scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-10-09 03:18:44 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 109

Merge remote-tracking branch 'origin/litellm_internal_staging' into litellm_decrease_anys_opus5_r4

Browse source 
# Conflicts:
#	basedpyright-code-budget.json
#	litellm/llms/openai/vector_stores/transformation.py
#	litellm/llms/ragflow/vector_stores/transformation.py
#	litellm/llms/xai/responses/transformation.py
#	litellm/proxy/guardrails/guardrail_hooks/model_armor/model_armor.py
#	ruff-strict-budget.json
#	type-discipline-budget.json
This commit is contained in:
mateo-berri 2026-09-03 06:44:40 +00:00
commit 5b96995b95
99 changed files with 6355 additions and 783 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

24
basedpyright-code-budget.json
View file

@ -1,9 +1,9 @@
{
"reportAny": {
"limit": 13431
"limit": 14074
},
"reportArgumentType": {
"limit": 2207
"limit": 2215
},
"reportAssignmentType": {
"limit": 319
@ -24,7 +24,7 @@
"limit": 19
},
"reportExplicitAny": {
"limit": 3371
"limit": 4124
},
"reportFunctionMemberAccess": {
"limit": 7
@ -48,16 +48,16 @@
"limit": 34
},
"reportInvalidTypeVarUse": {
"limit": 1
"limit": 2
},
"reportMatchNotExhaustive": {
"limit": 0
},
"reportMissingParameterType": {
"limit": 5570
"limit": 5601
},
"reportMissingTypeArgument": {
"limit": 15287
"limit": 15288
},
"reportMissingTypeStubs": {
"limit": 40
@ -90,7 +90,7 @@
"limit": 8
},
"reportReturnType": {
"limit": 180
"limit": 181
},
"reportTypedDictNotRequiredAccess": {
"limit": 24
@ -105,16 +105,16 @@
"limit": 109
},
"reportUnknownMemberType": {
"limit": 38306
"limit": 38324
},
"reportUnknownParameterType": {
"limit": 19589
"limit": 19625
},
"reportUnknownVariableType": {
"limit": 29846
"limit": 29861
},
"reportUnnecessaryCast": {
"limit": 110
"limit": 111
},
"reportUnnecessaryComparison": {
"limit": 687
@ -123,7 +123,7 @@
"limit": 4
},
"reportUnnecessaryIsInstance": {
"limit": 820
"limit": 819
},
"reportUntypedBaseClass": {
"limit": 0

188
litellm-rust/crates/core/src/providers/mistral/ocr/transformation.rs
View file

@ -186,68 +186,180 @@ mod tests {
use serde_json::json;
#[test]
fn supported_params_match_python_mistral_ocr_config() {
fn extract_header_is_a_supported_ocr_param() {
assert!(supported_ocr_params().contains(&"extract_header"));
}
#[test]
fn extract_footer_is_a_supported_ocr_param() {
assert!(supported_ocr_params().contains(&"extract_footer"));
}
#[test]
fn existing_ocr_params_remain_supported() {
for param in [
"pages",
"include_image_base64",
"image_limit",
"image_min_size",
"bbox_annotation_format",
"document_annotation_format",
] {
assert!(supported_ocr_params().contains(&param));
}
}
#[test]
fn map_ocr_params_forwards_extract_header() {
let params = json!({"extract_header": true});
assert_eq!(
supported_ocr_params(),
&[
"pages",
"include_image_base64",
"image_limit",
"image_min_size",
"bbox_annotation_format",
"document_annotation_format",
"document_annotation_prompt",
"extract_header",
"extract_footer",
"table_format",
"confidence_scores_granularity",
"include_blocks",
"id",
]
map_ocr_params(params.as_object().unwrap()),
params.as_object().unwrap().clone()
);
}
#[test]
fn map_ocr_params_forwards_extract_footer() {
let params = json!({"extract_footer": true});
assert_eq!(
map_ocr_params(params.as_object().unwrap()),
params.as_object().unwrap().clone()
);
}
#[test]
fn map_ocr_params_forwards_extract_header_and_footer() {
let params = json!({"extract_header": true, "extract_footer": false});
assert_eq!(
map_ocr_params(params.as_object().unwrap()),
params.as_object().unwrap().clone()
);
}
#[test]
fn map_ocr_params_drops_unknown_params() {
let params = json!({
"extract_header": true,
"unsupported_param": "value",
"pages": [0, 1]
});
let params = json!({"extract_header": true, "unsupported_param": "value"});
let mapped = map_ocr_params(params.as_object().unwrap());
assert_eq!(mapped.get("extract_header"), Some(&json!(true)));
assert_eq!(mapped.get("pages"), Some(&json!([0, 1])));
assert!(!mapped.contains_key("unsupported_param"));
}
#[test]
fn transform_ocr_request_builds_mistral_body() {
fn new_ocr_params_are_supported() {
for param in [
"table_format",
"confidence_scores_granularity",
"document_annotation_prompt",
"include_blocks",
"id",
] {
assert!(supported_ocr_params().contains(&param));
}
}
#[test]
fn map_ocr_params_forwards_new_ocr_params() {
for (param, value) in [
("table_format", json!("html")),
("confidence_scores_granularity", json!("word")),
(
"document_annotation_prompt",
json!("Extract all invoice line items"),
),
("include_blocks", json!(true)),
("id", json!("req-123")),
] {
let params = json!({param: value});
assert_eq!(
map_ocr_params(params.as_object().unwrap()),
params.as_object().unwrap().clone()
);
}
}
#[test]
fn transform_ocr_request_includes_each_optional_param() {
let document = json!({
"type": "document_url",
"document_url": "https://example.com/doc.pdf"
});
for (param, value) in [
("table_format", json!("html")),
("confidence_scores_granularity", json!("word")),
(
"document_annotation_prompt",
json!("Extract all invoice line items"),
),
("id", json!("req-123")),
("extract_header", json!(true)),
("include_blocks", json!(true)),
("pages", json!([0, 1])),
] {
let result = transform_ocr_request(
"mistral-ocr-latest",
document.clone(),
json!({param: value}).as_object().unwrap().clone(),
)
.expect("request should transform");
assert_eq!(result.data.get(param), Some(&value));
assert_eq!(result.data.get("model"), Some(&json!("mistral-ocr-latest")));
assert_eq!(result.data.get("document"), Some(&document));
assert_eq!(result.files, None);
}
}
#[test]
fn transform_ocr_request_includes_multiple_new_params() {
let document = json!({
"type": "document_url",
"document_url": "https://example.com/doc.pdf"
});
let optional_params = json!({
"include_image_base64": true,
"table_format": "html"
"table_format": "html",
"confidence_scores_granularity": "page",
"extract_header": true
})
.as_object()
.unwrap()
.clone();
let result = transform_ocr_request("mistral-ocr-latest", document.clone(), optional_params)
let result = transform_ocr_request("mistral-ocr-latest", document, optional_params)
.expect("request should transform");
assert_eq!(result.data.get("table_format"), Some(&json!("html")));
assert_eq!(
result.data,
json!({
"model": "mistral-ocr-latest",
"document": document,
"include_image_base64": true,
"table_format": "html"
})
result.data.get("confidence_scores_granularity"),
Some(&json!("page"))
);
assert_eq!(result.files, None);
assert_eq!(result.data.get("extract_header"), Some(&json!(true)));
}
#[test]
fn transform_ocr_response_preserves_blocks_and_confidence_scores() {
let blocks = json!([{"type": "title", "content": "Invoice"}]);
let confidence_scores = json!({"page": 0.98});
let response = json!({
"pages": [{"index": 0, "markdown": "# Invoice", "blocks": blocks, "confidence_scores": confidence_scores}],
"model": "mistral-ocr-4-0",
"usage_info": {"pages_processed": 1}
});
let result =
transform_ocr_response("mistral-ocr-4-0", response).expect("response should transform");
assert_eq!(result.pages[0].get("blocks"), Some(&blocks));
assert_eq!(
result.pages[0].get("confidence_scores"),
Some(&confidence_scores)
);
}
#[test]
fn transform_ocr_response_preserves_ocr4_page_fields() {
let response = json!({
"pages": [{"index": 0, "markdown": "table page", "tables": [{"rows": 2, "cols": 3}], "hyperlinks": ["https://example.com"], "header": "Acme Corp", "footer": "Page 1"}],
"model": "mistral-ocr-4-0",
"usage_info": {"pages_processed": 1}
});
let result = transform_ocr_response("mistral-ocr-4-0", response.clone())
.expect("response should transform");
assert_eq!(result.pages[0], response["pages"][0]);
}
#[test]

12
litellm/cost_calculator.py
View file

@ -4,6 +4,7 @@ import logging
import time
from collections.abc import Mapping, Sequence
from functools import lru_cache
from types import MappingProxyType
from typing import TYPE_CHECKING, Any, Final, Literal, cast
from httpx import Response
@ -1164,6 +1165,12 @@ def _store_cost_breakdown_in_logging_obj(
# Don't fail the main cost calculation if breakdown storage fails
def _without_provider_stated_cost(usage: Usage | None) -> Usage | None:
if usage is None or getattr(usage, "cost", None) is None:
return usage
return usage.model_copy(update=MappingProxyType({"cost": None}))
def completion_cost(
completion_response: object | None = None,
model: str | None = None,
@ -1243,7 +1250,10 @@ def completion_cost(
cache_creation_input_tokens: int | None = None
cache_read_input_tokens: int | None = None
audio_transcription_file_duration: float = 0.0
cost_per_token_usage_object: Final[Usage | None] = _get_usage_object(completion_response=completion_response)
provider_usage_object: Final = _get_usage_object(completion_response=completion_response)
cost_per_token_usage_object: Final[Usage | None] = (
_without_provider_stated_cost(provider_usage_object) if custom_pricing else provider_usage_object
)
rerank_billed_units: RerankBilledUnits | None = None
# Extract service_tier from optional_params if not provided directly

49
litellm/integrations/azure_storage/azure_storage.py
View file

@ -1,7 +1,9 @@
import asyncio
import os
import time
from collections.abc import Callable
from datetime import datetime, timedelta
from functools import cache
from typing import Final
from litellm._logging import verbose_logger
@ -19,21 +21,40 @@ from litellm.llms.custom_httpx.http_handler import (
get_async_httpx_client,
httpxSpecialProvider,
)
from litellm.secret_managers.get_azure_ad_token_provider import (
get_azure_ad_token_provider,
)
from litellm.types.secret_managers.get_azure_ad_token_provider import (
AzureCredentialType,
)
from litellm.types.utils import StandardLoggingPayload
AZURE_STORAGE_TOKEN_SCOPE: Final = "https://storage.azure.com/.default"
@cache
def _cached_credential_chain_token_provider() -> Callable[[], str]:
return get_azure_ad_token_provider(
azure_scope=AZURE_STORAGE_TOKEN_SCOPE,
azure_credential=AzureCredentialType.DefaultAzureCredential,
)
class AzureBlobStorageLogger(CustomBatchLogger):
def __init__(
self,
build_credential_chain_token_provider: Callable[
[], Callable[[], str]
] = _cached_credential_chain_token_provider,
**kwargs,
):
try:
verbose_logger.debug("AzureBlobStorageLogger: in init azure blob storage logger")
# Env Variables used for Azure Storage Authentication
self.tenant_id = os.getenv("AZURE_STORAGE_TENANT_ID")
self.client_id = os.getenv("AZURE_STORAGE_CLIENT_ID")
self.client_secret = os.getenv("AZURE_STORAGE_CLIENT_SECRET")
self.tenant_id = os.getenv("AZURE_STORAGE_TENANT_ID") or None
self.client_id = os.getenv("AZURE_STORAGE_CLIENT_ID") or None
self.client_secret = os.getenv("AZURE_STORAGE_CLIENT_SECRET") or None
self.azure_storage_account_key: str | None = os.getenv("AZURE_STORAGE_ACCOUNT_KEY")
# Required Env Variables for Azure Storage
@ -55,6 +76,9 @@ class AzureBlobStorageLogger(CustomBatchLogger):
# Internal variables used for Token based authentication
self.azure_auth_token: str | None = None # the Azure AD token to use for Azure Storage API requests
self.token_expiry: datetime | None = None # the expiry time of the currentAzure AD token
self._build_credential_chain_token_provider: Callable[[], Callable[[], str]] = (
build_credential_chain_token_provider
)
asyncio.create_task(self.periodic_flush())
self.flush_lock = asyncio.Lock()
@ -231,10 +255,15 @@ class AzureBlobStorageLogger(CustomBatchLogger):
"""
Wrapper to set self.azure_auth_token to a valid Azure AD token, refreshing if necessary
Refreshes the token when:
- Token is expired
- Token is not set
Without a service principal configured, the credential chain provider is read every
time; it caches internally and refreshes against the token's real expiry. The read runs
in a worker thread because the chain walk (IMDS probe, CLI subprocess) is blocking
"""
if self.tenant_id is None and self.client_id is None and self.client_secret is None:
token_provider: Final = self._build_credential_chain_token_provider()
self.azure_auth_token = await asyncio.to_thread(token_provider)
return
# Check if token needs refresh
if self._azure_ad_token_is_expired() or self.azure_auth_token is None:
verbose_logger.debug("Azure AD token needs refresh")
@ -273,13 +302,9 @@ class AzureBlobStorageLogger(CustomBatchLogger):
tenant_id=tenant_id,
client_id=client_id,
client_secret=client_secret,
scope="https://storage.azure.com/.default",
scope=AZURE_STORAGE_TOKEN_SCOPE,
)
token: Final = token_provider()
verbose_logger.debug("azure auth token %s", token)
return token
return token_provider()
def _azure_ad_token_is_expired(self):
"""

217
litellm/integrations/datadog/datadog_llm_obs.py
View file

@ -19,11 +19,13 @@ import httpx
import litellm
from litellm._logging import verbose_logger
from litellm._uuid import uuid
from litellm.constants import REDACTED_BY_LITELLM
from litellm.integrations.custom_batch_logger import CustomBatchLogger
from litellm.integrations.datadog.datadog_handler import (
get_datadog_base_url_from_env,
get_datadog_service,
get_datadog_tags,
normalize_datadog_tag_value,
)
from litellm.integrations.datadog.datadog_mock_client import (
create_mock_datadog_client,
@ -34,6 +36,7 @@ from litellm.litellm_core_utils.prompt_templates.common_utils import (
convert_content_list_to_str,
handle_any_messages_to_chat_completion_str_messages_conversion,
)
from litellm.litellm_core_utils.redact_messages import should_redact_message_logging
from litellm.litellm_core_utils.safe_json_dumps import safe_dumps
from litellm.litellm_core_utils.safe_json_loads import safe_json_loads
from litellm.llms.custom_httpx.http_handler import (
@ -43,6 +46,7 @@ from litellm.llms.custom_httpx.http_handler import (
from litellm.proxy.spend_tracking.savings import extract_cache_creation_tokens, extract_cache_read_tokens
from litellm.types.integrations.datadog_llm_obs import *
from litellm.types.utils import (
PROMPT_QUOTING_ROUTING_DECISION_FIELDS,
CallTypes,
StandardLoggingGuardrailInformation,
StandardLoggingPayload,
@ -52,6 +56,120 @@ from litellm.types.utils import (
_EMPTY_MAPPING: Final[Mapping[str, Any]] = MappingProxyType({})
_EMPTY_MESSAGE: Final[Message] = {"role": "", "content": ""}
_MAX_PARSED_TOOL_ARGUMENT_CHARS: Final = 256 * 1024
_SAFE_REDACTED_MESSAGE_ROLES: Final = frozenset(
{"agent", "assistant", "developer", "function", "model", "system", "tool", "user"}
)
_PROMPT_CARRYING_METADATA_FIELDS: Final = frozenset(
{
"routing_decision",
"requester_metadata",
"prompt_management_metadata",
"mcp_tool_call_metadata",
"vector_store_request_metadata",
}
)
_ROUTER_SPAN_FIELDS: Final[Mapping[str, str]] = MappingProxyType(
{
"tier": "router_tier",
"cause": "router_cause",
"score": "router_score",
"escalated": "router_escalated",
"signals": "router_signals",
"routed_model": "routed_model",
}
)
_ROUTER_DIMENSIONS: Final[tuple[str, ...]] = ("router_tier", "router_cause", "router_escalated", "routed_model")
_COST_DIMENSIONS: Final[tuple[str, ...]] = ("team", "user", "key_alias", "model_group", *_ROUTER_DIMENSIONS)
def _metadata_of(standard_logging_payload: StandardLoggingPayload) -> Mapping[str, Any]:
metadata: Final = standard_logging_payload.get("metadata")
return metadata or _EMPTY_MAPPING
def _router_span_fields(
standard_logging_payload: StandardLoggingPayload, redact_prompt_text: bool
) -> Mapping[str, object]:
"""Flatten the auto-router decision, omitting prompt-quoting fields when redaction is enabled."""
routing_decision: Final = _mapping_field(_metadata_of(standard_logging_payload), "routing_decision")
if not routing_decision:
return _EMPTY_MAPPING
escalated: Final = bool(routing_decision.get("escalated") or routing_decision.get("context_escalated"))
return MappingProxyType(
{
_ROUTER_SPAN_FIELDS[record_field]: value
for record_field, value in (*routing_decision.items(), ("escalated", escalated))
if record_field in _ROUTER_SPAN_FIELDS
and value is not None
and not (redact_prompt_text and record_field in PROMPT_QUOTING_ROUTING_DECISION_FIELDS)
}
)
def _metadata_without_prompt_carriers(standard_logging_metadata: Mapping[str, Any]) -> Mapping[str, Any]:
"""The metadata minus the records that quote prompts, tool arguments, tool results, or retrieved text."""
return MappingProxyType(
{
field: value
for field, value in standard_logging_metadata.items()
if field not in _PROMPT_CARRYING_METADATA_FIELDS
}
)
def _redact_messages(messages: Sequence[Message]) -> tuple[Message, ...]:
"""Each message's shape with its content replaced and tool payloads dropped; no message is invented."""
return tuple(
{
"role": role if isinstance(role, str) and role in _SAFE_REDACTED_MESSAGE_ROLES else "",
"content": REDACTED_BY_LITELLM,
}
for message in messages
for role in (message.get("role", ""),)
)
def _cost_dimension_tags(
standard_logging_payload: StandardLoggingPayload, router_fields: Mapping[str, object]
) -> tuple[str, ...]:
"""The dimensions LLM Obs breaks token and cost metrics down by, as span tags."""
metadata: Final = _metadata_of(standard_logging_payload)
dimensions: Final = (
("user", metadata.get("user_api_key_user_id")),
("key_alias", metadata.get("user_api_key_alias")),
("model_group", standard_logging_payload.get("model_group")),
*((dimension, router_fields.get(dimension)) for dimension in _ROUTER_DIMENSIONS),
)
return tuple(
f"{key}:{normalized}"
for key, value in dimensions
if value is not None and (normalized := normalize_datadog_tag_value(value)) != ""
)
def _declared_cost_tags(span_tags: Sequence[str]) -> tuple[str, ...]:
"""Declare only cost dimensions carrying a value on this span."""
present: Final = frozenset(key for tag in span_tags if (key := tag.partition(":")[0]) and tag.partition(":")[2])
return tuple(dimension for dimension in _COST_DIMENSIONS if dimension in present)
def _reasoning_output_tokens(usage_object: Mapping[str, Any] | None) -> float:
"""The provider's reasoning-token count, from either the chat or the responses spelling."""
if usage_object is None:
return 0.0
return next(
(
float(reasoning_tokens)
for details_field in ("completion_tokens_details", "output_tokens_details")
if isinstance(
reasoning_tokens := _mapping_field(usage_object, details_field).get("reasoning_tokens"), (int, float)
)
and not isinstance(reasoning_tokens, bool)
),
0.0,
)
def _mapping_field(source: Mapping[str, Any], key: str) -> Mapping[str, Any]:
@ -316,12 +434,12 @@ class DataDogLLMObsLogger(CustomBatchLogger):
dict_datadog_llm_obs_params: dict = {}
if litellm.datadog_llm_observability_params is not None:
if isinstance(litellm.datadog_llm_observability_params, DatadogLLMObsInitParams):
dict_datadog_llm_obs_params = litellm.datadog_llm_observability_params.model_dump()
dict_datadog_llm_obs_params = litellm.datadog_llm_observability_params.model_dump(exclude_unset=True)
elif isinstance(litellm.datadog_llm_observability_params, dict):
# only allow params that are of DatadogLLMObsInitParams
dict_datadog_llm_obs_params = DatadogLLMObsInitParams(
**litellm.datadog_llm_observability_params
).model_dump()
).model_dump(exclude_unset=True)
return dict_datadog_llm_obs_params
async def async_log_success_event(self, kwargs, response_obj, start_time, end_time):
@ -410,25 +528,40 @@ class DataDogLLMObsLogger(CustomBatchLogger):
if standard_logging_payload is None:
raise Exception("DataDogLLMObs: standard_logging_object is not set")
metadata: Final = kwargs.get("litellm_params", {}).get("metadata", {})
raw_metadata: Final = kwargs.get("litellm_params", {}).get("metadata", {})
metadata: Final = raw_metadata if isinstance(raw_metadata, dict) else {}
redact_payload: Final = self._payload_logging_is_off(kwargs)
input_meta: Final = InputMeta(messages=_to_dd_messages(standard_logging_payload["messages"]))
input_messages: Final = _to_dd_messages(standard_logging_payload.get("messages"))
output_messages: Final = self._get_response_messages(
standard_logging_payload=standard_logging_payload,
call_type=standard_logging_payload.get("call_type"),
)
input_meta: Final = InputMeta(messages=_redact_messages(input_messages) if redact_payload else input_messages)
output_meta: Final = OutputMeta(
messages=self._get_response_messages(
standard_logging_payload=standard_logging_payload,
call_type=standard_logging_payload.get("call_type"),
)
messages=_redact_messages(output_messages) if redact_payload else output_messages
)
error_info: Final = self._assemble_error_info(standard_logging_payload)
metadata_parent_id: str | None = None
if isinstance(metadata, dict):
metadata_parent_id = metadata.get("parent_id")
raw_parent_id: Final = metadata.get("parent_id")
metadata_parent_id: Final[str | None] = str(raw_parent_id) if raw_parent_id else None
tool_definitions: Final = _to_dd_tool_definitions(standard_logging_payload.get("model_parameters"))
tool_definitions: Final = (
() if redact_payload else _to_dd_tool_definitions(standard_logging_payload.get("model_parameters"))
)
span_kind: Final = self._get_datadog_span_kind(standard_logging_payload.get("call_type"), metadata_parent_id)
payload_metadata: Final = self._get_dd_llm_obs_payload_metadata(standard_logging_payload)
router_fields: Final = _router_span_fields(standard_logging_payload, redact_prompt_text=redact_payload)
span_tags: Final = [
*get_datadog_tags(standard_logging_object=standard_logging_payload),
*_cost_dimension_tags(standard_logging_payload, router_fields),
]
payload_metadata: Final = self._get_dd_llm_obs_payload_metadata(
standard_logging_payload,
router_fields=router_fields,
cost_tags=_declared_cost_tags(span_tags),
redact_prompt_text=redact_payload,
)
meta: Final[Meta] = {
"kind": span_kind,
@ -451,7 +584,7 @@ class DataDogLLMObsLogger(CustomBatchLogger):
duration=int((end_time - start_time).total_seconds() * 1e9),
metrics=metrics,
status="error" if error_info else "ok",
tags=get_datadog_tags(standard_logging_object=standard_logging_payload),
tags=span_tags,
)
apm_trace_id: Final = self._get_apm_trace_id()
@ -497,6 +630,13 @@ class DataDogLLMObsLogger(CustomBatchLogger):
)
return error_info
def _payload_logging_is_off(self, kwargs: Mapping[str, Any]) -> bool:
return (
bool(self.turn_off_message_logging)
or self.message_logging is not True
or should_redact_message_logging(dict(kwargs))
)
def _assemble_metrics(self, standard_logging_payload: StandardLoggingPayload) -> LLMMetrics:
"""
Build the span metrics, including the prompt-cache counts LLM Obs charts cache savings from.
@ -513,10 +653,11 @@ class DataDogLLMObsLogger(CustomBatchLogger):
total_cost: Final = float(standard_logging_payload.get("response_cost", 0))
time_to_first_token: Final = self._get_time_to_first_token_seconds(standard_logging_payload)
raw_usage: Final = (standard_logging_payload.get("metadata") or {}).get("usage_object")
raw_usage: Final = _metadata_of(standard_logging_payload).get("usage_object")
usage_object: Final = raw_usage if isinstance(raw_usage, dict) else None
cache_read: Final = float(extract_cache_read_tokens(usage_object))
cache_write: Final = float(extract_cache_creation_tokens(usage_object))
reasoning_output_tokens: Final = _reasoning_output_tokens(usage_object)
metrics: Final[LLMMetrics] = {
"input_tokens": prompt_tokens,
@ -533,6 +674,7 @@ class DataDogLLMObsLogger(CustomBatchLogger):
if cache_read or cache_write
else {}
),
**({"reasoning_output_tokens": reasoning_output_tokens} if reasoning_output_tokens else {}),
}
return metrics
@ -707,11 +849,21 @@ class DataDogLLMObsLogger(CustomBatchLogger):
# Default fallback for unknown or passthrough operations
return "llm"
def _get_dd_llm_obs_payload_metadata(self, standard_logging_payload: StandardLoggingPayload) -> dict[str, object]:
def _get_dd_llm_obs_payload_metadata(
self,
standard_logging_payload: StandardLoggingPayload,
router_fields: Mapping[str, object] | None = None,
cost_tags: Sequence[str] = (),
redact_prompt_text: bool = False,
) -> dict[str, object]:
"""
Fields to track in DD LLM Observability metadata from litellm standard logging payload
"""
_metadata: Final[dict[str, object]] = {
raw_metadata: Final = _metadata_of(standard_logging_payload)
standard_logging_metadata: Final = (
_metadata_without_prompt_carriers(raw_metadata) if redact_prompt_text else raw_metadata
)
return {
"model_name": standard_logging_payload.get("model", "unknown"),
"model_provider": standard_logging_payload.get("custom_llm_provider", "unknown"),
"id": standard_logging_payload.get("id", "unknown"),
@ -719,26 +871,21 @@ class DataDogLLMObsLogger(CustomBatchLogger):
"cache_hit": standard_logging_payload.get("cache_hit", "unknown"),
"cache_key": standard_logging_payload.get("cache_key", "unknown"),
"saved_cache_cost": standard_logging_payload.get("saved_cache_cost", 0),
"guardrail_information": standard_logging_payload.get("guardrail_information", None),
"guardrail_information": (
None if redact_prompt_text else standard_logging_payload.get("guardrail_information", None)
),
"is_streamed_request": self._get_stream_value_from_payload(standard_logging_payload),
"latency_metrics": dict(self._get_latency_metrics(standard_logging_payload)),
"spend_metrics": dict(self._get_spend_metrics(standard_logging_payload)),
**standard_logging_metadata,
**(router_fields or _EMPTY_MAPPING),
**(
{"_dd": {**_mapping_field(standard_logging_metadata, "_dd"), "cost_tags": list(cost_tags)}}
if cost_tags
else _EMPTY_MAPPING
),
}
#########################################################
# Add latency metrics to metadata
#########################################################
latency_metrics: Final = self._get_latency_metrics(standard_logging_payload)
_metadata.update({"latency_metrics": dict(latency_metrics)})
#########################################################
# Add spend metrics to metadata
#########################################################
spend_metrics: Final = self._get_spend_metrics(standard_logging_payload)
_metadata.update({"spend_metrics": dict(spend_metrics)})
_standard_logging_metadata: Final[dict] = dict(standard_logging_payload.get("metadata", {})) or {}
_metadata.update(_standard_logging_metadata)
return _metadata
def _get_latency_metrics(self, standard_logging_payload: StandardLoggingPayload) -> DDLLMObsLatencyMetrics:
"""
Get the latency metrics from the standard logging payload
@ -808,7 +955,7 @@ class DataDogLLMObsLogger(CustomBatchLogger):
spend_metrics["response_cost"] = standard_logging_payload.get("response_cost", 0.0)
# Get budget information from metadata
metadata: Final = standard_logging_payload.get("metadata", {})
metadata: Final = _metadata_of(standard_logging_payload)
# API key max budget
user_api_key_max_budget: Final = metadata.get("user_api_key_max_budget")

17
litellm/litellm_core_utils/streaming_handler.py
View file

@ -54,6 +54,7 @@ FUNCTION_CALL_ATTRIBUTE: Final = "function_call"
_SYNC_ITER_EXHAUSTED: Final = object()
_GCHUNK_FIELDS: Final[frozenset] = frozenset(GChunk.__annotations__)
_USAGE_COST_HEADER_PROVIDERS: Final[frozenset[str]] = frozenset({LlmProviders.OPENROUTER.value})
def _next_sync_or_exhausted(it: Any) -> object:
@ -1886,8 +1887,8 @@ class CustomStreamWrapper:
@staticmethod
def _resolve_provider_reported_cost(usage_cost: object) -> float | None:
"""
Providers report usage.cost either as a number or, for Perplexity, as a
breakdown object whose total lives under ``total_cost``.
Providers report usage.cost either as a number or as a breakdown object
whose total lives under ``total_cost``.
"""
if isinstance(usage_cost, bool):
return None
@ -1900,12 +1901,10 @@ class CustomStreamWrapper:
@staticmethod
def _propagate_usage_cost_to_hidden_params(
response: "ModelResponse",
custom_llm_provider: str | None,
) -> None:
"""
If the assembled response carries a provider-reported cost on
usage.cost, copy it into _hidden_params so litellm's cost
calculator uses it instead of a token-based estimate.
"""
if custom_llm_provider not in _USAGE_COST_HEADER_PROVIDERS:
return
_usage: Final[Usage | None] = getattr(response, "usage", None)
_cost: Final = CustomStreamWrapper._resolve_provider_reported_cost(getattr(_usage, "cost", None))
if _cost is not None:
@ -2020,7 +2019,7 @@ class CustomStreamWrapper:
response = self.model_response_creator()
if complete_streaming_response is not None:
self._propagate_usage_cost_to_hidden_params(complete_streaming_response)
self._propagate_usage_cost_to_hidden_params(complete_streaming_response, self.custom_llm_provider)
setattr(
response,
@ -2270,7 +2269,7 @@ class CustomStreamWrapper:
response: Final = self.model_response_creator()
if complete_streaming_response is not None:
self._propagate_usage_cost_to_hidden_params(complete_streaming_response)
self._propagate_usage_cost_to_hidden_params(complete_streaming_response, self.custom_llm_provider)
setattr(
response,

7
litellm/llms/azure_ai/vector_stores/transformation.py
View file

@ -24,7 +24,6 @@ from litellm.types.vector_stores import (
if TYPE_CHECKING:
from litellm.litellm_core_utils.litellm_logging import Logging as _LiteLLMLoggingObj
from litellm.router import Router
LiteLLMLoggingObj = _LiteLLMLoggingObj
else:
@ -121,11 +120,10 @@ class AzureAIVectorStoreConfig(BaseQueryEmbeddingVectorStoreConfig, BaseAzureLLM
litellm_logging_obj: LiteLLMLoggingObj,
litellm_params: Mapping[str, object],
extra_body: Mapping[str, object] | None = None,
router: Router | None = None,
embedding_executor: VectorStoreEmbeddingExecutor | None = None,
) -> tuple[str, dict[str, object]]:
query_text: Final = self.query_text(query)
query_vector: Final = self.embed_query(query_text, litellm_params, embedding_executor, router)
query_vector: Final = self.embed_query(query_text, litellm_params, embedding_executor)
return self._search_request(
vector_store_id,
query_text,
@ -145,11 +143,10 @@ class AzureAIVectorStoreConfig(BaseQueryEmbeddingVectorStoreConfig, BaseAzureLLM
litellm_logging_obj: LiteLLMLoggingObj,
litellm_params: Mapping[str, object],
extra_body: Mapping[str, object] | None = None,
router: Router | None = None,
embedding_executor: VectorStoreEmbeddingExecutor | None = None,
) -> tuple[str, dict[str, object]]:
query_text: Final = self.query_text(query)
query_vector: Final = await self.aembed_query(query_text, litellm_params, embedding_executor, router)
query_vector: Final = await self.aembed_query(query_text, litellm_params, embedding_executor)
return self._search_request(
vector_store_id,
query_text,

20
litellm/llms/base_llm/vector_store/transformation.py
View file

@ -99,12 +99,9 @@ class RouterVectorStoreEmbeddingExecutor:
)
return bool(resolved) or model in deployment_models
def _embeds_through_sdk(self, model: str, configuration: Mapping[str, object]) -> bool:
return bool(configuration) and not self._router_serves(model)
def embed(self, model: str, query: str, configuration: Mapping[str, object]) -> EmbeddingResponse:
embedding_kwargs: Final = self._embedding_kwargs(configuration)
if self._embeds_through_sdk(model, configuration):
if not self._router_serves(model):
return LiteLLMVectorStoreEmbeddingExecutor().embed(model, query, embedding_kwargs)
return self.router.embedding( # pyright: ignore[reportUnknownMemberType] # Router embedding input retains a legacy untyped list
model=model,
@ -114,7 +111,7 @@ class RouterVectorStoreEmbeddingExecutor:
async def aembed(self, model: str, query: str, configuration: Mapping[str, object]) -> EmbeddingResponse:
embedding_kwargs: Final = self._embedding_kwargs(configuration)
if self._embeds_through_sdk(model, configuration):
if not self._router_serves(model):
return await LiteLLMVectorStoreEmbeddingExecutor().aembed(model, query, embedding_kwargs)
return await self.router.aembedding( # pyright: ignore[reportUnknownMemberType] # Router embedding input retains a legacy untyped list
model=model,
@ -153,7 +150,6 @@ class BaseVectorStoreConfig:
litellm_logging_obj: LiteLLMLoggingObj,
litellm_params: dict,
extra_body: dict[str, Any] | None = None,
router: Router | None = None,
) -> tuple[str, dict]:
pass
@ -166,7 +162,6 @@ class BaseVectorStoreConfig:
litellm_logging_obj: LiteLLMLoggingObj,
litellm_params: dict,
extra_body: dict[str, Any] | None = None,
router: Router | None = None,
) -> tuple[str, dict]:
"""
Optional async version of transform_search_vector_store_request.
@ -182,7 +177,6 @@ class BaseVectorStoreConfig:
litellm_logging_obj=litellm_logging_obj,
litellm_params=litellm_params,
extra_body=extra_body,
router=router,
)
@abstractmethod
@ -271,7 +265,6 @@ class BaseQueryEmbeddingVectorStoreConfig(BaseVectorStoreConfig):
litellm_logging_obj: LiteLLMLoggingObj,
litellm_params: Mapping[str, object],
extra_body: Mapping[str, object] | None = None,
router: Router | None = None,
embedding_executor: VectorStoreEmbeddingExecutor | None = None,
) -> tuple[str, dict[str, object]]:
pass
@ -285,7 +278,6 @@ class BaseQueryEmbeddingVectorStoreConfig(BaseVectorStoreConfig):
litellm_logging_obj: LiteLLMLoggingObj,
litellm_params: Mapping[str, object],
extra_body: Mapping[str, object] | None = None,
router: Router | None = None,
embedding_executor: VectorStoreEmbeddingExecutor | None = None,
) -> tuple[str, dict[str, object]]:
return self.transform_search_vector_store_request(
@ -296,7 +288,6 @@ class BaseQueryEmbeddingVectorStoreConfig(BaseVectorStoreConfig):
litellm_logging_obj=litellm_logging_obj,
litellm_params=litellm_params,
extra_body=extra_body,
router=router,
embedding_executor=embedding_executor,
)
@ -338,11 +329,10 @@ class BaseQueryEmbeddingVectorStoreConfig(BaseVectorStoreConfig):
query_text: str,
litellm_params: Mapping[str, object],
embedding_executor: VectorStoreEmbeddingExecutor | None,
router: Router | None = None,
) -> Sequence[float]:
model: Final = self.query_embedding_model(litellm_params)
configuration: Final = self.query_embedding_configuration(litellm_params)
executor: Final = self.query_embedding_executor(embedding_executor, router)
executor: Final = self.query_embedding_executor(embedding_executor, None)
try:
response: Final = executor.embed(model, query_text, configuration)
except Exception as e:
@ -354,11 +344,10 @@ class BaseQueryEmbeddingVectorStoreConfig(BaseVectorStoreConfig):
query_text: str,
litellm_params: Mapping[str, object],
embedding_executor: VectorStoreEmbeddingExecutor | None,
router: Router | None = None,
) -> Sequence[float]:
model: Final = self.query_embedding_model(litellm_params)
configuration: Final = self.query_embedding_configuration(litellm_params)
executor: Final = self.query_embedding_executor(embedding_executor, router)
executor: Final = self.query_embedding_executor(embedding_executor, None)
try:
response: Final = await executor.aembed(model, query_text, configuration)
except Exception as e:
@ -408,7 +397,6 @@ class BaseDirectVectorStoreConfig(BaseVectorStoreConfig):
litellm_logging_obj: LiteLLMLoggingObj,
litellm_params: Mapping[str, object],
extra_body: Mapping[str, object] | None = None,
router: Router | None = None,
) -> NoReturn:
raise NotImplementedError("Direct vector store providers execute the search themselves; no HTTP request shape")

2
litellm/llms/bedrock/vector_stores/transformation.py
View file

@ -27,7 +27,6 @@ from litellm.types.vector_stores import (
if TYPE_CHECKING:
from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
from litellm.router import Router
else:
LiteLLMLoggingObj = Any
@ -197,7 +196,6 @@ class BedrockVectorStoreConfig(BaseVectorStoreConfig, BaseAWSLLM):
litellm_logging_obj: LiteLLMLoggingObj,
litellm_params: dict,
extra_body: dict[str, Any] | None = None,
router: "Router | None" = None,
) -> tuple[str, dict]:
if isinstance(query, list):
query = " ".join(query)

8
litellm/llms/custom_httpx/llm_http_handler.py
View file

@ -184,7 +184,6 @@ if TYPE_CHECKING:
AnthropicMessagesStreamingResponse,
)
from litellm.llms.base_llm.passthrough.transformation import BasePassthroughConfig
from litellm.router import Router
from litellm.types.llms.openai_evals import (
CancelEvalResponse,
CancelRunResponse,
@ -9709,7 +9708,6 @@ class BaseLLMHTTPHandler:
timeout: float | httpx.Timeout | None = None,
client: HTTPHandler | AsyncHTTPHandler | None = None,
_is_async: bool = False,
router: "Router | None" = None,
) -> VectorStoreSearchResponse:
if isinstance(vector_store_provider_config, BaseDirectVectorStoreConfig):
self._pre_call_direct_vector_store_search(
@ -9760,7 +9758,6 @@ class BaseLLMHTTPHandler:
litellm_logging_obj=logging_obj,
litellm_params=dict(litellm_params),
extra_body=extra_body,
router=router,
embedding_executor=embedding_executor,
)
else:
@ -9775,7 +9772,6 @@ class BaseLLMHTTPHandler:
litellm_logging_obj=logging_obj,
litellm_params=dict(litellm_params),
extra_body=extra_body,
router=router,
)
all_optional_params: Final[dict[str, object]] = dict(litellm_params)
all_optional_params.update(vector_store_search_optional_params or {})
@ -9828,7 +9824,6 @@ class BaseLLMHTTPHandler:
timeout: float | httpx.Timeout | None = None,
client: HTTPHandler | AsyncHTTPHandler | None = None,
_is_async: bool = False,
router: "Router | None" = None,
) -> VectorStoreSearchResponse | Coroutine[object, object, VectorStoreSearchResponse]:
if _is_async:
return self.async_vector_store_search_handler(
@ -9844,7 +9839,6 @@ class BaseLLMHTTPHandler:
extra_body=extra_body,
timeout=timeout,
client=client,
router=router,
)
if isinstance(vector_store_provider_config, BaseDirectVectorStoreConfig):
@ -9893,7 +9887,6 @@ class BaseLLMHTTPHandler:
litellm_logging_obj=logging_obj,
litellm_params=dict(litellm_params),
extra_body=extra_body,
router=router,
embedding_executor=embedding_executor,
)
else:
@ -9908,7 +9901,6 @@ class BaseLLMHTTPHandler:
litellm_logging_obj=logging_obj,
litellm_params=dict(litellm_params),
extra_body=extra_body,
router=router,
)
all_optional_params: Final[dict[str, object]] = dict(litellm_params)

2
litellm/llms/gemini/vector_stores/transformation.py
View file

@ -33,7 +33,6 @@ from litellm.types.vector_stores import (
if TYPE_CHECKING:
from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
from litellm.router import Router
else:
LiteLLMLoggingObj = Any
@ -169,7 +168,6 @@ class GeminiVectorStoreConfig(BaseVectorStoreConfig):
litellm_logging_obj: LiteLLMLoggingObj,
litellm_params: dict,
extra_body: Mapping[str, object] | None = None,
router: "Router | None" = None,
) -> tuple[str, dict]:
"""
Transform search request to Gemini's generateContent format.

