_model(self.model) if self.api_base is not None and self.api_key is None: raise ValueError( "opensource_classifier_config.api_base requires opensource_classifier_config.api_key: " diff --git a/litellm/router_strategy/complexity_router/jev_classifier.py b/litellm/router_strategy/complexity_router/jev_classifier.py index b8b3b4aadba..87bd031490a 100644 --- a/litellm/router_strategy/complexity_router/jev_classifier.py +++ b/litellm/router_strategy/complexity_router/jev_classifier.py @@ -19,7 +19,7 @@ from litellm.litellm_core_utils.internal_call_metadata import ( from litellm.litellm_core_utils.litellm_logging import Logging from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler -from litellm.llms.databricks.decisions.transformation import DATABRICKS_DECISIONS_CONFIG +from litellm.llms.databricks.decisions.transformation import DATABRICKS_DECISIONS_CONFIG, is_ai_decide_model from litellm.llms.laya.common_utils import laya_response_model from litellm.llms.pass_through.typesafe_logging_handler import TypeSafePassthroughLoggingHandler from litellm.router_strategy.complexity_router.config import DEFAULT_JEV_INSTRUCTIONS as _DEFAULT_JEV_INSTRUCTIONS @@ -33,6 +33,7 @@ if TYPE_CHECKING: JevProbability: TypeAlias = Annotated[float, Field(ge=0.0, le=1.0)] ClassifierProvider: TypeAlias = Literal["typesafe", "laya", "bespoke", "databricks"] DEFAULT_JEV_INSTRUCTIONS: Final = _DEFAULT_JEV_INSTRUCTIONS +_WITHOUT_MODEL: Final = MappingProxyType({"model": True}) class JevChoiceQuestion(LiteLLMBaseModel): @@ -181,7 +182,7 @@ class HttpJevClassifierClient: ) response: Final = await self._http_client.post( # pyright: ignore[reportUnknownMemberType] # AsyncHTTPHandler has a dynamic post signature self._request_url(request.model), - json=request.model_dump(mode="json"), + json=request.model_dump(mode="json", exclude=self._excluded_fields(request.model)), headers=MappingProxyType({**authorization, "Content-Type": "application/json"}), # pyright: ignore[reportArgumentType] # HTTP headers are not mutated by AsyncHTTPHandler timeout=timeout_s, ) @@ -194,6 +195,9 @@ class HttpJevClassifierClient: verbose_router_logger.warning("JEV response logging failed (%s)", type(exc).__name__) return TypeAdapter(JevSystemOneResponse).validate_python(normalized_body) + def _excluded_fields(self, model: str) -> Mapping[str, bool] | None: + return _WITHOUT_MODEL if self._provider == "databricks" and is_ai_decide_model(model) else None + def _request_url(self, model: str) -> str: if self._provider == "databricks": return DATABRICKS_DECISIONS_CONFIG.get_complete_url(self._api_base, model) diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index bf7f1db7a31..782ef2f2ff2 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -21177,6 +21177,18 @@ "output_cost_per_token": 0.0, "source": "https://docs.databricks.com/aws/en/machine-learning/foundation-model-apis/supported-models" }, + "databricks/ai_decide": { + "cache_creation_input_token_cost": 0.0, + "cache_read_input_token_cost": 0.0, + "input_cost_per_token": 0.0, + "litellm_provider": "databricks", + "metadata": { + "notes": "The ai_decide AI Function (Beta), called through POST /api/2.0/ai-functions/ai-decide. Its response carries no token usage, so LiteLLM records no spend for it." + }, + "mode": "evaluation", + "output_cost_per_token": 0.0, + "source": "https://docs.databricks.com/api/ai-functions/v1/ai-decide" + }, "dataforseo/search": { "input_cost_per_query": 0.003, "litellm_provider": "dataforseo", diff --git a/tests/integration/management/test_auto_router_databricks_classifier_writes.py b/tests/integration/management/test_auto_router_databricks_classifier_writes.py index e277eddf009..d954fd49d81 100644 --- a/tests/integration/management/test_auto_router_databricks_classifier_writes.py +++ b/tests/integration/management/test_auto_router_databricks_classifier_writes.py @@ -2,28 +2,34 @@ import asyncio import json import threading import uuid -from collections.abc import Callable, Mapping +from collections.abc import Callable, Mapping, Sequence from queue import SimpleQueue from typing import Final import httpx +import pytest from integration._support.client import Gateway, Scenario, eventually, object_value, string_value from integration._support.database import read_rows from integration._support.openai_wire import answering_model_discovery, chat_reply from integration._support.wire import Reply, Request, Wire, wire_server from pydantic import JsonValue +from litellm.router_strategy.complexity_router.config import DEFAULT_JEV_INSTRUCTIONS + +_MODEL: Final = "ai_decide" +_ROUTE: Final = "/api/2.0/ai-functions/ai-decide" _ENDPOINT: Final = "databricks-openjev-qwen35-4b" -_SECOND_ENDPOINT: Final = "databricks-openjev-qwen35-4b-canary" +_SECOND_WORKSPACE: Final = "/second-workspace" _API_KEY: Final = "synthetic-databricks-key" -_BARE_NAME_RULE: Final = ( - "must be the bare endpoint name (letters, digits, '-', '_' and '.', with no '/', '?', '#', spaces, or a " - "leading '.'), e.g. databricks-openjev-qwen35-4b" +_MODEL_RULE: Final = ( + "Databricks decisions model must be 'ai_decide' (the ai_decide AI Function) or a bare serving endpoint name " + "(letters, digits, '-', '_' and '.', with no '/', '?', '#', spaces, or a leading '.'), e.g. " + "databricks-openjev-qwen35-4b" ) _REJECTED_AT_LOAD: Final = "The router would drop this deployment at load time, so the write is rejected instead." _MODEL_REQUIRED: Final = ( - "opensource_classifier_config.model is required for provider 'databricks': the serving endpoint name, " - "e.g. databricks-openjev-qwen35-4b" + "opensource_classifier_config.model is required for provider 'databricks': 'ai_decide' or a serving endpoint " + "name, e.g. databricks-openjev-qwen35-4b" ) _BASE_NEEDS_KEY: Final = ( "opensource_classifier_config.api_base requires opensource_classifier_config.api_key: DATABRICKS_API_KEY or " @@ -32,17 +38,26 @@ _BASE_NEEDS_KEY: Final = ( _BLANK_KEY: Final = ( "opensource_classifier_config.api_key must be non-empty; omit it to use the provider environment key" ) -_CLASSIFIER_ANSWER: Final[dict[str, JsonValue]] = { - "answers": { - "tier": { - "type": "choice", - "choice": "COMPLEX", - "confidence": 0.91, - "probabilities": {"SIMPLE": 0.03, "MEDIUM": 0.06, "COMPLEX": 0.91}, - } - }, +_ANSWERS: Final[dict[str, JsonValue]] = { + "tier": { + "type": "choice", + "choice": "COMPLEX", + "confidence": 0.91, + "probabilities": {"SIMPLE": 0.03, "MEDIUM": 0.06, "COMPLEX": 0.91}, + } +} +_AI_DECIDE_ANSWER: Final[dict[str, JsonValue]] = {"response": {"answers": _ANSWERS}, "metadata": {"version": "1.0"}} +_ENDPOINT_ANSWER: Final[dict[str, JsonValue]] = { + "model": "/mosaicml/local_model", + "answers": _ANSWERS, "usage": {"input_tokens": 12, "output_tokens": 1}, } +_TIER_CRITERIA: Final[dict[str, JsonValue]] = { + "SIMPLE": "Greetings, chitchat, or short factual lookups with known answers", + "MEDIUM": "Everyday requests needing explanation, light reasoning, or minor technical work", + "COMPLEX": "Non-trivial code, architecture, multi-step work, or specialized domain depth", + "REASONING": "Open-ended analysis, proofs, tradeoffs, or tasks requiring careful thought", +} _ROWS_QUERY: Final = ( "SELECT request_id, status, metadata->>'internal_call_origin' AS origin " 'FROM "LiteLLM_SpendLogs" WHERE model_group = %s' @@ -50,7 +65,8 @@ _ROWS_QUERY: Final = ( def _classifier(request: Request) -> Reply: - return Reply(body=json.dumps(_CLASSIFIER_ANSWER).encode()) + answer: Final = _ENDPOINT_ANSWER if request.target.endswith("/invocations") else _AI_DECIDE_ANSWER + return Reply(body=json.dumps(answer).encode()) def _tier(text: str) -> Callable[[Request], Reply]: @@ -64,7 +80,7 @@ def _tier(text: str) -> Callable[[Request], Reply]: def _classifier_config(api_base: str, **overrides: JsonValue) -> dict[str, JsonValue]: return { "provider": "databricks", - "model": _ENDPOINT, + "model": _MODEL, "api_base": api_base, "api_key": _API_KEY, "timeout_ms": 20000, @@ -133,21 +149,39 @@ async def _burst(base_url: str, key: str, model: str, count: int) -> tuple[httpx ) -def _rows(router: str, *, want: int) -> tuple[dict[str, JsonValue], ...]: - return tuple(eventually(lambda: read_rows(_ROWS_QUERY, (router,)), lambda found: len(found) >= want, seconds=70)) +def _request_rows(rows: Sequence[dict[str, JsonValue]]) -> tuple[dict[str, JsonValue], ...]: + return tuple(row for row in rows if row["origin"] != "autorouter_classifier") + + +def _rows(router: str, *, requests: int) -> tuple[dict[str, JsonValue], ...]: + return tuple( + eventually( + lambda: read_rows(_ROWS_QUERY, (router,)), + lambda found: len(_request_rows(found)) >= requests, + seconds=70, + ) + ) + + +def test_validate_accepts_a_databricks_ai_decide_classifier_with_or_without_the_serving_endpoints_suffix( + gateway: Gateway, +) -> None: + for api_base in ( + "https://dbc-example.cloud.databricks.com", + "https://dbc-example.cloud.databricks.com/serving-endpoints", + ): + config: Final = _router_config(_classifier_config(api_base), "simple", "complex") + assert _validate(gateway, config) == {"valid": True, "error": None}, api_base def test_validate_accepts_a_databricks_classifier_with_a_bare_endpoint_name_of_any_length(gateway: Gateway) -> None: - config: Final = _router_config( - _classifier_config("https://dbc-example.cloud.databricks.com/serving-endpoints"), "simple", "complex" - ) - assert _validate(gateway, config) == {"valid": True, "error": None} - long_name: Final = _router_config( - _classifier_config("https://dbc-example.cloud.databricks.com/serving-endpoints", model="e" * 5000), - "simple", - "complex", - ) - assert _validate(gateway, long_name) == {"valid": True, "error": None} + for model in (_ENDPOINT, "e" * 5000): + config: Final = _router_config( + _classifier_config("https://dbc-example.cloud.databricks.com/serving-endpoints", model=model), + "simple", + "complex", + ) + assert _validate(gateway, config) == {"valid": True, "error": None}, model[:40] def test_validate_names_each_databricks_classifier_misconfiguration(gateway: Gateway) -> None: @@ -159,10 +193,10 @@ def test_validate_names_each_databricks_classifier_misconfiguration(gateway: Gat _without(base, "api_key"), f"at