fix(vector-store): carry request metadata into the Router executor built from the router kwarg

litellm/llms/base_llm/vector_store/transformation.py

@@ -62,6 +62,19 @@ class LiteLLMVectorStoreEmbeddingExecutor:
         )
 
 
+_REQUEST_METADATA: Final = TypeAdapter(dict[str, object])
+
+
+def vector_store_request_metadata(kwargs: Mapping[str, object]) -> Mapping[str, object]:
+    litellm_metadata: Final = kwargs.get("litellm_metadata")
+    if isinstance(litellm_metadata, dict):
+        return _REQUEST_METADATA.validate_python(litellm_metadata)
+    metadata: Final = kwargs.get("metadata")
+    if isinstance(metadata, dict):
+        return _REQUEST_METADATA.validate_python(metadata)
+    return MappingProxyType({})
+
+
 @dataclass(frozen=True, slots=True)
 class RouterVectorStoreEmbeddingExecutor:
     router: Router
@@ -312,11 +325,12 @@ class BaseQueryEmbeddingVectorStoreConfig(BaseVectorStoreConfig):
     def query_embedding_executor(
         embedding_executor: VectorStoreEmbeddingExecutor | None,
         router: Router | None,
+        request_metadata: Mapping[str, object] = MappingProxyType({}),
     ) -> VectorStoreEmbeddingExecutor:
         if embedding_executor is not None:
             return embedding_executor
         if router is not None:
-            return RouterVectorStoreEmbeddingExecutor(router=router, metadata=MappingProxyType({}))
+            return RouterVectorStoreEmbeddingExecutor(router=router, metadata=request_metadata)
         return LiteLLMVectorStoreEmbeddingExecutor()
 
     def embed_query(

litellm/router.py

@@ -87,6 +87,7 @@ from litellm.litellm_core_utils.sensitive_data_masker import (
 )
 from litellm.llms.base_llm.vector_store.transformation import (
     RouterVectorStoreEmbeddingExecutor,
+    vector_store_request_metadata,
 )
 from litellm.llms.openai_like.json_loader import JSONProviderRegistry
 from litellm.router_strategy.budget_limiter import RouterBudgetLimiting
@@ -6697,15 +6698,7 @@ class Router:
 
     @staticmethod
     def _vector_store_request_metadata(kwargs: Mapping[str, object]) -> Mapping[str, object]:
-        litellm_metadata: Final = kwargs.get("litellm_metadata")
-        if isinstance(litellm_metadata, dict):
-            return cast(  # cast-ok: isinstance validates the runtime dict boundary
-                "dict[str, object]", litellm_metadata
-            )
-        metadata: Final = kwargs.get("metadata")
-        if isinstance(metadata, dict):
-            return cast("dict[str, object]", metadata)  # cast-ok: isinstance validates the runtime dict boundary
-        return MappingProxyType({})
+        return vector_store_request_metadata(kwargs)
 
     async def _init_vector_store_api_endpoints(
         self,

litellm/vector_stores/main.py

@@ -18,6 +18,7 @@ from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLogging
 from litellm.llms.base_llm.vector_store.transformation import (
     BaseQueryEmbeddingVectorStoreConfig,
     VectorStoreEmbeddingExecutor,
+    vector_store_request_metadata,
 )
 from litellm.llms.custom_httpx.llm_http_handler import BaseLLMHTTPHandler
 from litellm.types.router import GenericLiteLLMParams
@@ -42,10 +43,14 @@ base_llm_http_handler = BaseLLMHTTPHandler()
 #################################################
 