7
litellm/llms/milvus/vector_stores/transformation.py
View file

@ -24,7 +24,6 @@ from litellm.types.vector_stores import (
if TYPE_CHECKING:
from litellm.litellm_core_utils.litellm_logging import Logging as _LiteLLMLoggingObj
from litellm.router import Router
LiteLLMLoggingObj = _LiteLLMLoggingObj
else:
@ -129,11 +128,10 @@ class MilvusVectorStoreConfig(BaseQueryEmbeddingVectorStoreConfig):
litellm_logging_obj: LiteLLMLoggingObj,
litellm_params: Mapping[str, object],
extra_body: Mapping[str, object] | None = None,
router: Router | None = None,
embedding_executor: VectorStoreEmbeddingExecutor | None = None,
) -> tuple[str, dict[str, object]]:
query_text: Final = self.query_text(query)
query_vector: Final = self.embed_query(query_text, litellm_params, embedding_executor, router)
query_vector: Final = self.embed_query(query_text, litellm_params, embedding_executor)
return self._search_request(
vector_store_id,
query_text,
@ -153,11 +151,10 @@ class MilvusVectorStoreConfig(BaseQueryEmbeddingVectorStoreConfig):
litellm_logging_obj: LiteLLMLoggingObj,
litellm_params: Mapping[str, object],
extra_body: Mapping[str, object] | None = None,
router: Router | None = None,
embedding_executor: VectorStoreEmbeddingExecutor | None = None,
) -> tuple[str, dict[str, object]]:
query_text: Final = self.query_text(query)
query_vector: Final = await self.aembed_query(query_text, litellm_params, embedding_executor, router)
query_vector: Final = await self.aembed_query(query_text, litellm_params, embedding_executor)
return self._search_request(
vector_store_id,
query_text,

2
litellm/llms/openai/vector_stores/transformation.py
View file

@ -21,7 +21,6 @@ from litellm.utils import add_openai_metadata
if TYPE_CHECKING:
from litellm.litellm_core_utils.litellm_logging import Logging as _LiteLLMLoggingObj
from litellm.router import Router
LiteLLMLoggingObj = _LiteLLMLoggingObj
else:
@ -100,7 +99,6 @@ class OpenAIVectorStoreConfig(BaseVectorStoreConfig):
litellm_logging_obj: LiteLLMLoggingObj,
litellm_params: dict,
extra_body: dict[str, object] | None = None,
router: "Router | None" = None,
) -> tuple[str, dict]:
encoded_vector_store_id: Final = encode_url_path_segment(vector_store_id, field_name="vector_store_id")
url: Final = f"{api_base}/{encoded_vector_store_id}/search"

2
litellm/llms/pg_vector/vector_stores/transformation.py
View file

@ -8,7 +8,6 @@ from litellm.types.vector_stores import VectorStoreSearchOptionalRequestParams
if TYPE_CHECKING:
from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
from litellm.router import Router
else:
LiteLLMLoggingObj = Any
@ -81,7 +80,6 @@ class PGVectorStoreConfig(OpenAIVectorStoreConfig):
litellm_logging_obj: LiteLLMLoggingObj,
litellm_params: dict,
extra_body: dict[str, Any] | None = None,
router: "Router | None" = None,
) -> tuple[str, dict]:
encoded_vector_store_id: Final = encode_url_path_segment(vector_store_id, field_name="vector_store_id")
url: Final = f"{api_base}/{encoded_vector_store_id}/search"

2
litellm/llms/ragflow/vector_stores/transformation.py
View file

@ -18,7 +18,6 @@ from litellm.types.vector_stores import (
if TYPE_CHECKING:
from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
from litellm.router import Router
else:
LiteLLMLoggingObj = Any
@ -94,7 +93,6 @@ class RAGFlowVectorStoreConfig(BaseVectorStoreConfig):
litellm_logging_obj: LiteLLMLoggingObj,
litellm_params: dict,
extra_body: Mapping[str, object] | None = None,
router: "Router | None" = None,
) -> tuple[str, dict]:
"""RAGFlow vector stores are management-only, search is not supported."""
raise NotImplementedError("RAGFlow vector stores support dataset management only, not search/retrieval")

209
litellm/llms/s3_vectors/vector_stores/transformation.py
View file

@ -1,9 +1,12 @@
from collections.abc import Mapping, Sequence
from typing import TYPE_CHECKING, Any, Final
import httpx
from litellm.caching._embedding_router import resolve_embedding_router
from litellm.llms.base_llm.vector_store.transformation import BaseVectorStoreConfig
from litellm.llms.base_llm.vector_store.transformation import (
BaseQueryEmbeddingVectorStoreConfig,
VectorStoreEmbeddingExecutor,
)
from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
from litellm.types.router import GenericLiteLLMParams
from litellm.types.vector_stores import (
@ -18,16 +21,18 @@ from litellm.types.vector_stores import (
if TYPE_CHECKING:
from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
from litellm.router import Router
else:
LiteLLMLoggingObj = Any
_DEFAULT_QUERY_EMBEDDING_MODEL: Final = "text-embedding-3-small"
_DEFAULT_TOP_K: Final = 5
class S3VectorsVectorStoreConfig(BaseVectorStoreConfig, BaseAWSLLM):
class S3VectorsVectorStoreConfig(BaseQueryEmbeddingVectorStoreConfig, BaseAWSLLM):
"""Vector store configuration for AWS S3 Vectors."""
def __init__(self) -> None:
BaseVectorStoreConfig.__init__(self)
BaseQueryEmbeddingVectorStoreConfig.__init__(self)
BaseAWSLLM.__init__(self)
def get_auth_credentials(self, litellm_params: dict) -> BaseVectorStoreAuthCredentials:
@ -59,141 +64,94 @@ class S3VectorsVectorStoreConfig(BaseVectorStoreConfig, BaseAWSLLM):
return headers
def get_complete_url(self, api_base: str | None, litellm_params: dict) -> str:
# Resolve region the same way the ingestion path does:
# dynamic param -> AWS_REGION_NAME -> AWS_REGION -> default (us-west-2)
aws_region_name: Final = self.get_aws_region_name_for_non_llm_api_calls(litellm_params.get("aws_region_name"))
return f"https://s3vectors.{aws_region_name}.api.aws"
def _resolve_query_embedding_router(self, embedding_model: str, router: "Router | None") -> "Router | None":
"""Return the router iff it serves ``embedding_model`` as a deployment."""
if router is None:
return None
model_list: Final = [
dict(m) for m in (router.get_model_list() or ())
] # mutable-ok: resolve_embedding_router requires list[dict]
return resolve_embedding_router(embedding_model=embedding_model, llm_router=router, llm_model_list=model_list)
@staticmethod
def query_embedding_model(litellm_params: Mapping[str, object]) -> str:
configured: Final = litellm_params.get("litellm_embedding_model") or litellm_params.get("embedding_model")
return configured if isinstance(configured, str) and configured else _DEFAULT_QUERY_EMBEDDING_MODEL
@staticmethod
def _query_target(vector_store_id: str, litellm_params: Mapping[str, object]) -> tuple[str, str]:
if ":" in vector_store_id:
bucket_name, index_name = vector_store_id.split(":", 1)
return bucket_name, index_name
bucket_name_from_params: Final = litellm_params.get("vector_bucket_name")
if not isinstance(bucket_name_from_params, str) or not bucket_name_from_params:
raise ValueError(
"vector_store_id must be in format 'bucket_name:index_name' for S3 Vectors, "
"or vector_bucket_name must be provided in litellm_params"
)
return bucket_name_from_params, vector_store_id
@staticmethod
def _query_request(
bucket_name: str,
index_name: str,
query_text: str,
query_vector: Sequence[float],
vector_store_search_optional_params: VectorStoreSearchOptionalRequestParams,
api_base: str,
litellm_logging_obj: LiteLLMLoggingObj,
) -> tuple[str, dict[str, object]]:
litellm_logging_obj.model_call_details["query"] = query_text
return f"{api_base}/QueryVectors", {
"vectorBucketName": bucket_name,
"indexName": index_name,
"queryVector": {"float32": list(query_vector)},
"topK": vector_store_search_optional_params.get("max_num_results", _DEFAULT_TOP_K),
"returnDistance": True,
"returnMetadata": True,
}
def transform_search_vector_store_request(
self,
vector_store_id: str,
query: str | list[str],
query: str | Sequence[str],
vector_store_search_optional_params: VectorStoreSearchOptionalRequestParams,
api_base: str,
litellm_logging_obj: LiteLLMLoggingObj,
litellm_params: dict,
extra_body: dict[str, Any] | None = None,
router: "Router | None" = None,
) -> tuple[str, dict]:
"""Sync version - generates embedding synchronously."""
# For S3 Vectors, vector_store_id should be in format: bucket_name:index_name
# If not in that format, try to construct it from litellm_params
bucket_name: str
index_name: str
if ":" in vector_store_id:
bucket_name, index_name = vector_store_id.split(":", 1)
else:
# Try to get bucket_name from litellm_params
bucket_name_from_params: Final = litellm_params.get("vector_bucket_name")
if not bucket_name_from_params or not isinstance(bucket_name_from_params, str):
raise ValueError(
"vector_store_id must be in format 'bucket_name:index_name' for S3 Vectors, "
"or vector_bucket_name must be provided in litellm_params"
)
bucket_name = bucket_name_from_params
index_name = vector_store_id
if isinstance(query, list):
query = " ".join(query)
# Generate embedding for the query
embedding_model: Final = litellm_params.get("embedding_model", "text-embedding-3-small")
embedding_router: Final = self._resolve_query_embedding_router(embedding_model=embedding_model, router=router)
import litellm as litellm_module
embedding_input: Final = [query] # mutable-ok: the embedding API takes list input
embedding_response: Final = (
embedding_router.embedding(model=embedding_model, input=embedding_input)
if embedding_router is not None
else litellm_module.embedding(model=embedding_model, input=embedding_input)
litellm_params: Mapping[str, object],
extra_body: Mapping[str, object] | None = None,
embedding_executor: VectorStoreEmbeddingExecutor | None = None,
) -> tuple[str, dict[str, object]]:
bucket_name, index_name = self._query_target(vector_store_id, litellm_params)
query_text: Final = self.query_text(query)
query_vector: Final = self.embed_query(query_text, litellm_params, embedding_executor)
return self._query_request(
bucket_name,
index_name,
query_text,
query_vector,
vector_store_search_optional_params,
api_base,
litellm_logging_obj,
)
query_embedding: Final = embedding_response.data[0]["embedding"]
url: Final = f"{api_base}/QueryVectors"
request_body: Final[dict[str, Any]] = {
"vectorBucketName": bucket_name,
"indexName": index_name,
"queryVector": {"float32": query_embedding},
"topK": vector_store_search_optional_params.get("max_num_results", 5), # Default to 5
"returnDistance": True,
"returnMetadata": True,
}
litellm_logging_obj.model_call_details["query"] = query
return url, request_body
async def atransform_search_vector_store_request(
self,
vector_store_id: str,
query: str | list[str],
query: str | Sequence[str],
vector_store_search_optional_params: VectorStoreSearchOptionalRequestParams,
api_base: str,
litellm_logging_obj: LiteLLMLoggingObj,
litellm_params: dict,
extra_body: dict[str, Any] | None = None,
router: "Router | None" = None,
) -> tuple[str, dict]:
"""Async version - generates embedding asynchronously."""
# For S3 Vectors, vector_store_id should be in format: bucket_name:index_name
# If not in that format, try to construct it from litellm_params
bucket_name: str
index_name: str
if ":" in vector_store_id:
bucket_name, index_name = vector_store_id.split(":", 1)
else:
# Try to get bucket_name from litellm_params
bucket_name_from_params: Final = litellm_params.get("vector_bucket_name")
if not bucket_name_from_params or not isinstance(bucket_name_from_params, str):
raise ValueError(
"vector_store_id must be in format 'bucket_name:index_name' for S3 Vectors, "
"or vector_bucket_name must be provided in litellm_params"
)
bucket_name = bucket_name_from_params
index_name = vector_store_id
if isinstance(query, list):
query = " ".join(query)
# Generate embedding for the query asynchronously
embedding_model: Final = litellm_params.get("embedding_model", "text-embedding-3-small")
embedding_router: Final = self._resolve_query_embedding_router(embedding_model=embedding_model, router=router)
import litellm as litellm_module
embedding_input: Final = [query] # mutable-ok: the embedding API takes list input
embedding_response: Final = (
await embedding_router.aembedding(model=embedding_model, input=embedding_input)
if embedding_router is not None
else await litellm_module.aembedding(model=embedding_model, input=embedding_input)
litellm_params: Mapping[str, object],
extra_body: Mapping[str, object] | None = None,
embedding_executor: VectorStoreEmbeddingExecutor | None = None,
) -> tuple[str, dict[str, object]]:
bucket_name, index_name = self._query_target(vector_store_id, litellm_params)
query_text: Final = self.query_text(query)
query_vector: Final = await self.aembed_query(query_text, litellm_params, embedding_executor)
return self._query_request(
bucket_name,
index_name,
query_text,
query_vector,
vector_store_search_optional_params,
api_base,
litellm_logging_obj,
)
query_embedding: Final = embedding_response.data[0]["embedding"]
url: Final = f"{api_base}/QueryVectors"
request_body: Final[dict[str, Any]] = {
"vectorBucketName": bucket_name,
"indexName": index_name,
"queryVector": {"float32": query_embedding},
"topK": vector_store_search_optional_params.get("max_num_results", 5), # Default to 5
"returnDistance": True,
"returnMetadata": True,
}
litellm_logging_obj.model_call_details["query"] = query
return url, request_body
def sign_request(
self,
@ -226,21 +184,13 @@ class S3VectorsVectorStoreConfig(BaseVectorStoreConfig, BaseAWSLLM):
if not source_text:
continue
# Extract file information from metadata
chunk_index = metadata.get("chunk_index", "0")
file_id = f"s3-vectors-chunk-{chunk_index}"
filename = metadata.get("filename", f"document-{chunk_index}")
# S3 Vectors returns distance, convert to similarity score (0-1)
# Lower distance = higher similarity
# We'll normalize using 1 / (1 + distance) to get a 0-1 score
distance = item.get("distance")
score = None
if distance is not None:
# Convert distance to similarity score between 0 and 1
# For cosine distance: similarity = 1 - distance
# For euclidean: use 1 / (1 + distance)
# Assuming cosine distance here
score = max(0.0, min(1.0, 1.0 - float(distance)))
results.append(
@ -265,7 +215,6 @@ class S3VectorsVectorStoreConfig(BaseVectorStoreConfig, BaseAWSLLM):
headers=response.headers,
)
# Vector store creation is not yet implemented
def transform_create_vector_store_request(
self,
vector_store_create_optional_params,

2
litellm/llms/vertex_ai/vector_stores/rag_api/transformation.py
View file

@ -21,7 +21,6 @@ from litellm.types.vector_stores import (
if TYPE_CHECKING:
from litellm.litellm_core_utils.litellm_logging import Logging as _LiteLLMLoggingObj
from litellm.router import Router
LiteLLMLoggingObj = _LiteLLMLoggingObj
else:
@ -162,7 +161,6 @@ class VertexVectorStoreConfig(BaseVectorStoreConfig, VertexBase):
litellm_logging_obj: LiteLLMLoggingObj,
litellm_params: dict,
extra_body: Mapping[str, object] | None = None,
router: "Router | None" = None,
) -> tuple[str, dict[str, object]]:
"""
Transform search request for Vertex AI RAG API

2
litellm/llms/vertex_ai/vector_stores/search_api/transformation.py
View file

@ -25,7 +25,6 @@ from litellm.types.vector_stores import (
if TYPE_CHECKING:
from litellm.litellm_core_utils.litellm_logging import Logging as _LiteLLMLoggingObj
from litellm.router import Router
LiteLLMLoggingObj = _LiteLLMLoggingObj
else:
@ -246,7 +245,6 @@ class VertexSearchAPIVectorStoreConfig(BaseVectorStoreConfig, VertexBase):
litellm_logging_obj: LiteLLMLoggingObj,
litellm_params: dict,
extra_body: Mapping[str, object] | None = None,
router: "Router | None" = None,
) -> tuple[str, dict[str, object]]:
"""
Transform a search request for the Vertex AI Search (Discovery Engine) API.

19
litellm/llms/xai/chat/transformation.py
View file

@ -1,4 +1,5 @@
from collections.abc import AsyncIterator, Iterator, Mapping
from types import MappingProxyType
from typing import Any, Final
import httpx
@ -11,7 +12,7 @@ from litellm.litellm_core_utils.prompt_templates.common_utils import (
filter_value_from_dict,
strip_name_from_messages,
)
from litellm.llms.xai.common_utils import XAIModelInfo
from litellm.llms.xai.common_utils import XAIModelInfo, xai_reported_cost_in_usd
from litellm.llms.xai.cost_calculator import (
apply_server_side_tool_usage_details_to_usage,
)
@ -30,6 +31,13 @@ from ...openai.chat.gpt_transformation import (
)
def _usage_restated_from_xai_ticks(usage: Usage | None) -> Usage | None:
reported_cost: Final = xai_reported_cost_in_usd(getattr(usage, "cost_in_usd_ticks", None))
if usage is None or reported_cost is None:
return None
return usage.model_copy(update=MappingProxyType({"cost": reported_cost}))
class XAIChatConfig(OpenAIGPTConfig):
@property
def custom_llm_provider(self) -> str | None:
@ -283,6 +291,9 @@ class XAIChatConfig(OpenAIGPTConfig):
self._fold_reasoning_tokens_into_completion(response)
self._normalize_openai_compatible_usage_totals(getattr(response, "usage", None))
restated_usage: Final = _usage_restated_from_xai_ticks(getattr(response, "usage", None))
if restated_usage is not None:
response.usage = restated_usage
return response
@staticmethod
@ -411,4 +422,8 @@ class XAIChatCompletionStreamingHandler(OpenAIChatCompletionStreamingHandler):
XAIChatConfig._fold_reasoning_tokens_into_completion(chunk["usage"])
XAIChatConfig._normalize_openai_compatible_usage_totals(chunk["usage"])
return super().chunk_parser(chunk)
parsed_chunk: Final = super().chunk_parser(chunk)
restated_usage: Final = _usage_restated_from_xai_ticks(getattr(parsed_chunk, "usage", None))
if restated_usage is not None:
parsed_chunk.usage = restated_usage
return parsed_chunk

11
litellm/llms/xai/common_utils.py
View file

@ -8,6 +8,17 @@ from litellm.secret_managers.main import get_secret_str
from litellm.types.llms.openai import AllMessageValues
from litellm.types.utils import ProviderSpecificModelInfo
USD_TICKS_PER_DOLLAR: Final = 10_000_000_000
def xai_reported_cost_in_usd(cost_in_usd_ticks: object) -> float | None:
"""xAI bills in ticks of a dollar: https://docs.x.ai/developers/cost-tracking"""
if not isinstance(cost_in_usd_ticks, int) or isinstance(cost_in_usd_ticks, bool):
return None
if cost_in_usd_ticks < 0:
return None
return cost_in_usd_ticks / USD_TICKS_PER_DOLLAR
class XAIModelInfo(BaseLLMModelInfo):
def get_provider_info(

20
litellm/llms/xai/cost_calculator.py
View file

@ -1,9 +1,11 @@
"""
Helper util for handling XAI-specific cost calculation
- Prefers the cost xAI reports on the response over recomputing it locally
- Uses the generic cost calculator which already handles tiered pricing correctly
- Handles XAI-specific reasoning token billing (billed as part of completion tokens)
"""
import math
from collections.abc import Mapping
from typing import TYPE_CHECKING, Final
@ -36,6 +38,17 @@ def apply_server_side_tool_usage_details_to_usage(usage: Usage, details: Mapping
usage.prompt_tokens_details = prompt_tokens_details # rebind-ok: write details onto caller usage
def _cost_reported_by_xai(usage: "Usage") -> float | None:
reported_cost: Final[object] = getattr(usage, "cost", None)
if not isinstance(reported_cost, (int, float)) or isinstance(reported_cost, bool):
return None
if not math.isfinite(reported_cost):
return None
if reported_cost < 0:
return None
return float(reported_cost)
def cost_per_token(model: str, usage: Usage) -> tuple[float, float]:
"""
Calculates the cost per token for a given XAI model, prompt tokens, and completion tokens.
@ -48,6 +61,10 @@ def cost_per_token(model: str, usage: Usage) -> tuple[float, float]:
Returns:
Tuple[float, float] - prompt_cost_in_usd, completion_cost_in_usd
"""
reported_cost: Final = _cost_reported_by_xai(usage)
if reported_cost is not None:
return 0.0, reported_cost
# XAI-specific completion cost: completion is billed as visible + reasoning
# tokens. Detect when the transformation layer already folded them so we
# don't double-count; fall back to raw xAI shape for callers that bypass
@ -112,6 +129,9 @@ def cost_per_web_search_request(usage: "Usage", model_info: "ModelInfo") -> floa
Per-call rate comes from model_info.search_context_cost_per_query when set,
otherwise the default xAI tools rate ($5 / 1k calls).
"""
if _cost_reported_by_xai(usage) is not None:
return 0.0
details: Final = getattr(usage, "server_side_tool_usage_details", None)
if not isinstance(details, Mapping):
return 0.0

68
litellm/llms/xai/responses/transformation.py
View file

@ -1,17 +1,44 @@
from collections.abc import Mapping
from typing import Final
from types import MappingProxyType
from typing import TYPE_CHECKING, Any, Final
import httpx
import litellm
from litellm._logging import verbose_logger
from litellm.constants import XAI_API_BASE
from litellm.exceptions import AuthenticationError
from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig
from litellm.llms.xai.common_utils import XAIModelInfo
from litellm.llms.xai.common_utils import XAIModelInfo, xai_reported_cost_in_usd
from litellm.secret_managers.main import get_secret_str
from litellm.types.llms.openai import ResponsesAPIOptionalRequestParams
from litellm.types.llms.openai import (
ResponseAPIUsage,
ResponseCompletedEvent,
ResponseFailedEvent,
ResponseIncompleteEvent,
ResponsesAPIOptionalRequestParams,
ResponsesAPIResponse,
ResponsesAPIStreamingResponse,
)
from litellm.types.router import GenericLiteLLMParams
from litellm.types.utils import LlmProviders
if TYPE_CHECKING:
from litellm.litellm_core_utils.litellm_logging import (
Logging as _LiteLLMLoggingObj,
)
LiteLLMLoggingObj = _LiteLLMLoggingObj
else:
LiteLLMLoggingObj = Any
def _usage_restated_from_xai_ticks(usage: ResponseAPIUsage | None) -> ResponseAPIUsage | None:
reported_cost: Final = xai_reported_cost_in_usd(getattr(usage, "cost_in_usd_ticks", None))
if usage is None or reported_cost is None:
return None
return usage.model_copy(update=MappingProxyType({"cost": reported_cost}))
class XAIResponsesAPIConfig(OpenAIResponsesAPIConfig):
"""
@ -250,6 +277,41 @@ class XAIResponsesAPIConfig(OpenAIResponsesAPIConfig):
return f"{api_base}/responses"
def transform_response_api_response(
self,
model: str,
raw_response: httpx.Response,
logging_obj: LiteLLMLoggingObj,
) -> ResponsesAPIResponse:
response: Final = super().transform_response_api_response(
model=model,
raw_response=raw_response,
logging_obj=logging_obj,
)
restated_usage: Final = _usage_restated_from_xai_ticks(response.usage)
if restated_usage is not None:
response.usage = restated_usage
return response
def transform_streaming_response(
self,
model: str,
parsed_chunk: dict, # mutable-ok: overrides the base class signature
logging_obj: LiteLLMLoggingObj,
) -> ResponsesAPIStreamingResponse:
event: Final = super().transform_streaming_response(
model=model,
parsed_chunk=parsed_chunk,
logging_obj=logging_obj,
)
if not isinstance(event, (ResponseCompletedEvent, ResponseIncompleteEvent, ResponseFailedEvent)):
return event
restated_usage: Final = _usage_restated_from_xai_ticks(event.response.usage)
if restated_usage is not None:
event.response.usage = restated_usage
return event
def supports_native_websocket(self) -> bool:
"""XAI does not support native WebSocket for Responses API"""
return False

12
litellm/main.py
View file

@ -8595,9 +8595,19 @@ def stream_chunk_builder_text_completion(chunks: list, messages: list | None = N
return TextCompletionResponse(**response)
_CALCULATOR_PRICED_REPORTED_COST_PROVIDERS: Final = frozenset({LlmProviders.XAI.value})
def _reported_cost_is_priced_by_calculator(logging_obj: Optional["Logging"]) -> bool:
if logging_obj is None:
return False
provider: Final[object] = logging_obj.model_call_details.get("custom_llm_provider")
return provider in _CALCULATOR_PRICED_REPORTED_COST_PROVIDERS
def _stream_builder_response_cost(response: ModelResponse, logging_obj: Optional["Logging"]) -> float | None:
usage_cost: Final = getattr(getattr(response, "usage", None), "cost", None)
if isinstance(usage_cost, (int, float)):
if isinstance(usage_cost, (int, float)) and not _reported_cost_is_priced_by_calculator(logging_obj):
return float(usage_cost)
if logging_obj is not None:
return None

97
litellm/proxy/_experimental/mcp_server/mcp_server_manager.py
View file

@ -820,20 +820,46 @@ def _should_strip_caller_authorization(
if not (mcp_server.is_oauth_passthrough or mcp_server.is_oauth_delegate):
return False
normalized_raw_headers: Final = {str(k).lower(): v for k, v in (raw_headers or {}).items() if isinstance(k, str)}
has_explicit_litellm_admission_header: Final = normalized_raw_headers.get("x-litellm-api-key") is not None
has_explicit_litellm_admission_header: Final = _has_explicit_litellm_admission_header(raw_headers)
if mcp_server.is_oauth_delegate:
return not has_explicit_litellm_admission_header
admission_consumed_authorization_as_litellm_key: Final = (
user_api_key_auth is not None
and bool(getattr(user_api_key_auth, "api_key", None))
and not has_explicit_litellm_admission_header
)
return admission_consumed_authorization_as_litellm_key or (
return _authorization_is_litellm_admission_credential(raw_headers, user_api_key_auth) or (
user_api_key_auth is None and not has_explicit_litellm_admission_header
)
LITELLM_VIRTUAL_KEY_PREFIX: Final = "sk-"
def _raw_header_value(raw_headers: Mapping[str, str] | None, name: str) -> str | None:
return next((v for k, v in (raw_headers or {}).items() if isinstance(k, str) and k.lower() == name), None)
def _has_explicit_litellm_admission_header(raw_headers: Mapping[str, str] | None) -> bool:
"""Admission only consumes a non-empty ``x-litellm-api-key``; an empty one falls back to ``Authorization``."""
return bool(_raw_header_value(raw_headers, "x-litellm-api-key"))
def _authorization_is_litellm_admission_credential(
raw_headers: Mapping[str, str] | None,
user_api_key_auth: UserAPIKeyAuth | None,
) -> bool:
"""True when ``Authorization`` carries the LiteLLM key admission validated.
That is the case when no usable ``x-litellm-api-key`` was sent, or when the client repeated the
same key in both headers.
"""
if user_api_key_auth is None or not user_api_key_auth.api_key:
return False
admission_header: Final = _raw_header_value(raw_headers, "x-litellm-api-key")
if not admission_header:
return True
authorization: Final = _raw_header_value(raw_headers, "authorization")
return authorization is not None and strip_auth_scheme(authorization, "Bearer") == strip_auth_scheme(
admission_header, "Bearer"
)
def _format_byok_openapi_auth_header(mcp_server: MCPServer, mcp_auth_header: str) -> str:
"""Format a raw BYOK credential for OpenAPI tool ``Authorization`` injection.
@ -3277,8 +3303,8 @@ class MCPServerManager:
#########################################################
@staticmethod
def _extract_bearer_token(
oauth2_headers: dict[str, str] | None,
raw_headers: dict[str, str] | None,
oauth2_headers: Mapping[str, str] | None,
raw_headers: Mapping[str, str] | None,
) -> str | None:
"""Extract the bare Bearer token from oauth2_headers or raw_headers.
@ -3298,10 +3324,29 @@ class MCPServerManager:
return auth_value
return None
@staticmethod
def _extract_subject_token(
oauth2_headers: Mapping[str, str] | None,
raw_headers: Mapping[str, str] | None,
user_api_key_auth: UserAPIKeyAuth | None,
) -> str | None:
"""The caller's upstream identity token, or ``None`` when the bearer is a LiteLLM key.
Rejects the key admission validated and, because virtual keys always carry the ``sk-`` prefix,
any other LiteLLM key a client puts in ``Authorization`` next to ``x-litellm-api-key``.
"""
if _authorization_is_litellm_admission_credential(raw_headers, user_api_key_auth):
return None
bearer: Final = MCPServerManager._extract_bearer_token(oauth2_headers, raw_headers)
if bearer is not None and bearer.startswith(LITELLM_VIRTUAL_KEY_PREFIX):
return None
return bearer
def _obo_subject_token(
self,
server: MCPServer,
raw_headers: dict[str, str] | None,
raw_headers: Mapping[str, str] | None,
user_api_key_auth: UserAPIKeyAuth | None,
) -> str | None:
"""The caller's bearer as the token_exchange (OBO) subject token, for that mode only.
@ -3311,7 +3356,7 @@ class MCPServerManager:
"""
if server.auth_type != MCPAuth.oauth2_token_exchange:
return None
return self._extract_bearer_token(None, raw_headers)
return self._extract_subject_token(None, raw_headers, user_api_key_auth)
def _build_stdio_env(
self,
@ -3566,6 +3611,7 @@ class MCPServerManager:
server: MCPServer,
oauth2_headers: dict[str, str] | None,
user_api_key_auth: UserAPIKeyAuth | None,
raw_headers: Mapping[str, str] | None = None,
) -> None:
"""Run the OBO exchange for a caller-supplied subject at the transport edge.
@ -3577,13 +3623,15 @@ class MCPServerManager:
"""
if server.auth_type != MCPAuth.oauth2_token_exchange:
return
subject_token: Final = self._extract_bearer_token(oauth2_headers, None)
if not subject_token:
if not self._extract_bearer_token(oauth2_headers, None):
return
resolved_server: Final = await self.ensure_oauth_metadata_discovered(server)
spec: Final = to_server_spec(resolved_server)
if spec is None or not isinstance(spec.config, TokenExchangeConfig):
return
subject_token: Final = self._extract_subject_token(oauth2_headers, raw_headers, user_api_key_auth)
if subject_token is None:
raise_token_exchange_challenge(resolved_server, root_path=get_server_root_path())
match await self._cred_provider.resolve_credentials(to_subject(user_api_key_auth, subject_token), spec):
case Ok(_):
return
@ -3851,7 +3899,7 @@ class MCPServerManager:
# token (mirrors the call path), not v1's deleted client_credentials fallback. Other modes
# never read the inbound bearer, so leave subject_token None to avoid forwarding it.
subject_token: Final = (
self._extract_bearer_token(oauth2_headers, raw_headers)
self._extract_subject_token(oauth2_headers, raw_headers, user_api_key_auth)
if server.auth_type == MCPAuth.oauth2_token_exchange
else None
)
@ -3931,6 +3979,7 @@ class MCPServerManager:
async def get_prompts_from_server(
self,
server: MCPServer,
user_api_key_auth: UserAPIKeyAuth | None,
mcp_auth_header: str | dict[str, str] | None = None,
extra_headers: dict[str, str] | None = None,
add_prefix: bool = True,
@ -3959,7 +4008,7 @@ class MCPServerManager:
extra_headers.update(server.static_headers)
stdio_env: Final = self._build_stdio_env(server, raw_headers)
subject_token: Final = self._obo_subject_token(server, raw_headers)
subject_token: Final = self._obo_subject_token(server, raw_headers, user_api_key_auth)
client = await self._create_mcp_client(
server=server,
@ -3982,6 +4031,7 @@ class MCPServerManager:
async def get_resources_from_server(
self,
server: MCPServer,
user_api_key_auth: UserAPIKeyAuth | None,
mcp_auth_header: str | dict[str, str] | None = None,
extra_headers: dict[str, str] | None = None,
add_prefix: bool = True,
@ -4001,7 +4051,7 @@ class MCPServerManager:
extra_headers.update(server.static_headers)
stdio_env: Final = self._build_stdio_env(server, raw_headers)
subject_token: Final = self._obo_subject_token(server, raw_headers)
subject_token: Final = self._obo_subject_token(server, raw_headers, user_api_key_auth)
client = await self._create_mcp_client(
server=server,
@ -4024,6 +4074,7 @@ class MCPServerManager:
async def get_resource_templates_from_server(
self,
server: MCPServer,
user_api_key_auth: UserAPIKeyAuth | None,
mcp_auth_header: str | dict[str, str] | None = None,
extra_headers: dict[str, str] | None = None,
add_prefix: bool = True,
@ -4043,7 +4094,7 @@ class MCPServerManager:
extra_headers.update(server.static_headers)
stdio_env: Final = self._build_stdio_env(server, raw_headers)
subject_token: Final = self._obo_subject_token(server, raw_headers)
subject_token: Final = self._obo_subject_token(server, raw_headers, user_api_key_auth)
client = await self._create_mcp_client(
server=server,
@ -4068,6 +4119,7 @@ class MCPServerManager:
async def read_resource_from_server(
self,
server: MCPServer,
user_api_key_auth: UserAPIKeyAuth | None,
url: AnyUrl,
mcp_auth_header: str | dict[str, str] | None = None,
extra_headers: dict[str, str] | None = None,
@ -4084,7 +4136,7 @@ class MCPServerManager:
extra_headers.update(server.static_headers)
stdio_env: Final = self._build_stdio_env(server, raw_headers)
subject_token: Final = self._obo_subject_token(server, raw_headers)
subject_token: Final = self._obo_subject_token(server, raw_headers, user_api_key_auth)
client: Final = await self._create_mcp_client(
server=server,
@ -4099,6 +4151,7 @@ class MCPServerManager:
async def get_prompt_from_server(
self,
server: MCPServer,
user_api_key_auth: UserAPIKeyAuth | None,
prompt_name: str,
arguments: dict[str, str] | None = None,
mcp_auth_header: str | dict[str, str] | None = None,
@ -4116,7 +4169,7 @@ class MCPServerManager:
extra_headers.update(server.static_headers)
stdio_env: Final = self._build_stdio_env(server, raw_headers)
subject_token: Final = self._obo_subject_token(server, raw_headers)
subject_token: Final = self._obo_subject_token(server, raw_headers, user_api_key_auth)
client: Final = await self._create_mcp_client(
server=server,
@ -5290,7 +5343,7 @@ class MCPServerManager:
MCPAuth.oauth2_token_exchange,
MCPAuth.oauth2_id_jag,
):
subject_token = self._extract_bearer_token(oauth2_headers, raw_headers)
subject_token = self._extract_subject_token(oauth2_headers, raw_headers, user_api_key_auth)
elif mcp_server.auth_type == MCPAuth.oauth2:
if mcp_server.has_client_credentials:
# For M2M OAuth servers, Authorization must come from token fetch.
@ -5638,7 +5691,7 @@ class MCPServerManager:
subject_token: str | None = None
if isinstance(spec.config, (TokenExchangeConfig, IdJagConfig)):
subject_token = self._extract_bearer_token(oauth2_headers, raw_headers)
subject_token = self._extract_subject_token(oauth2_headers, raw_headers, user_api_key_auth)
elif isinstance(spec.config, PassthroughConfig):
inbound_token, forwarded_headers = _take_forwarded_authorization(forwarded_headers)
per_server_token: Final = _passthrough_token_from_mcp_auth_header(mcp_auth_header)