opensource_classifier_config: Value error, {_BASE_NEEDS_KEY}", ), - ("slash in the name", {**base, "model": "a/b"}, f"Databricks serving endpoint name 'a/b' {_BARE_NAME_RULE}"), - ("space in the name", {**base, "model": "a b"}, f"Databricks serving endpoint name 'a b' {_BARE_NAME_RULE}"), - ("leading dot", {**base, "model": ".hidden"}, f"Databricks serving endpoint name '.hidden' {_BARE_NAME_RULE}"), - ("empty name", {**base, "model": ""}, f"Databricks serving endpoint name '' {_BARE_NAME_RULE}"), + ("slash in the name", {**base, "model": "a/b"}, _MODEL_RULE), + ("space in the name", {**base, "model": "a b"}, _MODEL_RULE), + ("leading dot", {**base, "model": ".hidden"}, _MODEL_RULE), + ("empty name", {**base, "model": ""}, _MODEL_RULE), ( "blank api_key", {**base, "api_key": ""}, @@ -185,8 +219,13 @@ def test_validate_names_each_databricks_classifier_misconfiguration(gateway: Gat assert error.endswith(_REJECTED_AT_LOAD), (label, error) -def test_test_routing_classifies_through_the_databricks_endpoint_without_calling_the_tier_it_picked( - gateway: Gateway, +@pytest.mark.parametrize( + ("model", "target", "cost"), + [(_MODEL, _ROUTE, {}), (_ENDPOINT, f"/{_ENDPOINT}/invocations", {"classifier_cost": 0.0})], + ids=["ai_decide", "serving_endpoint"], +) +def test_test_routing_classifies_through_databricks_without_calling_the_tier_it_picked( + gateway: Gateway, model: str, target: str, cost: dict[str, JsonValue] ) -> None: prompt: Final = f"design a distributed cache {uuid.uuid4().hex}" with ( @@ -197,7 +236,7 @@ def test_test_routing_classifies_through_the_databricks_endpoint_without_calling ): simple_model: Final = scenario.model(model="openai/simple-tier", api_base=simple.url) complex_model: Final = scenario.model(model="openai/complex-tier", api_base=complex_tier.url) - config: Final = _router_config(_classifier_config(judge.url), simple_model, complex_model) + config: Final = _router_config(_classifier_config(judge.url, model=model), simple_model, complex_model) response: Final = gateway.request( "POST", "/auto_router/test_routing", {"prompt": prompt, "complexity_router_config": config} ) @@ -218,23 +257,27 @@ def test_test_routing_classifies_through_the_databricks_endpoint_without_calling "tier-probability:MEDIUM=0.060000", "tier-probability:COMPLEX=0.910000", ], - "classifier_model": f"databricks/{_ENDPOINT}", - "classifier_cost": 0.0, + "classifier_model": f"databricks/{model}", + **cost, "classifier_probabilities": {"SIMPLE": 0.03, "MEDIUM": 0.06, "COMPLEX": 0.91}, "classifier_confidence": 0.91, "conversation_continuing": False, }, }, response.text (call,) = [request for request in judge.drain() if request.method == "POST"] - assert call.target == f"/{_ENDPOINT}/invocations", call.target + assert call.target == target, call.target assert call.headers.get("authorization") == f"Bearer {_API_KEY}", call.headers - body: Final = json.loads(call.body) - assert body["model"] == _ENDPOINT, body - assert body["state"] == f"\nClassify this message:\n{prompt}", body + assert json.loads(call.body) == { + "state": f"\nClassify this message:\n{prompt}", + **({} if model == _MODEL else {"model": model}), + "questions": { + "tier": {"type": "choice", "instructions": DEFAULT_JEV_INSTRUCTIONS, "criteria": _TIER_CRITERIA} + }, + }, call.body assert _judge_targets(simple) == () and _judge_targets(complex_tier) == () -def test_model_new_refuses_a_databricks_classifier_without_an_endpoint_name_and_stores_nothing( +def test_model_new_refuses_a_databricks_classifier_without_a_model_and_stores_nothing( gateway: Gateway, ) -> None: name: Final = f"router-{uuid.uuid4().hex[:10]}" @@ -261,7 +304,7 @@ def test_model_new_refuses_a_databricks_classifier_without_an_endpoint_name_and_ assert [entry for entry in map(object_value, listed) if entry["model_name"] == name] == [] -def test_model_update_refuses_dropping_the_endpoint_name_and_keeps_the_stored_classifier_serving( +def test_model_update_refuses_dropping_the_model_and_keeps_the_stored_classifier_serving( gateway: Gateway, ) -> None: with ( @@ -284,15 +327,15 @@ def test_model_update_refuses_dropping_the_endpoint_name_and_keeps_the_stored_cl assert _MODEL_REQUIRED in string_value(object_value(object_value(response.json())["error"])["message"]), ( response.text ) - assert _stored_classifier(gateway, identity)["model"] == _ENDPOINT + assert _stored_classifier(gateway, identity)["model"] == _MODEL chat: Final = gateway.request("POST", "/v1/chat/completions", _chat_body(name)) assert chat.status_code == 200, chat.text assert chat.json()["choices"][0]["message"]["content"] == "complex answer", chat.text assert chat.headers["x-litellm-complexity-router-cause"] == "jev_classifier", dict(chat.headers) - assert _judge_targets(judge) == (f"/{_ENDPOINT}/invocations",) + assert _judge_targets(judge) == (_ROUTE,) -async def test_renaming_the_endpoint_under_a_burst_moves_the_classifier_and_logs_every_request_once( +async def test_moving_the_workspace_under_a_burst_moves_the_classifier_and_logs_every_request_once( gateway: Gateway, ) -> None: base_url: Final = str(gateway.client.base_url) @@ -315,7 +358,9 @@ async def test_renaming_the_endpoint_under_a_burst_moves_the_classifier_and_logs complex_model: Final = scenario.model(model="openai/complex-tier", api_base=complex_tier.url) classifier: Final = _classifier_config(judge.url) name, identity = _create_router(scenario, _router_config(classifier, simple_model, complex_model)) - renamed: Final = _router_config({**classifier, "model": _SECOND_ENDPOINT}, simple_model, complex_model) + moved: Final = _router_config( + {**classifier, "api_base": f"{judge.url}{_SECOND_WORKSPACE}"}, simple_model, complex_model + ) async with asyncio.TaskGroup() as requests: burst: Final = requests.create_task(_burst(base_url, gateway.key, name, 24)) try: @@ -323,15 +368,15 @@ async def test_renaming_the_endpoint_under_a_burst_moves_the_classifier_and_logs patched: Final = gateway.request( "PATCH", f"/model/{identity}/update", - {"litellm_params": {"model": "auto_router/complexity_router", "complexity_router_config": renamed}}, + {"litellm_params": {"model": "auto_router/complexity_router", "complexity_router_config": moved}}, ) assert patched.status_code == 200, patched.text finally: release.set() during: Final = await burst assert [response.status_code for response in during] == [200] * 24, [response.text for response in during] - assert set(_judge_targets(judge)) <= {f"/{_ENDPOINT}/invocations", f"/{_SECOND_ENDPOINT}/invocations"} - assert _stored_classifier(gateway, identity)["model"] == _SECOND_ENDPOINT + assert set(_judge_targets(judge)) <= {_ROUTE, f"{_SECOND_WORKSPACE}{_ROUTE}"} + assert _stored_classifier(gateway, identity)["api_base"] == f"{judge.url}{_SECOND_WORKSPACE}" convergence_calls: Final[SimpleQueue[str]] = SimpleQueue() @@ -341,16 +386,16 @@ async def test_renaming_the_endpoint_under_a_burst_moves_the_classifier_and_logs convergence_calls.put(string_value(chat.json()["id"])) return _judge_targets(judge) - eventually(classifier_target, lambda targets: targets == (f"/{_SECOND_ENDPOINT}/invocations",), seconds=30) + eventually(classifier_target, lambda targets: targets == (f"{_SECOND_WORKSPACE}{_ROUTE}",), seconds=30) after: Final = await _burst(base_url, gateway.key, name, 6) assert [response.status_code for response in after] == [200] * 6, [response.text for response in after] - assert set(_judge_targets(judge)) == {f"/{_SECOND_ENDPOINT}/invocations"} + assert set(_judge_targets(judge)) == {f"{_SECOND_WORKSPACE}{_ROUTE}"} identities: Final = ( *tuple(string_value(response.json()["id"]) for response in (*during, *after)), *tuple(convergence_calls.get_nowait() for _ in range(convergence_calls.qsize())), ) assert len(set(identities)) == len(identities), identities - rows: Final = _rows(name, want=len(identities)) - request_rows: Final = [row for row in rows if row["origin"] != "autorouter_classifier"] + rows: Final = _rows(name, requests=len(identities)) + request_rows: Final = _request_rows(rows) assert sorted(string_value(row["request_id"]) for row in request_rows) == sorted(identities), request_rows assert all(row["status"] == "success" for row in rows), rows diff --git a/tests/integration/management/test_model_health_check.py b/tests/integration/management/test_model_health_check.py index ef8412a6ed1..8d77c7d1094 100644 --- a/tests/integration/management/test_model_health_check.py +++ b/tests/integration/management/test_model_health_check.py @@ -190,7 +190,7 @@ def test_evaluation_mode_health_check_of_hosted_vllm_sends_a_choice_probe(gatewa ] -_DATABRICKS_PROBE_REPLY: Final = JsonResponse( +_DATABRICKS_ENDPOINT_PROBE_REPLY: Final = JsonResponse( content_type="application/json", body={ "model": "databricks-openjev-qwen35-4b", @@ -204,7 +204,7 @@ def test_evaluation_mode_health_check_of_a_databricks_serving_endpoint_resolves_ gateway: Gateway, ) -> None: with gateway.scenario() as scenario: - handle: Final = register_scenario(f"health-decisions-{uuid.uuid4().hex[:12]}", _DATABRICKS_PROBE_REPLY) + handle: Final = register_scenario(f"health-decisions-{uuid.uuid4().hex[:12]}", _DATABRICKS_ENDPOINT_PROBE_REPLY) scenario.cleanups.callback(delete_scenario, handle) model: Final = scenario.model( model="databricks/databricks-openjev-qwen35-4b", @@ -228,3 +228,41 @@ def test_evaluation_mode_health_check_of_a_databricks_serving_endpoint_resolves_ }, ) ] + + +_DATABRICKS_AI_DECIDE_PROBE_REPLY: Final = JsonResponse( + content_type="application/json", + body={ + "response": {"answers": {"reachable": {"type": "noul", "probability": 1.0}}}, + "metadata": {"version": "1.0"}, + }, +) + + +def test_evaluation_mode_health_check_of_databricks_ai_decide_resolves_the_mode_from_the_cost_map( + gateway: Gateway, +) -> None: + with gateway.scenario() as scenario: + handle: Final = register_scenario(f"health-decisions-{uuid.uuid4().hex[:12]}", _DATABRICKS_AI_DECIDE_PROBE_REPLY) + scenario.cleanups.callback(delete_scenario, handle) + model: Final = scenario.model( + model="databricks/ai_decide", + api_base=handle.api_base(), + api_key="synthetic-databricks-key", + ) + listed: Final = gateway.request("GET", "/v2/model/info", params={"model": model}) + assert listed.status_code == 200, listed.text + assert [object_value(object_value(entry)["model_info"])["mode"] for entry in listed.json()["data"]] == [ + "evaluation" + ], listed.text + report: Final = _health_report(gateway, model) + assert (report["healthy_count"], report["unhealthy_count"]) == (1, 0), report + assert _probes_sent_to(gateway, handle) == [ + ( + f"/{handle.scenario_id}/api/2.0/ai-functions/ai-decide", + { + "state": os.environ.get("DEFAULT_HEALTH_CHECK_PROMPT", "test from litellm"), + "questions": {"reachable": {"type": "noul", "instructions": "Is the service reachable?"