 
-def _direct_vector_store_embedding_executor(value: object, router: "Router | None") -> VectorStoreEmbeddingExecutor:
+def _direct_vector_store_embedding_executor(
+    value: object, router: "Router | None", request_kwargs: Mapping[str, object]
+) -> VectorStoreEmbeddingExecutor:
     if value is not None and not isinstance(value, VectorStoreEmbeddingExecutor):
         raise TypeError("Invalid direct vector store embedding executor")
-    return BaseQueryEmbeddingVectorStoreConfig.query_embedding_executor(value, router)
+    return BaseQueryEmbeddingVectorStoreConfig.query_embedding_executor(
+        value, router, vector_store_request_metadata(request_kwargs)
+    )
 
 
 def mock_vector_store_search_response(
@@ -300,7 +305,7 @@ async def asearch(
     Async: Search a vector store for relevant chunks based on a query and file attributes filter.
     """
     embedding_executor: Final = _direct_vector_store_embedding_executor(
-        kwargs.pop("_direct_vector_store_embedding_executor", None), router
+        kwargs.pop("_direct_vector_store_embedding_executor", None), router, kwargs
     )
     local_vars: Final = {  # mutable-ok: exception logging requires a sanitized mutable snapshot
         key: value for key, value in locals().items() if key != "embedding_executor"
@@ -386,7 +391,7 @@ def search(
         VectorStoreSearchResponse containing the search results.
     """
     embedding_executor: Final = _direct_vector_store_embedding_executor(
-        kwargs.pop("_direct_vector_store_embedding_executor", None), router
+        kwargs.pop("_direct_vector_store_embedding_executor", None), router, kwargs
     )
     local_vars: Final = {  # mutable-ok: exception logging requires a sanitized mutable snapshot
         key: value for key, value in locals().items() if key != "embedding_executor"

tests/test_litellm/proxy/vector_store_endpoints/test_vector_store_endpoints.py

@@ -52,7 +52,7 @@ def _serialize_litellm_params(litellm_params):
 
 def test_direct_vector_store_embedding_executor_rejects_invalid_value():
     with pytest.raises(TypeError, match="Invalid direct vector store embedding executor"):
-        _direct_vector_store_embedding_executor(object(), None)
+        _direct_vector_store_embedding_executor(object(), None, {})
 
 
 def test_router_vector_store_search_injects_executor_and_request_metadata():

tests/vector_store_tests/test_milvus_vector_store.py

@@ -601,6 +601,52 @@ async def test_sdk_search_with_router_kwarg_resolves_bare_embedding_alias_async(
     _assert_alias_resolved(embedding_route, search_route, response)
 
 
+def _team_alias_router():
+    return Router(
+        model_list=[
+            {
+                "model_name": "team-a-embedder",
+                "litellm_params": {
+                    "model": "openai/text-embedding-3-small",
+                    "api_key": "deployment-key",
+                },
+                "model_info": {"team_id": "team-a", "team_public_model_name": "multilingual-e5-large"},
+            }
+        ]
+    )
+
+
+@pytest.mark.asyncio
+async def test_sdk_search_with_router_kwarg_resolves_team_alias_from_request_metadata(
+    respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch
+):
+    monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
+    monkeypatch.delenv("OPENAI_API_KEY", raising=False)
+    embedding_route = _mock_embedding_route(respx_mock)
+    search_route = _mock_search_route(respx_mock)
+
+    response = await litellm.vector_stores.asearch(
+        router=_team_alias_router(), metadata={"user_api_key_team_id": "team-a"}, **ALIAS_SEARCH_KWARGS
+    )
+
+    _assert_alias_resolved(embedding_route, search_route, response)
+
+
+@pytest.mark.asyncio
+async def test_sdk_search_with_router_kwarg_rejects_team_alias_without_team_metadata(
+    respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch
+):
+    monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
+    monkeypatch.delenv("OPENAI_API_KEY", raising=False)
+    embedding_route = _mock_embedding_route(respx_mock)
+    _mock_search_route(respx_mock)
+
+    with pytest.raises(litellm.APIConnectionError):
+        await litellm.vector_stores.asearch(router=_team_alias_router(), **ALIAS_SEARCH_KWARGS)
+
+    assert embedding_route.call_count == 0
+
+
 @pytest.mark.asyncio
 async def test_transform_uses_injected_executor_without_embedding_config(respx_mock: respx.MockRouter):
     executor = RecordingEmbeddingExecutor(ALIAS_EMBEDDING_RESPONSE)