9
litellm/proxy/_experimental/mcp_server/server.py
View file

@ -2189,6 +2189,7 @@ if MCP_AVAILABLE:
try:
prompts = await global_mcp_server_manager.get_prompts_from_server(
server=server,
user_api_key_auth=user_api_key_auth,
mcp_auth_header=server_auth_header,
extra_headers=extra_headers,
add_prefix=True, # Always add server prefix
@ -2242,6 +2243,7 @@ if MCP_AVAILABLE:
try:
resources = await global_mcp_server_manager.get_resources_from_server(
server=server,
user_api_key_auth=user_api_key_auth,
mcp_auth_header=server_auth_header,
extra_headers=extra_headers,
add_prefix=True, # Always add server prefix
@ -2293,6 +2295,7 @@ if MCP_AVAILABLE:
try:
resource_templates = await global_mcp_server_manager.get_resource_templates_from_server(
server=server,
user_api_key_auth=user_api_key_auth,
mcp_auth_header=server_auth_header,
extra_headers=extra_headers,
add_prefix=True, # Always add server prefix
@ -3211,6 +3214,7 @@ if MCP_AVAILABLE:
return await global_mcp_server_manager.get_prompt_from_server(
server=server,
user_api_key_auth=user_api_key_auth,
prompt_name=original_prompt_name,
arguments=arguments,
mcp_auth_header=server_auth_header,
@ -3261,6 +3265,7 @@ if MCP_AVAILABLE:
return await global_mcp_server_manager.read_resource_from_server(
server=server,
user_api_key_auth=user_api_key_auth,
url=url,
mcp_auth_header=server_auth_header,
extra_headers=extra_headers,
@ -3723,6 +3728,7 @@ if MCP_AVAILABLE:
user_api_key_auth: UserAPIKeyAuth | None,
client_ip: str | None,
allowed_server_ids: set[str] | None = None,
raw_headers: Mapping[str, str] | None = None,
) -> None:
"""Fail fast with HTTP 401 for MCP servers that need user auth but
didn't receive it on this request. Covers both gateway-managed OAuth2
@ -3867,6 +3873,7 @@ if MCP_AVAILABLE:
server=server,
oauth2_headers=oauth2_headers,
user_api_key_auth=user_api_key_auth,
raw_headers=raw_headers,
)
# Pass-through OAuth: when the admin has opted a server into
@ -4195,6 +4202,7 @@ if MCP_AVAILABLE:
user_api_key_auth=user_api_key_auth,
client_ip=_client_ip,
allowed_server_ids=toolset_allowed_server_ids,
raw_headers=raw_headers,
)
# Pre-flight auth check for pass-through servers. Must run after
@ -4518,6 +4526,7 @@ if MCP_AVAILABLE:
user_api_key_auth=user_api_key_auth,
client_ip=_sse_client_ip,
allowed_server_ids=toolset_allowed_server_ids,
raw_headers=raw_headers,
)
# Pre-flight auth check for pass-through servers: surface upstream

64
litellm/proxy/guardrails/anthropic_sse.py
View file

@ -13,6 +13,19 @@ from typing import Final
from litellm.types.utils import Choices, ModelResponse
_ANTHROPIC_EVENT_TYPES: Final = frozenset(
{
"message_start",
"message_delta",
"message_stop",
"content_block_start",
"content_block_delta",
"content_block_stop",
"ping",
"error",
}
)
def is_raw_sse_stream(all_chunks: Sequence[object]) -> bool:
return any(isinstance(chunk, (str, bytes)) for chunk in all_chunks)
@ -30,23 +43,43 @@ def _joined_sse_stream(all_chunks: Sequence[object]) -> str | None:
return None
def _anthropic_message_start(sse_stream: str) -> Mapping[str, object] | None:
def _parsed_sse_events(sse_stream: str) -> tuple[Mapping[str, object], ...]:
from litellm.proxy.pass_through_endpoints.llm_provider_handlers.anthropic_passthrough_logging_handler import (
AnthropicPassthroughLoggingHandler,
)
return tuple(
event_data
for event in AnthropicPassthroughLoggingHandler._split_sse_chunk_into_events(sse_stream) # pyright: ignore[reportPrivateUsage] # same parser the assembler uses
if (event_data := AnthropicPassthroughLoggingHandler._extract_sse_data(event)) is not None # pyright: ignore[reportPrivateUsage] # same parser the assembler uses; a private import beats forking SSE parsing
)
def _anthropic_message_start(sse_stream: str) -> Mapping[str, object] | None:
return next(
(
message
for event in AnthropicPassthroughLoggingHandler._split_sse_chunk_into_events(sse_stream) # pyright: ignore[reportPrivateUsage] # same parser the assembler uses
if (event_data := AnthropicPassthroughLoggingHandler._extract_sse_data(event)) is not None # pyright: ignore[reportPrivateUsage] # same parser the assembler uses; a private import beats forking SSE parsing
and event_data.get("type") == "message_start"
and isinstance(message := event_data.get("message"), dict)
for event_data in _parsed_sse_events(sse_stream)
if event_data.get("type") == "message_start" and isinstance(message := event_data.get("message"), dict)
),
None,
)
def is_anthropic_sse_stream(all_chunks: Sequence[object]) -> bool:
"""Whether raw SSE frames are Anthropic Messages events.
``is_raw_sse_stream`` only says the chunks are unparsed bytes, and ``/v1/messages`` is not the
only endpoint that streams those: the Google ``:streamGenerateContent`` route marks its own
stream raw too. Reading its frames as Anthropic ones would refuse the response in a wire format
its client cannot parse, so the surface is decided on the event types actually present.
"""
sse_stream: Final = _joined_sse_stream(all_chunks)
if sse_stream is None:
return False
return any(event.get("type") in _ANTHROPIC_EVENT_TYPES for event in _parsed_sse_events(sse_stream))
def assemble_anthropic_sse_stream(
all_chunks: Sequence[object], *, restore_identity: bool = False
) -> ModelResponse | None:
@ -111,6 +144,27 @@ def anthropic_sse_error_frames(message: str) -> tuple[bytes, ...]:
)
def is_sse_error_stream(all_chunks: Sequence[object]) -> bool:
"""Whether the buffered stream carries nothing but error frames.
post_call guardrails run in a chain, so a hook can be handed the terminal error frames an
earlier guardrail emitted when it blocked. Those carry no message to assemble, and replacing
them would hide the refusal the client is owed. Covers both wire forms a guardrail emits: the
Anthropic ``error`` event and the chat-completions ``{"error": ...}`` payload.
"""
if not all(isinstance(chunk, (str, bytes)) for chunk in all_chunks):
# A stream mixing typed chunks with an error frame still carries content to scan, and the
# frames-only join below would drop exactly the part that has to be scanned
return False
sse_stream: Final = _joined_sse_stream(all_chunks)
if sse_stream is None:
return False
events: Final = _parsed_sse_events(sse_stream)
return len(events) > 0 and all(
event.get("type") == "error" or isinstance(event.get("error"), Mapping) for event in events
)
def anthropic_sse_chunks_from_response(assembled: ModelResponse) -> tuple[bytes, ...]:
from litellm.llms.anthropic.experimental_pass_through.adapters.transformation import (
LiteLLMAnthropicMessagesAdapter,

479
litellm/proxy/guardrails/guardrail_hooks/model_armor/model_armor.py
View file

@ -1,4 +1,5 @@
from collections.abc import AsyncGenerator, Mapping, Sequence
from enum import Enum, auto
from typing import TYPE_CHECKING, Any, Final, Literal
import httpx
@ -27,12 +28,25 @@ from litellm.llms.custom_httpx.http_handler import (
)
from litellm.llms.vertex_ai.vertex_llm_base import VertexBase
from litellm.proxy._types import UserAPIKeyAuth
from litellm.proxy.guardrails.anthropic_sse import (
anthropic_sse_chunks_from_response,
anthropic_sse_error_frames,
assemble_anthropic_sse_stream,
is_anthropic_sse_stream,
is_raw_sse_stream,
is_sse_error_stream,
)
from litellm.proxy.guardrails.guardrail_hooks.model_armor.file_scanning import (
MODEL_ARMOR_MAX_FILE_SIZE_BYTES,
plan_file_scans,
)
from litellm.types.guardrails import GuardrailEventHooks, LitellmParams
from litellm.types.llms.openai import AllMessageValues
from litellm.types.llms.openai import (
AllMessageValues,
ChatCompletionToolCallChunk,
ResponsesAPIResponse,
ResponsesAPIStreamEvents,
)
from litellm.types.llms.vertex_ai import VERTEX_CREDENTIALS_TYPES
from litellm.types.utils import (
CallTypes,
@ -42,10 +56,33 @@ from litellm.types.utils import (
ModelResponse,
ModelResponseStream,
StandardLoggingGuardrailInformation,
TextCompletionResponse,
)
GUARDRAIL_NAME: Final = "model_armor"
# Only these carry the finished output; response.created carries an empty body
_RESPONSES_TERMINAL_EVENT_TYPES: Final = frozenset({"response.completed", "response.incomplete", "response.failed"})
# Every event whose ``delta`` is model output already on its way to the client. Read off the event
# enum rather than listed, so an event added there cannot quietly fall out of the scan
_RESPONSES_DELTA_EVENT_TYPES: Final = frozenset(
event.value for event in ResponsesAPIStreamEvents if event.value.endswith(".delta")
)
# What makes two delta events part of the same field of the turn, rather than two fields that merely
# streamed next to each other
_RESPONSES_DELTA_FIELD_ATTRS: Final = ("type", "item_id", "output_index", "content_index", "summary_index")
class _StreamSurface(Enum):
"""Wire format of a buffered streaming response, which decides how it is read and how it is refused."""
CHAT_COMPLETIONS = auto()
ANTHROPIC_MESSAGES = auto()
RESPONSES = auto()
OPAQUE_SSE = auto()
class ModelArmorAPIError(Exception):
"""Model Armor API failure (non-2xx), distinct from a content-block decision so
@ -323,19 +360,9 @@ class ModelArmorGuardrail(CustomGuardrail, VertexBase):
else:
return {"modelResponseData": {"byteItem": {"byteDataType": file_type, "byteData": base64_data}}}
def _should_block_content(self, armor_response: dict, allow_sanitization: bool = False) -> bool:
def _should_block_content(self, armor_response: Mapping[str, Any], allow_sanitization: bool = False) -> bool:
"""Check if Model Armor response indicates content should be blocked, including both inspectResult and deidentifyResult."""
sanitization_result: Final = armor_response.get("sanitizationResult", {})
filter_results: Final = sanitization_result.get("filterResults", {})
# filterResults can be a dict (named keys) or a list (array of filter result dicts)
filter_result_items = []
if isinstance(filter_results, dict):
filter_result_items = list(filter_results.values())
elif isinstance(filter_results, list):
filter_result_items = filter_results
for filt in filter_result_items:
for filt in self._filter_result_items(armor_response):
# Check RAI, PI/Jailbreak, Malicious URI, CSAM, Virus scan as before
if filt.get("raiFilterResult", {}).get("matchState") == "MATCH_FOUND":
return True
@ -359,22 +386,12 @@ class ModelArmorGuardrail(CustomGuardrail, VertexBase):
# Fallback dict code removed; all cases handled above
return False
def _get_sanitized_content(self, armor_response: dict) -> str | None:
def _get_sanitized_content(self, armor_response: Mapping[str, Any]) -> str | None:
"""
Get the sanitized content from a Model Armor response, if available.
Looks for sanitized text in deidentifyResult, and falls back to root-level fields if not found.
"""
result: Final = armor_response.get("sanitizationResult", {})
filter_results: Final = result.get("filterResults", {})
# filterResults can be a dict (single filter) or a list (multiple filters)
filters: Final = (
list(filter_results.values())
if isinstance(filter_results, dict)
else filter_results
if isinstance(filter_results, list)
else []
)
filters: Final = self._filter_result_items(armor_response)
# Prefer sanitized text from deidentifyResult if present
for filter_entry in filters:
@ -398,6 +415,61 @@ class ModelArmorGuardrail(CustomGuardrail, VertexBase):
# Fallback: if Model Armor put sanitized text at the root, use it
return armor_response.get("sanitizedText") or armor_response.get("text")
@staticmethod
def _filter_result_items(armor_response: Mapping[str, Any]) -> Sequence[Any]:
"""Every filter result in a scan response.
filterResults is a dict of named filters on most templates and a list on some, so both
shapes are flattened to the same list of filter entries.
"""
filter_results: Final = armor_response.get("sanitizationResult", {}).get("filterResults", {})
if isinstance(filter_results, dict):
return list(filter_results.values())
if isinstance(filter_results, list):
return filter_results
return []
def _has_deidentify_match(self, armor_response: Mapping[str, Any]) -> bool:
"""Whether an SDP de-identify filter matched, i.e. Model Armor owes this response a redaction."""
for filter_entry in self._filter_result_items(armor_response):
sdp = filter_entry.get("sdpFilterResult")
if sdp and sdp.get("deidentifyResult", {}).get("matchState") == "MATCH_FOUND":
return True
return False
def _resolve_streaming_outcome(
self,
armor_response: Mapping[str, Any],
assembled_response: object,
content: str,
) -> tuple[bool, str | None]:
"""Whether to block the buffered stream, and the rewrite to emit when it is not blocked.
A de-identify match only reaches here unblocked because masking is on, so the redaction it
stands for has to be both resolvable and emittable. Where it is neither, the buffered
original still carries what Model Armor matched on, so this fails closed instead of
releasing it.
"""
if self._should_block_content(armor_response, allow_sanitization=self.mask_response_content):
return True, None
if not self.mask_response_content:
return False, None
sanitized_content: Final = self._get_sanitized_content(armor_response)
if not sanitized_content:
# No rewrite to apply. Harmless unless a match is outstanding, in which case applying
# nothing would hand back the very content that matched
return self._has_deidentify_match(armor_response), None
if sanitized_content == content:
return False, None
if not isinstance(assembled_response, ModelResponse):
verbose_proxy_logger.warning(
"Model Armor: sanitized content cannot be re-emitted on this streaming endpoint, "
"blocking the response instead"
)
return True, None
return False, sanitized_content
@staticmethod
def _append_armor_response(existing: object, armor_response: Mapping[str, object]) -> object:
"""Accumulate scan responses so a later text scan does not drop an earlier file scan.
@ -832,6 +904,185 @@ class ModelArmorGuardrail(CustomGuardrail, VertexBase):
return response
@staticmethod
def _is_terminal_error_stream(all_chunks: Sequence[object]) -> bool:
"""Whether the buffered stream is only the refusal an earlier guardrail in the chain emitted.
post_call guardrails are composed, so this hook can be handed the terminal error items a
preceding one produced. They carry no message to scan, and replacing them would hide the
refusal the client is owed.
"""
if all(getattr(chunk, "type", None) == "error" for chunk in all_chunks):
return True
return is_sse_error_stream(all_chunks)
@staticmethod
def _classify_stream(all_chunks: Sequence[object]) -> _StreamSurface:
"""Wire format the buffered chunks belong to."""
if is_raw_sse_stream(all_chunks):
return (
_StreamSurface.ANTHROPIC_MESSAGES if is_anthropic_sse_stream(all_chunks) else _StreamSurface.OPAQUE_SSE
)
if any(
isinstance(event_type := getattr(chunk, "type", None), str) and event_type.startswith("response.")
for chunk in all_chunks
):
return _StreamSurface.RESPONSES
return _StreamSurface.CHAT_COMPLETIONS
@staticmethod
def _final_responses_api_response(all_chunks: Sequence[object]) -> ResponsesAPIResponse | None:
"""Response body carried by a terminal ``/v1/responses`` event.
A stream cut short before it completes has to read as unassembled rather than as a clean
empty response: ``response.created`` also carries a body, but an empty one, and scanning
that would release every buffered delta unscanned.
"""
return next(
(
body
for chunk in reversed(all_chunks)
if getattr(chunk, "type", None) in _RESPONSES_TERMINAL_EVENT_TYPES
and isinstance(body := getattr(chunk, "response", None), ResponsesAPIResponse)
),
None,
)
@staticmethod
def _responses_api_response_text(response: ResponsesAPIResponse) -> str:
"""Text to scan in a Responses API response, tool-call arguments included.
Tool calls are folded in because ``get_content_from_model_response`` folds them into what
the chat surface scans, and a Responses turn can carry its whole payload in them.
"""
from litellm.llms.openai.responses.guardrail_translation.handler import (
OpenAIResponsesHandler,
)
texts: Final[list[str]] = [] # mutable-ok: the shared extractor below appends into caller-owned lists
tool_calls: Final[list[ChatCompletionToolCallChunk]] = [] # mutable-ok: the same extractor's tool-call sink
handler: Final = OpenAIResponsesHandler()
for output_idx, output_item in enumerate(response.output or ()):
handler._extract_output_text_and_images( # pyright: ignore[reportPrivateUsage] # the shared Responses output extractor; forking it would duplicate per-item parsing
output_item=output_item,
output_idx=output_idx,
texts_to_check=texts,
images_to_check=[], # mutable-ok: the extractor's images sink, unused here
task_mappings=[], # mutable-ok: the extractor's task-mapping sink, unused here
tool_calls_to_check=tool_calls,
)
return "".join((*texts, *(json.dumps(tool_call) for tool_call in tool_calls)))
def _extract_streaming_content(self, assembled_response: object) -> str:
"""Text to scan from an assembled stream, for every endpoint shape this hook serves."""
if isinstance(assembled_response, ResponsesAPIResponse):
return self._responses_api_response_text(assembled_response)
return self._extract_content_from_response(assembled_response)
@staticmethod
def _responses_delta_field(chunk: object) -> tuple[str, ...]:
"""Which field of the turn a delta event belongs to."""
return tuple(str(getattr(chunk, attr, None)) for attr in _RESPONSES_DELTA_FIELD_ATTRS)
@staticmethod
def _responses_delta_field_texts(all_chunks: Sequence[object]) -> tuple[str, ...]:
"""Text each field of a ``/v1/responses`` turn has already spelled out in its delta events.
One field's deltas are joined as they streamed, since a finding can be split across them,
and separate fields stay apart, so a reasoning summary running into the visible answer
cannot spell out a finding that neither of them carries.
"""
deltas: Final = tuple(
(ModelArmorGuardrail._responses_delta_field(chunk), delta)
for chunk in all_chunks
if getattr(chunk, "type", None) in _RESPONSES_DELTA_EVENT_TYPES
and isinstance(delta := getattr(chunk, "delta", None), str)
)
return tuple(
"".join(delta for field, delta in deltas if field == streamed_field)
for streamed_field in dict.fromkeys(field for field, _ in deltas)
)
def _streaming_content_to_scan(
self,
assembled_response: object,
all_chunks: Sequence[object],
surface: _StreamSurface,
) -> str:
"""Text to scan for a buffered stream, which is everything the client is about to receive.
A ``/v1/responses`` stream also spells out reasoning summaries and tool-call arguments in
delta events that its terminal body never repeats, so every delta field the body does not
already carry is scanned after it.
"""
content: Final = self._extract_streaming_content(assembled_response)
if surface is not _StreamSurface.RESPONSES:
return content
unscanned: Final = tuple(text for text in self._responses_delta_field_texts(all_chunks) if text not in content)
return "\n".join(part for part in (content, *unscanned) if part)
@staticmethod
def _apply_sanitized_content(assembled_response: ModelResponse, sanitized_content: str) -> None:
"""Replace every non-empty choice message with the Model Armor sanitized text."""
for choice in assembled_response.choices:
if isinstance(choice, Choices) and choice.message.content:
choice.message.content = sanitized_content
@staticmethod
def _assemble_chat_completion_stream(
all_chunks: list[object], # mutable-ok: stream_chunk_builder only accepts a mutable list
) -> ModelResponse | TextCompletionResponse | None:
"""Assemble chat-completion chunks, returning ``None`` when they cannot be assembled."""
from litellm.main import stream_chunk_builder
try:
return stream_chunk_builder(chunks=all_chunks)
except Exception as exc:
verbose_proxy_logger.warning("Model Armor: chat-completion stream assembly failed (%s)", exc)
return None
def _assemble_stream(
self, all_chunks: Sequence[object], surface: _StreamSurface
) -> ModelResponse | TextCompletionResponse | ResponsesAPIResponse | None:
"""Assemble the buffered stream into the scannable response its surface produces."""
if surface is _StreamSurface.ANTHROPIC_MESSAGES:
return assemble_anthropic_sse_stream(all_chunks, restore_identity=True)
if surface is _StreamSurface.RESPONSES:
return self._final_responses_api_response(all_chunks)
if surface is _StreamSurface.OPAQUE_SSE:
return None
return self._assemble_chat_completion_stream(list(all_chunks))
@staticmethod
def _error_payload(exc: HTTPException) -> Mapping[str, object]:
"""Error object for a terminal stream item, carrying the status the frame would otherwise lose."""
detail: Final = exc.detail if isinstance(exc.detail, Mapping) else {"message": str(exc.detail)}
error_value: Final = detail.get("error", detail)
return {
**(dict(error_value) if isinstance(error_value, Mapping) else {"message": str(error_value)}),
"code": str(exc.status_code),
}
@staticmethod
def _build_responses_error_items(exc: HTTPException) -> Sequence[object] | None:
"""Responses API error events for a failure discovered after the stream started."""
from litellm.llms.openai.responses.guardrail_translation.handler import (
OpenAIResponsesHandler,
)
return OpenAIResponsesHandler().build_stream_error_items(exc, responses_so_far=None)
def _stream_error_items(self, exc: HTTPException, *, surface: _StreamSurface) -> Sequence[object]:
"""Frame a guardrail failure as terminal stream items in this endpoint's wire format."""
payload: Final = self._error_payload(exc)
if surface is _StreamSurface.ANTHROPIC_MESSAGES:
return anthropic_sse_error_frames(str(payload.get("message", "")))
if surface is _StreamSurface.RESPONSES and (responses_items := self._build_responses_error_items(exc)):
return responses_items
# Also the fallback when a surface cannot frame its own error: create_response() reads the
# status back out of this form, so the refusal keeps its code instead of arriving as a 200
return (f"data: {json.dumps({'error': payload})}\n\n",)
async def async_post_call_streaming_iterator_hook(
self,
user_api_key_dict: UserAPIKeyAuth,
@ -841,97 +1092,125 @@ class ModelArmorGuardrail(CustomGuardrail, VertexBase):
"""Process streaming response chunks."""
from litellm.llms.base_llm.base_model_iterator import MockResponseIterator
from litellm.main import stream_chunk_builder
from litellm.proxy.common_utils.callback_utils import (
add_guardrail_to_applied_guardrails_header,
)
# Collect all chunks
all_chunks: Final[list[ModelResponseStream]] = []
all_chunks: Final[list[Any]] = []
async for chunk in response:
all_chunks.append(chunk)
if not all_chunks or self._is_terminal_error_stream(all_chunks):
for chunk in all_chunks:
yield chunk
return
surface: Final = self._classify_stream(all_chunks)
# Build complete response
assembled_response: Final = stream_chunk_builder(chunks=all_chunks)
assembled_response: Final = self._assemble_stream(all_chunks, surface)
if isinstance(assembled_response, ModelResponse):
# Extract content
content: Final = self._extract_content_from_response(assembled_response)
if assembled_response is None:
if not self.optional_params.get("fail_on_error", True):
verbose_proxy_logger.warning(
"Model Armor: streamed response could not be assembled for scanning, "
"forwarding it unscanned because fail_on_error is disabled"
)
for chunk in all_chunks:
yield chunk
return
if content:
try:
# Check with Model Armor
armor_response: Final = await self.make_model_armor_request(
content=content,
source="model_response",
request_data=request_data,
)
# Forwarding an unscannable stream would silently disable the guardrail, so fail closed
add_guardrail_to_applied_guardrails_header(request_data=request_data, guardrail_name=self.guardrail_name)
for error_item in self._stream_error_items(
HTTPException(
status_code=500,
detail=f"{self.guardrail_name}: streamed response could not be assembled for scanning, blocking it",
),
surface=surface,
):
yield error_item
return
# Attach Model Armor response & status to this request's metadata to avoid race conditions
if isinstance(request_data, dict):
_, metadata = get_or_create_metadata_bucket(request_data)
metadata["_model_armor_response"] = self._build_logging_response(armor_response)
metadata["_model_armor_status"] = (
"blocked" if self._should_block_content(armor_response) else "success"
)
# Extract content
content: Final = self._streaming_content_to_scan(
assembled_response=assembled_response, all_chunks=all_chunks, surface=surface
)
# Add guardrail to applied_guardrails BEFORE potential blocking
# This ensures guardrail is recorded even when it blocks the request
from litellm.proxy.common_utils.callback_utils import (
add_guardrail_to_applied_guardrails_header,
)
if not content:
verbose_proxy_logger.debug("Model Armor: No text content in streaming response, skipping guardrail")
for chunk in all_chunks:
yield chunk
return
add_guardrail_to_applied_guardrails_header(
request_data=request_data, guardrail_name=self.guardrail_name
)
try:
# Check with Model Armor
armor_response: Final = await self.make_model_armor_request(
content=content,
source="model_response",
request_data=request_data,
)
# Check if blocked
if self._should_block_content(armor_response):
raise HTTPException(
status_code=400,
detail=self._build_block_error_detail(
"Streaming response blocked by Model Armor",
armor_response,
),
)
# Decide the outcome before recording it. Mirrors the non-streaming sibling: with
# masking on, a de-identify match is a redaction to apply rather than a refusal, but
# that only holds while the redaction can actually be delivered
blocked, sanitized_content = self._resolve_streaming_outcome(
armor_response=armor_response,
assembled_response=assembled_response,
content=content,
)
# Apply sanitization if enabled
if self.mask_response_content:
sanitized_content: Final = self._get_sanitized_content(armor_response)
if sanitized_content and sanitized_content != content:
# Update assembled response
for choice in assembled_response.choices:
if isinstance(choice, Choices):
if choice.message.content:
choice.message.content = sanitized_content
# Attach Model Armor response & status to this request's metadata to avoid race conditions
if isinstance(request_data, dict):
_, metadata = get_or_create_metadata_bucket(request_data)
metadata["_model_armor_response"] = self._build_logging_response(armor_response)
metadata["_model_armor_status"] = "blocked" if blocked else "success"
# Return sanitized stream
mock_response: Final = MockResponseIterator(model_response=assembled_response)
async for chunk in mock_response:
yield chunk
return
# Add guardrail to applied_guardrails BEFORE potential blocking
# This ensures guardrail is recorded even when it blocks the request
add_guardrail_to_applied_guardrails_header(request_data=request_data, guardrail_name=self.guardrail_name)
except ModelArmorAPIError as e:
if self.optional_params.get("fail_on_error", True):
error_obj = {"message": e.detail, "code": "500"}
yield f"data: {json.dumps({'error': error_obj})}\n\n"
return
except HTTPException as e:
# Yield error as SSE event so create_response() detects it and
# returns a proper JSON error response with the correct status code.
# (Raising from a generator hits create_response's generic except → 500.)
detail: Final = e.detail if isinstance(e.detail, dict) else {"message": str(e.detail)}
error_value: Final = detail.get("error", detail)
if isinstance(error_value, dict):
error_obj = dict(error_value)
else:
error_obj = {"message": str(error_value)}
error_obj["code"] = str(e.status_code)
yield f"data: {json.dumps({'error': error_obj})}\n\n"
if blocked:
raise HTTPException(
status_code=400,
detail=self._build_block_error_detail(
"Streaming response blocked by Model Armor",
armor_response,
),
)
if sanitized_content is not None and isinstance(assembled_response, ModelResponse):
self._apply_sanitized_content(assembled_response, sanitized_content)
# Return sanitized stream
if surface is _StreamSurface.ANTHROPIC_MESSAGES:
for sse_chunk in anthropic_sse_chunks_from_response(assembled_response):
yield sse_chunk
return
except Exception as e:
verbose_proxy_logger.error("Model Armor streaming error: %s", str(e), exc_info=True)
if self.optional_params.get("fail_on_error", True):
raise
else:
verbose_proxy_logger.debug("Model Armor: No text content in streaming response, skipping guardrail")
mock_response: Final = MockResponseIterator(model_response=assembled_response)
async for chunk in mock_response:
yield chunk
return
except ModelArmorAPIError as e:
if self.optional_params.get("fail_on_error", True):
for error_item in self._stream_error_items(
HTTPException(status_code=500, detail=e.detail), surface=surface
):
yield error_item
return
except HTTPException as e:
# Yield the error as a terminal stream item so create_response() detects it and returns
# a proper JSON error response with the correct status code. Raising from a generator
# instead hits create_response's generic except and becomes a 500.
for error_item in self._stream_error_items(e, surface=surface):
yield error_item
return
except Exception as e:
verbose_proxy_logger.error("Model Armor streaming error: %s", str(e), exc_info=True)
if self.optional_params.get("fail_on_error", True):
raise
# Return original chunks if no sanitization needed
for chunk in all_chunks:

6
litellm/proxy/management_endpoints/sso/saml_sso.py
View file

@ -443,7 +443,9 @@ class SAMLAuthHandler:
last_name: Final = SAMLAuthHandler._attribute_value(
attributes, "SAML_ATTRIBUTE_LAST_NAME", _LAST_NAME_ATTRIBUTE_CANDIDATES
)
role_value = SAMLAuthHandler._attribute_value(attributes, "SAML_ATTRIBUTE_ROLE", _ROLE_ATTRIBUTE_CANDIDATES)
role_values: Final = SAMLAuthHandler._attribute_values(
attributes, "SAML_ATTRIBUTE_ROLE", _ROLE_ATTRIBUTE_CANDIDATES
)
team_ids: Final = SAMLAuthHandler._attribute_values(
attributes, "SAML_ATTRIBUTE_TEAM_IDS", _TEAM_IDS_ATTRIBUTE_CANDIDATES
)
@ -464,7 +466,7 @@ class SAMLAuthHandler:
picture=None,
provider="saml",
team_ids=team_ids,
user_role=get_litellm_user_role(role_value) if role_value else None,
user_role=get_litellm_user_role(role_values),
)
except ValidationError as e:
raise HTTPException(

62
litellm/proxy/management_endpoints/types.py
View file

@ -4,12 +4,44 @@ Types for the management endpoints
Might include fastapi/proxy requirements.txt related imports
"""
from collections.abc import Iterable, Sequence
from typing import Any, Final, cast
from fastapi_sso.sso.base import OpenID
from litellm.proxy._types import LitellmUserRoles
# Ordered highest to lowest privilege
LITELLM_USER_ROLE_HIERARCHY: Final = (
LitellmUserRoles.PROXY_ADMIN,
LitellmUserRoles.PROXY_ADMIN_VIEW_ONLY,
LitellmUserRoles.INTERNAL_USER,
LitellmUserRoles.INTERNAL_USER_VIEW_ONLY,
)
def highest_privilege_role(roles: Iterable[LitellmUserRoles]) -> LitellmUserRoles | None:
"""
Pick the highest privilege role out of the roles an IdP asserted for one user.
IdPs do not guarantee ordering within a multi-valued role claim, so a user holding
several roles resolves to the most privileged one rather than whichever came first.
Roles the hierarchy does not rank (org_admin, team, customer) resolve by name to stay
deterministic.
Args:
roles: The roles resolved from the claim
Returns:
The highest privilege role, or None if `roles` is empty
"""
resolved: Final = frozenset(roles)
if not resolved:
return None
ranked: Final = next((role for role in LITELLM_USER_ROLE_HIERARCHY if role in resolved), None)
return ranked if ranked is not None else min(resolved, key=lambda role: role.value)
def is_valid_litellm_user_role(role_str: str) -> bool:
"""
@ -28,12 +60,22 @@ def is_valid_litellm_user_role(role_str: str) -> bool:
return False
def get_litellm_user_role(role_str) -> LitellmUserRoles | None:
def _role_from_claim_value(role_str: object) -> LitellmUserRoles | None:
if not isinstance(role_str, str):
return None
# Use _value2member_map_ for O(1) lookup, case-insensitive
result: Final = LitellmUserRoles._value2member_map_.get(role_str.lower())
return cast(LitellmUserRoles | None, result)
def get_litellm_user_role(role_str: object) -> LitellmUserRoles | None:
"""
Convert a string (or list of strings) to a LitellmUserRoles enum if valid (case-insensitive).
Handles list inputs since some SSO providers (e.g., Keycloak) return roles
as arrays like ["proxy_admin"] instead of plain strings.
as arrays like ["proxy_admin"] instead of plain strings. A claim carrying several
roles resolves to the highest privilege one, so a user does not lose access just
because the IdP listed a weaker role first.
Args:
role_str: String or list to convert (e.g., "proxy_admin", ["proxy_admin"])
@ -41,16 +83,12 @@ def get_litellm_user_role(role_str) -> LitellmUserRoles | None:
Returns:
LitellmUserRoles enum if valid, None otherwise
"""
try:
if isinstance(role_str, list):
if len(role_str) == 0:
return None
role_str = role_str[0]
# Use _value2member_map_ for O(1) lookup, case-insensitive
result: Final = LitellmUserRoles._value2member_map_.get(role_str.lower())
return cast(LitellmUserRoles | None, result)
except Exception:
return None
if isinstance(role_str, (list, tuple)):
entries: Final = cast(Sequence[object], role_str) # cast-ok: isinstance narrows the claim, not its elements
return highest_privilege_role(
role for role in (_role_from_claim_value(entry) for entry in entries) if role is not None
)
return _role_from_claim_value(role_str)
class CustomOpenID(OpenID):

19
litellm/proxy/management_endpoints/ui_sso.py
View file

@ -112,6 +112,7 @@ from litellm.proxy.management_endpoints.sso_helper_utils import (
)
from litellm.proxy.management_endpoints.team_endpoints import new_team, team_member_add
from litellm.proxy.management_endpoints.types import (
LITELLM_USER_ROLE_HIERARCHY,
CustomOpenID,
get_litellm_user_role,
is_valid_litellm_user_role,
@ -809,15 +810,6 @@ def normalize_email(email: str | None) -> str | None:
return email.lower() if isinstance(email, str) else email
# Ordered highest to lowest privilege
LITELLM_USER_ROLE_HIERARCHY: Final = (
LitellmUserRoles.PROXY_ADMIN,
LitellmUserRoles.PROXY_ADMIN_VIEW_ONLY,
LitellmUserRoles.INTERNAL_USER,
LitellmUserRoles.INTERNAL_USER_VIEW_ONLY,
)
def determine_role_from_groups(
user_groups: list[str],
role_mappings: "RoleMappings",
@ -4312,14 +4304,7 @@ class MicrosoftSSOHandler:
listed first. Roles the hierarchy does not rank (org_admin, team, customer)
resolve by name to stay deterministic
"""
resolved: Final = frozenset(
role for role in (get_litellm_user_role(role_str) for role_str in app_roles or ()) if role is not None
)
if not resolved:
return None
ranked: Final = next((role for role in LITELLM_USER_ROLE_HIERARCHY if role in resolved), None)
return ranked if ranked is not None else min(resolved, key=lambda role: role.value)
return get_litellm_user_role(tuple(app_roles or ()))
@staticmethod
def get_app_roles_from_id_token(id_token: str | None) -> list[str]:

26
litellm/proxy/public_endpoints/autorouter_presets.json
View file

@ -1,4 +1,26 @@
{
"1m_context": {
"label": "1M Context",
"description": "Routes across models with 1M-token context windows: Luna for simple queries, Terra for medium, Opus 5 for complex, Opus 5 at high thinking for reasoning.",
"complexity_router_config": {
"tiers": {
"SIMPLE": ["gpt-5.6-luna"],
"MEDIUM": ["gpt-5.6-terra"],
"COMPLEX": ["claude-opus-5"],
"REASONING": ["claude-opus-5"]
},
"tier_model_configs": {
"REASONING": [{ "model_name": "claude-opus-5", "litellm_params": { "reasoning_effort": "high" } }]
},
"classifier_type": "heuristic_v2",
"escalation_keywords": ["LITELLM ESCALATE"],
"classification_mode": "every_request",
"session_affinity": false,
"modality_routing": false,
"modality_pin_override": false,
"deployment_affinity": true
}
},
"anthropic_family": {
"label": "Anthropic Family",
"description": "Routes across the Claude model family: Haiku for simple queries, Sonnet for medium, Opus for complex, Opus at high thinking for reasoning.",
@ -17,6 +39,7 @@
"classification_mode": "every_request",
"session_affinity": false,
"modality_routing": false,
"modality_pin_override": false,
"deployment_affinity": true
}
},
@ -35,6 +58,7 @@
"classification_mode": "every_request",
"session_affinity": false,
"modality_routing": false,
"modality_pin_override": false,
"deployment_affinity": true
}
},
@ -63,6 +87,7 @@
"classification_mode": "every_request",
"session_affinity": false,
"modality_routing": false,
"modality_pin_override": false,
"deployment_affinity": true
}
},
@ -84,6 +109,7 @@
"classification_mode": "every_request",
"session_affinity": false,
"modality_routing": false,
"modality_pin_override": false,
"deployment_affinity": true
}
}