}}, + }, + ) + ] diff --git a/tests/integration/providers/test_databricks_decisions_wire.py b/tests/integration/providers/test_databricks_decisions_wire.py index 12065d8c370..24db0153bcd 100644 --- a/tests/integration/providers/test_databricks_decisions_wire.py +++ b/tests/integration/providers/test_databricks_decisions_wire.py @@ -14,6 +14,8 @@ from integration._support.upstream import ScenarioHandle, delete_scenario, regis from integration.cost_calculation.cost_tracking_case import JsonResponse from pydantic import JsonValue +_MODEL: Final = "ai_decide" +_ROUTE: Final = "/api/2.0/ai-functions/ai-decide" _ENDPOINT: Final = "databricks-openjev-qwen35-4b" _API_KEY: Final = "synthetic-databricks-key" _ENV_KEY: Final = "synthetic-databricks-env-key" @@ -38,9 +40,18 @@ _ANSWERS: Final[dict[str, JsonValue]] = { } _USAGE: Final[dict[str, JsonValue]] = {"input_tokens": 367, "output_tokens": 3} _ANSWER_BODY: Final[dict[str, JsonValue]] = {"model": _ENDPOINT, "answers": _ANSWERS, "usage": _USAGE} -_BARE_NAME_RULE: Final = ( - "must be the bare endpoint name (letters, digits, '-', '_' and '.', with no '/', '?', '#', spaces, or a " - "leading '.'), e.g. databricks-openjev-qwen35-4b" +_METADATA: Final[dict[str, JsonValue]] = {"version": "1.0"} +_AI_DECIDE_BODY: Final[dict[str, JsonValue]] = {"response": {"answers": _ANSWERS}, "metadata": _METADATA} +_SYSTEM_ONE_BODY: Final[dict[str, JsonValue]] = { + "model": None, + "answers": _ANSWERS, + "usage": {"input_tokens": 0, "output_tokens": 0}, + "metadata": _METADATA, +} +_MODEL_RULE: Final = ( + "Databricks decisions model must be 'ai_decide' (the ai_decide AI Function) or a bare serving endpoint name " + "(letters, digits, '-', '_' and '.', with no '/', '?', '#', spaces, or a leading '.'), e.g. " + "databricks-openjev-qwen35-4b" ) _MISSING_KEY: Final = "Missing API key for Decisions provider 'databricks'" _NON_BARE_NAMES: Final[tuple[tuple[str, str], ...]] = ( @@ -49,10 +60,17 @@ _NON_BARE_NAMES: Final[tuple[tuple[str, str], ...]] = ( ("leading-dot", ".hidden"), ("question-mark", "a?b"), ("hash", "a#b"), + ("route-path", "serving-endpoints/ai_decide"), ) -_DATABRICKS_ERRORS: Final[tuple[tuple[int, dict[str, JsonValue]], ...]] = ( - (400, {"error_code": "BAD_REQUEST", "message": "Invalid input: questions must not be empty"}), - (404, {"error_code": "RESOURCE_DOES_NOT_EXIST", "message": f"Endpoint with name '{_ENDPOINT}' does not exist."}), +_DATABRICKS_ERRORS: Final[tuple[tuple[str, str, int, dict[str, JsonValue]], ...]] = ( + (_MODEL, _ROUTE, 400, {"error_code": "BAD_REQUEST", "message": "Invalid input: questions must not be empty"}), + (_MODEL, _ROUTE, 403, {"error_code": "PERMISSION_DENIED", "message": "ai_decide is not enabled for this workspace"}), + ( + _ENDPOINT, + f"/{_ENDPOINT}/invocations", + 404, + {"error_code": "RESOURCE_DOES_NOT_EXIST", "message": f"Endpoint with name '{_ENDPOINT}' does not exist."}, + ), ) _SPEND_QUERY: Final = ( 'SELECT spend, status, call_type, model_group, custom_llm_provider, api_base FROM "LiteLLM_SpendLogs" ' @@ -60,6 +78,7 @@ _SPEND_QUERY: Final = ( ) _OWNED_PROXY_BUDGET: Final = int(2 * graceful_stop_seconds() + 120) _MEGABYTE_NAME: Final = "a" * 1_000_000 + "/" +_TENTH_MEGABYTE_NAME: Final = "a" * 100_000 + "/" def _number(value: JsonValue) -> float: @@ -92,6 +111,12 @@ def _upstream_calls(gateway: Gateway, handle: ScenarioHandle) -> list[dict[str, return _calls_to(_observed(gateway), handle) +def _timed_decide(gateway: Gateway, model: str) -> tuple[httpx.Response, float]: + started: Final = time.perf_counter() + response: Final = _decide(gateway, model) + return response, time.perf_counter() - started + + def _spend_row(call_id: str) -> dict[str, JsonValue]: rows: Final = eventually(lambda: read_rows(_SPEND_QUERY, (call_id,)), lambda found: len(found) == 1, seconds=70) return rows[0] @@ -116,7 +141,7 @@ def _router_config(classifier: Mapping[str, JsonValue], simple: str, complex_: s def _environment_classifier() -> dict[str, JsonValue]: - return {"provider": "databricks", "model": _ENDPOINT, "timeout_ms": 20000, "circuit_breaker_enabled": False} + return {"provider": "databricks", "model": _MODEL, "timeout_ms": 20000, "circuit_breaker_enabled": False} def _router(scenario: Scenario, classifier: Mapping[str, JsonValue]) -> str: @@ -155,8 +180,8 @@ def _assert_classified_by_the_complex_tier(response: httpx.Response, router: str response.headers["x-litellm-model-name"], response.headers["x-litellm-complexity-router-tier"], response.headers["x-litellm-complexity-router-cause"], - response.headers["x-litellm-classifier-cost"], - ) == (router, "openai/complex-tier", "COMPLEX", "jev_classifier", "0.0"), dict(response.headers) + ) == (router, "openai/complex-tier", "COMPLEX", "jev_classifier"), dict(response.headers) + assert "x-litellm-classifier-cost" not in response.headers, dict(response.headers) def _assert_classifier_call( @@ -165,14 +190,14 @@ def _assert_classifier_call( assert call["path"] == path, call assert call["authorization"] == authorization, call body: Final = object_value(call["body"]) - assert body["model"] == _ENDPOINT, body + assert "model" not in body, body assert prompt_marker in string_value(body["state"]), body assert sorted(object_value(object_value(body["questions"])["tier"])) == ["criteria", "instructions", "type"], body def _assert_refused_before_any_upstream_call(gateway: Gateway, model: str, message: str) -> None: with gateway.scenario() as scenario: - handle: Final = _register(scenario, _ANSWER_BODY) + handle: Final = _register(scenario, _AI_DECIDE_BODY) deployment: Final = scenario.model(model=model, api_base=handle.api_base(), api_key=_API_KEY) response: Final = _decide(gateway, deployment) assert response.status_code == 400, response.text @@ -186,12 +211,10 @@ def _assert_refused_before_any_upstream_call(gateway: Gateway, model: str, messa def test_a_deployment_naming_a_non_bare_endpoint_is_refused_before_any_upstream_call( gateway: Gateway, name: str ) -> None: - _assert_refused_before_any_upstream_call( - gateway, f"databricks/{name}", f"Databricks serving endpoint name {name!r} {_BARE_NAME_RULE}" - ) + _assert_refused_before_any_upstream_call(gateway, f"databricks/{name}", _MODEL_RULE) -def test_a_deployment_naming_no_endpoint_is_refused_before_any_upstream_call(gateway: Gateway) -> None: +def test_a_deployment_naming_no_model_is_refused_before_any_upstream_call(gateway: Gateway) -> None: _assert_refused_before_any_upstream_call(gateway, "databricks/", "A model name is required for the Decisions API") @@ -256,6 +279,81 @@ def test_an_openai_format_request_at_v1_decisions_reaches_the_serving_endpoint_a ) +def test_an_openai_format_request_at_v1_decisions_reaches_the_ai_decide_route_without_a_model_field( + gateway: Gateway, +) -> None: + with gateway.scenario() as scenario: + handle: Final = _register(scenario, _AI_DECIDE_BODY) + deployment: Final = scenario.model(model=f"databricks/{_MODEL}", api_base=handle.api_base(), api_key=_API_KEY) + response: Final = gateway.request( + "POST", + "/v1/decisions", + { + "model": deployment, + "input": _OPENAI_INPUT, + "questions": [_OPENAI_TIER_QUESTION], + }, + ) + assert response.status_code == 200, response.text + assert response.json() == { + "model": f"databricks/{_MODEL}", + "answers": [ + { + "type": "choice", + "name": "tier", + "choice": "COMPLEX", + "probabilities": [ + {"value": "SIMPLE", "probability": 0.03}, + {"value": "MEDIUM", "probability": 0.06}, + {"value": "COMPLEX", "probability": 0.91}, + ], + "confidence": 0.91, + } + ], + "usage": { + "input_tokens": 0, + "input_tokens_details": {"cached_tokens": 0, "cache_write_tokens": 0}, + "output_tokens": 0, + "output_tokens_details": {"reasoning_tokens": 0}, + "total_tokens": 0, + }, + }, response.text + assert response.headers["x-litellm-model-name"] == f"databricks/{_MODEL}", dict(response.headers) + (call,) = _upstream_calls(gateway, handle) + assert call["path"] == f"/{handle.scenario_id}{_ROUTE}", call + assert call["authorization"] == f"Bearer {_API_KEY}", call + assert call["body"] == { + "state": _OPENAI_INPUT, + "questions": {"tier": {"type": "choice", "instructions": "Which tier?", "criteria": _TIER_CRITERIA}}, + }, call + row: Final = _spend_row(response.headers["x-litellm-call-id"]) + assert ( + row["status"], + row["call_type"], + row["custom_llm_provider"], + row["model_group"], + row["api_base"], + _number(row["spend"]), + ) == ("success", "adecisions", "databricks", deployment, f"{handle.api_base()}{_ROUTE}", 0.0), row + + +def test_a_noul_probability_reaches_the_caller_as_the_system_one_noul_answer(gateway: Gateway) -> None: + with gateway.scenario() as scenario: + handle: Final = _register( + scenario, {"response": {"answers": {"defect": {"type": "noul", "probability": 0.8}}}, "metadata": _METADATA} + ) + deployment: Final = scenario.model(model=f"databricks/{_MODEL}", api_base=handle.api_base(), api_key=_API_KEY) + response: Final = gateway.request( + "POST", + "/v1/systemone", + {"model": deployment, "state": _STATE, "questions": {"defect": {"type": "noul", "instructions": "Bug?"}}}, + ) + assert response.status_code == 200, response.text + assert response.json()["answers"] == {"defect": {"type": "noul", "noul": 0.8}}, response.text + (call,) = _upstream_calls(gateway, handle) + assert call["body"] == {"state": _STATE, "questions": {"defect": {"type": "noul", "instructions": "Bug?"