12
litellm/router_strategy/complexity_router/README.md
View file

@ -187,6 +187,9 @@ model_list:
# Replace a routed model that cannot take image input (default: false)
modality_routing: true
# Let that replacement also override a kept session pin, for image turns only (default: false)
modality_pin_override: true
```
## Usage
@ -227,9 +230,16 @@ vision model sits below the decided tier gets the 400 and an actionable message
A same-tier re-pick keeps the decision's cause and adds `modality:image` to `signals`; a tier
change or default takeover records `cause: modality_escalation` with the displaced placement
(`modality_escalated_from:<TIER>` or `modality_displaced_default_model`). Escalations are never
pinned by session affinity, and a KEPT session pin bypasses the gate entirely: a session pinned
pinned by session affinity, and by default a KEPT session pin bypasses the gate: a session pinned
to a text-only model keeps it even when an image arrives.
Add `modality_pin_override: true` to lift that last exemption. The image turn is then re-placed
the same way every other decision is, and records `cause: modality_pin_override` whether or not
the tier moved, since the model left the pin either way. The pin itself is untouched: the session
affinity write happens upstream of the gate and stores the session's own model, so the next text
turn replays the original pin and the override is never pinned in its place. It does nothing
unless `modality_routing` is also on.
### Heuristic-first chaining
`classifier_type: heuristic_first` runs the local scorer on every request and only calls the LLM

23
litellm/router_strategy/complexity_router/complexity_router.py
View file

@ -751,7 +751,8 @@ def _decision_is_pinnable(decision: StandardLoggingRoutingDecision | None) -> bo
A modality escalation is transient the same way: it describes what this one call carries (an
image), not what the session's traffic looks like, and pinning it would hold every following
text turn on the vision-capable model the image forced.
text turn on the vision-capable model the image forced. A modality pin override is the same
fact on a session that already holds a pin, so it must not overwrite the pin it displaced.
"""
return decision is None or (
decision.get("cause")
@ -760,6 +761,7 @@ def _decision_is_pinnable(decision: StandardLoggingRoutingDecision | None) -> bo
"plan_mode",
"housekeeping",
"modality_escalation",
"modality_pin_override",
)
and not decision.get("context_escalated")
)
@ -2393,8 +2395,11 @@ class ComplexityRouter(CustomLogger):
"""Replace a routed model that cannot accept this request's image input.
The single modality owner, applied to the decided response at the hook's exits so every
routing path is covered uniformly. A KEPT session pin is exempt by design (its cause);
replacement picks and every other path are just responses. The re-placement walks
routing path is covered uniformly. A KEPT session pin is exempt by design (its cause)
unless modality_pin_override is set, in which case the image turn is re-placed and reported
as modality_pin_override while the stored pin, written upstream from the session's own
model, is left for the next text turn; replacement picks and every other path are just
responses. The re-placement walks
UPWARD-ONLY from the decision's tier (so a plan-mode floor can never be undercut), picks
through `_pick_model_for_tier` so routing plugins still apply, then falls to
default_model (never on plugin routers, and never on a plan-floored decision, since
@ -2407,7 +2412,11 @@ class ComplexityRouter(CustomLogger):
not self.config.modality_routing
or not resolved_messages
or response.model is None
or (decision is not None and decision.get("cause") == "session_affinity_pin")
or (
decision is not None
and decision.get("cause") == "session_affinity_pin"
and not self.config.modality_pin_override
)
or not request_contains_image_content(resolved_messages)
or self._model_accepts_image_input(response.model)
):
@ -2449,6 +2458,10 @@ class ComplexityRouter(CustomLogger):
self._restamp_adaptive_choice(request_kwargs, response.model, new_model)
same_tier: Final = capable is not None and decided == capable
base_cause: Final = (decision.get("cause") if decision is not None else None) or "default_fallback"
# Reaching here on a kept pin means modality_pin_override is on, since the guard above
# returns otherwise. The model moved off the pin even on a same-tier repick, so reporting
# the pin's own cause would claim the session's model served a request it did not.
displaced_pin: Final = base_cause == "session_affinity_pin"
displaced_default: Final = decided is None and response.model == self.config.default_model
markers: Final = (
"modality:image",
@ -2458,7 +2471,7 @@ class ComplexityRouter(CustomLogger):
old_signals: Final = tuple(decision.get("signals") or ()) if decision is not None else ()
new_decision: Final = self._build_routing_decision(
routed_model=new_model,
cause=base_cause if same_tier else "modality_escalation",
cause="modality_pin_override" if displaced_pin else (base_cause if same_tier else "modality_escalation"),
tier=new_tier,
score=decision.get("score") if decision is not None else None,
signals=(*old_signals, *markers),

15
litellm/router_strategy/complexity_router/config.py
View file

@ -883,7 +883,20 @@ class ComplexityRouterConfig(BaseModel):
"a routed model explicitly declared supports_vision false (deployment model_info "
"or the model cost map; unmapped names stay routable) is replaced by the nearest "
"HIGHER tier holding a capable model, then default_model, else a clear 400. A kept "
"session-affinity pin still wins even when an image arrives."
"session-affinity pin still wins even when an image arrives, unless "
"modality_pin_override is also enabled."
),
)
modality_pin_override: bool = Field(
default=False,
description=(
"Let modality_routing replace a kept session-affinity pin on the turns that carry an "
"image. Without this, a session pinned to a text-only model fails every image turn with "
"a provider 400, since the pin is exempt from the modality gate. When enabled, such a "
"turn routes to a capable model for that request only and the stored pin is left "
"untouched, so the next text turn replays the session's own model; the override is "
"reported as cause modality_pin_override and is never itself pinned. Inert unless "
"modality_routing is also enabled."
),
)

1
litellm/types/integrations/datadog_llm_obs.py
View file

@ -86,6 +86,7 @@ class LLMMetrics(TypedDict, total=False):
cache_read_input_tokens: ReadOnly[float]
cache_write_input_tokens: ReadOnly[float]
non_cached_input_tokens: ReadOnly[float]
reasoning_output_tokens: ReadOnly[float]
class LLMObsPayload(TypedDict, total=False):

4
litellm/types/utils.py
View file

@ -2879,6 +2879,10 @@ RoutingDecisionCause = Literal[
# routed model does not accept image input, so the nearest higher capable tier or
# default_model served instead. The displaced placement rides in signals.
"modality_escalation",
# modality_pin_override replaced a KEPT session-affinity pin for this request only: the turn
# carries an image the pinned model cannot accept. The stored pin is untouched, so the next
# text turn replays it. Distinct from "modality_escalation", which never displaces a pin.
"modality_pin_override",
"session_affinity_pin",
"session_affinity_escalation",
# classification_mode 'user_turn': the request is an agent loop's continuation turn (no new

1
litellm/vector_stores/main.py
View file

@ -482,7 +482,6 @@ def search(
timeout=timeout or request_timeout,
_is_async=_is_async,
client=kwargs.get("client"),
router=router,
)
return response

16
ruff-strict-budget.json
View file

@ -1,21 +1,21 @@
{
"ANN001": {
"limit": 2957
"limit": 2984
},
"ANN002": {
"limit": 71
},
"ANN003": {
"limit": 806
"limit": 809
},
"ANN201": {
"limit": 1981
"limit": 2000
},
"ANN202": {
"limit": 831
"limit": 835
},
"ANN204": {
"limit": 683
"limit": 693
},
"ANN205": {
"limit": 112
@ -24,7 +24,7 @@
"limit": 133
},
"ANN401": {
"limit": 119
"limit": 304
},
"ASYNC230": {
"limit": 11
@ -57,7 +57,7 @@
"limit": 3
},
"BLE001": {
"limit": 2917
"limit": 2916
},
"C401": {
"limit": 8
@ -231,7 +231,7 @@
"limit": 5
},
"TID251": {
"limit": 1034
"limit": 1071
},
"TRY002": {
"limit": 524

3
tests/logging_callback_tests/test_bedrock_knowledgebase_hook.py
View file

@ -370,12 +370,13 @@ async def test_bedrock_kb_request_body_has_transformed_filters(
custom_llm_provider,
litellm_params,
logging_obj,
embedding_executor=None,
extra_headers=None,
extra_body=None,
timeout=None,
client=None,
_is_async=False,
router: "litellm.Router | None" = None,
embedding_executor=None,
):
litellm_params_dict = (
litellm_params.model_dump(exclude_none=False)

13
tests/router_unit_tests/test_router_embedding_integration.py
View file

@ -187,7 +187,7 @@ class TestRouterEmbeddingIntegration:
assert _sent(store_route, 1) == ("Bearer store-key", "text-embedding-3-large", ["async query"])
@pytest.mark.asyncio
async def test_router_executor_rejects_unserved_models_without_explicit_config(
async def test_router_executor_embeds_unserved_models_through_the_sdk(
self, respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch
):
monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
@ -198,12 +198,13 @@ class TestRouterEmbeddingIntegration:
metadata={"user_api_key_team_id": "team-a"},
)
with pytest.raises(litellm.BadRequestError):
executor.embed("openai/text-embedding-3-large", "sync query", {})
with pytest.raises(litellm.BadRequestError):
await executor.aembed("openai/text-embedding-3-large", "async query", {})
sync_response = executor.embed("text-embedding-3-large", "sync query", {})
async_response = await executor.aembed("text-embedding-3-large", "async query", {})
assert openai_route.call_count == 0
assert sync_response.data[0]["embedding"] == QUERY_VECTOR
assert async_response.data[0]["embedding"] == QUERY_VECTOR
assert _sent(openai_route, 0) == ("Bearer env-key", "text-embedding-3-large", ["sync query"])
assert _sent(openai_route, 1) == ("Bearer env-key", "text-embedding-3-large", ["async query"])
def test_router_executor_routes_deployment_model_names_through_the_router(
self, respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch

9
tests/rust-python-harness/README.md
View file

@ -8,14 +8,17 @@ The matrix always has these SDK columns:
- `messages / amessages`
- `responses / aresponses`
- `count_tokens`
- `chat_completions / acompletion`
- `transcription / atranscription`
The harness has three deliberately broad test-strategy folders:
The harness has four deliberately broad test-strategy folders:
| Strategy | Folder |
| --- | --- |
| Public SDK parity over generated and recorded inputs | [`e2e_fuzz_tests/`](e2e_fuzz_tests/) |
| Focused tests of Rust-owned behavior | [`unit_tests_rust/`](unit_tests_rust/) |
| Isolated transform and Python-to-Rust helper coverage | [`validate_sub_methods/`](validate_sub_methods/) |
| Already-existing live-API SDK tests | [`existing_e2e_test_sdk/`](existing_e2e_test_sdk/) |
## Run it
@ -112,7 +115,7 @@ The initial end-to-end entries deliberately show `◐`: the repository has Rust
## Attach parity tests
Each of the three folders contains a concise `README.md` and a `strategy.json`. Add a pytest file or node ID to the appropriate SDK function's `selectors` list:
Each of the four folders contains a concise `README.md` and a `strategy.json`. Add a pytest file or node ID to the appropriate SDK function's `selectors` list:
```json
{
@ -123,7 +126,7 @@ Each of the three folders contains a concise `README.md` and a `strategy.json`.
}
```
Selectors use the same syntax as pytest. A file selector aggregates every test in the file; a node selector can target one test or parametrized family. The runner deduplicates selectors, so one test may intentionally prove more than one cell without executing twice.
Selectors use the same syntax as pytest. A file selector aggregates every test in the file; a node selector can target one test or parametrized family; a selector ending in `/` aggregates every test in that folder, recursively. The runner deduplicates selectors, so one test may intentionally prove more than one cell without executing twice.
Use these coverage values:

49
tests/rust-python-harness/cli.py
View file

@ -6,9 +6,10 @@ from collections.abc import Sequence
from pathlib import Path
from .catalog import load_catalog
from .models import HarnessCase, Strategy
from .models import SDK_FUNCTIONS, HarnessCase, Strategy
from .runner import run_pytest
from .ui import make_dashboard
from .strategies.unit_tests.mapping_validator import FunctionReport, build_function_report
REPO_ROOT = Path(__file__).resolve().parents[2]
COVERAGE_ROOT = REPO_ROOT / "target" / "rust-python-harness"
@ -40,9 +41,17 @@ def _parser() -> argparse.ArgumentParser:
action="append",
default=[],
dest="sdk_functions",
choices=("ocr", "messages", "responses", "count_tokens"),
choices=SDK_FUNCTIONS,
help="run only this SDK function",
)
parser.add_argument(
"--validate-ledger",
action="store_true",
help=(
"report Python<->Rust test-parity ledger gaps and drift instead of "
"running the dashboard; narrow with --function"
),
)
parser.add_argument(
"--plain",
action="store_true",
@ -100,7 +109,7 @@ def _interactive_filters(strategies: Sequence[Strategy]) -> tuple[set[str], set[
)
sdk_functions = _pick_values(
"SDK functions",
[(name, name) for name in ("ocr", "messages", "responses", "count_tokens")],
[(name, name) for name in SDK_FUNCTIONS],
)
return strategy_ids, sdk_functions
@ -131,6 +140,38 @@ def _print_catalog(strategies: Sequence[Strategy]) -> None:
print(f" {case.sdk_function:12} {case.coverage.value:14} {selectors}")
def _print_function_report(report: FunctionReport) -> None:
print(f"\n{report.sdk_function}")
if report.ledger is None or report.audit is None:
print(" no ledger yet")
return
ledger, audit = report.ledger, report.audit
print(
f" {ledger.mapped_count}/{ledger.total_count} python tests mapped to rust "
f"({ledger.percentage}%)"
)
print(f" {len(ledger.rust_only_tests)} rust-only tests with no python counterpart")
if audit.is_clean:
print(" ledger is in sync with the live test files")
return
for label, items in (
("ledger references a python test that no longer exists", audit.missing_python_tests),
("python test exists but is not tracked in the ledger", audit.stale_python_tests),
("ledger references a rust test that no longer exists", audit.missing_rust_tests),
("rust test exists but is not tracked in the ledger", audit.stale_rust_tests),
):
for item in items:
print(f" {label}: {item}")
def _validate_ledger(sdk_functions: set[str]) -> int:
functions = sdk_functions or set(SDK_FUNCTIONS)
reports = tuple(build_function_report(function) for function in sorted(functions))
for report in reports:
_print_function_report(report)
return 0 if all(report.is_clean for report in reports) else 1
def main(argv: Sequence[str] | None = None) -> int:
args = _parser().parse_args(argv)
if args.coverage and importlib.util.find_spec("pytest_cov") is None:
@ -138,6 +179,8 @@ def main(argv: Sequence[str] | None = None) -> int:
"--coverage requires the project's pytest-cov dependency; run with "
"`poetry run python -m tests.rust-python-harness --coverage`"
)
if args.validate_ledger:
return _validate_ledger(set(args.sdk_functions))
strategies = load_catalog()
if args.list:
_print_catalog(strategies)

4
tests/rust-python-harness/e2e_fuzz_tests/strategy.json
View file

@ -7,6 +7,8 @@
"ocr": {"coverage": "partial", "selectors": ["tests/test_litellm/ocr/test_rust_bridge.py"], "note": "Bridge coverage exists; frozen-oracle fuzz parity is still being added."},
"messages": {"coverage": "partial", "selectors": ["tests/test_litellm/anthropic_interface/test_rust_bridge_messages.py"], "note": "Bridge coverage exists; frozen-oracle fuzz parity is still being added."},
"responses": {"coverage": "partial", "selectors": ["tests/test_litellm/responses/test_rust_bridge_websocket.py"], "note": "Covers the websocket bridge; full responses parity is still being added."},
"count_tokens": {"coverage": "planned", "selectors": [], "note": "No Rust count_tokens parity test is present yet."}
"count_tokens": {"coverage": "planned", "selectors": [], "note": "No Rust count_tokens parity test is present yet."},
"chat_completions": {"coverage": "partial", "selectors": ["tests/test_litellm/rust_bridge/test_chat_completions.py"], "note": "Bridge coverage exists; frozen-oracle fuzz parity is still being added."},
"transcription": {"coverage": "partial", "selectors": ["tests/test_litellm/test_audio_transcription_rust_bridge.py"], "note": "Bridge coverage exists; frozen-oracle fuzz parity is still being added."}
}
}

3
tests/rust-python-harness/existing_e2e_test_sdk/README.md Normal file
View file

@ -0,0 +1,3 @@
# Existing e2e SDK tests
Wires already-existing live-API SDK tests into the matrix instead of writing new parity tests. Selectors point at real test files and folders, such as `tests/ocr_tests/`, rather than individual node IDs, so future tests added to those folders are picked up automatically.

14
tests/rust-python-harness/existing_e2e_test_sdk/strategy.json Normal file
View file

@ -0,0 +1,14 @@
{
"order": 40,
"id": "existing_e2e_test_sdk",
"label": "Existing e2e SDK tests",
"description": "Wire already-existing live-API SDK tests into the matrix instead of writing new parity tests.",
"functions": {
"ocr": {"coverage": "partial", "selectors": ["tests/ocr_tests/"], "note": "Existing live OCR provider tests; not yet a frozen Rust/Python oracle comparison."},
"messages": {"coverage": "planned", "selectors": []},
"responses": {"coverage": "planned", "selectors": []},
"count_tokens": {"coverage": "planned", "selectors": []},
"chat_completions": {"coverage": "partial", "selectors": ["tests/llm_translation/test_anthropic_completion.py", "tests/llm_translation/test_bedrock_completion.py"], "note": "Existing live chat completion tests for providers with confirmed Rust bridge regressions."},
"transcription": {"coverage": "partial", "selectors": ["tests/audio_tests/test_whisper.py"], "note": "Existing live Whisper transcription test."}
}
}

2
tests/rust-python-harness/models.py
View file

@ -33,7 +33,7 @@ class ConfidenceLevel(str, Enum):
LOW = "LOW"
SDK_FUNCTIONS = ("ocr", "messages", "responses", "count_tokens")
SDK_FUNCTIONS = ("ocr", "messages", "responses", "count_tokens", "chat_completions", "transcription")
@dataclass(frozen=True)

2
tests/rust-python-harness/runner.py
View file

@ -15,6 +15,8 @@ UpdateCallback = Callable[[HarnessRun], None]
def selector_matches_node(selector: str, nodeid: str) -> bool:
normalized_selector = selector.replace("\\", "/")
normalized_nodeid = nodeid.replace("\\", "/")
if normalized_selector.endswith("/"):
return normalized_nodeid.startswith(normalized_selector)
if "::" in normalized_selector:
return normalized_nodeid == normalized_selector or normalized_nodeid.startswith(
f"{normalized_selector}["

0
tests/rust-python-harness/shared/__init__.py Normal file
View file

0
tests/rust-python-harness/shared/parity/__init__.py Normal file
View file

136
tests/rust-python-harness/shared/parity/ledger.py Normal file
View file

@ -0,0 +1,136 @@
from __future__ import annotations
import json
from dataclasses import dataclass
from pathlib import Path
from typing import Any
@dataclass(frozen=True, slots=True)
class LedgerEntry:
python_file: str
python_test: str
status: str
rust_file: str
rust_test: str
justification: str
reason: str
@dataclass(frozen=True, slots=True)
class RustOnlyEntry:
rust_file: str
rust_test: str
reason: str
@dataclass(frozen=True, slots=True)
class TestLedger:
sdk_function: str
python_scope: tuple[str, ...]
rust_scope: tuple[str, ...]
entries: tuple[LedgerEntry, ...]
rust_only_tests: tuple[RustOnlyEntry, ...]
@property
def mapped_count(self) -> int:
return sum(1 for entry in self.entries if entry.status == "mapped")
@property
def total_count(self) -> int:
return len(self.entries)
@property
def percentage(self) -> float:
if self.total_count == 0:
return 0.0
return round(100.0 * self.mapped_count / self.total_count, 1)
def _require_string(value: Any, field: str, source: Path) -> str:
if not isinstance(value, str) or not value.strip():
raise ValueError(f"{source}: {field} must be a non-empty string")
return value
def _require_string_list(value: Any, field: str, source: Path) -> tuple[str, ...]:
if not isinstance(value, list) or not all(isinstance(item, str) and item for item in value):
raise ValueError(f"{source}: {field} must be a list of non-empty strings")
return tuple(value)
def _load_entry(data: Any, index: int, source: Path) -> LedgerEntry:
if not isinstance(data, dict):
raise ValueError(f"{source}: entries[{index}] must be an object")
python_file = _require_string(data.get("python_file"), f"entries[{index}].python_file", source)
python_test = _require_string(data.get("python_test"), f"entries[{index}].python_test", source)
status = data.get("status")
if status not in ("mapped", "unmapped"):
raise ValueError(f"{source}: entries[{index}].status must be 'mapped' or 'unmapped'")
if status == "mapped":
rust_file = _require_string(data.get("rust_file"), f"entries[{index}].rust_file", source)
rust_test = _require_string(data.get("rust_test"), f"entries[{index}].rust_test", source)
justification = _require_string(
data.get("justification"), f"entries[{index}].justification", source
)
return LedgerEntry(
python_file=python_file,
python_test=python_test,
status=status,
rust_file=rust_file,
rust_test=rust_test,
justification=justification,
reason="",
)
reason = _require_string(data.get("reason"), f"entries[{index}].reason", source)
return LedgerEntry(
python_file=python_file,
python_test=python_test,
status=status,
rust_file="",
rust_test="",
justification="",
reason=reason,
)
def _load_rust_only_entry(data: Any, index: int, source: Path) -> RustOnlyEntry:
if not isinstance(data, dict):
raise ValueError(f"{source}: rust_only_tests[{index}] must be an object")
return RustOnlyEntry(
rust_file=_require_string(data.get("rust_file"), f"rust_only_tests[{index}].rust_file", source),
rust_test=_require_string(data.get("rust_test"), f"rust_only_tests[{index}].rust_test", source),
reason=_require_string(data.get("reason"), f"rust_only_tests[{index}].reason", source),
)
def load_ledger(path: Path) -> TestLedger:
with path.open(encoding="utf-8") as stream:
data = json.load(stream)
sdk_function = _require_string(data.get("sdk_function"), "sdk_function", path)
python_scope = _require_string_list(data.get("python_scope"), "python_scope", path)
rust_scope = _require_string_list(data.get("rust_scope"), "rust_scope", path)
entries_data = data.get("entries")
if not isinstance(entries_data, list):
raise ValueError(f"{path}: entries must be a list")
entries = tuple(
_load_entry(entry, index, path) for index, entry in enumerate(entries_data)
)
rust_only_data = data.get("rust_only_tests")
if not isinstance(rust_only_data, list):
raise ValueError(f"{path}: rust_only_tests must be a list")
rust_only_tests = tuple(
_load_rust_only_entry(entry, index, path) for index, entry in enumerate(rust_only_data)
)
return TestLedger(
sdk_function=sdk_function,
python_scope=python_scope,
rust_scope=rust_scope,
entries=entries,
rust_only_tests=rust_only_tests,
)

0
tests/rust-python-harness/strategies/__init__.py Normal file
View file

0
tests/rust-python-harness/strategies/unit_tests/__init__.py Normal file
View file

1084
tests/rust-python-harness/strategies/unit_tests/ledgers/ocr/ocr_test_ledger.json Normal file
View file

File diff suppressed because it is too large Load diff

101
tests/rust-python-harness/strategies/unit_tests/mapping_validator.py Normal file
View file

@ -0,0 +1,101 @@
from __future__ import annotations
from dataclasses import dataclass
from pathlib import Path
from ...shared.parity.ledger import TestLedger, load_ledger
from .python_runner import enumerate_python_tests
from .rust_runner import enumerate_rust_tests
REPO_ROOT = Path(__file__).resolve().parents[4]
LEDGER_ROOT = Path(__file__).parent / "ledgers"
def ledger_path_for(sdk_function: str) -> Path:
return LEDGER_ROOT / sdk_function / f"{sdk_function}_test_ledger.json"
@dataclass(frozen=True, slots=True)
class AuditReport:
missing_python_tests: tuple[str, ...]
stale_python_tests: tuple[str, ...]
missing_rust_tests: tuple[str, ...]
stale_rust_tests: tuple[str, ...]
@property
def is_clean(self) -> bool:
return not (
self.missing_python_tests
or self.stale_python_tests
or self.missing_rust_tests
or self.stale_rust_tests
)
def _ledger_python_tests_by_file(ledger: TestLedger) -> dict[str, set[str]]:
grouping: dict[str, set[str]] = {path: set() for path in ledger.python_scope}
for entry in ledger.entries:
grouping.setdefault(entry.python_file, set()).add(entry.python_test)
return grouping
def _ledger_rust_tests_by_file(ledger: TestLedger) -> dict[str, set[str]]:
grouping: dict[str, set[str]] = {path: set() for path in ledger.rust_scope}
for entry in ledger.entries:
if entry.status == "mapped":
grouping.setdefault(entry.rust_file, set()).add(entry.rust_test)
for rust_only in ledger.rust_only_tests:
grouping.setdefault(rust_only.rust_file, set()).add(rust_only.rust_test)
return grouping
def audit_ledger(ledger: TestLedger, repo_root: Path = REPO_ROOT) -> AuditReport:
missing_python: list[str] = []
stale_python: list[str] = []
for python_file, ledger_tests in _ledger_python_tests_by_file(ledger).items():
actual_tests = enumerate_python_tests(repo_root, python_file)
for missing in sorted(ledger_tests - actual_tests):
missing_python.append(f"{python_file}:{missing}")
for stale in sorted(actual_tests - ledger_tests):
stale_python.append(f"{python_file}:{stale}")
missing_rust: list[str] = []
stale_rust: list[str] = []
for rust_file, ledger_tests in _ledger_rust_tests_by_file(ledger).items():
actual_tests = enumerate_rust_tests(repo_root, rust_file)
for missing in sorted(ledger_tests - actual_tests):
missing_rust.append(f"{rust_file}:{missing}")
for stale in sorted(actual_tests - ledger_tests):
stale_rust.append(f"{rust_file}:{stale}")
return AuditReport(
missing_python_tests=tuple(missing_python),
stale_python_tests=tuple(stale_python),
missing_rust_tests=tuple(missing_rust),
stale_rust_tests=tuple(stale_rust),
)
@dataclass(frozen=True, slots=True)
class FunctionReport:
sdk_function: str
ledger: TestLedger | None
audit: AuditReport | None
@property
def has_ledger(self) -> bool:
return self.ledger is not None
@property
def is_clean(self) -> bool:
return self.audit is None or self.audit.is_clean
def build_function_report(sdk_function: str, repo_root: Path = REPO_ROOT) -> FunctionReport:
path = ledger_path_for(sdk_function)
if not path.exists():
return FunctionReport(sdk_function=sdk_function, ledger=None, audit=None)
ledger = load_ledger(path)
return FunctionReport(
sdk_function=sdk_function, ledger=ledger, audit=audit_ledger(ledger, repo_root)
)

22
tests/rust-python-harness/strategies/unit_tests/python_runner.py Normal file
View file

@ -0,0 +1,22 @@
from __future__ import annotations
import ast
from pathlib import Path
def enumerate_python_tests(repo_root: Path, relative_path: str) -> frozenset[str]:
source = (repo_root / relative_path).read_text(encoding="utf-8")
tree = ast.parse(source, filename=relative_path)
module_level: list[str] = []
for node in ast.iter_child_nodes(tree):
if isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)) and node.name.startswith("test_"):
module_level.append(node.name)
elif isinstance(node, ast.ClassDef):
for child in ast.iter_child_nodes(node):
if isinstance(child, (ast.FunctionDef, ast.AsyncFunctionDef)) and child.name.startswith(
"test_"
):
module_level.append(f"{node.name}::{child.name}")
return frozenset(module_level)

13
tests/rust-python-harness/strategies/unit_tests/rust_runner.py Normal file
View file

@ -0,0 +1,13 @@
from __future__ import annotations
import re
from pathlib import Path
_RUST_TEST_PATTERN = re.compile(
r"#\[(?:test|tokio::test)\][^\n]*\n(?:[^\n]*\n)*?\s*(?:async\s+)?fn\s+(\w+)\s*\("
)
def enumerate_rust_tests(repo_root: Path, relative_path: str) -> frozenset[str]:
source = (repo_root / relative_path).read_text(encoding="utf-8")
return frozenset(match.group(1) for match in _RUST_TEST_PATTERN.finditer(source))

2
tests/rust-python-harness/ui.py
View file

@ -119,7 +119,7 @@ class RichDashboard(AbstractContextManager["RichDashboard"]):
table = Table(box=box.ROUNDED, expand=True, title="Strategy × SDK function")
table.add_column("Strategy", ratio=3)
for label in ("ocr/aocr", "messages", "responses", "count_tokens"):
for label in SDK_FUNCTIONS:
table.add_column(label, justify="center", ratio=1)
for strategy in self.strategies:
cells = []

4
tests/rust-python-harness/unit_tests_rust/strategy.json
View file

@ -7,6 +7,8 @@
"ocr": {"coverage": "planned", "selectors": []},
"messages": {"coverage": "planned", "selectors": []},
"responses": {"coverage": "planned", "selectors": []},
"count_tokens": {"coverage": "planned", "selectors": []}
"count_tokens": {"coverage": "planned", "selectors": []},
"chat_completions": {"coverage": "planned", "selectors": []},
"transcription": {"coverage": "planned", "selectors": []}
}
}

4
tests/rust-python-harness/validate_sub_methods/strategy.json
View file

@ -7,6 +7,8 @@
"ocr": {"coverage": "planned", "selectors": []},
"messages": {"coverage": "planned", "selectors": []},
"responses": {"coverage": "planned", "selectors": []},
"count_tokens": {"coverage": "planned", "selectors": []}
"count_tokens": {"coverage": "planned", "selectors": []},
"chat_completions": {"coverage": "planned", "selectors": []},
"transcription": {"coverage": "planned", "selectors": []}
}
}

241
tests/test_litellm/integrations/azure_storage/test_azure_storage.py
View file

@ -1,10 +1,15 @@
import asyncio
import sys
import threading
from unittest.mock import AsyncMock, MagicMock, patch
import pytest
from litellm.integrations.azure_storage.azure_storage import AzureBlobStorageLogger
from litellm.integrations.azure_storage.azure_storage import (
AzureBlobStorageLogger,
_cached_credential_chain_token_provider,
)
from litellm.types.secret_managers.get_azure_ad_token_provider import AzureCredentialType
from litellm.types.utils import StandardLoggingPayload
@ -25,6 +30,26 @@ def mock_gov_env_vars(mock_env_vars, monkeypatch):
monkeypatch.setenv("AZURE_STORAGE_ENDPOINT_SUFFIX", "core.usgovcloudapi.net")
@pytest.fixture
def workload_identity_env_vars(monkeypatch):
monkeypatch.setenv("AZURE_STORAGE_ACCOUNT_NAME", "test-account")
monkeypatch.setenv("AZURE_STORAGE_FILE_SYSTEM", "test-container")
for unset in (
"AZURE_STORAGE_TENANT_ID",
"AZURE_STORAGE_CLIENT_ID",
"AZURE_STORAGE_CLIENT_SECRET",
"AZURE_STORAGE_ACCOUNT_KEY",
"AZURE_STORAGE_ENDPOINT_SUFFIX",
"AZURE_CLIENT_SECRET",
"AZURE_CREDENTIAL",
"AZURE_SCOPE",
):
monkeypatch.delenv(unset, raising=False)
monkeypatch.setenv("AZURE_CLIENT_ID", "workload-identity-client-id")
monkeypatch.setenv("AZURE_TENANT_ID", "workload-identity-tenant-id")
monkeypatch.setenv("AZURE_FEDERATED_TOKEN_FILE", "/var/run/secrets/azure/tokens/azure-identity-token")
@pytest.mark.asyncio
async def test_async_upload_payload_to_azure_blob_storage(mock_env_vars):
"""
@ -32,17 +57,12 @@ async def test_async_upload_payload_to_azure_blob_storage(mock_env_vars):
a payload to Azure Blob Storage using the 3-step process (create, append, flush).
"""
with (
patch(
"litellm.integrations.azure_storage.azure_storage.get_async_httpx_client"
) as mock_get_client,
patch(
"litellm.llms.azure.common_utils.get_azure_ad_token_from_entra_id"
) as mock_get_token,
patch("litellm.integrations.azure_storage.azure_storage.get_async_httpx_client") as mock_get_client,
patch("litellm.integrations.azure_storage.azure_storage.get_azure_ad_token_from_entra_id") as mock_get_token,
):
# Create mock HTTP client
mock_http_client = AsyncMock()
mock_response = AsyncMock()
mock_response.raise_for_status = AsyncMock()
mock_response = MagicMock()
mock_http_client.put.return_value = mock_response
mock_http_client.patch.return_value = mock_response
mock_get_client.return_value = mock_http_client
@ -79,9 +99,7 @@ async def test_async_upload_payload_to_azure_blob_storage(mock_env_vars):
put_call_args = mock_http_client.put.call_args
assert put_call_args[0][0] == f"{expected_base_url}?resource=file"
assert put_call_args[1]["headers"]["x-ms-version"] is not None
assert (
put_call_args[1]["headers"]["Authorization"] == "Bearer mock-azure-ad-token"
)
assert put_call_args[1]["headers"]["Authorization"] == "Bearer mock-azure-ad-token"
# Step 2: Append data
assert mock_http_client.patch.call_count == 2 # Called for append and flush
@ -89,9 +107,7 @@ async def test_async_upload_payload_to_azure_blob_storage(mock_env_vars):
assert append_call[0][0] == f"{expected_base_url}?action=append&position=0"
assert append_call[1]["headers"]["x-ms-version"] is not None
assert append_call[1]["headers"]["Content-Type"] == "application/json"
assert (
append_call[1]["headers"]["Authorization"] == "Bearer mock-azure-ad-token"
)
assert append_call[1]["headers"]["Authorization"] == "Bearer mock-azure-ad-token"
assert "test-log-id-123" in append_call[1]["data"]
# Step 3: Flush data
@ -110,9 +126,7 @@ async def test_async_upload_payload_uses_configured_endpoint_suffix(mock_gov_env
AZURE_STORAGE_ENDPOINT_SUFFIX must reach the Entra-ID REST upload path so a
sovereign-cloud account is addressed instead of the commercial dfs host.
"""
with patch(
"litellm.integrations.azure_storage.azure_storage.get_async_httpx_client"
) as mock_get_client:
with patch("litellm.integrations.azure_storage.azure_storage.get_async_httpx_client") as mock_get_client:
mock_http_client = AsyncMock()
mock_response = MagicMock()
mock_http_client.put.return_value = mock_response
@ -127,17 +141,10 @@ async def test_async_upload_payload_uses_configured_endpoint_suffix(mock_gov_env
await logger.async_upload_payload_to_azure_blob_storage(test_payload)
expected_base_url = (
"https://test-account.dfs.core.usgovcloudapi.net/test-container/gov-log-id.json"
)
expected_base_url = "https://test-account.dfs.core.usgovcloudapi.net/test-container/gov-log-id.json"
assert mock_http_client.put.call_args[0][0] == f"{expected_base_url}?resource=file"
assert (
mock_http_client.patch.call_args_list[0][0][0]
== f"{expected_base_url}?action=append&position=0"
)
assert mock_http_client.patch.call_args_list[1][0][0].startswith(
f"{expected_base_url}?action=flush"
)
assert mock_http_client.patch.call_args_list[0][0][0] == f"{expected_base_url}?action=append&position=0"
assert mock_http_client.patch.call_args_list[1][0][0].startswith(f"{expected_base_url}?action=flush")
@pytest.mark.asyncio
@ -148,9 +155,7 @@ async def test_service_client_uses_configured_endpoint_suffix(mock_gov_env_vars)
"""
fake_aio_module = MagicMock()
with patch.dict(
sys.modules, {"azure.storage.filedatalake.aio": fake_aio_module}
):
with patch.dict(sys.modules, {"azure.storage.filedatalake.aio": fake_aio_module}):
logger = AzureBlobStorageLogger()
await logger.get_service_client()
@ -160,14 +165,180 @@ async def test_service_client_uses_configured_endpoint_suffix(mock_gov_env_vars)
)
@pytest.mark.asyncio
async def test_upload_authenticates_through_the_credential_chain_under_workload_identity(
workload_identity_env_vars,
):
build_provider = MagicMock(return_value=lambda: "workload-identity-token")
with patch( # test-quality-ok: REST client is created inside the method; assert emitted request headers
"litellm.integrations.azure_storage.azure_storage.get_async_httpx_client"
) as mock_get_client:
mock_http_client = AsyncMock()
mock_http_client.put.return_value = MagicMock()
mock_http_client.patch.return_value = MagicMock()
mock_get_client.return_value = mock_http_client
logger = AzureBlobStorageLogger(build_credential_chain_token_provider=build_provider)
await logger.async_upload_payload_to_azure_blob_storage({"id": "wif-log-id"})
build_provider.assert_called_once_with()
assert logger.azure_auth_token == "workload-identity-token"
sent_headers = [mock_http_client.put.call_args[1]["headers"]] + [
call[1]["headers"] for call in mock_http_client.patch.call_args_list
]
assert len(sent_headers) == 3
assert all(headers["Authorization"] == "Bearer workload-identity-token" for headers in sent_headers)
def test_default_chain_provider_is_storage_scoped_and_built_once_per_process():
_cached_credential_chain_token_provider.cache_clear()
with (
patch( # test-quality-ok: assert the default factory's fixed scope and credential type without constructing Azure SDK credentials
"litellm.integrations.azure_storage.azure_storage.get_azure_ad_token_provider",
return_value=lambda: "chain-token",
) as mock_builder
):
first = _cached_credential_chain_token_provider()
second = _cached_credential_chain_token_provider()
_cached_credential_chain_token_provider.cache_clear()
assert first is second
assert first() == "chain-token"
mock_builder.assert_called_once_with(
azure_scope="https://storage.azure.com/.default",
azure_credential=AzureCredentialType.DefaultAzureCredential,
)
@pytest.mark.asyncio
async def test_chain_tokens_are_read_from_the_provider_on_every_refresh(
workload_identity_env_vars,
):
provider = MagicMock(side_effect=["chain-token-1", "chain-token-2"])
logger = AzureBlobStorageLogger(build_credential_chain_token_provider=MagicMock(return_value=provider))
await logger.set_valid_azure_ad_token()
first_token = logger.azure_auth_token
await logger.set_valid_azure_ad_token()
assert first_token == "chain-token-1"
assert logger.azure_auth_token == "chain-token-2"
assert provider.call_count == 2
@pytest.mark.asyncio
async def test_chain_token_read_yields_to_the_event_loop(workload_identity_env_vars):
"""
The chain walk is blocking I/O (IMDS probe, CLI subprocess), so reading the provider
inline would stall every request on the worker. Prove other coroutines run during the read.
"""
loop_was_free = threading.Event()
def provider() -> str:
if not loop_was_free.wait(timeout=5):
raise TimeoutError("the event loop never ran the observer while the token was being read")
return "chain-token"
async def observer():
loop_was_free.set()
logger = AzureBlobStorageLogger(build_credential_chain_token_provider=MagicMock(return_value=provider))
observer_task = asyncio.create_task(observer())
await logger.set_valid_azure_ad_token()
await observer_task
assert logger.azure_auth_token == "chain-token"
@pytest.mark.asyncio
async def test_empty_string_service_principal_vars_still_use_the_credential_chain(
workload_identity_env_vars, monkeypatch
):
for name in ("AZURE_STORAGE_TENANT_ID", "AZURE_STORAGE_CLIENT_ID", "AZURE_STORAGE_CLIENT_SECRET"):
monkeypatch.setenv(name, "")
logger = AzureBlobStorageLogger(
build_credential_chain_token_provider=MagicMock(return_value=lambda: "workload-identity-token")
)
await logger.set_valid_azure_ad_token()
assert logger.azure_auth_token == "workload-identity-token"
@pytest.mark.asyncio
async def test_client_secret_auth_still_uses_the_storage_scoped_service_principal(mock_env_vars):
build_provider = MagicMock()
with (
patch( # test-quality-ok: assert the storage scope passed to the shared token factory without making an external auth call
"litellm.integrations.azure_storage.azure_storage.get_azure_ad_token_from_entra_id",
return_value=lambda: "client-secret-token",
) as mock_entra_id
):
logger = AzureBlobStorageLogger(build_credential_chain_token_provider=build_provider)
await logger.set_valid_azure_ad_token()
assert logger.azure_auth_token == "client-secret-token"
build_provider.assert_not_called()
assert mock_entra_id.call_args.kwargs == {
"tenant_id": "test-tenant-id",
"client_id": "test-client-id",
"client_secret": "test-client-secret",
"scope": "https://storage.azure.com/.default",
}
@pytest.mark.parametrize(
"missing_var",
["AZURE_STORAGE_TENANT_ID", "AZURE_STORAGE_CLIENT_ID", "AZURE_STORAGE_CLIENT_SECRET"],
)
@pytest.mark.asyncio
async def test_partially_configured_service_principal_still_names_the_missing_variable(
mock_env_vars, monkeypatch, missing_var
):
monkeypatch.delenv(missing_var)
build_provider = MagicMock()
logger = AzureBlobStorageLogger(build_credential_chain_token_provider=build_provider)
with pytest.raises(ValueError, match=f"Missing required environment variable: {missing_var}"):
await logger.set_valid_azure_ad_token()
build_provider.assert_not_called()
@pytest.mark.asyncio
async def test_account_key_auth_never_requests_a_token(workload_identity_env_vars, monkeypatch):
monkeypatch.setenv("AZURE_STORAGE_ACCOUNT_KEY", "dGVzdC1rZXk=")
file_client = MagicMock()
file_client.create_file = AsyncMock()
file_client.append_data = AsyncMock()
file_client.flush_data = AsyncMock()
directory_client = MagicMock()
directory_client.exists = AsyncMock(return_value=True)
directory_client.get_file_client = MagicMock(return_value=file_client)
file_system_client = MagicMock()
file_system_client.get_directory_client = MagicMock(return_value=directory_client)
service_client = MagicMock()
service_client.get_file_system_client = MagicMock(return_value=file_system_client)
fake_aio_module = MagicMock()
fake_aio_module.DataLakeServiceClient = MagicMock(return_value=service_client)
build_provider = MagicMock()
with patch.dict(sys.modules, {"azure.storage.filedatalake.aio": fake_aio_module}):
logger = AzureBlobStorageLogger(build_credential_chain_token_provider=build_provider)
await logger.async_upload_payload_to_azure_blob_storage({"id": "account-key-log-id"})
build_provider.assert_not_called()
assert logger.azure_auth_token is None
file_client.flush_data.assert_awaited_once()
assert fake_aio_module.DataLakeServiceClient.call_args.kwargs["credential"] == "dGVzdC1rZXk="
@pytest.mark.asyncio
async def test_service_client_defaults_to_commercial_endpoint(mock_env_vars):
"""Unset AZURE_STORAGE_ENDPOINT_SUFFIX keeps the pre-existing commercial host"""
fake_aio_module = MagicMock()
with patch.dict(
sys.modules, {"azure.storage.filedatalake.aio": fake_aio_module}
):
with patch.dict(sys.modules, {"azure.storage.filedatalake.aio": fake_aio_module}):
logger = AzureBlobStorageLogger()
await logger.get_service_client()

390
tests/test_litellm/integrations/datadog/test_datadog_llm_obs.py
View file

@ -17,6 +17,7 @@ from unittest.mock import patch
import pytest
import litellm
from litellm.integrations.datadog.datadog_llm_obs import DataDogLLMObsLogger
from litellm.litellm_core_utils.safe_json_dumps import safe_dumps
@ -55,15 +56,22 @@ def build_payload(
response_message: dict[str, Any] | None = None,
usage_object: dict[str, Any] | None = None,
model_parameters: dict[str, Any] | None = None,
metadata: dict[str, Any] | None = None,
model_group: str | None = None,
prompt_tokens: int = 4447,
) -> dict[str, Any]:
standard_logging_metadata: dict[str, Any] = {
**(metadata or {}),
**({"usage_object": usage_object} if usage_object is not None else {}),
}
return {
"standard_logging_object": {
"call_type": "acompletion",
"messages": [{"role": "user", "content": "hi"}] if messages is NOT_GIVEN else messages,
"response": {"choices": [{"message": response_message or {"role": "assistant", "content": "hello"}}]},
"model_parameters": model_parameters or {},
"metadata": {"usage_object": usage_object} if usage_object is not None else {},
"metadata": standard_logging_metadata,
"model_group": model_group,
"prompt_tokens": prompt_tokens,
"completion_tokens": 507,
"total_tokens": prompt_tokens + 507,
@ -244,6 +252,43 @@ def test_no_cache_keys_when_the_provider_reports_no_caching(logger: DataDogLLMOb
assert "non_cached_input_tokens" not in payload["metrics"]
def test_reasoning_tokens_are_reported_as_span_metrics(logger: DataDogLLMObsLogger) -> None:
payload = build(logger, usage_object={"completion_tokens_details": {"reasoning_tokens": 128}})
assert payload["metrics"]["reasoning_output_tokens"] == 128.0
def test_responses_reasoning_tokens_are_reported_as_span_metrics(logger: DataDogLLMObsLogger) -> None:
payload = build(logger, usage_object={"output_tokens_details": {"reasoning_tokens": 64}})
assert payload["metrics"]["reasoning_output_tokens"] == 64.0
def test_zero_reasoning_tokens_are_not_reported(logger: DataDogLLMObsLogger) -> None:
payload = build(logger, usage_object={"completion_tokens_details": {"reasoning_tokens": 0}})
assert "reasoning_output_tokens" not in payload["metrics"]
def test_reasoning_tokens_come_from_the_spelling_that_reports_them(logger: DataDogLLMObsLogger) -> None:
"""A chat-details mapping without the count must not shadow the responses spelling that has it."""
payload = build(
logger,
usage_object={
"completion_tokens_details": {"accepted_prediction_tokens": 5},
"output_tokens_details": {"reasoning_tokens": 64},
},
)
assert payload["metrics"]["reasoning_output_tokens"] == 64.0
def test_boolean_reasoning_tokens_are_not_a_count(logger: DataDogLLMObsLogger) -> None:
payload = build(logger, usage_object={"completion_tokens_details": {"reasoning_tokens": True}})
assert "reasoning_output_tokens" not in payload["metrics"]
def test_tool_definitions_are_sent_on_meta(logger: DataDogLLMObsLogger) -> None:
payload = build(logger, model_parameters={"tools": [TOOL_DEFINITION]})
@ -256,6 +301,340 @@ def test_tool_definitions_are_sent_on_meta(logger: DataDogLLMObsLogger) -> None:
]
def test_cost_tags_include_present_categories_and_dimensions(logger: DataDogLLMObsLogger) -> None:
payload = build(
logger,
metadata={
"user_api_key_user_id": "User 42",
"user_api_key_alias": "Primary Key",
"team_alias": "Platform",
"routing_decision": {
"tier": "premium",
"cause": "high_complexity",
"score": 0.91,
"escalated": True,
"signals": ["long prompt"],
"routed_model": "openai/gpt-5",
},
},
model_group="premium-models",
)
assert payload["tags"][-8:] == [
"team:platform",
"user:user_42",
"key_alias:primary_key",
"model_group:premium-models",
"router_tier:premium",
"router_cause:high_complexity",
"router_escalated:true",
"routed_model:openai/gpt-5",
]
assert payload["meta"]["metadata"]["_dd"]["cost_tags"] == [
"team",
"user",
"key_alias",
"model_group",
"router_tier",
"router_cause",
"router_escalated",
"routed_model",
]
def test_missing_cost_tag_values_are_not_declared(logger: DataDogLLMObsLogger) -> None:
payload = build(logger, metadata={"team_alias": "Platform"})
assert payload["meta"]["metadata"]["_dd"]["cost_tags"] == ["team"]
assert not any(tag.startswith(("user:", "key_alias:", "model_group:")) for tag in payload["tags"])
def test_values_that_normalize_to_empty_are_not_tagged_or_declared(logger: DataDogLLMObsLogger) -> None:
payload = build(logger, metadata={"user_api_key_user_id": "___", "user_api_key_alias": "!!!"}, model_group="tier-1")
assert not any(tag in ("user:", "key_alias:") for tag in payload["tags"])
assert payload["meta"]["metadata"]["_dd"]["cost_tags"] == ["model_group"]
def test_a_valueless_tag_from_the_shared_builder_is_not_declared(logger: DataDogLLMObsLogger) -> None:
"""The team tag comes from the shared builder, which emits it bare when the alias normalizes away."""
payload = build(logger, metadata={"team_alias": "!!!"}, model_group="tier-1")
assert "team:" in payload["tags"]
assert payload["meta"]["metadata"]["_dd"]["cost_tags"] == ["model_group"]
def test_router_fields_are_flattened(logger: DataDogLLMObsLogger) -> None:
payload = build(
logger,
metadata={
"routing_decision": {
"tier": "premium",
"cause": "high_complexity",
"score": 0.91,
"escalated": True,
"signals": ["secret prompt text"],
"routed_model": "openai/gpt-5",
}
},
model_group="premium-models",
)
assert payload["meta"]["metadata"]["router_tier"] == "premium"
assert payload["meta"]["metadata"]["router_cause"] == "high_complexity"
assert payload["meta"]["metadata"]["router_score"] == 0.91
assert payload["meta"]["metadata"]["router_escalated"] is True
assert payload["meta"]["metadata"]["router_signals"] == ["secret prompt text"]
assert payload["meta"]["metadata"]["routed_model"] == "openai/gpt-5"
def test_a_context_escalated_route_reports_as_escalated(logger: DataDogLLMObsLogger) -> None:
"""The router records a size-driven escalation under its own key, and it is still an escalation."""
payload = build(logger, metadata={"routing_decision": {"tier": "premium", "context_escalated": True}})
assert payload["meta"]["metadata"]["router_escalated"] is True
assert "router_escalated:true" in payload["tags"]
def test_a_routed_request_that_did_not_escalate_reports_false(logger: DataDogLLMObsLogger) -> None:
"""Without this the escalation dimension is absent on ordinary traffic, so nothing can group by it."""
payload = build(logger, metadata={"routing_decision": {"tier": "simple", "cause": "heuristic_scorer"}})
assert payload["meta"]["metadata"]["router_escalated"] is False
assert "router_escalated:false" in payload["tags"]
assert "router_escalated" in payload["meta"]["metadata"]["_dd"]["cost_tags"]
def test_a_request_that_never_reached_a_router_has_no_router_fields(logger: DataDogLLMObsLogger) -> None:
payload = build(logger, model_group="premium-models")
assert "router_escalated" not in payload["meta"]["metadata"]
assert not any(tag.startswith("router_") for tag in payload["tags"])
def test_redacted_payload_keeps_metrics_and_removes_sensitive_fields(logger: DataDogLLMObsLogger) -> None:
payload = build_payload(
messages=[{"role": "user", "content": "secret prompt"}],
response_message={"role": "assistant", "content": "secret response"},
usage_object={"prompt_tokens_details": {"cached_tokens": 128}},
metadata={"routing_decision": {"tier": "premium", "signals": ["secret prompt text"]}},
model_parameters={"tools": [TOOL_DEFINITION]},
)
with patch.dict(os.environ, {"DD_API_KEY": "k", "DD_SITE": "us5.datadoghq.com"}, clear=True):
with patch("asyncio.create_task"):
redacted_logger = DataDogLLMObsLogger(turn_off_message_logging=True)
redacted_payload = redacted_logger.redact_standard_logging_payload_from_model_call_details(payload)
result = json.loads(
safe_dumps(
redacted_logger.create_llm_obs_payload(
redacted_payload, datetime(2026, 9, 1, 12, 0, 0), datetime(2026, 9, 1, 12, 0, 2)
)
)
)
assert result["meta"]["input"]["messages"][0]["content"] == "redacted-by-litellm"
assert result["meta"]["output"]["messages"][0]["content"] == "redacted-by-litellm"
assert result["meta"]["metadata"]["router_tier"] == "premium"
assert "router_signals" not in result["meta"]["metadata"]
assert "routing_decision" not in result["meta"]["metadata"]
assert "tool_definitions" not in result["meta"]
assert result["metrics"]["cache_read_input_tokens"] == 128.0
assert result["metrics"]["total_cost"] == 0.02
def test_redaction_drops_the_routing_record_carried_in_metadata(logger: DataDogLLMObsLogger) -> None:
"""The whole routing record rides along in metadata, so dropping the flat copy alone leaks the prompt."""
with patch.dict(os.environ, {"DD_API_KEY": "k", "DD_SITE": "us5.datadoghq.com"}, clear=True):
with patch("asyncio.create_task"):
redacted_logger = DataDogLLMObsLogger(turn_off_message_logging=True)
result = json.loads(
safe_dumps(
redacted_logger.create_llm_obs_payload(
build_payload(
metadata={
"routing_decision": {
"tier": "premium",
"cause": "keyword_rule",
"signals": ["secret prompt text"],
"matched_keyword": "secret keyword",
"escalation_keyword": "secret escalation",
}
}
),
datetime(2026, 9, 1, 12, 0, 0),
datetime(2026, 9, 1, 12, 0, 2),
)
)
)
assert "routing_decision" not in result["meta"]["metadata"]
assert result["meta"]["metadata"]["router_tier"] == "premium"
assert result["meta"]["metadata"]["router_cause"] == "keyword_rule"
assert "secret" not in safe_dumps(result["meta"]["metadata"])
def test_a_failure_span_redacts_its_messages(logger: DataDogLLMObsLogger) -> None:
"""The redaction hook only runs on success, so the failure span has to redact for itself."""
failed = build_payload(messages=[{"role": "user", "content": "secret prompt"}])
failed["standard_logging_object"]["status"] = "failure"
failed["standard_logging_object"]["response"] = None
failed["standard_logging_object"]["error_information"] = {"error_message": "boom", "error_class": "BadRequestError"}
with patch.dict(os.environ, {"DD_API_KEY": "k", "DD_SITE": "us5.datadoghq.com"}, clear=True):
with patch("asyncio.create_task"):
redacted_logger = DataDogLLMObsLogger(turn_off_message_logging=True)
result = json.loads(
safe_dumps(
redacted_logger.create_llm_obs_payload(
failed, datetime(2026, 9, 1, 12, 0, 0), datetime(2026, 9, 1, 12, 0, 2)
)
)
)
assert result["meta"]["input"]["messages"] == [{"role": "user", "content": "redacted-by-litellm"}]
assert result["meta"]["output"]["messages"] == []
assert result["status"] == "error"
def test_excluding_messages_from_the_logging_payload_still_ships_the_span(logger: DataDogLLMObsLogger) -> None:
"""`standard_logging_payload_excluded_fields` deletes the key, and a span with no prompt is still a span."""
payload = build_payload()
del payload["standard_logging_object"]["messages"]
span = json.loads(
safe_dumps(
logger.create_llm_obs_payload(payload, datetime(2026, 9, 1, 12, 0, 0), datetime(2026, 9, 1, 12, 0, 2))
)
)
assert span["meta"]["input"]["messages"] == []
assert span["metrics"]["total_cost"] == 0.02
def test_an_explicit_redaction_setting_survives_the_global_params(logger: DataDogLLMObsLogger) -> None:
"""Global params carry defaults for keys the operator never set, and those must not win."""
with patch.dict(os.environ, {"DD_API_KEY": "k", "DD_SITE": "us5.datadoghq.com"}, clear=True):
with patch("asyncio.create_task"):
with patch.object( # test-quality-ok: the ctor reads this module global with no injection seam
litellm, "datadog_llm_observability_params", {}
):
configured_logger = DataDogLLMObsLogger(
turn_off_message_logging=True
) # test-quality-ok: verifies ctor setting
assert configured_logger.turn_off_message_logging is True
def _redacting_logger(
**kwargs: Any,
) -> DataDogLLMObsLogger: # test-quality-ok: shared test factory accepts init variants
with patch.dict(os.environ, {"DD_API_KEY": "k", "DD_SITE": "us5.datadoghq.com"}, clear=True):
with patch("asyncio.create_task"):
return DataDogLLMObsLogger(**kwargs)
def _span_json(logger_under_test: DataDogLLMObsLogger, payload: dict[str, Any]) -> dict[str, Any]:
span = logger_under_test.create_llm_obs_payload(
payload, datetime(2026, 9, 1, 12, 0, 0), datetime(2026, 9, 1, 12, 0, 2)
)
return json.loads(safe_dumps(span))
def test_redaction_keeps_the_conversation_shape_without_its_content() -> None:
"""Roles and message count survive so the trace stays legible; contents and tool payloads do not."""
result = _span_json(
_redacting_logger(turn_off_message_logging=True),
build_payload(
messages=[
{"role": "user", "content": "secret prompt"},
{"role": "assistant", "content": None, "tool_calls": [ASSISTANT_TOOL_CALL]},
],
response_message={"role": "assistant", "content": "secret response"},
),
)
assert result["meta"]["input"]["messages"] == [
{"role": "user", "content": "redacted-by-litellm"},
{"role": "assistant", "content": "redacted-by-litellm"},
]
assert result["meta"]["output"]["messages"] == [{"role": "assistant", "content": "redacted-by-litellm"}]
def test_redaction_drops_unrecognized_and_malformed_message_roles() -> None:
"""Caller-controlled role values must not bypass redaction or crash span creation."""
result = _span_json(
_redacting_logger(turn_off_message_logging=True),
build_payload(
messages=[
{"role": "SECRET-39402", "content": "hello"},
{"role": ["SECRET-39402"], "content": "hello"},
{"role": {"secret": "SECRET-39402"}, "content": "hello"},
{"role": "agent", "content": "hello"},
]
),
)
assert result["meta"]["input"]["messages"] == [
{"role": "", "content": "redacted-by-litellm"},
{"role": "", "content": "redacted-by-litellm"},
{"role": "", "content": "redacted-by-litellm"},
{"role": "agent", "content": "redacted-by-litellm"},
]
assert "SECRET-39402" not in safe_dumps(result)
def test_the_deprecated_message_logging_flag_engages_the_same_redaction() -> None:
"""The platform redacts for `message_logging is not True`, so this callback's own gate must agree."""
result = _span_json(
_redacting_logger(message_logging=False),
build_payload(
messages=[{"role": "user", "content": "secret prompt"}],
model_parameters={"tools": [TOOL_DEFINITION]},
metadata={"routing_decision": {"tier": "premium", "signals": ["secret prompt text"]}},
),
)
assert result["meta"]["input"]["messages"] == [{"role": "user", "content": "redacted-by-litellm"}]
assert "tool_definitions" not in result["meta"]
assert "routing_decision" not in result["meta"]["metadata"]
def test_a_truthy_redaction_setting_redacts_like_the_shared_hook() -> None:
"""The shared hook redacts on truthiness, so a config-provided string must not half-redact the span."""
result = _span_json(
_redacting_logger(turn_off_message_logging="yes"),
build_payload(messages=[{"role": "user", "content": "secret prompt"}]),
)
assert result["meta"]["input"]["messages"] == [{"role": "user", "content": "redacted-by-litellm"}]
def test_redaction_drops_every_prompt_carrying_metadata_record(logger: DataDogLLMObsLogger) -> None:
"""Tool arguments, retrieved text, and the guardrail's copy of the request ride in metadata records too."""
sensitive_metadata: dict[str, Any] = {
"requester_metadata": {"note": "secret prompt text"},
"prompt_management_metadata": {"prompt_id": "p1", "prompt_variables": {"topic": "secret"}},
"mcp_tool_call_metadata": {"name": "search", "arguments": {"query": "secret"}},
"vector_store_request_metadata": [{"query": "secret"}],
}
def sensitive_payload() -> dict[str, Any]:
payload = build_payload(metadata=sensitive_metadata)
payload["standard_logging_object"]["guardrail_information"] = [
{"guardrail_name": "g", "guardrail_request": {"messages": [{"content": "secret prompt"}]}}
]
return payload
redacted = _span_json(_redacting_logger(turn_off_message_logging=True), sensitive_payload())
unredacted = _span_json(logger, sensitive_payload())
assert "secret" not in safe_dumps(redacted["meta"]["metadata"])
for record in sensitive_metadata:
assert record not in redacted["meta"]["metadata"]
assert record in unredacted["meta"]["metadata"]
assert redacted["meta"]["metadata"]["guardrail_information"] is None
assert unredacted["meta"]["metadata"]["guardrail_information"] is not None
def test_tool_definitions_accept_the_bare_anthropic_shape(logger: DataDogLLMObsLogger) -> None:
"""The Anthropic surface declares tools unwrapped, with input_schema instead of parameters."""
payload = build(
@ -272,6 +651,15 @@ def test_meta_omits_tool_definitions_when_no_tools_were_offered(logger: DataDogL
assert "tool_definitions" not in build(logger)["meta"]
def test_a_ddtrace_integer_parent_id_is_forwarded_as_its_string(logger: DataDogLLMObsLogger) -> None:
"""ddtrace hands span ids as ints; dropping them detaches the span from its APM trace."""
kwargs = build_payload()
kwargs["litellm_params"]["metadata"]["parent_id"] = 8675309
start = datetime(2026, 9, 1, 12, 0, 0)
span = json.loads(safe_dumps(logger.create_llm_obs_payload(kwargs, start, start + timedelta(seconds=2))))
assert span["parent_id"] == "8675309"
def test_unparseable_tool_arguments_are_preserved_rather_than_dropped(logger: DataDogLLMObsLogger) -> None:
"""A truncated argument string is still the only record of what the model tried to call."""
payload = build(

91
tests/test_litellm/litellm_core_utils/test_streaming_handler.py
View file

@ -21,6 +21,7 @@ from litellm.litellm_core_utils.streaming_handler import (
from litellm.types.utils import (
CompletionTokensDetailsWrapper,
Delta,
ModelResponse,
ModelResponseStream,
PromptTokensDetailsWrapper,
StandardLoggingPayload,
@ -1750,7 +1751,7 @@ def test_openrouter_streaming_cost_propagates_to_hidden_params():
assert complete_response.usage.cost == 0.00025
# Use the real propagation method from CustomStreamWrapper
CustomStreamWrapper._propagate_usage_cost_to_hidden_params(complete_response)
CustomStreamWrapper._propagate_usage_cost_to_hidden_params(complete_response, "openrouter")
assert "additional_headers" in complete_response._hidden_params
assert (
@ -1769,14 +1770,12 @@ def test_openrouter_streaming_cost_propagates_to_hidden_params():
assert provider_cost == 0.00025
def test_perplexity_streaming_dict_cost_propagates_to_hidden_params():
"""
Regression: Perplexity reports usage.cost as a breakdown object, which used to
blow up the end of the stream with
`float() argument must be a string or a real number, not 'dict'`.
"""
def test_perplexity_streaming_dict_cost_bills_through_its_own_calculator():
import litellm
from litellm.cost_calculator import get_response_cost_from_hidden_params
from litellm.cost_calculator import (
get_response_cost_from_hidden_params,
response_cost_calculator,
)
chunks = [
ModelResponseStream(
@ -1828,13 +1827,81 @@ def test_perplexity_streaming_dict_cost_propagates_to_hidden_params():
assert complete_response is not None
CustomStreamWrapper._propagate_usage_cost_to_hidden_params(complete_response)
CustomStreamWrapper._propagate_usage_cost_to_hidden_params(complete_response, "perplexity")
assert (
get_response_cost_from_hidden_params(complete_response._hidden_params)
== 0.00503
assert get_response_cost_from_hidden_params(complete_response._hidden_params) is None
assert response_cost_calculator(
response_object=complete_response,
model="perplexity/sonar",
custom_llm_provider="perplexity",
call_type="completion",
optional_params={},
) == pytest.approx(0.00503)
def test_openai_compatible_streaming_cost_is_priced_from_the_cost_map():
import litellm
from litellm.cost_calculator import (
get_response_cost_from_hidden_params,
response_cost_calculator,
)
model = "openai/streams-cost-in-nanodollars"
litellm.register_model(
{
model: {
"input_cost_per_token": 1e-6,
"output_cost_per_token": 2e-6,
"litellm_provider": "openai",
"mode": "chat",
}
}
)
complete_response = ModelResponse(
id="chatcmpl-openai-compatible",
model=model,
choices=[],
usage=Usage(completion_tokens=5, prompt_tokens=10, total_tokens=15, cost=3_144_000),
)
CustomStreamWrapper._propagate_usage_cost_to_hidden_params(complete_response, "openai")
assert get_response_cost_from_hidden_params(complete_response._hidden_params) is None
assert response_cost_calculator(
response_object=complete_response,
model=model,
custom_llm_provider="openai",
call_type="completion",
optional_params={},
) == pytest.approx(2e-5)
def test_xai_streaming_reported_cost_still_takes_the_margin(monkeypatch):
import litellm
from litellm.cost_calculator import (
get_response_cost_from_hidden_params,
response_cost_calculator,
)
complete_response = ModelResponse(
id="chatcmpl-xai",
model="grok-4-latest",
choices=[],
usage=Usage(completion_tokens=353, prompt_tokens=198, total_tokens=551, cost=0.0009956),
)
CustomStreamWrapper._propagate_usage_cost_to_hidden_params(complete_response, "xai")
assert get_response_cost_from_hidden_params(complete_response._hidden_params) is None
monkeypatch.setattr(litellm, "cost_margin_config", {"xai": 0.5})
assert response_cost_calculator(
response_object=complete_response,
model="xai/grok-4-latest",
custom_llm_provider="xai",
call_type="completion",
optional_params={},
) == pytest.approx(0.0009956 * 1.5)
def test_provider_reported_cost_ignores_unusable_shapes():
assert CustomStreamWrapper._resolve_provider_reported_cost(None) is None

47
tests/test_litellm/llms/base_llm/files/test_azure_blob_storage_backend.py
View file

@ -27,6 +27,20 @@ def mock_gov_env_vars(mock_env_vars, monkeypatch):
monkeypatch.setenv("AZURE_STORAGE_ENDPOINT_SUFFIX", GOV_SUFFIX)
@pytest.fixture
def credential_chain_env_vars(monkeypatch):
monkeypatch.setenv("AZURE_STORAGE_ACCOUNT_NAME", "test-account")
monkeypatch.setenv("AZURE_STORAGE_FILE_SYSTEM", "test-container")
for name in (
"AZURE_STORAGE_TENANT_ID",
"AZURE_STORAGE_CLIENT_ID",
"AZURE_STORAGE_CLIENT_SECRET",
"AZURE_STORAGE_ACCOUNT_KEY",
"AZURE_STORAGE_ENDPOINT_SUFFIX",
):
monkeypatch.delenv(name, raising=False)
def _make_backend() -> AzureBlobStorageBackend:
backend = AzureBlobStorageBackend()
backend.azure_auth_token = "mock-azure-ad-token"
@ -42,6 +56,29 @@ def _mock_upload_client() -> AsyncMock:
return client
@pytest.mark.asyncio
async def test_upload_file_with_credential_chain(credential_chain_env_vars):
client = _mock_upload_client()
build_provider = MagicMock(return_value=lambda: "workload-identity-token")
with patch( # test-quality-ok: the backend creates its REST client internally; assert the emitted authorization header
"litellm.llms.custom_httpx.http_handler.get_async_httpx_client", return_value=client
):
backend = AzureBlobStorageBackend(build_credential_chain_token_provider=build_provider)
storage_url = await backend.upload_file(
file_content=b"hello",
filename="report.json",
content_type="application/json",
path_prefix="logs",
file_naming_strategy="original_filename",
)
build_provider.assert_called_once_with()
assert storage_url == "https://test-account.blob.core.windows.net/test-container/logs/report.json"
assert client.put.call_args[1]["headers"]["Authorization"] == "Bearer workload-identity-token"
assert client.patch.call_count == 2
@pytest.mark.parametrize(
"env_fixture, expected_suffix",
[("mock_env_vars", "core.windows.net"), ("mock_gov_env_vars", GOV_SUFFIX)],
@ -125,10 +162,7 @@ async def test_download_file_accepts_url_persisted_before_the_suffix_was_set(moc
)
assert content == b"file-bytes"
assert (
client.get.call_args[0][0]
== f"https://test-account.blob.{GOV_SUFFIX}/test-container/logs/report.json"
)
assert client.get.call_args[0][0] == f"https://test-account.blob.{GOV_SUFFIX}/test-container/logs/report.json"
@pytest.mark.parametrize(
@ -178,10 +212,7 @@ async def test_download_file_drops_query_string_from_the_stored_url(mock_env_var
"https://test-account.blob.core.windows.net/test-container/logs/report.json?sig=redacted&se=2026"
)
assert (
client.get.call_args[0][0]
== "https://test-account.blob.core.windows.net/test-container/logs/report.json"
)
assert client.get.call_args[0][0] == "https://test-account.blob.core.windows.net/test-container/logs/report.json"
@pytest.mark.parametrize(

301
tests/test_litellm/llms/s3_vectors/vector_stores/test_s3_vectors_transformation.py
View file

@ -1,40 +1,70 @@
from collections.abc import Mapping
from unittest.mock import AsyncMock, MagicMock, Mock, patch
import httpx
import pytest
from litellm.llms.base_llm.vector_store.transformation import (
RouterVectorStoreEmbeddingExecutor,
)
from litellm.llms.s3_vectors.vector_stores.transformation import (
S3VectorsVectorStoreConfig,
)
from litellm.types.utils import EmbeddingResponse
from litellm.types.vector_stores import VectorStoreSearchResponse
QUERY_VECTOR = [0.1, 0.2, 0.3]
def _mock_router(model_names, sync=False):
"""Router mock serving the given embedding model names."""
router = MagicMock()
router.get_model_list.return_value = [{"model_name": name} for name in model_names]
embedding_response = Mock(data=[{"embedding": [0.1, 0.2, 0.3]}])
if sync:
router.embedding = MagicMock(return_value=embedding_response)
else:
router.aembedding = AsyncMock(return_value=embedding_response)
return router
def _embedding_response(vector):
return EmbeddingResponse(data=[{"embedding": vector, "index": 0, "object": "embedding"}])
class _RecordingExecutor:
def __init__(self, vector=QUERY_VECTOR):
self.vector = vector
self.calls = []
def embed(self, model: str, query: str, configuration: Mapping[str, object]) -> EmbeddingResponse:
self.calls.append((model, query, dict(configuration)))
return _embedding_response(self.vector)
async def aembed(self, model: str, query: str, configuration: Mapping[str, object]) -> EmbeddingResponse:
self.calls.append((model, query, dict(configuration)))
return _embedding_response(self.vector)
def _logging_obj():
logging_obj = Mock()
logging_obj.model_call_details = {}
return logging_obj
def _search_kwargs(**overrides):
kwargs = {
"vector_store_id": "test-bucket:test-index",
"query": "test query",
"vector_store_search_optional_params": {},
"api_base": "https://s3vectors.us-west-2.api.aws",
"litellm_logging_obj": _logging_obj(),
"litellm_params": {},
"extra_body": None,
}
kwargs.update(overrides)
return kwargs
class TestS3VectorsVectorStoreConfig:
def test_init(self):
"""Test that S3VectorsVectorStoreConfig initializes correctly"""
config = S3VectorsVectorStoreConfig()
assert config is not None
def test_get_supported_openai_params(self):
"""Test that supported OpenAI params are returned"""
config = S3VectorsVectorStoreConfig()
params = config.get_supported_openai_params("test-model")
assert "max_num_results" in params
def test_get_complete_url(self):
"""Test URL generation for S3 Vectors"""
config = S3VectorsVectorStoreConfig()
litellm_params = {"aws_region_name": "us-west-2"}
url = config.get_complete_url(None, litellm_params)
@ -57,180 +87,170 @@ class TestS3VectorsVectorStoreConfig:
assert url == "https://s3vectors.eu-west-1.api.aws"
def test_get_complete_url_invalid_region_format(self):
"""Invalid region format raises"""
config = S3VectorsVectorStoreConfig()
with pytest.raises(ValueError, match="Invalid AWS region format"):
config.get_complete_url(None, {"aws_region_name": "Bad_Region!"})
def test_transform_search_request(self):
"""Full request-body transformation with a router-injected embedding"""
config = S3VectorsVectorStoreConfig()
mock_logging_obj = Mock()
mock_logging_obj.model_call_details = {}
router = _mock_router(["text-embedding-3-small"], sync=True)
logging_obj = _logging_obj()
executor = _RecordingExecutor()
url, request_body = config.transform_search_vector_store_request(
vector_store_id="test-bucket:test-index",
query="test query",
vector_store_search_optional_params={"max_num_results": 7},
api_base="https://s3vectors.us-west-2.api.aws",
litellm_logging_obj=mock_logging_obj,
litellm_params={},
extra_body=None,
router=router,
**_search_kwargs(
vector_store_search_optional_params={"max_num_results": 7},
litellm_logging_obj=logging_obj,
embedding_executor=executor,
)
)
assert url == "https://s3vectors.us-west-2.api.aws/QueryVectors"
assert request_body == {
"vectorBucketName": "test-bucket",
"indexName": "test-index",
"queryVector": {"float32": [0.1, 0.2, 0.3]},
"queryVector": {"float32": QUERY_VECTOR},
"topK": 7,
"returnDistance": True,
"returnMetadata": True,
}
assert mock_logging_obj.model_call_details["query"] == "test query"
assert executor.calls == [("text-embedding-3-small", "test query", {})]
assert logging_obj.model_call_details["query"] == "test query"
@pytest.mark.parametrize(
("litellm_params", "expected_model"),
[
({}, "text-embedding-3-small"),
({"embedding_model": ""}, "text-embedding-3-small"),
({"embedding_model": "my-embedding-model"}, "my-embedding-model"),
({"litellm_embedding_model": "shared-key-model"}, "shared-key-model"),
(
{"litellm_embedding_model": "shared-key-model", "embedding_model": "legacy-alias"},
"shared-key-model",
),
],
)
def test_query_embedding_model_accepts_embedding_model_alias(self, litellm_params, expected_model):
assert S3VectorsVectorStoreConfig.query_embedding_model(litellm_params) == expected_model
@pytest.mark.asyncio
async def test_atransform_search_uses_router_for_virtual_model(self):
"""Regression: router-served embedding models must resolve via the router,
not a bare litellm.aembedding call (which has no deployment credentials)."""
async def test_atransform_search_embeds_alias_and_store_config_through_executor(self):
config = S3VectorsVectorStoreConfig()
mock_logging_obj = Mock()
mock_logging_obj.model_call_details = {}
router = _mock_router(["my-embedding-model"])
executor = _RecordingExecutor(vector=[0.4, 0.5])
with patch("litellm.aembedding", new=AsyncMock()) as mock_bare_aembedding: # test-quality-ok: guards that the bare-embedding path is not taken; dispatch seam is the behavior under test
url, request_body = await config.atransform_search_vector_store_request(
vector_store_id="test-bucket:test-index",
query="test query",
vector_store_search_optional_params={},
api_base="https://s3vectors.us-west-2.api.aws",
litellm_logging_obj=mock_logging_obj,
litellm_params={"embedding_model": "my-embedding-model"},
extra_body=None,
router=router,
_, request_body = await config.atransform_search_vector_store_request(
**_search_kwargs(
query=["test", "query"],
litellm_params={
"embedding_model": "my-embedding-model",
"litellm_embedding_config": {"api_key": "store-key"},
},
embedding_executor=executor,
)
)
router.aembedding.assert_awaited_once_with(model="my-embedding-model", input=["test query"])
mock_bare_aembedding.assert_not_awaited()
assert request_body["queryVector"]["float32"] == [0.1, 0.2, 0.3]
assert request_body["topK"] == 5 # default
@pytest.mark.asyncio
async def test_atransform_search_falls_back_when_router_does_not_serve_model(self):
"""Router present but embedding_model is not a router deployment ->
bare litellm.aembedding keeps working (provider-prefixed + env creds stores)."""
config = S3VectorsVectorStoreConfig()
mock_logging_obj = Mock()
mock_logging_obj.model_call_details = {}
router = _mock_router(["some-other-model"])
mock_bare = AsyncMock(return_value=Mock(data=[{"embedding": [0.4, 0.5]}]))
with patch("litellm.aembedding", new=mock_bare): # test-quality-ok: stubs the bare-embedding fallback whose request body the test asserts on
_, request_body = await config.atransform_search_vector_store_request(
vector_store_id="test-bucket:test-index",
query="test query",
vector_store_search_optional_params={},
api_base="https://s3vectors.us-west-2.api.aws",
litellm_logging_obj=mock_logging_obj,
litellm_params={"embedding_model": "azure/text-embedding-3-small"},
extra_body=None,
router=router,
)
mock_bare.assert_awaited_once_with(model="azure/text-embedding-3-small", input=["test query"])
router.aembedding.assert_not_awaited()
assert executor.calls == [("my-embedding-model", "test query", {"api_key": "store-key"})]
assert request_body["queryVector"]["float32"] == [0.4, 0.5]
assert request_body["topK"] == 5
@pytest.mark.asyncio
async def test_atransform_search_without_router_uses_bare_embedding(self):
"""Backward compat: no router -> bare litellm.aembedding as before"""
async def test_atransform_search_router_executor_carries_request_metadata(self):
"""Regression (LIT-6750): a bare Router alias resolves through the Router with the request's
team metadata on the embedding call, so the embedding is attributed to the calling key and team."""
config = S3VectorsVectorStoreConfig()
mock_logging_obj = Mock()
mock_logging_obj.model_call_details = {}
router = MagicMock()
router.aembedding = AsyncMock(return_value=_embedding_response(QUERY_VECTOR))
request_metadata = {"user_api_key_team_id": "team-a", "user_api_key": "hashed-key"}
mock_bare = AsyncMock(return_value=Mock(data=[{"embedding": [0.6, 0.7]}]))
with patch("litellm.aembedding", new=mock_bare): # test-quality-ok: stubs the bare-embedding fallback whose request body the test asserts on
_, request_body = await config.atransform_search_vector_store_request(
vector_store_id="test-bucket:test-index",
query="test query",
vector_store_search_optional_params={},
api_base="https://s3vectors.us-west-2.api.aws",
litellm_logging_obj=mock_logging_obj,
litellm_params={},
extra_body=None,
_, request_body = await config.atransform_search_vector_store_request(
**_search_kwargs(
litellm_params={"embedding_model": "team-embeddings"},
embedding_executor=RouterVectorStoreEmbeddingExecutor(router=router, metadata=request_metadata),
)
)
router.aembedding.assert_awaited_once_with(
model="team-embeddings", input=["test query"], metadata=request_metadata
)
assert request_body["queryVector"]["float32"] == QUERY_VECTOR
@pytest.mark.asyncio
async def test_atransform_search_default_model_falls_back_to_the_sdk(self):
"""Regression (LIT-6750): a store that never named an embedding model keeps working on a proxy
whose model list has no text-embedding-3-small, embedding through the SDK instead of erroring."""
config = S3VectorsVectorStoreConfig()
router = MagicMock()
router.get_model_list.return_value = [
{"model_name": "team-embeddings", "litellm_params": {"model": "openai/text-embedding-3-small"}}
]
router.resolved_litellm_models.return_value = []
router.aembedding = AsyncMock(side_effect=AssertionError("unserved model must not reach the Router"))
request_metadata = {"user_api_key_team_id": "team-a"}
mock_bare = AsyncMock(return_value=_embedding_response(QUERY_VECTOR))
with patch("litellm.aembedding", new=mock_bare): # test-quality-ok: stubs the bare-embedding fallback whose call the test asserts on
_, request_body = await config.atransform_search_vector_store_request(
**_search_kwargs(
embedding_executor=RouterVectorStoreEmbeddingExecutor(router=router, metadata=request_metadata)
)
)
mock_bare.assert_awaited_once_with(
model="text-embedding-3-small", input=["test query"], metadata=request_metadata
)
assert request_body["queryVector"]["float32"] == QUERY_VECTOR
@pytest.mark.asyncio
async def test_atransform_search_without_executor_uses_bare_embedding(self):
config = S3VectorsVectorStoreConfig()
mock_bare = AsyncMock(return_value=_embedding_response([0.6, 0.7]))
with patch("litellm.aembedding", new=mock_bare): # test-quality-ok: stubs the bare-embedding fallback whose request body the test asserts on
_, request_body = await config.atransform_search_vector_store_request(**_search_kwargs())
mock_bare.assert_awaited_once_with(model="text-embedding-3-small", input=["test query"])
assert request_body["queryVector"]["float32"] == [0.6, 0.7]
def test_transform_search_uses_router_for_virtual_model_sync(self):
"""Sync twin: router-served embedding model resolves via router.embedding"""
def test_transform_search_without_executor_uses_bare_embedding_sync(self):
config = S3VectorsVectorStoreConfig()
mock_logging_obj = Mock()
mock_logging_obj.model_call_details = {}
router = _mock_router(["my-embedding-model"], sync=True)
with patch("litellm.embedding", new=MagicMock()) as mock_bare_embedding: # test-quality-ok: guards that the bare-embedding path is not taken; dispatch seam is the behavior under test
_, request_body = config.transform_search_vector_store_request(
vector_store_id="test-bucket:test-index",
query="test query",
vector_store_search_optional_params={},
api_base="https://s3vectors.us-west-2.api.aws",
litellm_logging_obj=mock_logging_obj,
litellm_params={"embedding_model": "my-embedding-model"},
extra_body=None,
router=router,
)
router.embedding.assert_called_once_with(model="my-embedding-model", input=["test query"])
mock_bare_embedding.assert_not_called()
assert request_body["queryVector"]["float32"] == [0.1, 0.2, 0.3]
def test_transform_search_without_router_uses_bare_embedding_sync(self):
"""Sync twin: no router -> bare litellm.embedding as before"""
config = S3VectorsVectorStoreConfig()
mock_logging_obj = Mock()
mock_logging_obj.model_call_details = {}
mock_bare = MagicMock(return_value=Mock(data=[{"embedding": [0.8, 0.9]}]))
mock_bare = MagicMock(return_value=_embedding_response([0.8, 0.9]))
with patch("litellm.embedding", new=mock_bare): # test-quality-ok: stubs the bare-embedding fallback whose request body the test asserts on
_, request_body = config.transform_search_vector_store_request(
vector_store_id="test-bucket:test-index",
query="test query",
vector_store_search_optional_params={},
api_base="https://s3vectors.us-west-2.api.aws",
litellm_logging_obj=mock_logging_obj,
litellm_params={},
extra_body=None,
**_search_kwargs(litellm_params={"embedding_model": "my-embedding-model"})
)
mock_bare.assert_called_once_with(model="text-embedding-3-small", input=["test query"])
mock_bare.assert_called_once_with(model="my-embedding-model", input=["test query"])
assert request_body["queryVector"]["float32"] == [0.8, 0.9]
def test_transform_search_request_invalid_vector_store_id(self):
"""Test that invalid vector_store_id format raises error"""
config = S3VectorsVectorStoreConfig()
mock_logging_obj = Mock()
mock_logging_obj.model_call_details = {}
executor = _RecordingExecutor()
with pytest.raises(
ValueError,
match="vector_store_id must be in format 'bucket_name:index_name'",
):
config.transform_search_vector_store_request(
vector_store_id="invalid-format",
query="test query",
vector_store_search_optional_params={},
api_base="https://s3vectors.us-west-2.api.aws",
litellm_logging_obj=mock_logging_obj,
litellm_params={},
extra_body=None,
**_search_kwargs(vector_store_id="invalid-format", embedding_executor=executor)
)
assert executor.calls == []
def test_transform_search_request_bucket_from_litellm_params(self):
config = S3VectorsVectorStoreConfig()
_, request_body = config.transform_search_vector_store_request(
**_search_kwargs(
vector_store_id="only-index",
litellm_params={"vector_bucket_name": "params-bucket"},
embedding_executor=_RecordingExecutor(),
)
)
assert request_body["vectorBucketName"] == "params-bucket"
assert request_body["indexName"] == "only-index"
def test_transform_search_response(self):
"""Test search response transformation"""
config = S3VectorsVectorStoreConfig()
mock_logging_obj = Mock()
mock_logging_obj.model_call_details = {"query": "test query"}
@ -239,7 +259,7 @@ class TestS3VectorsVectorStoreConfig:
mock_response.json.return_value = {
"vectors": [
{
"distance": 0.05, # S3 Vectors returns distance, not score
"distance": 0.05,
"metadata": {
"source_text": "This is test content",
"chunk_index": "0",
@ -258,23 +278,18 @@ class TestS3VectorsVectorStoreConfig:
mock_response.status_code = 200
mock_response.headers = {}
result = config.transform_search_vector_store_response(
mock_response, mock_logging_obj
)
result = config.transform_search_vector_store_response(mock_response, mock_logging_obj)
# VectorStoreSearchResponse is a TypedDict, so check structure instead of isinstance
assert result["object"] == "vector_store.search_results.page"
assert result["search_query"] == "test query"
assert len(result["data"]) == 2
# Score should be 1 - distance (cosine similarity)
assert result["data"][0]["score"] == 0.95 # 1 - 0.05
assert result["data"][0]["score"] == 0.95
assert result["data"][0]["content"][0]["text"] == "This is test content"
assert result["data"][0]["filename"] == "test.pdf"
assert result["data"][1]["score"] == 0.85 # 1 - 0.15
assert result["data"][1]["score"] == 0.85
assert result["data"][1]["content"][0]["text"] == "More test content"
def test_map_openai_params(self):
"""Test OpenAI parameter mapping"""
config = S3VectorsVectorStoreConfig()
non_default_params = {"max_num_results": 5}
optional_params = {}

96
tests/test_litellm/llms/xai/responses/test_xai_responses_transformation.py
View file

@ -7,12 +7,13 @@ transformations for the Responses API.
Source: litellm/llms/xai/responses/transformation.py
"""
from unittest.mock import MagicMock
from unittest.mock import MagicMock, Mock
import httpx
import pytest
import litellm
from litellm.llms.xai.cost_calculator import cost_per_token
from litellm.llms.xai.responses.transformation import XAIResponsesAPIConfig
from litellm.responses.utils import ResponseAPILoggingUtils
from litellm.types.llms.openai import (
@ -400,3 +401,94 @@ class TestXAIResponsesWebSearchBilling:
bridged = ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(event.response.usage)
assert getattr(bridged, "server_side_tool_usage_details") == self._TOOL_DETAILS
class TestXAIResponsesReportedCost:
"""xAI reports what it charged; the transformation moves it to where litellm bills from.
``ResponseAPILoggingUtils`` copies ``usage.cost`` onto the chat Usage that cost
tracking prices, so restating ``cost_in_usd_ticks`` there is what makes /v1/responses
bill the reported figure. At 10^10 ticks to the dollar, 37756000 ticks is $0.0037756.
"""
@staticmethod
def _response_body(usage: dict) -> dict:
return {
"id": "resp_xai",
"object": "response",
"created_at": 0,
"model": "grok-4-latest",
"status": "completed",
"output": [],
"parallel_tool_calls": False,
"tool_choice": "auto",
"tools": [],
"usage": usage,
}
def _transformed_usage(self, usage: dict) -> ResponseAPIUsage | None:
raw_response = httpx.Response(status_code=200, json=self._response_body(usage))
response = XAIResponsesAPIConfig().transform_response_api_response(
model="grok-4-latest",
raw_response=raw_response,
logging_obj=Mock(),
)
return response.usage
def test_reported_cost_reaches_the_cost_calculator(self):
usage = self._transformed_usage(
{
"input_tokens": 100,
"output_tokens": 200,
"total_tokens": 300,
"cost_in_usd_ticks": 37756000,
}
)
assert usage.cost == 0.0037756
chat_usage = ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(usage)
assert cost_per_token(model="grok-4-latest", usage=chat_usage) == (0.0, 0.0037756)
def test_streamed_reported_cost_reaches_the_cost_calculator(self):
event = XAIResponsesAPIConfig().transform_streaming_response(
model="grok-4-latest",
parsed_chunk={
"type": "response.completed",
"sequence_number": 7,
"response": self._response_body(
{
"input_tokens": 100,
"output_tokens": 200,
"total_tokens": 300,
"cost_in_usd_ticks": 37756000,
}
),
},
logging_obj=Mock(),
)
assert isinstance(event, ResponseCompletedEvent)
chat_usage = ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(event.response.usage)
assert cost_per_token(model="grok-4-latest", usage=chat_usage) == (0.0, 0.0037756)
def test_usage_without_a_reported_cost_is_left_alone(self):
usage = self._transformed_usage(
{"input_tokens": 100, "output_tokens": 200, "total_tokens": 300}
)
assert usage.cost is None
def test_negative_reported_cost_is_not_carried(self):
"""A caller who can set api_base must not be able to report negative spend."""
usage = self._transformed_usage(
{
"input_tokens": 100,
"output_tokens": 200,
"total_tokens": 300,
"cost_in_usd_ticks": -37756000,
}
)
assert usage.cost is None

119
tests/test_litellm/llms/xai/test_xai_chat_transformation.py
View file

@ -1,9 +1,14 @@
from unittest.mock import Mock
import httpx
import pytest
import litellm
from litellm.llms.xai.chat.transformation import XAIChatConfig
from litellm.llms.xai.chat.transformation import (
XAIChatCompletionStreamingHandler,
XAIChatConfig,
)
from litellm.llms.xai.cost_calculator import cost_per_token
from litellm.types.utils import (
CompletionTokensDetailsWrapper,
ModelResponse,
@ -195,3 +200,113 @@ class TestXAIChatWebSearchBilling:
)
assert with_search - without_search == pytest.approx(3 * 5.0 / 1000.0)
class TestXAIReportedCost:
"""xAI reports what it charged; the transformation moves it to where litellm bills from.
``cost`` is the field litellm already carries a provider stated cost in, so restating
``cost_in_usd_ticks`` there is what lets ``llms/xai/cost_calculator.py`` bill the
reported figure. At 10^10 ticks to the dollar, 37756000 ticks is $0.0037756.
"""
@staticmethod
def _transformed_usage(usage: dict) -> Usage:
raw_response = httpx.Response(
status_code=200,
json={
"id": "chatcmpl-xai",
"object": "chat.completion",
"created": 0,
"model": "grok-4-latest",
"choices": [
{
"index": 0,
"message": {"role": "assistant", "content": "hi"},
"finish_reason": "stop",
}
],
"usage": usage,
},
)
response = XAIChatConfig().transform_response(
model="grok-4-latest",
raw_response=raw_response,
model_response=ModelResponse(),
logging_obj=Mock(),
request_data={},
messages=[{"role": "user", "content": "hi"}],
optional_params={},
litellm_params={},
encoding=None,
)
return response.usage
def test_reported_cost_reaches_the_cost_calculator(self):
usage = self._transformed_usage(
{
"prompt_tokens": 100,
"completion_tokens": 200,
"total_tokens": 300,
"cost_in_usd_ticks": 37756000,
}
)
assert usage.cost == 0.0037756
assert cost_per_token(model="grok-4-latest", usage=usage) == (0.0, 0.0037756)
def test_usage_without_a_reported_cost_is_left_alone(self):
usage = self._transformed_usage(
{"prompt_tokens": 100, "completion_tokens": 200, "total_tokens": 300}
)
assert getattr(usage, "cost", None) is None
def test_negative_reported_cost_is_not_carried(self):
"""A caller who can set api_base must not be able to report negative spend."""
usage = self._transformed_usage(
{
"prompt_tokens": 100,
"completion_tokens": 200,
"total_tokens": 300,
"cost_in_usd_ticks": -37756000,
}
)
assert getattr(usage, "cost", None) is None
def test_streamed_reported_cost_survives_chunk_aggregation(self):
"""Streamed spend only matches if the conversion happens on the chunk.
Chunk aggregation rebuilds usage from the fields it models plus ``cost``, so a
chunk still carrying only ``cost_in_usd_ticks`` loses the reported amount.