}}} + + def test_a_wildcard_deployment_sends_each_request_to_the_named_endpoint_and_bills_each_once(gateway: Gateway) -> None: with gateway.scenario() as scenario: handle: Final = _register(scenario, _ANSWER_BODY) @@ -289,12 +387,43 @@ def test_a_wildcard_deployment_sends_each_request_to_the_named_endpoint_and_bill ) +def test_a_wildcard_deployment_sends_each_ai_decide_request_to_the_route_and_logs_each_once(gateway: Gateway) -> None: + with gateway.scenario() as scenario: + handle: Final = _register(scenario, _AI_DECIDE_BODY) + prefix: Final = _wildcard_deployment(scenario, api_base=handle.api_base(), api_key=_API_KEY) + model: Final = f"{prefix}/{_MODEL}" + responses: Final = tuple(_decide(gateway, model) for _ in range(2)) + for response in responses: + assert response.status_code == 200, response.text + assert response.json() == _SYSTEM_ONE_BODY, response.text + assert response.headers["x-litellm-model-group"] == model, dict(response.headers) + assert response.headers["x-litellm-model-name"] == f"databricks/{_MODEL}", dict(response.headers) + call_ids: Final = tuple(response.headers["x-litellm-call-id"] for response in responses) + assert len(set(call_ids)) == 2, call_ids + calls: Final = _upstream_calls(gateway, handle) + assert len(calls) == 2, calls + for call in calls: + assert call["path"] == f"/{handle.scenario_id}{_ROUTE}", call + assert call["authorization"] == f"Bearer {_API_KEY}", call + assert call["body"] == {"state": _STATE, "questions": _QUESTIONS}, call + for call_id in call_ids: + row: Final = _spend_row(call_id) + assert ( + row["status"], + row["call_type"], + row["custom_llm_provider"], + row["model_group"], + row["api_base"], + _number(row["spend"]), + ) == ("success", "asystemone", "databricks", model, f"{handle.api_base()}{_ROUTE}", 0.0), row + + @pytest.mark.parametrize("model", (5, ["a"]), ids=("int", "list")) def test_a_non_string_model_is_refused_at_the_gateway_without_an_upstream_call( gateway: Gateway, model: JsonValue ) -> None: with gateway.scenario() as scenario: - handle: Final = _register(scenario, _ANSWER_BODY) + handle: Final = _register(scenario, _AI_DECIDE_BODY) _wildcard_deployment(scenario, api_base=handle.api_base(), api_key=_API_KEY) response: Final = gateway.request( "POST", "/v1/systemone", {"model": model, "state": _STATE, "questions": _QUESTIONS} @@ -328,18 +457,20 @@ def test_a_five_kilobyte_endpoint_name_reaches_the_upstream_and_its_not_found_an ), row -@pytest.mark.parametrize("status,body", _DATABRICKS_ERRORS, ids=("400", "404")) +@pytest.mark.parametrize( + "name,path,status,body", _DATABRICKS_ERRORS, ids=("ai_decide-400", "ai_decide-403", "serving_endpoint-404") +) def test_a_databricks_error_keeps_its_status_and_message_and_bills_nothing( - gateway: Gateway, status: int, body: dict[str, JsonValue] + gateway: Gateway, name: str, path: str, status: int, body: dict[str, JsonValue] ) -> None: with gateway.scenario() as scenario: handle: Final = _register(scenario, body, status=status) - model: Final = scenario.model(model=f"databricks/{_ENDPOINT}", api_base=handle.api_base(), api_key=_API_KEY) + model: Final = scenario.model(model=f"databricks/{name}", api_base=handle.api_base(), api_key=_API_KEY) response: Final = _decide(gateway, model) assert response.status_code == status, response.text assert string_value(body["message"]) in response.text, response.text (call,) = _upstream_calls(gateway, handle) - assert call["path"] == f"/{handle.scenario_id}/{_ENDPOINT}/invocations", call + assert call["path"] == f"/{handle.scenario_id}{path}", call row: Final = _spend_row(response.headers["x-litellm-call-id"]) assert (row["status"], row["call_type"], row["model_group"], _number(row["spend"])) == ( "failure", @@ -353,8 +484,8 @@ def test_a_deployment_without_a_key_and_no_environment_key_is_refused_before_any gateway: Gateway, ) -> None: with gateway.scenario() as scenario: - handle: Final = _register(scenario, _ANSWER_BODY) - model: Final = scenario.model(model=f"databricks/{_ENDPOINT}", api_base=handle.api_base(), api_key=None) + handle: Final = _register(scenario, _AI_DECIDE_BODY) + model: Final = scenario.model(model=f"databricks/{_MODEL}", api_base=handle.api_base(), api_key=None) response: Final = _decide(gateway, model) assert response.status_code == 401, response.text assert _MISSING_KEY in response.text, response.text @@ -366,9 +497,9 @@ def test_the_environment_key_and_base_serve_a_deployment_a_classifier_and_a_team gateway: Gateway, tmp_path: Path ) -> None: with gateway.scenario() as rig_scenario: - handle: Final = _register(rig_scenario, _ANSWER_BODY) - environment_base: Final = f"{handle.api_base()}/serving-endpoints" - environment_path: Final = f"/{handle.scenario_id}/serving-endpoints/{_ENDPOINT}/invocations" + handle: Final = _register(rig_scenario, _AI_DECIDE_BODY) + environment_base: Final = f"{handle.api_base()}/environment-workspace/serving-endpoints" + environment_path: Final = f"/{handle.scenario_id}/environment-workspace{_ROUTE}" environment: Final = { "DATABRICKS_API_KEY": _ENV_KEY, "DATABRICKS_TOKEN": _ENV_TOKEN, @@ -376,14 +507,14 @@ def test_the_environment_key_and_base_serve_a_deployment_a_classifier_and_a_team } with owned_proxy_process(gateway, tmp_path, environment) as owned, owned.gateway.scenario() as scenario: candidate: Final = owned.gateway - with_base: Final = scenario.model(model=f"databricks/{_ENDPOINT}", api_base=handle.api_base(), api_key=None) - bare: Final = scenario.model(model=f"databricks/{_ENDPOINT}", api_base=None, api_key=None) + with_base: Final = scenario.model(model=f"databricks/{_MODEL}", api_base=handle.api_base(), api_key=None) + bare: Final = scenario.model(model=f"databricks/{_MODEL}", api_base=None, api_key=None) decided: Final = tuple(_decide(candidate, model) for model in (with_base, bare)) for response in decided: assert response.status_code == 200, response.text - assert response.json() == _ANSWER_BODY, response.text + assert response.json() == _SYSTEM_ONE_BODY, response.text assert [(call["path"], call["authorization"]) for call in _upstream_calls(candidate, handle)] == [ - (f"/{handle.scenario_id}/{_ENDPOINT}/invocations", f"Bearer {_ENV_KEY}"), + (f"/{handle.scenario_id}{_ROUTE}", f"Bearer {_ENV_KEY}"), (environment_path, f"Bearer {_ENV_KEY}"), ] for response, model in zip(decided, (with_base, bare)): @@ -391,7 +522,7 @@ def test_the_environment_key_and_base_serve_a_deployment_a_classifier_and_a_team assert (row["status"], row["model_group"], _number(row["spend"])) == ("success", model, 0.0), row assert ( _spend_row(decided[1].headers["x-litellm-call-id"])["api_base"] - == f"{environment_base}/{_ENDPOINT}/invocations" + == f"{handle.api_base()}/environment-workspace{_ROUTE}" ) router: Final = _router(scenario, _environment_classifier()) @@ -435,7 +566,7 @@ def test_the_environment_key_and_base_serve_a_deployment_a_classifier_and_a_team "opensource_classifier_config" ] ) - assert (stored_classifier["provider"], stored_classifier["model"]) == ("databricks", _ENDPOINT), entry + assert (stored_classifier["provider"], stored_classifier["model"]) == ("databricks", _MODEL), entry assert "api_key" not in stored_classifier and "api_base" not in stored_classifier, entry member_chat: Final = _chat(candidate, team_router, key=member_key) _assert_classified_by_the_complex_tier(member_chat, team_router) @@ -453,21 +584,21 @@ def test_the_environment_token_serves_a_deployment_and_a_classifier_when_no_envi gateway: Gateway, tmp_path: Path ) -> None: with gateway.scenario() as rig_scenario: - handle: Final = _register(rig_scenario, _ANSWER_BODY) - environment_path: Final = f"/{handle.scenario_id}/serving-endpoints/{_ENDPOINT}/invocations" + handle: Final = _register(rig_scenario, _AI_DECIDE_BODY) + environment_path: Final = f"/{handle.scenario_id}/environment-workspace{_ROUTE}" environment: Final = { "DATABRICKS_TOKEN": _ENV_TOKEN, - "DATABRICKS_API_BASE": f"{handle.api_base()}/serving-endpoints", + "DATABRICKS_API_BASE": f"{handle.api_base()}/environment-workspace/serving-endpoints", } with ( owned_proxy_process(gateway, tmp_path, environment, remove_environment=("DATABRICKS_API_KEY",)) as owned, owned.gateway.scenario() as scenario, ): candidate: Final = owned.gateway - bare: Final = scenario.model(model=f"databricks/{_ENDPOINT}", api_base=None, api_key=None) + bare: Final = scenario.model(model=f"databricks/{_MODEL}", api_base=None, api_key=None) decided: Final = _decide(candidate, bare) assert decided.status_code == 200, decided.text - assert decided.json() == _ANSWER_BODY, decided.text + assert decided.json() == _SYSTEM_ONE_BODY, decided.text (decision_call,) = _upstream_calls(candidate, handle) assert (decision_call["path"], decision_call["authorization"]) == (environment_path, f"Bearer {_ENV_TOKEN}") assert _spend_row(decided.headers["x-litellm-call-id"])["status"] == "success" @@ -483,15 +614,16 @@ def test_the_environment_token_serves_a_deployment_and_a_classifier_when_no_envi ) -def test_a_megabyte_endpoint_name_is_refused_in_linear_time_while_liveliness_stays_fast(gateway: Gateway) -> None: +def test_a_megabyte_model_name_is_refused_in_linear_time_while_liveliness_stays_up(gateway: Gateway) -> None: with gateway.scenario() as scenario: - handle: Final = _register(scenario, _ANSWER_BODY) + handle: Final = _register(scenario, _AI_DECIDE_BODY) prefix: Final = _wildcard_deployment(scenario, api_base=handle.api_base(), api_key=_API_KEY) model: Final = f"{prefix}/{_MEGABYTE_NAME}" anonymous: Final = gateway.client.post( "/v1/systemone", json={"model": model, "state": _STATE, "questions": _QUESTIONS} ) assert anonymous.status_code == 401, anonymous.text[:300] + tenth_response, tenth_elapsed = _timed_decide(gateway, f"{prefix}/{_TENTH_MEGABYTE_NAME}") liveliness: Final[list[tuple[int, float]]] = [] stop: Final = threading.Event() @@ -503,16 +635,15 @@ def test_a_megabyte_endpoint_name_is_refused_in_linear_time_while_liveliness_sta poller: Final = threading.Thread(target=poll) poller.start() - started: Final = time.perf_counter() - response: Final = _decide(gateway, model) - elapsed: Final = time.perf_counter() - started + response, elapsed = _timed_decide(gateway, model) stop.set() poller.join() - assert elapsed < 3, elapsed + assert tenth_response.status_code == 400, tenth_response.text[:300] + assert elapsed < 25 * tenth_elapsed, (elapsed, tenth_elapsed) assert liveliness