"""
handler = XAIChatCompletionStreamingHandler(
streaming_response=iter([]), sync_stream=True
)
parsed = handler.chunk_parser(
{
"id": "chatcmpl-xai",
"object": "chat.completion.chunk",
"created": 0,
"model": "grok-4-latest",
"choices": [],
"usage": {
"prompt_tokens": 100,
"completion_tokens": 200,
"total_tokens": 300,
"cost_in_usd_ticks": 37756000,
},
}
)
assert parsed.usage.cost == 0.0037756
assembled = litellm.stream_chunk_builder(chunks=[parsed])
assert assembled.usage.cost == 0.0037756
assert cost_per_token(model="grok-4-latest", usage=assembled.usage) == (
0.0,
0.0037756,
)

184
tests/test_litellm/llms/xai/test_xai_cost_calculator.py
View file

@ -7,7 +7,10 @@ import os
import litellm
from litellm.types.utils import (
Choices,
CompletionTokensDetailsWrapper,
Message,
ModelResponse,
PromptTokensDetailsWrapper,
Usage,
)
@ -361,6 +364,145 @@ class TestXAICostCalculator:
response_object=object(), usage=usage
)
def test_reported_cost_is_preferred_over_token_math(self):
"""The amount xAI reported, carried on usage.cost by the transformation, is billed.
It lands entirely on completion cost because xAI does not split its total by
direction, the same shape the perplexity calculator returns.
"""
usage = Usage(
prompt_tokens=100,
completion_tokens=200,
total_tokens=300,
cost=0.0037756,
)
prompt_cost, completion_cost = cost_per_token(model="grok-4-latest", usage=usage)
assert prompt_cost == 0.0
assert math.isclose(completion_cost, 0.0037756, rel_tol=1e-10)
def test_reported_cost_suppresses_web_search_surcharge(self):
"""The reported total already covers server-side tool calls.
Without the suppression these 3 searches would be billed a second time on
top of the total xAI already charged.
"""
usage = Usage(
prompt_tokens=100,
completion_tokens=50,
total_tokens=150,
prompt_tokens_details=PromptTokensDetailsWrapper(
text_tokens=100,
web_search_requests=3,
),
cost=0.0037756,
)
assert cost_per_web_search_request(usage=usage, model_info={}) == 0.0
def test_web_search_surcharge_suppressed_through_the_dispatcher(self):
"""The suppression has to hold on the path cost tracking actually uses.
Legacy behaviour stays intact when xAI reports no cost.
"""
from litellm.llms import get_cost_for_web_search_request
usage = Usage(prompt_tokens=100, completion_tokens=50, total_tokens=150)
setattr(usage, "server_side_tool_usage_details", {"web_search_calls": 3})
assert get_cost_for_web_search_request("xai", usage, {}) > 0.0
reported = Usage(
prompt_tokens=100, completion_tokens=50, total_tokens=150, cost=0.0037756
)
setattr(reported, "server_side_tool_usage_details", {"web_search_calls": 3})
assert get_cost_for_web_search_request("xai", reported, {}) == 0.0
def test_no_reported_cost_falls_back_to_token_math(self):
"""Absent the provider figure, nothing changes for existing callers."""
usage = Usage(prompt_tokens=100, completion_tokens=200, total_tokens=300)
prompt_cost, completion_cost = cost_per_token(model="grok-4-latest", usage=usage)
assert prompt_cost > 0.0
assert completion_cost > 0.0
def test_malformed_reported_cost_falls_back_to_token_math(self):
"""A junk value must not fail the request, fall back to calculating."""
usage = Usage(prompt_tokens=100, completion_tokens=200, total_tokens=300)
setattr(usage, "cost", "not-a-number")
prompt_cost, completion_cost = cost_per_token(model="grok-4-latest", usage=usage)
assert prompt_cost > 0.0
assert completion_cost > 0.0
def test_boolean_reported_cost_falls_back_to_token_math(self):
"""True is an int in python and would otherwise be billed as $1."""
usage = Usage(prompt_tokens=100, completion_tokens=200, total_tokens=300)
setattr(usage, "cost", True)
prompt_cost, completion_cost = cost_per_token(model="grok-4-latest", usage=usage)
assert prompt_cost > 0.0
assert completion_cost > 0.0
assert completion_cost != 1.0
def test_negative_reported_cost_is_rejected(self):
"""A negative amount must never reach spend tracking.
A caller who can set api_base controls the response body, so trusting a
negative figure would let them subtract from their own recorded spend and
slip past a budget. Fall back to token pricing instead, and keep charging
the web search surcharge, since no trustworthy total was reported.
"""
usage = Usage(
prompt_tokens=100,
completion_tokens=200,
total_tokens=300,
cost=-0.0037756,
)
setattr(usage, "server_side_tool_usage_details", {"web_search_calls": 3})
prompt_cost, completion_cost = cost_per_token(model="grok-4-latest", usage=usage)
assert prompt_cost > 0.0
assert completion_cost > 0.0
assert cost_per_web_search_request(usage=usage, model_info={}) > 0.0
def test_non_finite_reported_cost_is_rejected(self):
"""NaN compares false against every budget threshold.
Usage stores a provider supplied cost without validating it, so a caller who
controls the response body could report NaN and leave spend >= max_budget
false for the life of the key rather than mispricing one request. The
infinities are refused alongside it. Fall back to token pricing and keep
charging the web search surcharge, since no trustworthy total was reported.
"""
for reported_cost in (float("nan"), float("inf"), float("-inf")):
usage = Usage(
prompt_tokens=100,
completion_tokens=200,
total_tokens=300,
cost=reported_cost,
)
setattr(usage, "server_side_tool_usage_details", {"web_search_calls": 3})
prompt_cost, completion_cost = cost_per_token(model="grok-4-latest", usage=usage)
assert math.isfinite(prompt_cost), reported_cost
assert math.isfinite(completion_cost), reported_cost
assert prompt_cost > 0.0, reported_cost
assert completion_cost > 0.0, reported_cost
assert cost_per_web_search_request(usage=usage, model_info={}) > 0.0, reported_cost
def test_zero_reported_cost_is_honoured(self):
"""A reported zero is a real answer, not a missing value."""
usage = Usage(prompt_tokens=100, completion_tokens=200, total_tokens=300, cost=0.0)
assert cost_per_token(model="grok-4-latest", usage=usage) == (0.0, 0.0)
def test_grok_4_20_beta_reasoning_cost_calculation(self):
"""Test cost calculation for grok-4.20-beta-0309-reasoning model."""
usage = Usage(prompt_tokens=100, completion_tokens=200, total_tokens=300)
@ -437,6 +579,48 @@ class TestXAICostCalculator:
assert math.isclose(prompt_cost, expected_prompt_cost, rel_tol=1e-10)
assert math.isclose(completion_cost, expected_completion_cost, rel_tol=1e-10)
def test_custom_pricing_beats_the_reported_cost(self):
response = ModelResponse(
id="chatcmpl-xai",
model="grok-4-latest",
choices=[Choices(index=0, message=Message(role="assistant", content="x"), finish_reason="stop")],
usage=Usage(prompt_tokens=198, completion_tokens=353, total_tokens=551, cost=0.0009956),
)
billed = litellm.completion_cost(
completion_response=response,
model="xai/grok-4-latest",
custom_llm_provider="xai",
custom_cost_per_token={"input_cost_per_token": 0.001, "output_cost_per_token": 0.001},
custom_pricing=True,
)
assert math.isclose(billed, 0.551, rel_tol=1e-10)
def test_deployment_custom_pricing_beats_the_reported_cost(self, monkeypatch):
deployment_id = "xai-deployment-priced-by-the-operator"
monkeypatch.setitem(
litellm.model_cost,
deployment_id,
{"input_cost_per_token": 0.001, "output_cost_per_token": 0.001, "litellm_provider": "xai", "mode": "chat"},
)
response = ModelResponse(
id="chatcmpl-xai",
model="grok-4-latest",
choices=[Choices(index=0, message=Message(role="assistant", content="x"), finish_reason="stop")],
usage=Usage(prompt_tokens=198, completion_tokens=353, total_tokens=551, cost=0.0009956),
)
billed = litellm.completion_cost(
completion_response=response,
model="xai/grok-4-latest",
custom_llm_provider="xai",
custom_pricing=True,
router_model_id=deployment_id,
)
assert math.isclose(billed, 0.551, rel_tol=1e-10)
class TestXAIWebSearchCostHelpers:
"""Focused coverage for web_search / tool-usage helpers in cost_calculator.py."""

6
tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server.py
View file

@ -914,6 +914,7 @@ async def test_mcp_get_prompt_success():
)
mock_manager.get_prompt_from_server.assert_awaited_once_with(
server=server,
user_api_key_auth=user_api_key_auth,
prompt_name="hello",
arguments={"foo": "bar"},
mcp_auth_header={"Authorization": "token"},
@ -976,6 +977,7 @@ async def test_mcp_read_resource_success():
)
mock_manager.read_resource_from_server.assert_awaited_once_with(
server=server,
user_api_key_auth=user_api_key_auth,
url="https://example.com/resource",
mcp_auth_header={"Authorization": "token"},
extra_headers={"X-Test": "1"},
@ -8268,7 +8270,9 @@ class TestOboPreflightScopedToAllowedServers:
_, preflight = await self._run(requested, allowed=[requested], user_api_key_auth=key)
preflight.assert_awaited_once_with(server=requested, oauth2_headers=self.SUBJECT_HEADERS, user_api_key_auth=key)
preflight.assert_awaited_once_with(
server=requested, oauth2_headers=self.SUBJECT_HEADERS, user_api_key_auth=key, raw_headers=None
)
@pytest.mark.asyncio

311
tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server_manager.py
View file

@ -30,6 +30,7 @@ from mcp.types import (
TextResourceContents,
)
from mcp.types import Tool as MCPTool
from pydantic import AnyUrl
from litellm.constants import MCP_METADATA_TIMEOUT
from litellm.proxy._experimental.mcp_server.mcp_server_manager import (
@ -2270,7 +2271,9 @@ class TestMCPServerManager:
"""prompts/list on an OBO server must exchange the caller's bearer, not connect with none."""
server = self._token_exchange_server("te-prompts")
st = await self._capture_subject_token(
lambda m: m.get_prompts_from_server(server=server, raw_headers={"authorization": "Bearer subj-jwt"})
lambda m: m.get_prompts_from_server(
server=server, user_api_key_auth=None, raw_headers={"authorization": "Bearer subj-jwt"}
)
)
assert st == "subj-jwt"
@ -2279,7 +2282,9 @@ class TestMCPServerManager:
"""resources/list on an OBO server must exchange the caller's bearer."""
server = self._token_exchange_server("te-resources")
st = await self._capture_subject_token(
lambda m: m.get_resources_from_server(server=server, raw_headers={"authorization": "Bearer subj-jwt"})
lambda m: m.get_resources_from_server(
server=server, user_api_key_auth=None, raw_headers={"authorization": "Bearer subj-jwt"}
)
)
assert st == "subj-jwt"
@ -2290,6 +2295,7 @@ class TestMCPServerManager:
st = await self._capture_subject_token(
lambda m: m.read_resource_from_server(
server=server,
user_api_key_auth=None,
url="https://up.example.com/r",
raw_headers={"authorization": "Bearer subj-jwt"},
)
@ -2307,7 +2313,9 @@ class TestMCPServerManager:
auth_type=MCPAuth.none,
)
st = await self._capture_subject_token(
lambda m: m.get_prompts_from_server(server=server, raw_headers={"authorization": "Bearer subj-jwt"})
lambda m: m.get_prompts_from_server(
server=server, user_api_key_auth=None, raw_headers={"authorization": "Bearer subj-jwt"}
)
)
assert st is None
@ -3254,7 +3262,7 @@ class TestMCPServerManager:
new_callable=AsyncMock,
return_value=mock_client,
):
prompts = await manager.get_prompts_from_server(server, add_prefix=True)
prompts = await manager.get_prompts_from_server(server, user_api_key_auth=None, add_prefix=True)
mock_client.list_prompts.assert_awaited_once()
assert len(prompts) == 1
@ -3289,6 +3297,7 @@ class TestMCPServerManager:
):
result = await manager.get_prompt_from_server(
server=server,
user_api_key_auth=None,
prompt_name="hello",
arguments={"tone": "casual"},
)
@ -3334,6 +3343,7 @@ class TestMCPServerManager:
):
result = await manager.get_resources_from_server(
server=server,
user_api_key_auth=None,
mcp_auth_header="auth",
extra_headers={"X-Test": "1"},
add_prefix=True,
@ -3391,6 +3401,7 @@ class TestMCPServerManager:
):
result = await manager.get_resource_templates_from_server(
server=server,
user_api_key_auth=None,
mcp_auth_header="auth",
extra_headers=None,
add_prefix=False,
@ -3441,6 +3452,7 @@ class TestMCPServerManager:
) as mock_create_client:
result = await manager.read_resource_from_server(
server=server,
user_api_key_auth=None,
url="https://example.com/resource",
mcp_auth_header="auth",
extra_headers={"X-Test": "1"},
@ -11006,3 +11018,294 @@ class TestOpenApiHandlerRelaysUpstreamAuth:
assert result.isError is True
assert "upstream returned HTTP 503" in result.content[0].text
class TestLitellmAdmissionKeyIsNeverTheSubjectToken:
"""The bearer that admitted the request as a LiteLLM key must not be sent to the IdP as the
RFC 8693 subject_token (or ID-JAG assertion). Only ``x-litellm-api-key`` disambiguates: with it
present, ``Authorization`` is the caller's own identity token and is exchanged as before."""
_ADMISSION_KEY: Final = "sk-litellm-virtual-key"
_USER_TOKEN: Final = "user-idp-jwt"
@staticmethod
def _token_exchange_server(server_id: str) -> MCPServer:
return MCPServer(
server_id=server_id,
name=f"{server_id}-server",
url="https://up.example.com/mcp",
transport=MCPTransport.http,
auth_type=MCPAuth.oauth2_token_exchange,
token_exchange_endpoint="https://idp.example.com/token",
client_id="cid",
client_secret="csec",
)
@staticmethod
def _id_jag_server(server_id: str) -> MCPServer:
return MCPServer(
server_id=server_id,
name=f"{server_id}-server",
url="https://up.example.com/mcp",
transport=MCPTransport.http,
auth_type=MCPAuth.oauth2_id_jag,
client_id="cid",
client_secret="csec",
token_exchange_endpoint="https://idp.example.com/token",
id_jag_resource_token_endpoint="https://resource-as.example.com/token",
)
@staticmethod
def _recording_provider() -> MagicMock:
from litellm.proxy._experimental.mcp_server.outbound_credentials.httpx_auth import StaticHeaderAuth
from litellm.proxy._experimental.mcp_server.outbound_credentials.result import Ok
provider: Final = MagicMock()
provider.resolve_credentials = AsyncMock(
return_value=Ok(StaticHeaderAuth("Bearer MINTED", header_name="Authorization"))
)
return provider
@staticmethod
def _subjects_seen_by(provider: MagicMock) -> list[str | None]:
return [
call.args[0].inbound_token.get_secret_value() if call.args[0].inbound_token else None
for call in provider.resolve_credentials.call_args_list
]
@staticmethod
def _manager_with_recording_client() -> MCPServerManager:
manager: Final = MCPServerManager()
client: Final = AsyncMock()
client.call_tool = AsyncMock(return_value=CallToolResult(content=[], isError=False))
client.list_prompts = AsyncMock(return_value=[])
client.read_resource = AsyncMock(return_value=ReadResourceResult(contents=[]))
manager._create_mcp_client = AsyncMock(return_value=client)
return manager
@staticmethod
def _subject_token_given_to_client(manager: MCPServerManager) -> str | None:
return manager._create_mcp_client.call_args.kwargs["subject_token"]
async def _call_tool_subject(self, server: MCPServer, oauth2_headers, raw_headers, user_api_key_auth):
manager: Final = self._manager_with_recording_client()
await manager._call_regular_mcp_tool(
mcp_server=server,
original_tool_name="tool",
arguments={},
tasks=[],
mcp_auth_header=None,
mcp_server_auth_headers=None,
oauth2_headers=oauth2_headers,
raw_headers=raw_headers,
proxy_logging_obj=None,
user_api_key_auth=user_api_key_auth,
)
return self._subject_token_given_to_client(manager)
@pytest.mark.asyncio
@pytest.mark.parametrize("auth_type", [MCPAuth.oauth2_token_exchange, MCPAuth.oauth2_id_jag])
async def test_tools_call_with_only_the_litellm_key_has_no_subject(self, auth_type):
server = (
self._token_exchange_server("te-call")
if auth_type == MCPAuth.oauth2_token_exchange
else self._id_jag_server("jag-call")
)
subject_token = await self._call_tool_subject(
server,
oauth2_headers={"Authorization": f"Bearer {self._ADMISSION_KEY}"},
raw_headers={"authorization": f"Bearer {self._ADMISSION_KEY}"},
user_api_key_auth=UserAPIKeyAuth(api_key="hashed-key", user_id="alice"),
)
assert subject_token is None
@pytest.mark.asyncio
async def test_rest_tools_call_with_only_the_litellm_key_has_no_subject(self):
"""The REST facade passes no oauth2_headers; the bearer is reached through raw_headers only."""
subject_token = await self._call_tool_subject(
self._token_exchange_server("te-rest"),
oauth2_headers=None,
raw_headers={"Authorization": f"Bearer {self._ADMISSION_KEY}"},
user_api_key_auth=UserAPIKeyAuth(api_key="hashed-key", user_id="alice"),
)
assert subject_token is None
@pytest.mark.asyncio
async def test_tools_call_exchanges_the_user_token_when_x_litellm_api_key_admits(self):
subject_token = await self._call_tool_subject(
self._token_exchange_server("te-split"),
oauth2_headers={"Authorization": f"Bearer {self._USER_TOKEN}"},
raw_headers={
"X-LiteLLM-API-Key": f"Bearer {self._ADMISSION_KEY}",
"authorization": f"Bearer {self._USER_TOKEN}",
},
user_api_key_auth=UserAPIKeyAuth(api_key="hashed-key", user_id="alice"),
)
assert subject_token == self._USER_TOKEN
@pytest.mark.asyncio
async def test_tools_call_with_an_empty_x_litellm_api_key_has_no_subject(self):
"""Admission ignores an empty ``x-litellm-api-key`` and validates ``Authorization`` instead."""
subject_token = await self._call_tool_subject(
self._token_exchange_server("te-empty-header"),
oauth2_headers={"Authorization": f"Bearer {self._ADMISSION_KEY}"},
raw_headers={"x-litellm-api-key": "", "authorization": f"Bearer {self._ADMISSION_KEY}"},
user_api_key_auth=UserAPIKeyAuth(api_key="hashed-key", user_id="alice"),
)
assert subject_token is None
@pytest.mark.asyncio
async def test_tools_call_with_the_same_litellm_key_in_both_headers_has_no_subject(self):
subject_token = await self._call_tool_subject(
self._token_exchange_server("te-same-key"),
oauth2_headers={"Authorization": f"Bearer {self._ADMISSION_KEY}"},
raw_headers={
"x-litellm-api-key": self._ADMISSION_KEY,
"authorization": f"Bearer {self._ADMISSION_KEY}",
},
user_api_key_auth=UserAPIKeyAuth(api_key="hashed-key", user_id="alice"),
)
assert subject_token is None
@pytest.mark.asyncio
async def test_tools_call_with_a_different_litellm_key_in_authorization_has_no_subject(self):
"""A second ``sk-`` virtual key next to ``x-litellm-api-key`` is still a gateway credential."""
subject_token = await self._call_tool_subject(
self._token_exchange_server("te-second-key"),
oauth2_headers={"Authorization": "Bearer sk-another-virtual-key"},
raw_headers={
"x-litellm-api-key": f"Bearer {self._ADMISSION_KEY}",
"authorization": "Bearer sk-another-virtual-key",
},
user_api_key_auth=UserAPIKeyAuth(api_key="hashed-key", user_id="alice"),
)
assert subject_token is None
@pytest.mark.asyncio
async def test_tools_call_exchanges_the_bearer_when_jwt_admission_left_api_key_unset(self):
subject_token = await self._call_tool_subject(
self._token_exchange_server("te-jwt"),
oauth2_headers={"Authorization": f"Bearer {self._USER_TOKEN}"},
raw_headers={"authorization": f"Bearer {self._USER_TOKEN}"},
user_api_key_auth=UserAPIKeyAuth(api_key=None, user_id="alice"),
)
assert subject_token == self._USER_TOKEN
@pytest.mark.asyncio
async def test_tools_list_with_only_the_litellm_key_has_no_subject(self):
manager: Final = self._manager_with_recording_client()
manager._fetch_tools_with_timeout = AsyncMock(return_value=[])
await manager._get_tools_from_server(
server=self._token_exchange_server("te-list-key"),
oauth2_headers={"Authorization": f"Bearer {self._ADMISSION_KEY}"},
raw_headers={"authorization": f"Bearer {self._ADMISSION_KEY}"},
user_api_key_auth=UserAPIKeyAuth(api_key="hashed-key", user_id="alice"),
)
assert self._subject_token_given_to_client(manager) is None
@pytest.mark.asyncio
async def test_prompts_list_with_only_the_litellm_key_has_no_subject(self):
manager: Final = self._manager_with_recording_client()
await manager.get_prompts_from_server(
server=self._token_exchange_server("te-prompts-key"),
user_api_key_auth=UserAPIKeyAuth(api_key="hashed-key", user_id="alice"),
raw_headers={"authorization": f"Bearer {self._ADMISSION_KEY}"},
)
assert self._subject_token_given_to_client(manager) is None
@pytest.mark.asyncio
async def test_resource_read_with_only_the_litellm_key_has_no_subject(self):
manager: Final = self._manager_with_recording_client()
await manager.read_resource_from_server(
server=self._token_exchange_server("te-read-key"),
user_api_key_auth=UserAPIKeyAuth(api_key="hashed-key", user_id="alice"),
url=AnyUrl("file:///notes.txt"),
raw_headers={"authorization": f"Bearer {self._ADMISSION_KEY}"},
)
assert self._subject_token_given_to_client(manager) is None
@pytest.mark.asyncio
async def test_resource_read_exchanges_the_user_token_when_x_litellm_api_key_admits(self):
manager: Final = self._manager_with_recording_client()
await manager.read_resource_from_server(
server=self._token_exchange_server("te-read-split"),
user_api_key_auth=UserAPIKeyAuth(api_key="hashed-key", user_id="alice"),
url=AnyUrl("file:///notes.txt"),
raw_headers={
"x-litellm-api-key": f"Bearer {self._ADMISSION_KEY}",
"authorization": f"Bearer {self._USER_TOKEN}",
},
)
assert self._subject_token_given_to_client(manager) == self._USER_TOKEN
@pytest.mark.asyncio
async def test_openapi_call_never_hands_the_litellm_key_to_the_exchanger(self):
provider: Final = self._recording_provider()
manager = MCPServerManager(cred_provider=provider)
server = MCPServer(
server_id="te-openapi",
name="te_openapi",
server_name="te_openapi",
url=None,
transport=MCPTransport.http,
auth_type=MCPAuth.oauth2_token_exchange,
token_exchange_endpoint="https://idp.example.com/token",
client_id="cid",
client_secret="csec",
spec_path="https://api.example.com/openapi.json",
)
user_auth = UserAPIKeyAuth(api_key="hashed-key", user_id="alice")
await manager.resolve_openapi_upstream_auth(
mcp_server=server,
oauth2_headers={"Authorization": f"Bearer {self._ADMISSION_KEY}"},
raw_headers={"authorization": f"Bearer {self._ADMISSION_KEY}"},
mcp_auth_header=None,
user_api_key_auth=user_auth,
forwarded_headers=None,
)
await manager.resolve_openapi_upstream_auth(
mcp_server=server,
oauth2_headers={"Authorization": f"Bearer {self._USER_TOKEN}"},
raw_headers={
"x-litellm-api-key": f"Bearer {self._ADMISSION_KEY}",
"authorization": f"Bearer {self._USER_TOKEN}",
},
mcp_auth_header=None,
user_api_key_auth=user_auth,
forwarded_headers=None,
)
assert self._subjects_seen_by(provider) == [None, self._USER_TOKEN]
@pytest.mark.asyncio
async def test_preflight_challenges_instead_of_exchanging_the_litellm_key(self):
provider: Final = self._recording_provider()
manager = MCPServerManager(cred_provider=provider)
with pytest.raises(HTTPException) as exc_info:
await manager.preflight_token_exchange(
server=self._token_exchange_server("te-preflight-key"),
oauth2_headers={"Authorization": f"Bearer {self._ADMISSION_KEY}"},
user_api_key_auth=UserAPIKeyAuth(api_key="hashed-key", user_id="alice"),
raw_headers={"authorization": f"Bearer {self._ADMISSION_KEY}"},
)
assert exc_info.value.status_code == 401
headers = exc_info.value.headers or {}
assert "resource_metadata" in (headers.get("WWW-Authenticate") or headers.get("www-authenticate") or "")
assert self._subjects_seen_by(provider) == []
@pytest.mark.asyncio
async def test_preflight_exchanges_the_user_token_when_x_litellm_api_key_admits(self):
provider: Final = self._recording_provider()
manager = MCPServerManager(cred_provider=provider)
await manager.preflight_token_exchange(
server=self._token_exchange_server("te-preflight-split"),
oauth2_headers={"Authorization": f"Bearer {self._USER_TOKEN}"},
user_api_key_auth=UserAPIKeyAuth(api_key="hashed-key", user_id="alice"),
raw_headers={
"x-litellm-api-key": f"Bearer {self._ADMISSION_KEY}",
"authorization": f"Bearer {self._USER_TOKEN}",
},
)
assert self._subjects_seen_by(provider) == [self._USER_TOKEN]

9
tests/test_litellm/proxy/guardrails/guardrail_hooks/test_crowdstrike_aidr.py
View file

@ -1703,8 +1703,8 @@ async def _guard_calls_for_stream(handler: CrowdStrikeAIDRHandler, chunk_texts:
@pytest.mark.parametrize(
("configured", "expected_calls"),
[
({}, 3),
({"streaming_sampling_rate": 2}, 6),
({}, 2),
({"streaming_sampling_rate": 2}, 5),
({"streaming_end_of_stream_only": True}, 1),
({"streaming_end_of_stream_only": True, "streaming_sampling_rate": 2}, 1),
],
@ -1712,7 +1712,10 @@ async def _guard_calls_for_stream(handler: CrowdStrikeAIDRHandler, chunk_texts:
async def test_streaming_params_from_config_control_output_scan_cadence(
configured: dict[str, object], expected_calls: int
) -> None:
"""10 chunks: default samples at 5 and 10 plus the final pass, rate 2 samples 5 times plus final, end-of-stream scans once."""
"""10 chunks: default samples at 5 and 10, rate 2 samples 5 times, end-of-stream scans once.
The final pass is skipped because chunk 10 already scanned the complete output.
"""
handler = _initialize_from_config(mode="post_call", **configured)
assert await _guard_calls_for_stream(handler, list("ABCDEFGHIJ")) == expected_calls

1151
tests/test_litellm/proxy/guardrails/guardrail_hooks/test_model_armor.py
View file

File diff suppressed because it is too large Load diff

33
tests/test_litellm/proxy/management_endpoints/test_saml_sso.py
View file

@ -516,6 +516,39 @@ async def test_team_ids_extracted_from_groups_attribute(saml_env_idp_initiated):
assert result.team_ids == ["team-a", "team-b"]
@pytest.mark.asyncio
@pytest.mark.parametrize(
"roles",
[
["internal_user", "proxy_admin_viewer"],
["proxy_admin_viewer", "internal_user"],
],
)
async def test_multi_valued_role_attribute_resolves_to_highest_privilege(saml_env_idp_initiated, roles):
"""An assertion carrying several roles must not depend on the order the IdP emitted them in."""
key_pem, cert_pem = saml_env_idp_initiated
resp = _build_signed_response(
key_pem,
cert_pem,
attributes={
"email": ["dave@example.com"],
"role": roles,
},
)
result = await _acs(_b64(resp), _shared_cache())
assert result.user_role == LitellmUserRoles.PROXY_ADMIN_VIEW_ONLY
@pytest.mark.asyncio
async def test_assertion_without_role_attribute_has_no_user_role(saml_env_idp_initiated):
key_pem, cert_pem = saml_env_idp_initiated
resp = _build_signed_response(key_pem, cert_pem, attributes={"email": ["erin@example.com"]})
result = await _acs(_b64(resp), _shared_cache())
assert result.user_role is None
@pytest.mark.asyncio
async def test_build_login_redirect_targets_idp_and_caches_request_id(saml_env):
cache = DualCache()

129
tests/test_litellm/proxy/management_endpoints/test_ui_sso.py
View file

@ -6598,13 +6598,94 @@ def test_get_litellm_user_role_with_invalid_role():
assert result is None
def test_get_litellm_user_role_with_list_multiple_roles():
"""Test that get_litellm_user_role takes the first element from a multi-element list."""
@pytest.mark.parametrize(
"role_claim",
[
["proxy_admin", "internal_user"],
["internal_user", "proxy_admin"],
],
)
def test_get_litellm_user_role_picks_highest_privilege_regardless_of_order(role_claim):
"""A multi-valued role claim resolves to the most privileged role, not the first one listed."""
from litellm.proxy._types import LitellmUserRoles
from litellm.proxy.management_endpoints.types import get_litellm_user_role
result = get_litellm_user_role(["proxy_admin", "internal_user"])
assert result == LitellmUserRoles.PROXY_ADMIN
assert get_litellm_user_role(role_claim) == LitellmUserRoles.PROXY_ADMIN
@pytest.mark.parametrize(
"role_claim",
[
["proxy_admin_viewer", "internal_user"],
["internal_user", "proxy_admin_viewer"],
],
)
def test_get_litellm_user_role_keeps_org_spend_visibility_for_mixed_roles(role_claim):
"""
Regression for LIT-6077: a user holding both proxy_admin_viewer and internal_user kept
losing org-level spend visibility whenever the IdP happened to list internal_user first.
"""
from litellm.proxy._types import LitellmUserRoles
from litellm.proxy.management_endpoints.types import get_litellm_user_role
assert get_litellm_user_role(role_claim) == LitellmUserRoles.PROXY_ADMIN_VIEW_ONLY
def test_get_litellm_user_role_ignores_unrecognised_entries():
"""Roles LiteLLM does not know about are skipped rather than swallowing the whole claim."""
from litellm.proxy._types import LitellmUserRoles
from litellm.proxy.management_endpoints.types import get_litellm_user_role
assert get_litellm_user_role(["some_idp_group", "internal_user"]) == LitellmUserRoles.INTERNAL_USER
assert get_litellm_user_role(["some_idp_group", "another_group"]) is None
def test_get_litellm_user_role_list_lookup_is_case_insensitive():
from litellm.proxy._types import LitellmUserRoles
from litellm.proxy.management_endpoints.types import get_litellm_user_role
assert get_litellm_user_role(["INTERNAL_USER", "Proxy_Admin"]) == LitellmUserRoles.PROXY_ADMIN
@pytest.mark.parametrize(
"role_claim",
[
["org_admin", "team"],
["team", "org_admin"],
],
)
def test_get_litellm_user_role_is_deterministic_for_unranked_roles(role_claim):
"""Roles outside the privilege hierarchy still resolve the same way in either claim order."""
from litellm.proxy._types import LitellmUserRoles
from litellm.proxy.management_endpoints.types import get_litellm_user_role
assert get_litellm_user_role(role_claim) == LitellmUserRoles.ORG_ADMIN
@pytest.mark.parametrize(
"role_claim",
[
["org_admin", "internal_user"],
["internal_user", "org_admin"],
],
)
def test_get_litellm_user_role_prefers_a_ranked_role_over_an_unranked_one(role_claim):
"""
org_admin, team and customer sit outside the privilege ladder, so a claim mixing one of
them with a ranked role settles on the ranked role in either order. Same rule the Entra
app_roles and role_mappings paths already follow.
"""
from litellm.proxy._types import LitellmUserRoles
from litellm.proxy.management_endpoints.types import get_litellm_user_role
assert get_litellm_user_role(role_claim) == LitellmUserRoles.INTERNAL_USER
def test_get_litellm_user_role_returns_none_for_non_string_claims():
from litellm.proxy.management_endpoints.types import get_litellm_user_role
assert get_litellm_user_role(None) is None
assert get_litellm_user_role({"role": "proxy_admin"}) is None
# ============================================================================
@ -6654,6 +6735,46 @@ def test_process_sso_jwt_access_token_extracts_role_from_access_token():
assert result.user_role == LitellmUserRoles.PROXY_ADMIN
@pytest.mark.parametrize(
"role_claim",
[
["internal_user", "proxy_admin_viewer"],
["proxy_admin_viewer", "internal_user"],
],
)
def test_process_sso_jwt_access_token_resolves_highest_privilege_role(role_claim):
"""
The generic SSO access-token path must land on the same role for a user whose role
claim holds several roles, whichever order the IdP emitted them in.
"""
import jwt as pyjwt
from litellm.proxy._types import LitellmUserRoles
access_token_str = pyjwt.encode(
{"sub": "user-123", "email": "mixed@test.com", "litellm_role": role_claim},
"secret",
algorithm="HS256",
)
result = CustomOpenID(
id="user-123",
email="mixed@test.com",
display_name="Mixed Role User",
team_ids=[],
user_role=None,
)
with patch.dict(os.environ, {"GENERIC_USER_ROLE_ATTRIBUTE": "litellm_role"}):
process_sso_jwt_access_token(
access_token_str=access_token_str,
sso_jwt_handler=None,
result=result,
role_mappings=None,
)
assert result.user_role == LitellmUserRoles.PROXY_ADMIN_VIEW_ONLY
def test_process_sso_jwt_access_token_does_not_override_existing_role():
"""
Test that process_sso_jwt_access_token does NOT override a role that was

10
tests/test_litellm/proxy/public_endpoints/test_public_endpoints.py
View file

@ -1104,6 +1104,16 @@ def test_get_autorouter_presets_local_mode_serves_bundled_catalog(
assert response.status_code == 200
payload = response.json()
assert "anthropic_family" in payload
assert payload["1m_context"]["complexity_router_config"]["classifier_type"] == "heuristic_v2"
assert payload["1m_context"]["complexity_router_config"]["tiers"] == {
"SIMPLE": ["gpt-5.6-luna"],
"MEDIUM": ["gpt-5.6-terra"],
"COMPLEX": ["claude-opus-5"],
"REASONING": ["claude-opus-5"],
}
assert payload["1m_context"]["complexity_router_config"]["tier_model_configs"] == {
"REASONING": [{"model_name": "claude-opus-5", "litellm_params": {"reasoning_effort": "high"}}]
}
for preset in payload.values():
assert isinstance(preset["label"], str)
assert isinstance(preset["description"], str)

68
tests/test_litellm/router_strategy/test_complexity_router.py
View file

@ -10797,6 +10797,9 @@ class TestModalityRouting:
("custom_tiers_walk", "premium-model", "modality_escalation"),
("pin_kept_bypasses", "text-cheap", "session_affinity_pin"),
("pin_replacement_gated", "vision-big", "modality_escalation"),
("pin_override_escalates", "vision-mid", "modality_pin_override"),
("pin_override_same_tier", "vision-cheap", "modality_pin_override"),
("pin_override_inert_without_modality_routing", "text-cheap", "session_affinity_pin"),
("adaptive_pick_rewritten", "vision-mid", "modality_escalation"),
],
)
@ -10847,7 +10850,7 @@ class TestModalityRouting:
messages = [
{"role": "user", "content": [{"type": "text", "text": "quick lookup: what is this?"}, IMG_PART]}
]
elif path in ("pin_kept_bypasses", "pin_replacement_gated"):
elif path.startswith(("pin_kept", "pin_replacement", "pin_override")):
cache = AsyncMock()
cache.async_get_cache = AsyncMock(return_value={"model": "text-cheap", "tier": "SIMPLE"})
mock_router_instance.cache = cache
@ -10859,6 +10862,13 @@ class TestModalityRouting:
messages = [
{"role": "user", "content": [{"type": "text", "text": "LITELLM ESCALATE describe this"}, IMG_PART]}
]
elif path == "pin_override_same_tier":
config["modality_pin_override"] = True
config["tiers"]["SIMPLE"] = ["text-cheap", "vision-cheap"]
vision["vision-cheap"] = True
elif path == "pin_override_inert_without_modality_routing":
config["modality_routing"] = False
config["modality_pin_override"] = path.startswith("pin_override")
elif path == "adaptive_pick_rewritten":
config["adaptive"] = True
mock_router_instance.model_list = []
@ -11025,4 +11035,60 @@ class TestModalityRouting:
from litellm.router_strategy.complexity_router.complexity_router import _decision_is_pinnable
assert _decision_is_pinnable({"cause": "modality_escalation"}) is False
assert _decision_is_pinnable({"cause": "modality_pin_override"}) is False
assert _decision_is_pinnable({"cause": "heuristic_scorer"}) is True
@pytest.mark.asyncio
async def test_pin_override_serves_the_image_turn_without_repinning(self, mock_router_instance):
"""The override is for one request: the session keeps the model it was pinned to."""
cache = AsyncMock()
cache.async_get_cache = AsyncMock(return_value={"model": "text-cheap", "tier": "SIMPLE"})
mock_router_instance.cache = cache
router = self._router(
mock_router_instance,
{
"tiers": dict(self.BASE_TIERS),
"modality_routing": True,
"modality_pin_override": True,
"session_affinity": True,
},
dict(self.BASE_VISION),
)
request_kwargs = {"metadata": {"session_id": "s1"}}
image_turn = await router.async_pre_routing_hook(
model="m", request_kwargs=request_kwargs, messages=self.IMAGE_MESSAGE
)
assert image_turn.model == "vision-mid"
assert image_turn.routing_decision["cause"] == "modality_pin_override"
assert "modality_escalated_from:SIMPLE" in image_turn.routing_decision["signals"]
assert cache.async_set_cache.await_args.kwargs["value"] == {"model": "text-cheap", "tier": "SIMPLE"}
text_turn = await router.async_pre_routing_hook(
model="m", request_kwargs={"metadata": {"session_id": "s1"}}, messages=[{"role": "user", "content": "hi"}]
)
assert text_turn.model == "text-cheap"
assert text_turn.routing_decision["cause"] == "session_affinity_pin"
@pytest.mark.asyncio
async def test_pin_override_with_no_capable_model_rejects_and_keeps_the_pin(self, mock_router_instance):
"""The clear 400 replaces the provider's, and a rejected turn must not cost the session its pin."""
cache = AsyncMock()
cache.async_get_cache = AsyncMock(return_value={"model": "text-cheap", "tier": "SIMPLE"})
mock_router_instance.cache = cache
router = self._router(
mock_router_instance,
{
"tiers": {"SIMPLE": "text-cheap", "COMPLEX": "text-big"},
"modality_routing": True,
"modality_pin_override": True,
"session_affinity": True,
},
{"text-cheap": False, "text-big": False},
)
with pytest.raises(litellm.BadRequestError, match="no model"):
await router.async_pre_routing_hook(
model="m", request_kwargs={"metadata": {"session_id": "s1"}}, messages=self.IMAGE_MESSAGE
)
assert cache.async_set_cache.await_args.kwargs["value"] == {"model": "text-cheap", "tier": "SIMPLE"}

30
tests/test_litellm/test_main.py
View file

@ -3131,9 +3131,9 @@ def test_stream_chunk_builder_prices_proxy_alias_via_model_map():
assert response._hidden_params["response_cost"] == pytest.approx(expected_cost)
def _stream_builder_logging_obj() -> LiteLLMLogging:
def _stream_builder_logging_obj(model: str = "gpt-4o", custom_llm_provider: str = "openai") -> LiteLLMLogging:
logging_obj: Final = LiteLLMLogging(
model="gpt-4o",
model=model,
messages=[{"role": "user", "content": "hi"}],
stream=True,
call_type="completion",
@ -3142,10 +3142,11 @@ def _stream_builder_logging_obj() -> LiteLLMLogging:
function_id="test-function-id",
)
logging_obj.update_environment_variables(
model="gpt-4o",
model=model,
user=None,
optional_params={},
litellm_params={"custom_llm_provider": "openai"},
litellm_params={"custom_llm_provider": custom_llm_provider},
custom_llm_provider=custom_llm_provider,
)
return logging_obj
@ -3237,3 +3238,24 @@ def test_stream_chunk_builder_prices_alias_from_openai_sdk_usage_chunk():
assert response.usage.completion_tokens == 60
assert getattr(response.usage, "cost", None) == pytest.approx(0.000704)
assert response._hidden_params["response_cost"] == pytest.approx(0.000704)
def test_stream_chunk_builder_leaves_xai_reported_cost_to_the_calculator(monkeypatch: pytest.MonkeyPatch):
monkeypatch.setattr(litellm, "cost_margin_config", {"xai": 0.5})
usage_chunk: Final = _stream_builder_text_chunk("grok-4", "")
usage_chunk.usage = Usage(prompt_tokens=5, completion_tokens=2, total_tokens=7, cost=0.42)
chunks: Final = [
_stream_builder_text_chunk("grok-4", "Hello "),
_stream_builder_text_chunk("grok-4", "world.", finish_reason="stop"),
usage_chunk,
]
logging_obj: Final = _stream_builder_logging_obj(model="grok-4", custom_llm_provider="xai")
response: Final = litellm.stream_chunk_builder(
chunks=chunks, messages=[{"role": "user", "content": "hi"}], logging_obj=logging_obj
)
assert response is not None
assert getattr(response.usage, "cost", None) == pytest.approx(0.42)
assert response._hidden_params.get("response_cost") is None
assert logging_obj._response_cost_calculator(result=response) == pytest.approx(0.63)

25
tests/test_litellm/vector_stores/test_main.py
View file

@ -2,14 +2,17 @@
Tests for litellm/vector_stores/main.py.
Pins the router threading contract for vector store search: the router is an
explicit named parameter that reaches the HTTP handler, and it must never leak
into litellm_params/kwargs where logging would model_dump() it (the #19550
serialization trap).
explicit named parameter that reaches the HTTP handler wrapped in the embedding
executor, and it must never leak into litellm_params/kwargs where logging would
model_dump() it (the #19550 serialization trap).
"""
from unittest.mock import MagicMock, patch
import litellm.vector_stores.main as vector_stores_main
from litellm.llms.base_llm.vector_store.transformation import (
RouterVectorStoreEmbeddingExecutor,
)
from litellm.vector_stores.main import search
MOCK_SEARCH_RESPONSE = {
@ -19,17 +22,18 @@ MOCK_SEARCH_RESPONSE = {
}
def test_search_threads_router_to_handler():
"""search() must pass its router param through to the HTTP handler"""
def test_search_wraps_router_into_the_handler_embedding_executor():
"""search() hands the HTTP handler a Router-backed embedding executor carrying the
request metadata, and no bare router kwarg (LIT-6750)"""
mock_router = MagicMock()
logger = MagicMock()
with (
patch( # test-quality-ok: stubs provider config resolution; the seam under test is the router kwarg threading
patch( # test-quality-ok: stubs provider config resolution; the seam under test is the executor threading
"litellm.vector_stores.main.ProviderConfigManager.get_provider_vector_stores_config",
return_value=MagicMock(),
),
patch.object( # test-quality-ok: the handler call is the observable boundary for the router kwarg contract
patch.object( # test-quality-ok: the handler call is the observable boundary for the executor contract
vector_stores_main.base_llm_http_handler,
"vector_store_search_handler",
return_value=MOCK_SEARCH_RESPONSE,
@ -41,11 +45,16 @@ def test_search_threads_router_to_handler():
custom_llm_provider="s3_vectors",
router=mock_router,
litellm_logging_obj=logger,
litellm_metadata={"user_api_key_team_id": "team-a"},
)
assert response == MOCK_SEARCH_RESPONSE
mock_handler.assert_called_once()
assert mock_handler.call_args.kwargs["router"] is mock_router
assert "router" not in mock_handler.call_args.kwargs
executor = mock_handler.call_args.kwargs["embedding_executor"]
assert isinstance(executor, RouterVectorStoreEmbeddingExecutor)
assert executor.router is mock_router
assert dict(executor.metadata) == {"user_api_key_team_id": "team-a"}
def test_search_router_not_in_litellm_params():

71
tests/test_rust_python_harness.py
View file

@ -8,14 +8,24 @@ import pytest
catalog = importlib.import_module("tests.rust-python-harness.catalog")
cli = importlib.import_module("tests.rust-python-harness.cli")
ledger_module = importlib.import_module("tests.rust-python-harness.shared.parity.ledger")
mapping_validator = importlib.import_module(
"tests.rust-python-harness.strategies.unit_tests.mapping_validator"
)
models = importlib.import_module("tests.rust-python-harness.models")
runner = importlib.import_module("tests.rust-python-harness.runner")
ui = importlib.import_module("tests.rust-python-harness.ui")
load_catalog = catalog.load_catalog
load_ledger = ledger_module.load_ledger
ledger_path_for = mapping_validator.ledger_path_for
REPO_ROOT = mapping_validator.REPO_ROOT
audit_ledger = mapping_validator.audit_ledger
build_function_report = mapping_validator.build_function_report
_pick_values = cli._pick_values
_coverage_pytest_args = cli._coverage_pytest_args
_select = cli._select
_validate_ledger = cli._validate_ledger
CaseResult = models.CaseResult
Coverage = models.Coverage
HarnessCase = models.HarnessCase
@ -56,13 +66,14 @@ def _manifest() -> dict[str, object]:
}
def test_should_load_the_three_harness_strategies_in_order() -> None:
def test_should_load_the_four_harness_strategies_in_order() -> None:
strategies = load_catalog()
assert [strategy.id for strategy in strategies] == [
"e2e_fuzz_tests",
"unit_tests_rust",
"validate_sub_methods",
"existing_e2e_test_sdk",
]
assert all(
tuple(case.sdk_function for case in strategy.cases) == SDK_FUNCTIONS
@ -94,6 +105,8 @@ def test_should_reject_a_manifest_missing_an_sdk_function(tmp_path: Path) -> Non
True,
),
("tests/test_parity.py::test_one", "tests/test_parity.py::test_two", False),
("tests/ocr_tests/", "tests/ocr_tests/test_ocr_mistral.py::test_one", True),
("tests/ocr_tests/", "tests/other_tests/test_ocr_mistral.py::test_one", False),
],
)
def test_should_match_pytest_file_and_node_selectors(
@ -113,6 +126,13 @@ def test_should_only_return_selectors_whose_files_exist(tmp_path: Path) -> None:
assert runnable_selectors((case,), tmp_path) == ("tests/test_parity.py",)
def test_should_treat_an_existing_folder_selector_as_runnable(tmp_path: Path) -> None:
(tmp_path / "tests" / "ocr_tests").mkdir(parents=True)
case = _case(selectors=("tests/ocr_tests/",))
assert runnable_selectors((case,), tmp_path) == ("tests/ocr_tests/",)
def test_should_mark_planned_and_not_applicable_cases_without_running() -> None:
planned = CaseResult(case=_case(coverage=Coverage.PLANNED))
not_applicable = CaseResult(case=_case(coverage=Coverage.NOT_APPLICABLE))
@ -229,8 +249,53 @@ def test_should_report_confidence_for_each_sdk_section() -> None:
}
assert scores["responses"].verified_strategies == 1
assert scores["responses"].required_strategies == 3
assert scores["responses"].percentage == 33
assert scores["responses"].required_strategies == 4
assert scores["responses"].percentage == 25
assert scores["responses"].level.value == "MEDIUM"
assert scores["count_tokens"].percentage == 0
assert scores["count_tokens"].level.value == "LOW"
def test_should_report_no_ledger_for_a_function_without_one() -> None:
report = build_function_report("messages", repo_root=REPO_ROOT)
assert report.has_ledger is False