and all(status == 200 for status, _ in liveliness), liveliness - assert max(seconds for _, seconds in liveliness) < 1, liveliness + assert max(seconds for _, seconds in liveliness) < max(1.0, elapsed / 2), (elapsed, liveliness) assert response.status_code == 400, response.text[:300] - assert _BARE_NAME_RULE in response.text, response.text[-400:] + assert _MODEL_RULE in response.text, response.text[:400] assert _upstream_calls(gateway, handle) == [] row: Final = _spend_row(response.headers["x-litellm-call-id"]) assert (row["status"], row["call_type"], _number(row["spend"])) == ("failure", "asystemone", 0.0), row diff --git a/tests/integration/providers/test_decisions_openai_format_wire.py b/tests/integration/providers/test_decisions_openai_format_wire.py index 1104a6151db..64387b293ce 100644 --- a/tests/integration/providers/test_decisions_openai_format_wire.py +++ b/tests/integration/providers/test_decisions_openai_format_wire.py @@ -88,6 +88,17 @@ _USAGE: Final[dict[str, JsonValue]] = { "output_tokens_details": {"reasoning_tokens": 0}, "total_tokens": _INPUT_TOKENS + _OUTPUT_TOKENS, } +_NO_USAGE: Final[dict[str, JsonValue]] = { + "input_tokens": 0, + "input_tokens_details": {"cached_tokens": 0, "cache_write_tokens": 0}, + "output_tokens": 0, + "output_tokens_details": {"reasoning_tokens": 0}, + "total_tokens": 0, +} +_AI_DECIDE_ANSWERS: Final[dict[str, JsonValue]] = { + **_SYSTEM_ONE_ANSWERS, + "defect": {"type": "noul", "probability": 0.93}, +} _SDK_QUESTIONS: Final = TypeAdapter(list[dict[str, object]]).validate_python(_QUESTIONS) _SPEND_QUERY: Final = ( "SELECT spend, status, call_type, model_group, custom_llm_provider, api_base, prompt_tokens, completion_tokens, " @@ -105,11 +116,14 @@ class _Provider: wraps_result: bool cost_map_key: str | None speaks_openai: bool = False + ai_decide: bool = False provider_reported_cost: float | None = None def upstream_body(self) -> dict[str, JsonValue]: if self.speaks_openai: return {"model": self.body_model, "input": _INPUT, "questions": _QUESTIONS} + if self.ai_decide: + return {"state": _INPUT, "questions": _SYSTEM_ONE_QUESTIONS} return {"model": self.body_model, "state": _INPUT, "questions": _SYSTEM_ONE_QUESTIONS} def _with_reported_cost(self, usage: dict[str, JsonValue]) -> dict[str, JsonValue]: @@ -120,6 +134,8 @@ class _Provider: def upstream_reply(self) -> dict[str, JsonValue]: if self.speaks_openai: return {"model": self.body_model, "answers": _ANSWERS, "usage": self._with_reported_cost(_USAGE)} + if self.ai_decide: + return {"response": {"answers": _AI_DECIDE_ANSWERS}, "metadata": {"version": "1.0"}} answer: Final[dict[str, JsonValue]] = { "model": self.body_model, "answers": _SYSTEM_ONE_ANSWERS, @@ -128,8 +144,13 @@ class _Provider: return {"result": answer, "success": True} if self.wraps_result else answer def litellm_response(self) -> dict[str, JsonValue]: + if self.ai_decide: + return {"model": self.model, "answers": _ANSWERS, "usage": _NO_USAGE} return {"model": self.body_model, "answers": _ANSWERS, "usage": _USAGE} + def billed_tokens(self) -> tuple[int, int]: + return (0, 0) if self.ai_decide else (_INPUT_TOKENS, _OUTPUT_TOKENS) + _PROVIDERS: Final = ( _Provider( @@ -165,6 +186,16 @@ _PROVIDERS: Final = ( "cloudflare/@cf/cloudflare/clef", ), _Provider("hosted_vllm", "hosted_vllm/Qwen/Qwen3-0.6B", "/v1/systemone", "Qwen/Qwen3-0.6B", None, False, None), + _Provider( + "databricks", + "databricks/ai_decide", + "/api/2.0/ai-functions/ai-decide", + "ai_decide", + _API_KEY, + False, + None, + ai_decide=True, + ), _Provider( "databricks", "databricks/databricks-openjev-qwen35-4b", @@ -234,7 +265,7 @@ def _response_cost(response: httpx.Response) -> float: def _provider_id(provider: _Provider) -> str: - return provider.name + return provider.model def _number(value: JsonValue) -> float: @@ -358,8 +389,7 @@ def test_each_provider_gets_its_own_path_key_and_body_and_is_billed(gateway: Gat provider.name, model, f"{handle.api_base()}{provider.path}", - _INPUT_TOKENS, - _OUTPUT_TOKENS, + *provider.billed_tokens(), ) assert math.isclose(_number(row["spend"]), expected_spend, rel_tol=1e-9), row diff --git a/tests/integration/routing/test_complexity_router_databricks_classifier.py b/tests/integration/routing/test_complexity_router_databricks_classifier.py index 25611b4ec08..419b786bfc1 100644 --- a/tests/integration/routing/test_complexity_router_databricks_classifier.py +++ b/tests/integration/routing/test_complexity_router_databricks_classifier.py @@ -31,9 +31,10 @@ from pydantic import JsonValue, TypeAdapter from litellm.proxy.common_utils.encrypt_decrypt_utils import decrypt_if_encrypted_with from litellm.router_strategy.complexity_router.config import DEFAULT_JEV_INSTRUCTIONS -_ENDPOINT: Final = "databricks-openjev-qwen35-4b" +_MODEL: Final = "ai_decide" +_ROUTE: Final = "/api/2.0/ai-functions/ai-decide" _API_KEY: Final = "synthetic-databricks-key" -_CLASSIFIER_MODEL: Final = f"databricks/{_ENDPOINT}" +_CLASSIFIER_MODEL: Final = f"databricks/{_MODEL}" _TYPESAFE_MODEL: Final = "jev-1.13.0" _TYPESAFE_KEY: Final = "synthetic-typesafe-key" _JSON_OBJECT: Final = TypeAdapter(dict[str, JsonValue]) @@ -48,10 +49,12 @@ _TIER_CRITERIA: Final[dict[str, JsonValue]] = { } _PROBABILITIES: Final[dict[str, JsonValue]] = {"SIMPLE": 0.03, "MEDIUM": 0.06, "COMPLEX": 0.91} _USAGE: Final[dict[str, JsonValue]] = {"input_tokens": 12, "output_tokens": 1} -_ANSWER: Final[dict[str, JsonValue]] = { - "answers": {"tier": {"type": "choice", "choice": "COMPLEX", "confidence": 0.91, "probabilities": _PROBABILITIES}}, - "usage": _USAGE, +_ANSWERS: Final[dict[str, JsonValue]] = { + "tier": {"type": "choice", "choice": "COMPLEX", "confidence": 0.91, "probabilities": _PROBABILITIES} } +_METADATA: Final[dict[str, JsonValue]] = {"version": "1.0"} +_ANSWER: Final[dict[str, JsonValue]] = {"response": {"answers": _ANSWERS}, "metadata": _METADATA} +_TYPESAFE_ANSWER: Final[dict[str, JsonValue]] = {"model": _TYPESAFE_MODEL, "answers": _ANSWERS, "usage": _USAGE} _ROWS_QUERY: Final = ( "SELECT request_id, status, model, call_type, cache_hit, custom_llm_provider, spend, " "metadata->>'internal_call_origin' AS origin " @@ -155,7 +158,7 @@ def _wires(classifier: Callable[[Request], Reply] | None = None) -> Iterator[tup def _classifier_config(judge_url: str, **overrides: JsonValue) -> dict[str, JsonValue]: return { "provider": "databricks", - "model": _ENDPOINT, + "model": _MODEL, "api_base": judge_url, "api_key": _API_KEY, "timeout_ms": 20000, @@ -209,25 +212,25 @@ def _state(prompt: str, system: str | None) -> str: def _assert_classifier_call( - call: Request, *, prompt: str, system: str | None = None, endpoint: str = _ENDPOINT, key: str = _API_KEY + call: Request, *, prompt: str, system: str | None = None, key: str = _API_KEY, endpoint: str | None = None ) -> None: - assert call.target == f"/{endpoint}/invocations", call.target + assert call.target == (_ROUTE if endpoint is None else f"/{endpoint}/invocations"), call.target assert call.headers.get("authorization") == f"Bearer {key}", call.headers assert json.loads(call.body) == { "state": _state(prompt, system), - "model": endpoint, + **({} if endpoint is None else {"model": endpoint}), "questions": {"tier": {"type": "choice", "instructions": DEFAULT_JEV_INSTRUCTIONS, "criteria": _TIER_CRITERIA}}, }, call.body -def _assert_classified(headers: Mapping[str, str], *, router: str, cost: str | None = "0.0") -> None: +def _assert_classified(headers: Mapping[str, str], *, router: str) -> None: assert ( headers.get("x-litellm-model-group"), headers.get("x-litellm-model-name"), headers.get("x-litellm-complexity-router-tier"), headers.get("x-litellm-complexity-router-cause"), headers.get("x-litellm-classifier-cost"), - ) == (router, "openai/complex-tier", "COMPLEX", "jev_classifier", cost), dict(headers) + ) == (router, "openai/complex-tier", "COMPLEX", "jev_classifier", None), dict(headers) def _assert_fell_back(headers: Mapping[str, str], *, router: str) -> None: @@ -622,10 +625,19 @@ def test_messages_stream_over_httpx_is_classified_and_logged_under_the_message_s _assert_logged_once(_rows(routed.name, want=2), route="/v1/messages", request_id=identity) -def test_a_non_catalog_endpoint_name_is_classified_without_a_cost_header_and_logged_at_zero(gateway: Gateway) -> None: +def test_a_serving_endpoint_name_is_classified_through_its_invocations_route_and_logged_at_zero( + gateway: Gateway, +) -> None: prompt: Final = _prompt() endpoint: Final = "my-jev-endpoint" - with _wires() as (judge, simple, complex_), gateway.scenario() as scenario: + with ( + _wires(_answering({"model": "/mosaicml/local_model", "answers": _ANSWERS, "usage": _USAGE})) as ( + judge, + simple, + complex_, + ), + gateway.scenario() as scenario, + ): routed: Final = _deploy(scenario, judge, simple, complex_, _classifier_config(judge.url, model=endpoint)) response: Final = gateway.request( "POST", "/v1/chat/completions", _body("/v1/chat/completions", routed.name, prompt, stream=False) @@ -633,7 +645,7 @@ def test_a_non_catalog_endpoint_name_is_classified_without_a_cost_header_and_log assert response.status_code == 200, response.text identity, text = _identity_and_text("/v1/chat/completions", _JSON_OBJECT.validate_json(response.content)) assert text == _COMPLEX_TEXT, response.text - _assert_classified(response.headers, router=routed.name, cost=None) + _assert_classified(response.headers, router=routed.name) (classifier_call,) = _posts(judge) _assert_classifier_call(classifier_call, prompt=prompt, endpoint=endpoint) assert len(_posts(complex_)) == 1 and _posts(simple) == () @@ -686,7 +698,7 @@ _FAILURES: Final = ( _Failure("401", _failing(401), {}, "classifier_error", "MaskedHTTPStatusError", 0), _Failure( "malformed", - _answering({"answers": {"tier": {"type": "choice", "choice": "COMPLEX", "confidence": 0.91}}, "usage": _USAGE}), + _answering({"response": {"answers": {"tier": {"type": "choice", "choice": "COMPLEX", "confidence": 0.91}}}}), {}, "invalid_response", "ValidationError", @@ -696,15 +708,17 @@ _FAILURES: Final = ( "unknown-tier", _answering( { - "answers": { - "tier": { - "type": "choice", - "choice": "GALAXY", - "confidence": 0.91, - "probabilities": {"GALAXY": 0.91}, + "response": { + "answers": { + "tier": { + "type": "choice", + "choice": "GALAXY", + "confidence": 0.91, + "probabilities": {"GALAXY": 0.91}, + } } }, - "usage": _USAGE, + "metadata": _METADATA, } ), {}, @@ -793,7 +807,7 @@ def test_the_default_jev_provider_still_posts_to_systemone_and_bills_from_the_co prompt: Final = _prompt() expected_cost: Final = _typesafe_cost() with ( - _wires(_answering({**_ANSWER, "model": _TYPESAFE_MODEL})) as (judge, simple, complex_), + _wires(_answering(_TYPESAFE_ANSWER)) as (judge, simple, complex_), gateway.scenario() as scenario, ): classifier: Final[dict[str, JsonValue]] = { @@ -848,7 +862,7 @@ async def test_a_classifier_outage_mid_traffic_falls_back_and_recovers_with_ever before_prompts: Final = tuple(_prompt() for _ in calls) before: Final = await _burst(base_url, gateway.key, routed.name, calls, before_prompts) assert [item.cause for item in before] == ["jev_classifier"] * 12, before - assert [item.classifier_cost for item in before] == ["0.0"] * 12, before + assert [item.classifier_cost for item in before] == [None] * 12, before assert [item.text for item in before] == [_COMPLEX_TEXT] * 12, before assert sorted(json.loads(call.body)["state"] for call in _posts(judge)) == sorted( _state(prompt, None) for prompt in before_prompts diff --git a/tests/unit/decisions/test_main.py b/tests/unit/decisions/test_main.py index 81f5b60a036..e1b2028fa30 100644 --- a/tests/unit/decisions/test_main.py +++ b/tests/unit/decisions/test_main.py @@ -78,7 +78,7 @@ _STRANDS_RESPONSE: Final[Mapping[str, object]] = { "latency_ms": 3722.17, } _DATABRICKS_ENDPOINT: Final = "databricks-openjev-qwen35-4b" -_DATABRICKS_RESPONSE: Final[Mapping[str, object]] = { +_DATABRICKS_ENDPOINT_RESPONSE: Final[Mapping[str, object]] = { "model": "/mosaicml/local_model", "answers": { "is_defect": {"type": "noul", "noul": 0.9}, @@ -91,6 +91,21 @@ _DATABRICKS_RESPONSE: Final[Mapping[str, object]] = { }, "usage": {"input_tokens": 214, "output_tokens": 0}, } +_DATABRICKS_ROUTE: Final = "https://workspace.example/api/2.0/ai-functions/ai-decide" +_DATABRICKS_RESPONSE: Final[Mapping[str, object]] = { + "response": { + "answers": { + "is_defect": {"type": "noul", "probability": 0.9}, + "sentiment": { + "type": "choice", + "choice": "positive", + "confidence": 0.8, + "probabilities": {"positive": 0.8, "negative": 0.2}, + }, + } + }, + "metadata": {"version": "1.0"}, +} _PROVIDERS: Final[tuple[tuple[str, str, str, str], ...]] = ( ( "perplexity", @@ -1093,7 +1108,7 @@ async def test_databricks_requires_api_base_before_http( with pytest.raises(litellm.BadRequestError, match="DATABRICKS_API_BASE"): await litellm.asystemone( - model=f"databricks/{_DATABRICKS_ENDPOINT}", + model="databricks/ai_decide", state="review", questions={"is_defect": {"type": "noul", "instructions": "Is this a defect?"}}, ) @@ -1113,7 +1128,7 @@ async def test_databricks_rejects_a_serving_endpoint_name_that_would_rewrite_the monkeypatch.setenv("DATABRICKS_API_BASE", "https://workspace.example/serving-endpoints") monkeypatch.setenv("DATABRICKS_API_KEY", "dapi-key") - with pytest.raises(litellm.BadRequestError, match="bare endpoint name"): + with pytest.raises(litellm.BadRequestError, match="bare serving endpoint name"): await litellm.asystemone( model=f"databricks/{endpoint_name}", state="review", @@ -1128,13 +1143,13 @@ async def test_databricks_requires_a_key_before_http( monkeypatch: pytest.MonkeyPatch, respx_mock: respx.MockRouter, ) -> None: - monkeypatch.setenv("DATABRICKS_API_BASE", "https://workspace.example/serving-endpoints") + monkeypatch.setenv("DATABRICKS_API_BASE", "https://workspace.example") monkeypatch.delenv("DATABRICKS_API_KEY", raising=False) monkeypatch.delenv("DATABRICKS_TOKEN", raising=False) with pytest.raises(litellm.AuthenticationError, match="Missing API key for Decisions provider 'databricks'"): await litellm.asystemone( - model=f"databricks/{_DATABRICKS_ENDPOINT}", + model="databricks/ai_decide", state="review", questions={"is_defect": {"type": "noul", "instructions": "Is this a defect?"}}, ) @@ -1156,7 +1171,7 @@ async def test_databricks_posts_the_jev_body_to_the_serving_endpoint_invocations monkeypatch.setenv("DATABRICKS_TOKEN", "dapi-token") route: Final = respx_mock.post( f"https://workspace.example/serving-endpoints/{_DATABRICKS_ENDPOINT}/invocations" - ).respond(json=_DATABRICKS_RESPONSE) + ).respond(json=_DATABRICKS_ENDPOINT_RESPONSE) questions: Final = { "is_defect": {"type": "noul", "instructions": "Is this a defect?"}, "sentiment": {"type": "choice", "criteria": {"positive": None, "negative": "unhappy"}}, @@ -1183,28 +1198,68 @@ async def test_databricks_posts_the_jev_body_to_the_serving_endpoint_invocations assert response.hidden_params["model"] == f"databricks/{_DATABRICKS_ENDPOINT}" +@pytest.mark.asyncio +@pytest.mark.parametrize( + "api_base", + [ + "https://workspace.example", + "https://workspace.example/", + "https://workspace.example/serving-endpoints", + "https://workspace.example/serving-endpoints/", + ], +) +async def test_databricks_posts_the_systemone_body_without_model_to_the_ai_decide_route( + monkeypatch: pytest.MonkeyPatch, + respx_mock: respx.MockRouter, + api_base: str, +) -> None: + monkeypatch.setenv("DATABRICKS_API_BASE", api_base) + monkeypatch.delenv("DATABRICKS_API_KEY", raising=False) + monkeypatch.setenv("DATABRICKS_TOKEN", "dapi-token") + route: Final = respx_mock.post(_DATABRICKS_ROUTE).respond(json=_DATABRICKS_RESPONSE) + questions: Final = { + "is_defect": {"type": "noul", "instructions": "Is this a defect?"}, + "sentiment": {"type": "choice", "criteria": {"positive": None, "negative": "unhappy"}}, + } + + response: Final = await litellm.adecisions( + model="databricks/ai_decide", + state={"ticket": "export hangs"}, + questions=questions, + ) + + assert route.called + sent: Final = respx_mock.calls[0].request + assert sent.headers["authorization"] == "Bearer dapi-token" + assert json.loads(sent.content) == {"state": {"ticket": "export hangs"}, "questions": questions} + assert response.model is None + assert response.answers["is_defect"] == NoulAnswer(type="noul", noul=0.9) + assert isinstance(response.answers["sentiment"], ChoiceAnswer) + assert response.model_extra == {"metadata": {"version": "1.0"}} + assert response.hidden_params["custom_llm_provider"] == "databricks" + assert response.hidden_params["model"] == "databricks/ai_decide" + + @pytest.mark.asyncio async def test_databricks_router_deployment_sends_its_own_key_to_its_own_base( monkeypatch: pytest.MonkeyPatch, respx_mock: respx.MockRouter, ) -> None: - monkeypatch.setenv("DATABRICKS_API_BASE", "https://other-workspace.example/serving-endpoints") + monkeypatch.setenv("DATABRICKS_API_BASE", "https://other-workspace.example") monkeypatch.setenv("DATABRICKS_API_KEY", "env-key-stays-home") router: Final = litellm.Router( model_list=[ { "model_name": "decider", "litellm_params": { - "model": f"databricks/{_DATABRICKS_ENDPOINT}", - "api_base": "https://workspace.example/serving-endpoints", + "model": "databricks/ai_decide", + "api_base": "https://workspace.example", "api_key": "deployment-key", }, } ] ) - route: Final = respx_mock.post( - f"https://workspace.example/serving-endpoints/{_DATABRICKS_ENDPOINT}/invocations" - ).respond(json=_DATABRICKS_RESPONSE) + route: Final = respx_mock.post(_DATABRICKS_ROUTE).respond(json=_DATABRICKS_RESPONSE) response: Final = await router.asystemone( model="decider", diff --git a/tests/unit/llms/databricks/decisions/test_transformation.py b/tests/unit/llms/databricks/decisions/test_transformation.py index e6de1d1a635..f6ea3444db1 100644 --- a/tests/unit/llms/databricks/decisions/test_transformation.py +++ b/tests/unit/llms/databricks/decisions/test_transformation.py @@ -3,27 +3,84 @@ from typing import Final import pytest from litellm.llms.databricks.decisions.transformation import DatabricksDecisionsConfig +from litellm.types.decisions import ( + DecisionsIRPredicateAnswer, + DecisionsIRPredicateQuestion, + DecisionsIRRequest, + DecisionsIRState, +) -_BASE: Final = "https://workspace.example/serving-endpoints" +_HOST: Final = "https://workspace.example" +_BASE: Final = f"{_HOST}/serving-endpoints" +_AI_DECIDE_ROUTE: Final = f"{_HOST}/api/2.0/ai-functions/ai-decide" _CONFIG: Final = DatabricksDecisionsConfig() +_REQUEST: Final = DecisionsIRRequest( + input=DecisionsIRState(state="export hangs"), + questions=(DecisionsIRPredicateQuestion(name="is_defect", instructions="Is this a defect?", criteria=None),), +) +_QUESTIONS: Final = {"is_defect": {"type": "noul", "instructions": "Is this a defect?"}} @pytest.mark.parametrize("api_base", [_BASE, f"{_BASE}/"]) -def test_the_complete_url_is_the_serving_endpoint_invocations_route(api_base: str) -> None: +def test_a_serving_endpoint_name_is_called_on_its_invocations_route(api_base: str) -> None: assert _CONFIG.get_complete_url(api_base, "openjev") == f"{_BASE}/openjev/invocations" +@pytest.mark.parametrize("api_base", [_HOST, f"{_HOST}/", _BASE, f"{_BASE}/"]) +def test_ai_decide_is_called_on_the_workspace_ai_function_route(api_base: str) -> None: + assert _CONFIG.get_complete_url(api_base, "ai_decide") == _AI_DECIDE_ROUTE + + @pytest.mark.parametrize( "name", ["", "serving-endpoints/openjev", "openjev?x=1", "openjev#frag", ".", "..", ".openjev", "open jev", "openjev%2Fx"], ) def test_a_name_that_would_rewrite_the_url_is_rejected_before_any_request(name: str) -> None: - with pytest.raises(ValueError, match="bare endpoint name"): + with pytest.raises(ValueError, match="bare serving endpoint name"): _ = _CONFIG.get_complete_url(_BASE, name) - with pytest.raises(ValueError, match="bare endpoint name"): + with pytest.raises(ValueError, match="bare serving endpoint name"): _ = _CONFIG.canonical_model(name) +def test_the_rejection_does_not_echo_the_rejected_name() -> None: + name: Final = "openjev?token=" + "x" * 4096 + with pytest.raises(ValueError, match="bare serving endpoint name") as raised: + _ = _CONFIG.canonical_model(name) + assert "openjev?token=" not in str(raised.value) + + +def test_a_serving_endpoint_request_keeps_the_model_the_endpoint_expects() -> None: + body: Final = _CONFIG.transform_decisions_request("openjev", _REQUEST, "databricks") + assert body == {"model": "openjev", "state": "export hangs", "questions": _QUESTIONS} + + +def test_the_ai_decide_request_omits_model_because_the_route_rejects_unexpected_parameters() -> None: + body: Final = _CONFIG.transform_decisions_request("ai_decide", _REQUEST, "databricks") + assert body == {"state": "export hangs", "questions": _QUESTIONS} + + +def test_the_ai_decide_envelope_is_unwrapped_and_noul_probability_becomes_the_noul_answer() -> None: + payload: Final = { + "response": {"answers": {"is_defect": {"type": "noul", "probability": 0.9}}}, + "metadata": {"version": "1.0"}, + } + parsed: Final = _CONFIG.parse_response(payload, _REQUEST) + assert parsed.answers == (DecisionsIRPredicateAnswer(probability=0.9),) + assert parsed.model is None + assert dict(parsed.extra) == {"metadata": {"version": "1.0"}} + + +def test_a_serving_endpoint_answer_is_parsed_as_sent() -> None: + payload: Final = { + "model": "openjev", + "answers": {"is_defect": {"type": "noul", "noul": 0.25}}, + "usage": {"input_tokens": 12, "output_tokens": 0}, + } + parsed: Final = _CONFIG.parse_response(payload, _REQUEST) + assert parsed.answers == (DecisionsIRPredicateAnswer(probability=0.25),) + assert parsed.model == "openjev" + + def test_connection_prefers_the_configured_values_over_the_environment(monkeypatch: pytest.MonkeyPatch) -> None: monkeypatch.setenv("DATABRICKS_API_BASE", "https://env.example/serving-endpoints") monkeypatch.setenv("DATABRICKS_API_KEY", "dapi-env") @@ -54,3 +111,10 @@ def test_classifier_response_accounts_the_serving_endpoint_instead_of_the_contai assert normalized["model"] == "openjev" assert normalized["usage"] == body["usage"] assert body["model"] == "/mosaicml/local_model" + + +def test_classifier_response_unwraps_the_ai_decide_envelope_and_accounts_the_configured_model() -> None: + choice: Final = {"type": "choice", "choice": "simple", "confidence": 0.9, "probabilities": {"simple": 0.9}} + body: Final = {"response": {"answers": {"tier": choice}}, "metadata": {"version": "1.0"}} + normalized: Final = _CONFIG.classifier_response(body, "ai_decide") + assert dict(normalized) == {"answers": {"tier": choice}, "metadata": {"version": "1.0"}, "model": "ai_decide"} diff --git a/tests/unit/proxy/decisions_endpoints/test_endpoints.py b/tests/unit/proxy/decisions_endpoints/test_endpoints.py index 87b2b64ad8d..81d7f8eb7ac 100644 --- a/tests/unit/proxy/decisions_endpoints/test_endpoints.py +++ b/tests/unit/proxy/decisions_endpoints/test_endpoints.py @@ -304,6 +304,53 @@ def test_proxy_systemone_dispatches_databricks_serving_endpoint( assert upstream.calls[0].request.headers["authorization"] == "Bearer dapi-deployment" +def test_proxy_systemone_dispatches_databricks_ai_decide( + client: TestClient, + monkeypatch: pytest.MonkeyPatch, + respx_mock: respx.MockRouter, +) -> None: + monkeypatch.delenv("DATABRICKS_API_BASE", raising=False) + monkeypatch.delenv("DATABRICKS_API_KEY", raising=False) + monkeypatch.delenv("DATABRICKS_TOKEN", raising=False) + router: Final = litellm.Router( + model_list=[ + { + "model_name": "databricks-decider", + "litellm_params": { + "model": "databricks/ai_decide", + "api_base": "https://workspace.example/serving-endpoints", + "api_key": "dapi-deployment", + }, + } + ] + ) + monkeypatch.setattr(litellm.proxy.proxy_server, "llm_router", router) + upstream: Final = respx_mock.post("https://workspace.example/api/2.0/ai-functions/ai-decide").respond( + json={ + "response": {"answers": {"is_defect": {"type": "noul", "probability": 0.9}}}, + "metadata": {"version": "1.0"}, + } + ) + + response: Final = client.post( + "/v1/systemone", + json={ + "model": "databricks-decider", + "state": {"source": "proxy-test"}, + "questions": {"is_defect": {"type": "noul", "instructions": "Is this a defect?"}}, + }, + ) + + assert response.status_code == 200, response.text + assert response.json()["answers"] == _RESPONSE["answers"] + assert upstream.called + assert json.loads(upstream.calls[0].request.content) == { + "state": {"source": "proxy-test"}, + "questions": {"is_defect": {"type": "noul", "instructions": "Is this a defect?"}}, + } + assert upstream.calls[0].request.headers["authorization"] == "Bearer dapi-deployment" + + def test_proxy_decisions_without_model_uses_the_proxy_default_model( client: TestClient, monkeypatch: pytest.MonkeyPatch, diff --git a/tests/unit/proxy/management_helpers/test_auto_router_permissions.py b/tests/unit/proxy/management_helpers/test_auto_router_permissions.py index 16ddf009b0b..e161ea3ecaf 100644 --- a/tests/unit/proxy/management_helpers/test_auto_router_permissions.py +++ b/tests/unit/proxy/management_helpers/test_auto_router_permissions.py @@ -146,8 +146,8 @@ def test_tier_config_is_normalized_and_unknown_router_extras_are_rejected() -> N ({"provider": "laya", "model": "english", "api_key": "sk-member"}, "api_key"), ({"provider": "bespoke", "model": "nimble-latest", "api_base": "https://collector.invalid"}, "api_base"), ({"provider": "bespoke", "model": "nimble-latest", "api_key": "sk-member"}, "api_key"), - ({"provider": "databricks", "model": "my-openjev", "api_base": "https://collector.invalid"}, "opensource_classifier_config"), - ({"provider": "databricks", "model": "my-openjev", "api_key": "dapi-member"}, "api_key"), + ({"provider": "databricks", "model": "ai_decide", "api_base": "https://collector.invalid"}, "opensource_classifier_config"), + ({"provider": "databricks", "model": "ai_decide", "api_key": "dapi-member"}, "api_key"), ], ) @pytest.mark.parametrize("legacy", [False, True]) @@ -168,7 +168,13 @@ def test_members_cannot_move_the_jev_classifier_off_the_proxys_typesafe_account( @pytest.mark.parametrize( ("provider", "model"), - [("typesafe", "jev-preview"), ("laya", "english"), ("bespoke", "nimble-latest"), ("databricks", "my-openjev")], + [ + ("typesafe", "jev-preview"), + ("laya", "english"), + ("bespoke", "nimble-latest"), + ("databricks", "ai_decide"), + ("databricks", "my-openjev"), + ], ) @pytest.mark.parametrize("legacy", [False, True]) def test_members_can_still_tune_the_jev_classifier(provider: str, model: str, legacy: bool) -> None: diff --git a/tests/unit/router_strategy/complexity_router/test_jev_classifier.py b/tests/unit/router_strategy/complexity_router/test_jev_classifier.py index c1c96588004..f7fbac21304 100644 --- a/tests/unit/router_strategy/complexity_router/test_jev_classifier.py +++ b/tests/unit/router_strategy/complexity_router/test_jev_classifier.py @@ -458,6 +458,7 @@ def test_jev_config_requires_classifier_config() -> None: ("laya", "english", "laya"), ("bespoke", "nimble-latest", "bespoke"), ("databricks", "databricks-openjev-qwen35-4b", "databricks"), + ("databricks", "ai_decide", "databricks"), ], ) def test_classifier_aliases_load_and_serialize_one_canonical_config( @@ -540,14 +541,18 @@ async def test_oss_routes_with_its_own_credentials_and_accounts_the_checkpoint( assert recorder.calls[0]["response_cost"] == pytest.approx(0.31) -def test_databricks_requires_the_bare_serving_endpoint_name_as_the_model() -> None: +@pytest.mark.parametrize("model", ["ai_decide", "my-openjev", "databricks-openjev-qwen35-4b"]) +def test_databricks_accepts_ai_decide_or_a_bare_serving_endpoint_name(model: str) -> None: + assert JevClassifierConfig.model_validate({"provider": "databricks", "model": model}).model == model + + +def test_databricks_requires_a_model_that_cannot_rewrite_the_url() -> None: with pytest.raises(ValueError, match="model is required for provider 'databricks'"): JevClassifierConfig.model_validate({"provider": "databricks"}) - with pytest.raises(ValueError, match="bare endpoint name"): + with pytest.raises(ValueError, match="bare serving endpoint name"): JevClassifierConfig.model_validate({"provider": "databricks", "model": "serving-endpoints/my-openjev"}) - with pytest.raises(ValueError, match="bare endpoint name"): + with pytest.raises(ValueError, match="bare serving endpoint name"): JevClassifierConfig.model_validate({"provider": "databricks", "model": ".."}) - assert JevClassifierConfig.model_validate({"provider": "databricks", "model": "my-openjev"}).model == "my-openjev" @pytest.mark.asyncio @@ -608,6 +613,53 @@ async def test_databricks_routes_through_the_serving_endpoint_and_accounts_the_e assert recorder.calls[0]["response_cost"] == pytest.approx(0.31) +@pytest.mark.asyncio +@pytest.mark.parametrize("configured_key", [False, True]) +@pytest.mark.parametrize("legacy", [False, True]) +@pytest.mark.parametrize("api_base", ["https://workspace.test", "https://workspace.test/serving-endpoints/"]) +async def test_databricks_routes_through_ai_decide_without_a_model_field_and_accounts_ai_decide( + monkeypatch: pytest.MonkeyPatch, configured_key: bool, legacy: bool, api_base: str +) -> None: + monkeypatch.setenv("TYPESAFE_API_KEY", "never-send-typesafe-key") + monkeypatch.setenv("DATABRICKS_API_BASE", api_base) + monkeypatch.delenv("DATABRICKS_API_KEY", raising=False) + monkeypatch.setenv("DATABRICKS_TOKEN", "dapi-env-token") + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + recorder: Final = _UsageRecorder("databricks/ai_decide") + monkeypatch.setattr(litellm, "_async_success_callback", [recorder]) + router: Final = ComplexityRouter( + "databricks-route", + litellm.Router(model_list=[]), + { + "classifier_type": "jev" if legacy else "oss_classifier", + "jev_classifier_config" if legacy else "opensource_classifier_config": { + "provider": "databricks", + "model": "ai_decide", + **({"api_key": "dapi-configured", "api_base": api_base} if configured_key else {}), + }, + "tiers": {"SIMPLE": "cheap"}, + }, + derive_savings_baseline=False, + ) + with respx.mock(assert_all_called=True) as upstream: + route: Final = upstream.post("https://workspace.test/api/2.0/ai-functions/ai-decide").respond( + 200, + json={"response": {"answers": {"tier": _answer().model_dump()}}, "metadata": {"version": "1.0"}}, + ) + outcome: Final = await router.aclassify("choose a tier") + await GLOBAL_LOGGING_WORKER.flush() + + assert outcome.cause == "jev_classifier" + assert outcome.jev_verdict is not None + assert (outcome.jev_verdict.provider, outcome.jev_verdict.model) == ("databricks", "ai_decide") + assert