assert report.is_clean is True
def test_should_report_ocr_ledger_stats_and_a_clean_audit() -> None:
ledger = load_ledger(ledger_path_for("ocr"))
report = build_function_report("ocr", repo_root=REPO_ROOT)
assert report.has_ledger is True
assert report.ledger.mapped_count == ledger.mapped_count
assert report.ledger.total_count == ledger.total_count
assert report.is_clean is True
def test_should_scope_validate_ledger_to_the_requested_function(
capsys: pytest.CaptureFixture[str],
) -> None:
exit_code = _validate_ledger({"messages"})
captured = capsys.readouterr()
assert exit_code == 0
assert "messages" in captured.out
assert "no ledger yet" in captured.out
assert "ocr" not in captured.out
def test_should_have_every_python_and_rust_ocr_test_accounted_for_in_the_ledger() -> None:
ledger = load_ledger(ledger_path_for("ocr"))
report = audit_ledger(ledger, repo_root=REPO_ROOT)
assert report.is_clean, (
"\nOCR test-parity ledger is out of sync with the live test files.\n"
f"Ledger references a Python test that no longer exists: {list(report.missing_python_tests)}\n"
f"Python test exists but is not tracked in the ledger: {list(report.stale_python_tests)}\n"
f"Ledger references a Rust test that no longer exists: {list(report.missing_rust_tests)}\n"
f"Rust test exists but is not tracked in the ledger: {list(report.stale_rust_tests)}\n"
)

12
type-discipline-budget.json
View file

@ -1,9 +1,9 @@
{
"LIT001": {
"limit": 22192
"limit": 22330
},
"LIT002": {
"limit": 26762
"limit": 26763
},
"LIT003": {
"limit": 261
@ -15,7 +15,7 @@
"limit": 0
},
"LIT006": {
"limit": 1035
"limit": 1038
},
"LIT007": {
"limit": 0
@ -27,12 +27,12 @@
"limit": 0
},
"LIT010": {
"limit": 16480
"limit": 16477
},
"LIT011": {
"limit": 5518
"limit": 5519
},
"LIT012": {
"limit": 4488
"limit": 4489
}
}

3
ui/litellm-dashboard/eslint-suppressions.json
View file

@ -1399,9 +1399,6 @@
},
"prefer-const": {
"count": 2
},
"react-hooks/set-state-in-effect": {
"count": 1
}
},
"src/components/TeamsPage/teamTableColumns.tsx": {

202
ui/litellm-dashboard/src/components/Teams.test.tsx
View file

@ -15,6 +15,7 @@ import {
teamCreateCall,
} from "./networking";
import Teams from "./Teams";
import { chooseSelectOption } from "../../tests/test-utils";
const can = vi.fn();
vi.mock("@/app/(dashboard)/hooks/useCan", () => ({
@ -1488,3 +1489,204 @@ describe("Teams - the exact bytes the create call sends", () => {
expect(teamCreateCall).not.toHaveBeenCalled();
});
});
describe("Teams - the create form keeps the organization and models picks while it is open", () => {
const ORGS = [
{ organization_id: "org-1", organization_alias: "Org 1", models: [], members: [] },
{ organization_id: "org-2", organization_alias: "Org 2", models: [], members: [] },
];
const orgField = () => screen.getByRole("combobox", { name: /organization/i });
const modelsField = () => screen.getByTestId("create-team-models-select");
const openCreateModal = async () => {
act(() => {
fireEvent.click(screen.getAllByRole("button", { name: /create team/i })[0]);
});
await screen.findByLabelText(/team name/i);
};
beforeEach(() => {
vi.clearAllMocks();
mockTeamInfoView.mockClear();
vi.mocked(fetchAvailableModelsForTeamOrKey).mockResolvedValue(["gpt-4", "gpt-3.5-turbo"]);
vi.mocked(fetchMCPAccessGroups).mockResolvedValue([]);
vi.mocked(getGuardrailsList).mockResolvedValue({ guardrails: [] });
vi.mocked(getDefaultTeamSettings).mockResolvedValue({ values: {} });
mockUseOrganizations.mockReturnValue({ data: ORGS });
});
it("keeps both picks when the organizations list comes back changed from a refetch", async () => {
const user = userEvent.setup();
renderWithQueryClient(<Teams accessToken="test-token" userID="user-123" userRole="Admin" />);
await openCreateModal();
await chooseSelectOption(user, orgField(), /Org 1/);
fireEvent.change(modelsField(), { target: { value: "gpt-4" } });
mockUseOrganizations.mockReturnValue({ data: ORGS.map((org) => ({ ...org, spend: 1 })) });
fireEvent.click(screen.getByText("Additional Settings"));
expect(orgField()).toHaveValue("Org 1");
expect(modelsField()).toHaveValue("gpt-4");
});
it("keeps models picked before the available models finish loading", async () => {
let resolveModels: (models: string[]) => void = () => {};
vi.mocked(fetchAvailableModelsForTeamOrKey).mockReturnValue(
new Promise<string[]>((resolve) => {
resolveModels = resolve;
}),
);
renderWithQueryClient(<Teams accessToken="test-token" userID="user-123" userRole="Admin" />);
await openCreateModal();
fireEvent.change(modelsField(), { target: { value: "gpt-4" } });
await act(async () => {
resolveModels(["gpt-4", "gpt-3.5-turbo"]);
});
expect(modelsField()).toHaveValue("gpt-4");
});
it("clears the models pick when the organization is changed, since models are org scoped", async () => {
const user = userEvent.setup();
renderWithQueryClient(<Teams accessToken="test-token" userID="user-123" userRole="Admin" />);
await openCreateModal();
await chooseSelectOption(user, orgField(), /Org 1/);
fireEvent.change(modelsField(), { target: { value: "gpt-4" } });
await chooseSelectOption(user, orgField(), /Org 2/);
await waitFor(() => expect(orgField()).toHaveValue("Org 2"));
expect(modelsField()).toHaveValue("");
});
it("keeps the models pick when the same organization is chosen again", async () => {
const user = userEvent.setup();
renderWithQueryClient(<Teams accessToken="test-token" userID="user-123" userRole="Admin" />);
await openCreateModal();
await chooseSelectOption(user, orgField(), /Org 1/);
fireEvent.change(modelsField(), { target: { value: "gpt-4" } });
await chooseSelectOption(user, orgField(), /Org 1/);
expect(orgField()).toHaveValue("Org 1");
expect(modelsField()).toHaveValue("gpt-4");
});
it("still preselects the only organization an org admin can create teams in", async () => {
mockUseOrganizations.mockReturnValue({
data: [
{
organization_id: "org-1",
organization_alias: "Org 1",
models: [],
members: [{ user_id: "user-123", user_role: "org_admin" }],
},
],
});
renderWithQueryClient(<Teams accessToken="test-token" userID="user-123" userRole="Internal User" />);
await openCreateModal();
expect(orgField()).toHaveValue("Org 1");
expect(orgField()).toBeDisabled();
});
it("leaves an org admin able to pick when their admin orgs narrow to one while the form is open", async () => {
const orgAdminOrgs = [
{
organization_id: "org-1",
organization_alias: "Org 1",
models: [],
members: [{ user_id: "user-123", user_role: "org_admin" }],
},
{
organization_id: "org-2",
organization_alias: "Org 2",
models: [],
members: [{ user_id: "user-123", user_role: "org_admin" }],
},
];
mockUseOrganizations.mockReturnValue({ data: orgAdminOrgs });
renderWithQueryClient(<Teams accessToken="test-token" userID="user-123" userRole="Internal User" />);
await openCreateModal();
expect(orgField()).toHaveValue("");
mockUseOrganizations.mockReturnValue({ data: [orgAdminOrgs[0]] });
fireEvent.click(screen.getByText("Additional Settings"));
expect(orgField()).toBeEnabled();
});
it("refuses to create the team in an organization the admin has lost access to", async () => {
const user = userEvent.setup();
const orgAdminOrgs = ORGS.map((org) => ({ ...org, members: [{ user_id: "user-123", user_role: "org_admin" }] }));
mockUseOrganizations.mockReturnValue({ data: orgAdminOrgs });
renderWithQueryClient(<Teams accessToken="test-token" userID="user-123" userRole="Internal User" />);
await openCreateModal();
fireEvent.change(screen.getByTestId("team-name-input"), { target: { value: "Revoked Team" } });
await chooseSelectOption(user, orgField(), /Org 1/);
mockUseOrganizations.mockReturnValue({ data: [orgAdminOrgs[1]] });
fireEvent.click(screen.getByText("Additional Settings"));
const submitButtons = screen.getAllByRole("button", { name: /create team/i });
fireEvent.click(submitButtons[submitButtons.length - 1]);
await screen.findByText(/no longer create teams in this organization/i);
expect(teamCreateCall).not.toHaveBeenCalled();
});
it("lets the admin switch to the one organization left after losing access to their pick", async () => {
const user = userEvent.setup();
const orgAdminOrgs = ORGS.map((org) => ({ ...org, members: [{ user_id: "user-123", user_role: "org_admin" }] }));
mockUseOrganizations.mockReturnValue({ data: orgAdminOrgs });
const createdTeam = {
team_id: "new-team-1",
team_alias: "Recovered Team",
models: [],
organization_id: "org-2",
keys: [],
members_with_roles: [],
spend: 0,
};
vi.mocked(teamCreateCall).mockResolvedValue(createdTeam);
renderWithQueryClient(<Teams accessToken="test-token" userID="user-123" userRole="Internal User" />);
await openCreateModal();
fireEvent.change(screen.getByTestId("team-name-input"), { target: { value: "Recovered Team" } });
await chooseSelectOption(user, orgField(), /Org 1/);
mockUseOrganizations.mockReturnValue({ data: [orgAdminOrgs[1]] });
fireEvent.click(screen.getByText("Additional Settings"));
expect(orgField()).toBeEnabled();
await chooseSelectOption(user, orgField(), /Org 2/);
const submitButtons = screen.getAllByRole("button", { name: /create team/i });
fireEvent.click(submitButtons[submitButtons.length - 1]);
await waitFor(() =>
expect(teamCreateCall).toHaveBeenCalledWith(
"test-token",
expect.objectContaining({ team_alias: "Recovered Team", organization_id: "org-2" }),
),
);
});
it("starts the form clean again when the modal is closed and reopened", async () => {
const user = userEvent.setup();
renderWithQueryClient(<Teams accessToken="test-token" userID="user-123" userRole="Admin" />);
await openCreateModal();
await chooseSelectOption(user, orgField(), /Org 1/);
fireEvent.change(modelsField(), { target: { value: "gpt-4" } });
fireEvent.click(screen.getByRole("button", { name: /^close$/i }));
await waitFor(() => expect(screen.queryByLabelText(/team name/i)).not.toBeInTheDocument());
await openCreateModal();
expect(orgField()).toHaveValue("");
expect(modelsField()).toHaveValue("");
});
});

74
ui/litellm-dashboard/src/components/Teams.tsx
View file

@ -208,7 +208,6 @@ const Teams: React.FC<TeamProps> = ({ accessToken, userID, userRole, premiumUser
const queryClient = useQueryClient();
const refreshTeams = () => queryClient.invalidateQueries({ queryKey: teamsTableKeys.all });
const [currentOrg] = useState<Organization | null>(null);
const [currentOrgForCreateTeam, setCurrentOrgForCreateTeam] = useState<Organization | null>(null);
const isOrgAdmin = userRole !== "Admin";
const [additionalSettingsOpen, setAdditionalSettingsOpen] = useState(false);
@ -216,17 +215,33 @@ const Teams: React.FC<TeamProps> = ({ accessToken, userID, userRole, premiumUser
const [agentSettingsOpen, setAgentSettingsOpen] = useState(false);
const [searchToolSettingsOpen, setSearchToolSettingsOpen] = useState(false);
const adminOrgs = useMemo(
() => getAdminOrganizations(userRole, userID, organizations),
[userRole, userID, organizations],
);
const teamCreateSchema = useMemo(
() =>
teamCreateFieldsSchema.superRefine((values, ctx) => {
if (isOrgAdmin && !values.organization_id) {
ctx.addIssue({ code: "custom", message: SUPPRESSED_BY_DESCRIPTION, path: ["organization_id"] });
}
const organizationIsStillPickable =
values.organization_id == null ||
organizations == null ||
adminOrgs.some((org) => org.organization_id === values.organization_id);
if (!organizationIsStillPickable) {
ctx.addIssue({
code: "custom",
message: "You can no longer create teams in this organization",
path: ["organization_id"],
});
}
if (additionalSettingsOpen && !isParsableJson(values.secret_manager_settings)) {
ctx.addIssue({ code: "custom", message: SUPPRESSED_BY_DESCRIPTION, path: ["secret_manager_settings"] });
}
}),
[isOrgAdmin, additionalSettingsOpen],
[isOrgAdmin, additionalSettingsOpen, adminOrgs, organizations],
);
const form = useZodForm(teamCreateSchema, { defaultValues: EMPTY_TEAM_CREATE_VALUES });
@ -264,28 +279,6 @@ const Teams: React.FC<TeamProps> = ({ accessToken, userID, userRole, premiumUser
? `Default: ${getBudgetDurationLabel(defaultBudgetDuration)} (${defaultBudgetDuration})`
: "n/a";
useEffect(() => {
form.setValue("models", []);
}, [currentOrgForCreateTeam, userModels]);
// Handle organization preselection when modal opens
useEffect(() => {
if (isTeamModalVisible) {
const adminOrgs = getAdminOrganizations(userRole, userID, organizations);
// Org admins must scope a team to an org, so with exactly one we preselect it.
// Proxy admins can create org-less teams, so the field stays optional regardless of org count.
if (isOrgAdmin && adminOrgs.length === 1) {
const org = adminOrgs[0];
form.setValue("organization_id", org.organization_id);
setCurrentOrgForCreateTeam(org);
} else {
form.setValue("organization_id", currentOrg?.organization_id || null);
setCurrentOrgForCreateTeam(currentOrg);
}
}
}, [isTeamModalVisible, isOrgAdmin, userRole, userID, organizations, currentOrg]);
// Add this useEffect to fetch guardrails
useEffect(() => {
const fetchGuardrails = async () => {
@ -320,6 +313,26 @@ const Teams: React.FC<TeamProps> = ({ accessToken, userID, userRole, premiumUser
if (canViewPolicies) fetchPolicies();
}, [accessToken, canViewPolicies]);
const openCreateTeamModal = () => {
// Org admins must scope a team to an org, so with exactly one we preselect it.
// Proxy admins can create org-less teams, so the field stays optional regardless of org count.
if (isOrgAdmin && adminOrgs.length === 1) {
form.setValue("organization_id", adminOrgs[0].organization_id);
}
setIsTeamModalVisible(true);
};
const selectCreateTeamOrganization = (
next: string,
currentOrganizationId: string | null,
onChange: (organizationId: string | null) => void,
) => {
const nextOrganizationId = next === "" ? null : next;
if (nextOrganizationId === currentOrganizationId) return;
onChange(nextOrganizationId);
form.setValue("models", []);
};
const resetCreateForm = () => {
form.reset(EMPTY_TEAM_CREATE_VALUES);
setAdditionalSettingsOpen(false);
@ -636,7 +649,7 @@ const Teams: React.FC<TeamProps> = ({ accessToken, userID, userRole, premiumUser
subtitle="Manage teams, members, and their access to models and budgets"
primaryAction={
canCreateOrManageTeams(userRole, userID, organizations) ? (
<UIButton onClick={() => setIsTeamModalVisible(true)} data-testid="create-team-button">
<UIButton onClick={openCreateTeamModal} data-testid="create-team-button">
<Plus className="size-4" />
Create Team
</UIButton>
@ -683,9 +696,9 @@ const Teams: React.FC<TeamProps> = ({ accessToken, userID, userRole, premiumUser
)}
</FormField>
{(() => {
const adminOrgs = getAdminOrganizations(userRole, userID, organizations);
const isSingleOrg = adminOrgs.length === 1;
const hasNoOrgs = adminOrgs.length === 0;
const soleOrganizationId = isSingleOrg ? adminOrgs[0].organization_id ?? null : null;
return (
<>
@ -715,18 +728,13 @@ const Teams: React.FC<TeamProps> = ({ accessToken, userID, userRole, premiumUser
label: org.organization_alias ?? "",
sublabel: org.organization_id ?? "",
}))}
disabled={isOrgAdmin && isSingleOrg}
disabled={isOrgAdmin && soleOrganizationId !== null && value === soleOrganizationId}
allowClear={!isOrgAdmin}
placeholder={
hasNoOrgs ? "No organizations available" : "Search or select an Organization"
}
emptyText="No organizations available"
onValueChange={(next) => {
onChange(next === "" ? null : next);
setCurrentOrgForCreateTeam(
adminOrgs.find((org) => org.organization_id === next) ?? null,
);
}}
onValueChange={(next) => selectCreateTeamOrganization(next, value ?? null, onChange)}
/>
)}
</FormField>

38
ui/litellm-dashboard/src/components/add_model/ComplexityRouterConfig.test.tsx
View file

@ -880,6 +880,44 @@ describe("ComplexityRouterConfig modality panel", () => {
expect(screen.getByRole("switch", { name: "Route image requests to vision-capable models" })).toBeChecked();
});
// The backend ignores modality_pin_override unless modality_routing is on, so offering it while
// image routing is off would let an operator save a flag that does nothing.
it("disables the pin-override switch while image routing is off", () => {
const onChange = vi.fn();
renderWithProviders(<ComplexityRouterConfig {...baseProps} onChange={onChange} />);
fireEvent.click(screen.getByText("Advanced: Modality Routing"));
const override = screen.getByRole("switch", { name: "Override session pin for image requests" });
expect(override).toHaveAttribute("aria-disabled", "true");
fireEvent.click(override);
expect(onChange).not.toHaveBeenCalled();
});
it("writes modality_pin_override through onChange once image routing is on", () => {
const onChange = vi.fn();
const value = { ...defaultValue, modality_routing: true };
renderWithProviders(<ComplexityRouterConfig {...baseProps} value={value} onChange={onChange} />);
fireEvent.click(screen.getByText("Advanced: Modality Routing"));
const override = screen.getByRole("switch", { name: "Override session pin for image requests" });
expect(override).not.toBeChecked();
fireEvent.click(override);
expect(onChange).toHaveBeenCalledWith({ ...value, modality_pin_override: true });
});
it("renders a stored modality_pin_override=true as on", () => {
renderWithProviders(
<ComplexityRouterConfig
{...baseProps}
value={{ ...defaultValue, modality_routing: true, modality_pin_override: true }}
/>,
);
fireEvent.click(screen.getByText("Advanced: Modality Routing"));
expect(screen.getByRole("switch", { name: "Override session pin for image requests" })).toBeChecked();
});
});
describe("ComplexityRouterConfig affinity panel", () => {

1
ui/litellm-dashboard/src/components/add_model/ComplexityRouterConfig.tsx
View file

@ -410,6 +410,7 @@ export interface ComplexityRouterConfigValue {
classification_mode?: ClassificationMode;
session_affinity?: boolean;
modality_routing?: boolean;
modality_pin_override?: boolean;
deployment_affinity?: boolean;
/** Plan-mode floor as a tier ROW ID, unset meaning off. The wire carries the row's name. */
plan_mode_min_tier?: string;

50
ui/litellm-dashboard/src/components/add_model/ModalityRoutingControls.tsx
View file

@ -7,20 +7,36 @@ import type { ComplexityRouterConfigValue } from "./ComplexityRouterConfig";
export const ModalityRoutingControls: React.FC<{
value: ComplexityRouterConfigValue;
onChange: (value: ComplexityRouterConfigValue) => void;
}> = ({ value, onChange }) => (
<>
<div className="flex items-center gap-2 mb-2">
<Switch
checked={value.modality_routing ?? false}
onCheckedChange={(modalityRouting) => onChange({ ...value, modality_routing: modalityRouting })}
aria-label="Route image requests to vision-capable models"
/>
<strong className="font-semibold">Route image requests to vision-capable models</strong>
</div>
<span className="block text-xs text-muted-foreground">
Replaces a routed model that cannot take image input with the nearest higher tier that can, then the default
model, instead of failing with a provider 400. Only models explicitly declared supports_vision false are replaced,
and a kept session pin still wins.
</span>
</>
);
}> = ({ value, onChange }) => {
const modalityRouting = value.modality_routing ?? false;
return (
<>
<div className="flex items-center gap-2 mb-2">
<Switch
checked={modalityRouting}
onCheckedChange={(nextModalityRouting) => onChange({ ...value, modality_routing: nextModalityRouting })}
aria-label="Route image requests to vision-capable models"
/>
<strong className="font-semibold">Route image requests to vision-capable models</strong>
</div>
<span className="block text-xs mb-3 text-muted-foreground">
Replaces a routed model that cannot take image input with the nearest higher tier that can, then the default
model, instead of failing with a provider 400. Only models explicitly declared supports_vision false are
replaced, and a kept session pin still wins unless you turn on the override below.
</span>
<div className="flex items-center gap-2 mb-2">
<Switch
checked={value.modality_pin_override ?? false}
onCheckedChange={(modalityPinOverride) => onChange({ ...value, modality_pin_override: modalityPinOverride })}
disabled={!modalityRouting}
aria-label="Override session pin for image requests"
/>
<strong className="font-semibold">Override session pin for image requests</strong>
</div>
<span className="block text-xs text-muted-foreground">
Route an image turn to a capable model even when the session is pinned to one that cannot take images. The pin
is kept, so the next text turn goes back to it. Needs image routing turned on.
</span>
</>
);
};

56
ui/litellm-dashboard/src/components/add_model/add_auto_router_tab.test.tsx
View file

@ -590,6 +590,44 @@ describe("AddAutoRouterTab", () => {
});
});
it("writes both modality flags as false into the create payload when the panel stays untouched", async () => {
const user = userEvent.setup();
vi.mocked(getMissingTiersError).mockReturnValue(null);
renderWithProviders(<Harness />);
fireEvent.change(screen.getByPlaceholderText(/smart_router/i), { target: { value: "modality-router" } });
await user.click(screen.getByRole("button", { name: /add auto router/i }));
await waitFor(() => expect(handleAddAutoRouterSubmit).toHaveBeenCalled());
expect(vi.mocked(handleAddAutoRouterSubmit).mock.calls.at(-1)?.[0].complexity_router_config).toMatchObject({
modality_routing: false,
modality_pin_override: false,
});
});
it("carries the pin override through to the create payload once image routing unlocks it", async () => {
const user = userEvent.setup();
vi.mocked(getMissingTiersError).mockReturnValue(null);
renderWithProviders(<Harness />);
fireEvent.change(screen.getByPlaceholderText(/smart_router/i), { target: { value: "modality-router" } });
expandDetailedConfiguration();
await user.click(screen.getByText("Advanced: Modality Routing"));
await user.click(await screen.findByRole("switch", { name: "Route image requests to vision-capable models" }));
await user.click(await screen.findByRole("switch", { name: "Override session pin for image requests" }));
await user.click(screen.getByRole("button", { name: /add auto router/i }));
await waitFor(() => expect(handleAddAutoRouterSubmit).toHaveBeenCalled());
expect(vi.mocked(handleAddAutoRouterSubmit).mock.calls.at(-1)?.[0].complexity_router_config).toMatchObject({
modality_routing: true,
modality_pin_override: true,
});
});
// Custom is the escape hatch, not the headline choice, so it's listed after every bundled preset
// rather than first.
it("lists Custom Configuration after the bundled presets", () => {
@ -598,7 +636,14 @@ describe("AddAutoRouterTab", () => {
const labels = visibleOptions().map((option) => option.querySelector(".font-medium")?.textContent);
expect(labels).toEqual(["Anthropic Family", "Gemini Family", "Lite", "OpenAI Family", "Custom Configuration"]);
expect(labels).toEqual([
"1M Context",
"Anthropic Family",
"Gemini Family",
"Lite",
"OpenAI Family",
"Custom Configuration",
]);
});
describe("routing test", () => {
@ -1022,7 +1067,14 @@ describe("AddAutoRouterTab", () => {
expect(isOptionDisabled(optionByLabel("Anthropic Family")!)).toBe(false);
});
const labels = visibleOptions().map((option) => option.querySelector(".font-medium")?.textContent);
expect(labels).toEqual(["Anthropic Family", "Gemini Family", "Lite", "OpenAI Family", "Custom Configuration"]);
expect(labels).toEqual([
"Anthropic Family",
"1M Context",
"Gemini Family",
"Lite",
"OpenAI Family",
"Custom Configuration",
]);
});
it.each([

1
ui/litellm-dashboard/src/components/add_model/add_auto_router_tab.tsx
View file

@ -358,6 +358,7 @@ const AddAutoRouterTab: React.FC<AddAutoRouterTabProps> = ({
classifierFallback: complexityRouterConfig.classifier_fallback,
sessionAffinity: complexityRouterConfig.session_affinity ?? DEFAULT_SESSION_AFFINITY,
modalityRouting: complexityRouterConfig.modality_routing ?? false,
modalityPinOverride: complexityRouterConfig.modality_pin_override ?? false,
deploymentAffinity: complexityRouterConfig.deployment_affinity ?? DEFAULT_DEPLOYMENT_AFFINITY,
customTechnicalKeywords,
keywordTierRules,

9
ui/litellm-dashboard/src/components/add_model/build_complexity_router_config.test.ts
View file

@ -56,6 +56,7 @@ describe("buildComplexityRouterConfig", () => {
session_affinity: false,
deployment_affinity: true,
modality_routing: false,
modality_pin_override: false,
escalation_keywords: ["LITELLM ESCALATE"],
};
expect(config).toEqual(expected);
@ -270,6 +271,14 @@ describe("buildComplexityRouterConfig", () => {
expect(buildComplexityRouterConfig({ ...baseParams, modalityRouting: false }).modality_routing).toBe(false);
});
it("writes modality_pin_override explicitly both ways, so the stored config never relies on the backend default", () => {
expect(buildComplexityRouterConfig({ ...baseParams, modalityPinOverride: true }).modality_pin_override).toBe(true);
expect(buildComplexityRouterConfig(baseParams).modality_pin_override).toBe(false);
expect(buildComplexityRouterConfig({ ...baseParams, modalityPinOverride: false }).modality_pin_override).toBe(
false,
);
});
it("writes session_affinity=true so turning the toggle on overrides the backend's off-by-default", () => {
const config = buildComplexityRouterConfig({ ...baseParams, sessionAffinity: true });
expect(config.session_affinity).toBe(true);

4
ui/litellm-dashboard/src/components/add_model/build_complexity_router_config.ts
View file

@ -111,6 +111,7 @@ export interface BuildComplexityRouterConfigParams {
classificationMode: ClassificationMode | undefined;
sessionAffinity: boolean;
modalityRouting?: boolean;
modalityPinOverride?: boolean;
deploymentAffinity: boolean;
customTechnicalKeywords: string[];
keywordTierRules: KeywordTierRule[];
@ -169,6 +170,7 @@ export interface ComplexityRouterConfigPayload {
session_affinity: boolean;
deployment_affinity: boolean;
modality_routing: boolean;
modality_pin_override: boolean;
custom_technical_keywords?: string[];
keyword_tier_rules?: { keywords: string[]; tier: KeywordTierRule["tier"] }[];
semantic_keyword_matching?: boolean;
@ -409,6 +411,7 @@ export const buildComplexityRouterConfig = ({
classificationMode,
sessionAffinity,
modalityRouting,
modalityPinOverride,
deploymentAffinity,
customTechnicalKeywords,
keywordTierRules,
@ -472,6 +475,7 @@ export const buildComplexityRouterConfig = ({
session_affinity: sessionAffinity,
deployment_affinity: deploymentAffinity,
modality_routing: modalityRouting ?? false,
modality_pin_override: modalityPinOverride ?? false,
...(customTechnicalKeywords.length > 0 && { custom_technical_keywords: customTechnicalKeywords }),
...(cleanedKeywordTierRules.length > 0 && { keyword_tier_rules: cleanedKeywordTierRules }),
escalation_keywords: cleanedEscalationKeywords,

26
ui/litellm-dashboard/src/components/edit_auto_router/build_updated_complexity_router_config.test.ts
View file

@ -257,6 +257,30 @@ describe("buildUpdatedComplexityRouterConfig session affinity", () => {
});
});
describe("buildUpdatedComplexityRouterConfig modality pin override", () => {
it("writes modality_pin_override explicitly both ways", () => {
expect(
buildUpdatedComplexityRouterConfig(STORED, { ...FORM_VALUE, modality_pin_override: true }).modality_pin_override,
).toBe(true);
expect(
buildUpdatedComplexityRouterConfig(STORED, { ...FORM_VALUE, modality_pin_override: false }).modality_pin_override,
).toBe(false);
});
it("re-asserts the backend's off-by-default when the form value is absent, rather than dropping the key", () => {
const result = buildUpdatedComplexityRouterConfig({ ...STORED, modality_pin_override: true }, FORM_VALUE);
expect(result.modality_pin_override).toBe(false);
});
it("round-trips a stored modality_pin_override=true through hydrate then save", () => {
const stored = { ...STORED, modality_routing: true, modality_pin_override: true };
const hydrated = hydrateComplexityRouterConfig(stored, undefined);
expect(hydrated.modality_pin_override).toBe(true);
expect(buildUpdatedComplexityRouterConfig(stored, hydrated).modality_pin_override).toBe(true);
});
});
describe("buildUpdatedComplexityRouterConfig classification mode", () => {
it("round-trips a stored user_turn through hydrate then save", () => {
const stored = { ...STORED, classification_mode: "user_turn" };
@ -505,6 +529,8 @@ describe("managed keys survive an untouched open-and-save", () => {
classifier_fallback: "default_model",
classification_mode: "user_turn",
session_affinity: true,
modality_routing: true,
modality_pin_override: true,
deployment_affinity: false,
adaptive: true,
adaptive_weights: { quality: 0.4, cost: 0.6 },

22
ui/litellm-dashboard/src/components/edit_auto_router/edit_auto_router_modal.test.ts
View file

@ -51,6 +51,7 @@ const expectedClassifiedTierConfig = {
session_affinity: false,
deployment_affinity: true,
modality_routing: false,
modality_pin_override: false,
adaptive: true,
adaptive_weights: { quality: 0.4, cost: 0.6 },
adaptive_eligible: "classified_tier",
@ -74,6 +75,7 @@ const expectedAdaptiveDisabledConfig = {
session_affinity: false,
deployment_affinity: true,
modality_routing: false,
modality_pin_override: false,
};
describe("buildUpdatedComplexityRouterConfig", () => {
@ -109,6 +111,26 @@ describe("buildUpdatedComplexityRouterConfig", () => {
expect(disabled.modality_routing).toBe(false);
});
it("hydrates a stored modality_pin_override into form state and defaults absent to off", () => {
expect(
hydrateComplexityRouterConfig({ ...storedConfig, modality_pin_override: true }, null).modality_pin_override,
).toBe(true);
expect(hydrateComplexityRouterConfig(storedConfig, null).modality_pin_override).toBe(false);
});
it("round-trips modality_pin_override explicitly in both directions", () => {
const enabled = buildUpdatedComplexityRouterConfig(storedConfig, {
...classifiedTierValue,
modality_pin_override: true,
});
expect(enabled.modality_pin_override).toBe(true);
const disabled = buildUpdatedComplexityRouterConfig(
{ ...storedConfig, modality_pin_override: true },
{ ...classifiedTierValue, modality_pin_override: false },
);
expect(disabled.modality_pin_override).toBe(false);
});
it("includes return_raw_model_name only when enabled", () => {
const updatedConfig = buildUpdatedComplexityRouterConfig(storedConfig, {
...classifiedTierValue,

44
ui/litellm-dashboard/src/components/edit_auto_router/edit_auto_router_modal.test.tsx
View file

@ -549,6 +549,50 @@ describe("EditAutoRouterModal deployment affinity", () => {
await waitFor(() => expect(modelPatchUpdateCall).toHaveBeenCalled());
expect(savedConfig().deployment_affinity).toBe(false);
});
// modality_pin_override is a managed key, so the modal rewrites it from form state on save. A
// hydration gap would silently turn a stored override off on the next untouched save.
it("shows a stored modality_pin_override=true as on and preserves it through an untouched save", async () => {
const user = userEvent.setup();
renderWithStoredConfig({ ...STORED_CONFIG, modality_routing: true, modality_pin_override: true });
await user.click(await screen.findByText("Advanced: Modality Routing"));
expect(await screen.findByRole("switch", { name: "Override session pin for image requests" })).toBeChecked();
await user.click(screen.getByRole("button", { name: /save changes/i }));
await waitFor(() => expect(modelPatchUpdateCall).toHaveBeenCalled());
expect(savedConfig().modality_pin_override).toBe(true);
});
it("persists turning the modality pin override on", async () => {
const user = userEvent.setup();
renderWithStoredConfig({ ...STORED_CONFIG, modality_routing: true });
await user.click(await screen.findByText("Advanced: Modality Routing"));
await user.click(await screen.findByRole("switch", { name: "Override session pin for image requests" }));
await user.click(screen.getByRole("button", { name: /save changes/i }));
await waitFor(() => expect(modelPatchUpdateCall).toHaveBeenCalled());
expect(savedConfig().modality_pin_override).toBe(true);
});
it("writes modality_pin_override=false for a stored config that never carried the key", async () => {
const user = userEvent.setup();
renderWithStoredConfig(STORED_CONFIG);
await user.click(await screen.findByText("Advanced: Modality Routing"));
expect(await screen.findByRole("switch", { name: "Override session pin for image requests" })).toHaveAttribute(
"aria-disabled",
"true",
);
await user.click(screen.getByRole("button", { name: /save changes/i }));
await waitFor(() => expect(modelPatchUpdateCall).toHaveBeenCalled());
expect(savedConfig().modality_pin_override).toBe(false);
});
});
describe("EditAutoRouterModal custom classifier prompt and fallback", () => {

5
ui/litellm-dashboard/src/components/edit_auto_router/edit_auto_router_modal.tsx
View file

@ -104,6 +104,7 @@ export interface StoredComplexityRouterConfig {
reasoning_override_min_score?: unknown;
session_affinity?: unknown;
modality_routing?: unknown;
modality_pin_override?: unknown;
deployment_affinity?: unknown;
adaptive?: boolean;
adaptive_weights?: AdaptiveRouterWeights;
@ -181,6 +182,8 @@ export const hydrateComplexityRouterConfig = (
session_affinity:
typeof parsedConfig.session_affinity === "boolean" ? parsedConfig.session_affinity : DEFAULT_SESSION_AFFINITY,
modality_routing: typeof parsedConfig.modality_routing === "boolean" ? parsedConfig.modality_routing : false,
modality_pin_override:
typeof parsedConfig.modality_pin_override === "boolean" ? parsedConfig.modality_pin_override : false,
deployment_affinity:
typeof parsedConfig.deployment_affinity === "boolean"
? parsedConfig.deployment_affinity
@ -221,6 +224,7 @@ export const MANAGED_COMPLEXITY_ROUTER_KEYS = new Set([
"classification_mode",
"session_affinity",
"modality_routing",
"modality_pin_override",
"deployment_affinity",
"adaptive",
"adaptive_weights",
@ -318,6 +322,7 @@ export const buildUpdatedComplexityRouterConfig = (
classifierFallback: value.classifier_fallback,
sessionAffinity: value.session_affinity ?? DEFAULT_SESSION_AFFINITY,
modalityRouting: value.modality_routing ?? false,
modalityPinOverride: value.modality_pin_override ?? false,
deploymentAffinity: value.deployment_affinity ?? DEFAULT_DEPLOYMENT_AFFINITY,
customTechnicalKeywords: customTechnicalKeywords ?? [],
keywordTierRules: keywordMatching?.keywordTierRules ?? [],

6
ui/litellm-dashboard/src/components/view_logs/LogDetailsDrawer/RoutingDecisionCard.test.tsx
View file

@ -186,6 +186,12 @@ describe("RoutingDecisionCard", () => {
expect(screen.queryByText("housekeeping")).not.toBeInTheDocument();
});
it("labels a modality pin override instead of showing the raw cause token", () => {
render(<RoutingDecisionCard decision={{ ...heuristic, cause: "modality_pin_override" }} />);
expect(screen.getByText("Overrode session pin for image input")).toBeInTheDocument();
expect(screen.queryByText("modality_pin_override")).not.toBeInTheDocument();
});
it("labels a modality escalation instead of showing the raw cause token", () => {
render(<RoutingDecisionCard decision={{ ...heuristic, cause: "modality_escalation" }} />);
expect(screen.getByText("Escalated for image input")).toBeInTheDocument();

1
ui/litellm-dashboard/src/components/view_logs/LogDetailsDrawer/RoutingDecisionCard.tsx
View file

@ -93,6 +93,7 @@ const CONSTANT_CAUSE_LABELS: Record<string, string> = {
session_affinity_escalation: "Escalated from session pin",
user_turn_continuation: "Continuation turn, classifier skipped",
modality_escalation: "Escalated for image input",
modality_pin_override: "Overrode session pin for image input",
quality_tier: "Quality tier mapping",
bandit: "Adaptive bandit",
default_fallback: "Default model, no route matched",

51
ui/litellm-dashboard/src/lib/autorouter_presets.test.ts
View file

@ -1,9 +1,8 @@
import { describe, it, expect } from "vitest";
import bundledPresets from "../../../../litellm/proxy/public_endpoints/autorouter_presets.json";
import { BUNDLED_PRESETS_RESPONSE } from "../../tests/mocks/autoRouterPresets";
import {
hydratePresets,
AutoRouterPreset,
AutoRouterPresetsResponse,
getRequiredModelsInPreset,
getMissingModelsInPreset,
getRequiredModels,
@ -21,14 +20,20 @@ import { DEFAULT_ESCALATION_KEYWORDS } from "@/components/add_model/EscalationKe
const groupsOnly = (models: Iterable<string>) => buildModelAvailability(models, []);
// Hydrated from the real bundled catalog so a catalog edit flows into these expectations.
const PRESETS = hydratePresets(bundledPresets as AutoRouterPresetsResponse);
const PRESETS = hydratePresets(BUNDLED_PRESETS_RESPONSE);
const getAllPresets = (): AutoRouterPreset[] => PRESETS;
const getPresetByKey = (key: string): AutoRouterPreset | undefined => PRESETS.find((p) => p.key === key);
describe("autorouter_presets", () => {
it("hydrates exactly the bundled presets", () => {
const presets = getAllPresets();
expect(presets.map((p) => p.label).sort()).toEqual(["Anthropic Family", "Gemini Family", "Lite", "OpenAI Family"]);
expect(presets.map((p) => p.label).sort()).toEqual([
"1M Context",
"Anthropic Family",
"Gemini Family",
"Lite",
"OpenAI Family",
]);
// Every preset carries all four fields the UI relies on; a JSON typo dropping one fails here.
for (const p of presets) {
expect(p).toMatchObject({ key: expect.any(String), label: expect.any(String), description: expect.any(String) });
@ -148,6 +153,25 @@ describe("autorouter_presets", () => {
expect(withoutFlag.complexityRouterConfig.modality_routing).toBe(false);
});
it("carries a preset's modality_pin_override into the prefilled form state", () => {
const preset = getPresetByKey("anthropic_family")!;
const withFlag = { ...preset.complexity_router_config, modality_routing: true, modality_pin_override: true };
const prefill = buildPresetPrefill(withFlag, groupsOnly(getRequiredModelsInPreset(preset)));
expect(prefill.complexityRouterConfig.modality_pin_override).toBe(true);
const withoutFlag = buildPresetPrefill(
preset.complexity_router_config,
groupsOnly(getRequiredModelsInPreset(preset)),
);
expect(withoutFlag.complexityRouterConfig.modality_pin_override).toBe(false);
});
it("ships every bundled preset with both modality flags written out, since the payload type requires them", () => {
for (const preset of getAllPresets()) {
expect(preset.complexity_router_config.modality_routing, preset.key).toBe(false);
expect(preset.complexity_router_config.modality_pin_override, preset.key).toBe(false);
}
});
it("prefills the anthropic preset's effort through to tier_model_params", () => {
const preset = getPresetByKey("anthropic_family")!;
const prefill = buildPresetPrefill(preset.complexity_router_config, groupsOnly(getRequiredModelsInPreset(preset)));
@ -183,6 +207,25 @@ describe("autorouter_presets", () => {
});
});
it("pins the 1M context preset to Luna, Terra, and Opus at high thinking", () => {
const preset = getPresetByKey("1m_context")!;
const expectedTiers = {
SIMPLE: ["gpt-5.6-luna"],
MEDIUM: ["gpt-5.6-terra"],
COMPLEX: ["claude-opus-5"],
REASONING: ["claude-opus-5"],
};
expect(preset.complexity_router_config.classifier_type).toBe("heuristic_v2");
expect(preset.complexity_router_config.tiers).toEqual(expectedTiers);
expect(preset.complexity_router_config.tier_model_configs).toEqual({
REASONING: [{ model_name: "claude-opus-5", litellm_params: { reasoning_effort: "high" } }],
});
const prefill = buildPresetPrefill(preset.complexity_router_config, groupsOnly(getRequiredModelsInPreset(preset)));
expect(prefill.complexityRouterConfig.tier_model_params).toEqual({
REASONING: { "claude-opus-5": { reasoning_effort: "high" } },
});
});
it("pins the gemini preset to concrete model ids, never Google's hot-swapping -latest aliases", () => {
const gemini = getPresetByKey("gemini_family")!;
const config = gemini.complexity_router_config;

1
ui/litellm-dashboard/src/lib/autorouter_presets.ts
View file

@ -286,6 +286,7 @@ export const buildPresetPrefill = (
session_affinity: config.session_affinity ?? DEFAULT_SESSION_AFFINITY,
deployment_affinity: config.deployment_affinity ?? DEFAULT_DEPLOYMENT_AFFINITY,
modality_routing: config.modality_routing ?? false,
modality_pin_override: config.modality_pin_override ?? false,
adaptive: config.adaptive,
adaptive_weights: config.adaptive_weights,
tier_distance_penalty: config.tier_distance_penalty,

10
ui/litellm-dashboard/src/lib/http/schema.d.ts generated vendored
View file

@ -34740,9 +34740,15 @@ export interface components {
* @default 0.5
*/
match_threshold: number;
/**
* Modality Pin Override
* @description Let modality_routing replace a kept session-affinity pin on the turns that carry an image. Without this, a session pinned to a text-only model fails every image turn with a provider 400, since the pin is exempt from the modality gate. When enabled, such a turn routes to a capable model for that request only and the stored pin is left untouched, so the next text turn replays the session's own model; the override is reported as cause modality_pin_override and is never itself pinned. Inert unless modality_routing is also enabled.
* @default false
*/
modality_pin_override: boolean;
/**
* Modality Routing
* @description Route image-bearing requests only to models that can accept image input. The classifier reads text alone, so an image request whose text classifies cheap otherwise lands on a text-only model and fails with a provider 400. When enabled, a routed model explicitly declared supports_vision false (deployment model_info or the model cost map; unmapped names stay routable) is replaced by the nearest HIGHER tier holding a capable model, then default_model, else a clear 400. A kept session-affinity pin still wins even when an image arrives.
* @description Route image-bearing requests only to models that can accept image input. The classifier reads text alone, so an image request whose text classifies cheap otherwise lands on a text-only model and fails with a provider 400. When enabled, a routed model explicitly declared supports_vision false (deployment model_info or the model cost map; unmapped names stay routable) is replaced by the nearest HIGHER tier holding a capable model, then default_model, else a clear 400. A kept session-affinity pin still wins even when an image arrives, unless modality_pin_override is also enabled.
* @default false
*/
modality_routing: boolean;
@ -35945,7 +35951,7 @@ export interface components {
* Cause
* @enum {string}
*/
cause?: "heuristic_scorer" | "heuristic_v2" | "reasoning_override" | "llm_classifier" | "heuristic_first_short_circuit" | "hybrid_short_circuit" | "classifier_plugin" | "classifier_fallback" | "default_model_fallback" | "literal_keyword_match" | "semantic_keyword_match" | "plan_mode" | "housekeeping" | "modality_escalation" | "session_affinity_pin" | "session_affinity_escalation" | "user_turn_continuation" | "default_fallback" | "keyword" | "quality_tier" | "bandit";
cause?: "heuristic_scorer" | "heuristic_v2" | "reasoning_override" | "llm_classifier" | "heuristic_first_short_circuit" | "hybrid_short_circuit" | "classifier_plugin" | "classifier_fallback" | "default_model_fallback" | "literal_keyword_match" | "semantic_keyword_match" | "plan_mode" | "housekeeping" | "modality_escalation" | "modality_pin_override" | "session_affinity_pin" | "session_affinity_escalation" | "user_turn_continuation" | "default_fallback" | "keyword" | "quality_tier" | "bandit";
/** Classifier Cost */
classifier_cost?: number;
/** Classifier Model */

9
ui/litellm-dashboard/tests/mocks/autoRouterPresets.ts
View file

@ -1,10 +1,15 @@
import { readFileSync } from "fs";
import { resolve } from "path";
import { vi } from "vitest";
import bundledPresets from "../../../../litellm/proxy/public_endpoints/autorouter_presets.json";
import { hydratePresets, type AutoRouterPresetsResponse } from "@/lib/autorouter_presets";
// Derived from the real bundled catalog so a preset edit there flows into test expectations
// instead of redding on a stale copy. Exported as vi.fn so a test can override the query state.
export const BUNDLED_PRESETS = hydratePresets(bundledPresets as AutoRouterPresetsResponse);
const CATALOG_PATH = resolve(__dirname, "../../../../litellm/proxy/public_endpoints/autorouter_presets.json");
export const BUNDLED_PRESETS_RESPONSE = JSON.parse(readFileSync(CATALOG_PATH, "utf8")) as AutoRouterPresetsResponse;
export const BUNDLED_PRESETS = hydratePresets(BUNDLED_PRESETS_RESPONSE);
export const LOADED_PRESETS_QUERY = {
data: BUNDLED_PRESETS,