outcome.jev_verdict.label == "SIMPLE" + assert outcome.classifier_cost is None + sent: Final = route.calls.last.request + assert sent.headers.get("authorization") == ("Bearer dapi-configured" if configured_key else "Bearer dapi-env-token") + assert set(json.loads(sent.content)) == {"state", "questions"} + assert len(recorder.calls) == 1 + + def test_databricks_without_a_base_or_key_fails_at_client_build(monkeypatch: pytest.MonkeyPatch) -> None: monkeypatch.delenv("DATABRICKS_API_BASE", raising=False) monkeypatch.delenv("DATABRICKS_API_KEY", raising=False) @@ -618,7 +670,7 @@ def test_databricks_without_a_base_or_key_fails_at_client_build(monkeypatch: pyt litellm.Router(model_list=[]), { "classifier_type": "oss_classifier", - "opensource_classifier_config": {"provider": "databricks", "model": "databricks-openjev-qwen35-4b"}, + "opensource_classifier_config": {"provider": "databricks", "model": "ai_decide"}, "tiers": {"SIMPLE": "cheap"}, }, derive_savings_baseline=False, @@ -665,17 +717,12 @@ def test_jev_api_base_without_its_own_key_is_rejected_so_the_environment_key_sta def test_databricks_api_base_without_its_own_key_is_rejected_so_the_workspace_token_stays_home() -> None: with pytest.raises(ValueError, match="DATABRICKS_API_KEY or DATABRICKS_TOKEN is only sent to DATABRICKS_API_BASE"): JevClassifierConfig.model_validate( - {"provider": "databricks", "model": "my-openjev", "api_base": "https://collector.invalid/serving-endpoints"} + {"provider": "databricks", "model": "ai_decide", "api_base": "https://collector.invalid"} ) paired: Final = JevClassifierConfig.model_validate( - { - "provider": "databricks", - "model": "my-openjev", - "api_base": "https://workspace.invalid/serving-endpoints", - "api_key": "dapi-own", - } + {"provider": "databricks", "model": "ai_decide", "api_base": "https://workspace.invalid", "api_key": "dapi-own"} ) - assert (paired.api_base, paired.api_key) == ("https://workspace.invalid/serving-endpoints", "dapi-own") + assert (paired.api_base, paired.api_key) == ("https://workspace.invalid", "dapi-own") @pytest.mark.parametrize( diff --git a/ui/litellm-dashboard/src/components/add_model/ClassifierTypeRadios.tsx b/ui/litellm-dashboard/src/components/add_model/ClassifierTypeRadios.tsx index 9deb53753bd..6ed46e2a27d 100644 --- a/ui/litellm-dashboard/src/components/add_model/ClassifierTypeRadios.tsx +++ b/ui/litellm-dashboard/src/components/add_model/ClassifierTypeRadios.tsx @@ -55,9 +55,7 @@ const ClassifierTypeRadios: React.FC = ({ value, clas OSS Classifier{" "} - - uses Jev, Laya, Bespoke, or a Databricks endpoint to decide the tier - + uses Jev, Laya, Bespoke, or Databricks to decide the tier diff --git a/ui/litellm-dashboard/src/components/add_model/JevClassifierConfig.integration.test.tsx b/ui/litellm-dashboard/src/components/add_model/JevClassifierConfig.integration.test.tsx index 3480a8e922a..f829a2f2e41 100644 --- a/ui/litellm-dashboard/src/components/add_model/JevClassifierConfig.integration.test.tsx +++ b/ui/litellm-dashboard/src/components/add_model/JevClassifierConfig.integration.test.tsx @@ -125,7 +125,7 @@ describe("JEV classifier editor", () => { expect(screen.getByLabelText("Classifier Model")).toHaveValue("jev-latest"); await chooseSelectOption(userEvent, screen.getByRole("combobox", { name: "Decision model" }), label); if (provider === "bespoke") expect(screen.getByLabelText("Classifier Model")).toHaveTextContent("nimble-latest"); - if (provider === "databricks") expect(screen.getByLabelText("Classifier Model")).toHaveValue(""); + if (provider === "databricks") expect(screen.getByLabelText("Classifier Model")).toHaveValue("ai_decide"); if (typesTheModel) { fireEvent.change(screen.getByLabelText("Classifier Model"), { target: { value: model } }); } else { diff --git a/ui/litellm-dashboard/src/components/add_model/JevClassifierConfig.tsx b/ui/litellm-dashboard/src/components/add_model/JevClassifierConfig.tsx index 31f29dd9e9d..4b3cfa21f9d 100644 --- a/ui/litellm-dashboard/src/components/add_model/JevClassifierConfig.tsx +++ b/ui/litellm-dashboard/src/components/add_model/JevClassifierConfig.tsx @@ -15,7 +15,7 @@ const providerDescriptions = { bespoke: "Uses Bespoke Nimble with your configured tiers. Set BESPOKE_API_BASE on the gateway to connect your Nimble server.", databricks: - "Uses a Databricks ai_decide serving endpoint with your configured tiers. Set DATABRICKS_API_BASE and DATABRICKS_API_KEY on the gateway, and enter the serving endpoint name as the classifier model.", + "Uses the Databricks ai_decide AI Function, or a decision serving endpoint you name, with your configured tiers. Set DATABRICKS_API_BASE to https:///serving-endpoints and DATABRICKS_API_KEY on the gateway.", }; export default function JevClassifierConfig({ @@ -59,7 +59,7 @@ export default function JevClassifierConfig({ value={config.model} placeholder={ config.provider === "databricks" - ? "Serving endpoint name, e.g. databricks-openjev-qwen35-4b" + ? "ai_decide, or a serving endpoint name such as databricks-openjev-qwen35-4b" : undefined } onChange={(event) => update({ model: event.target.value })} diff --git a/ui/litellm-dashboard/src/components/add_model/build_auto_router_routing_test_request.test.ts b/ui/litellm-dashboard/src/components/add_model/build_auto_router_routing_test_request.test.ts index 9ce99d39564..5cfedba7b1e 100644 --- a/ui/litellm-dashboard/src/components/add_model/build_auto_router_routing_test_request.test.ts +++ b/ui/litellm-dashboard/src/components/add_model/build_auto_router_routing_test_request.test.ts @@ -54,6 +54,8 @@ describe("buildAutoRouterRoutingTestRequest", () => { ["bespoke", "bespokelabs/Bespoke-Nimble-9B", "json"], ["databricks", "databricks-openjev-qwen35-4b", "object"], ["databricks", "databricks-openjev-qwen35-4b", "json"], + ["databricks", "ai_decide", "object"], + ["databricks", "ai_decide", "json"], ])("probes saved %s/%s %s configuration with custom tiers and team context", (provider, model, format) => { const config = { classifier_type: "oss_classifier", diff --git a/ui/litellm-dashboard/src/components/add_model/build_complexity_router_config.test.ts b/ui/litellm-dashboard/src/components/add_model/build_complexity_router_config.test.ts index 23dedbbc860..89d419d0487 100644 --- a/ui/litellm-dashboard/src/components/add_model/build_complexity_router_config.test.ts +++ b/ui/litellm-dashboard/src/components/add_model/build_complexity_router_config.test.ts @@ -163,6 +163,7 @@ describe("buildComplexityRouterConfig", () => { ["bespoke", "nimble"], ["bespoke", "bespokelabs/Bespoke-Nimble-9B"], ["databricks", "databricks-openjev-qwen35-4b"], + ["databricks", "ai_decide"], ["jev", "custom-jev-model"], [undefined, "custom-jev-model"], ] as const)("accepts %s model %s before saving or testing", (provider, model) => { diff --git a/ui/litellm-dashboard/src/components/add_model/jev_classifier_config.test.ts b/ui/litellm-dashboard/src/components/add_model/jev_classifier_config.test.ts index 109e780cc68..06298817073 100644 --- a/ui/litellm-dashboard/src/components/add_model/jev_classifier_config.test.ts +++ b/ui/litellm-dashboard/src/components/add_model/jev_classifier_config.test.ts @@ -27,11 +27,11 @@ describe("hydrateOssClassifier", () => { }); describe("jevClassifierConfigSchema", () => { - it("starts a Databricks classifier with an empty endpoint name that does not pass until typed", () => { + it("starts a Databricks classifier on ai_decide and still takes a typed serving endpoint name", () => { const fresh = defaultJevClassifierConfig("databricks"); - expect(fresh).toEqual({ provider: "databricks", model: "", timeout_ms: 3000 }); - expect(jevClassifierConfigSchema.safeParse(fresh).success).toBe(false); + expect(fresh).toEqual({ provider: "databricks", model: "ai_decide", timeout_ms: 3000 }); + expect(jevClassifierConfigSchema.safeParse(fresh).success).toBe(true); expect(jevClassifierConfigSchema.safeParse({ ...fresh, model: "my-openjev" }).success).toBe(true); }); diff --git a/ui/litellm-dashboard/src/components/add_model/jev_classifier_config.ts b/ui/litellm-dashboard/src/components/add_model/jev_classifier_config.ts index a74247f2060..f29921e5e77 100644 --- a/ui/litellm-dashboard/src/components/add_model/jev_classifier_config.ts +++ b/ui/litellm-dashboard/src/components/add_model/jev_classifier_config.ts @@ -13,12 +13,13 @@ export const isOssClassifierProvider = (value: unknown): value is OssClassifierP const DEFAULT_TIMEOUT_MS = 3000; const SERVING_ENDPOINT_NAME = /^[A-Za-z0-9_-][A-Za-z0-9._-]*$/; +const DATABRICKS_AI_DECIDE_MODEL = "ai_decide"; export const fixedClassifierModels = (provider: OssClassifierProvider | undefined): readonly string[] | undefined => provider === "laya" || provider === "bespoke" ? OSS_CLASSIFIER_MODELS[provider] : undefined; const defaultClassifierModel = (provider: OssClassifierProvider | undefined): string => - fixedClassifierModels(provider)?.[0] ?? (provider === "databricks" ? "" : "jev-latest"); + fixedClassifierModels(provider)?.[0] ?? (provider === "databricks" ? DATABRICKS_AI_DECIDE_MODEL : "jev-latest"); const isSupportedClassifierModel = (provider: OssClassifierProvider | undefined, model: string): boolean => { if (provider === "databricks") return SERVING_ENDPOINT_NAME.test(model); @@ -49,7 +50,7 @@ export const jevClassifierConfigSchema = z .refine( (config) => config.deployment_name !== undefined || isSupportedClassifierModel(config.provider, config.model), { - error: "Select a supported classifier model, or enter the bare Databricks serving endpoint name", + error: "Select a supported classifier model, or enter ai_decide or a bare Databricks serving endpoint name", path: ["model"], }, ); @@ -57,9 +58,7 @@ export const jevClassifierConfigSchema = z export type JevClassifierConfig = z.infer; export const defaultJevClassifierConfig = (provider: OssClassifierProvider = "jev"): JevClassifierConfig => - provider === "databricks" - ? { provider, model: "", timeout_ms: DEFAULT_TIMEOUT_MS } - : jevClassifierConfigSchema.parse({ provider }); + jevClassifierConfigSchema.parse({ provider }); export const hydrateOssClassifier = (config: { classifier_type?: ClassifierType | "oss_classifier";