fix(proxy): honor request model on vector store routes

Vector store create, list, retrieve, update, delete and search never read the
`?model=` query parameter or the `x-litellm-model` header, so with no managed
registry entry and no body model the request fell through to the SDK default
provider (openai) instead of the configured deployment.

Resolve the hint into data["model"] when nothing else established routing, and
add /vector_stores to the auth model-routing markers so the hinted model is
checked against key/team model access.

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/proxy/auth/auth_utils.py

@@ -1326,6 +1326,7 @@ _MODEL_ROUTING_HEADER_OR_QUERY_ROUTE_MARKERS = (
     "/batches",
     "/skills",
     "/evals",
+    "/vector_stores",
 )
 _MODEL_ROUTING_QUERY_TARGET_MODEL_ROUTE_MARKERS = (
     "/files",

litellm/proxy/vector_store_endpoints/endpoints.py

@@ -6,6 +6,7 @@ from litellm.integrations.vector_store_integrations.vector_store_pre_call_hook i
     LiteLLM_ManagedVectorStore,
 )
 from litellm.proxy._types import CommonProxyErrors, UserAPIKeyAuth
+from litellm.proxy.auth.auth_utils import MODEL_ROUTING_HEADER_NAME
 from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
 from litellm.proxy.common_request_processing import ProxyBaseLLMRequestProcessing
 from litellm.proxy.utils import jsonify_object
@@ -88,6 +89,35 @@ async def _update_request_data_with_litellm_managed_vector_store_registry(
     return data
 
 
+_EXPLICIT_ROUTING_KEYS = (
+    "model",
+    "custom_llm_provider",
+    "litellm_credential_name",
+    "api_key",
+    "api_base",
+)
+
+
+def _apply_model_routing_hint(data: Dict, request: Request) -> Dict:
+    """
+    Resolve ``?model=`` / ``x-litellm-model`` into ``data["model"]`` so the router
+    picks the requested deployment.
+
+    Routing already established by the request body or by the managed vector store
+    registry wins, so the hint only applies when nothing else selected a deployment.
+    The hint is authorized against the key/team model allowlists by
+    ``get_model_from_request``, which reads the same query param and header.
+    """
+    if any(data.get(key) for key in _EXPLICIT_ROUTING_KEYS):
+        return data
+
+    model_hint = request.query_params.get("model") or request.headers.get(MODEL_ROUTING_HEADER_NAME)
+    if not model_hint:
+        return data
+
+    return {**data, "model": model_hint}
+
+
 @router.post(
     "/v1/vector_stores/{vector_store_id:path}/search",
     dependencies=[Depends(user_api_key_auth)],
@@ -130,6 +160,7 @@ async def vector_store_search(
     data = await _update_request_data_with_litellm_managed_vector_store_registry(
         data=data, vector_store_id=vector_store_id, user_api_key_dict=user_api_key_dict
     )
+    data = _apply_model_routing_hint(data=data, request=request)
 
     # The managed_vector_stores pre-call hook will handle:
     # 1. Decoding managed vector store IDs
@@ -244,6 +275,8 @@ async def vector_store_create(
 
         return response
 
+    data = _apply_model_routing_hint(data=data, request=request)
+
     processor = ProxyBaseLLMRequestProcessing(data=data)
     try:
         return await processor.base_process_llm_request(
@@ -306,6 +339,7 @@ async def vector_store_retrieve(
     data = await _update_request_data_with_litellm_managed_vector_store_registry(
         data=data, vector_store_id=vector_store_id, user_api_key_dict=user_api_key_dict
     )
+    data = _apply_model_routing_hint(data=data, request=request)
 
     processor = ProxyBaseLLMRequestProcessing(data=data)
     try:
@@ -376,6 +410,7 @@ async def vector_store_list(
         data["limit"] = limit
     if order is not None:
         data["order"] = order
+    data = _apply_model_routing_hint(data=data, request=request)
 
     processor = ProxyBaseLLMRequestProcessing(data=data)
     try:
@@ -442,6 +477,7 @@ async def vector_store_update(
     data = await _update_request_data_with_litellm_managed_vector_store_registry(
         data=data, vector_store_id=vector_store_id, user_api_key_dict=user_api_key_dict
     )
+    data = _apply_model_routing_hint(data=data, request=request)
 
     processor = ProxyBaseLLMRequestProcessing(data=data)
     try:
@@ -505,6 +541,7 @@ async def vector_store_delete(
     data = await _update_request_data_with_litellm_managed_vector_store_registry(
         data=data, vector_store_id=vector_store_id, user_api_key_dict=user_api_key_dict
     )
+    data = _apply_model_routing_hint(data=data, request=request)
 
     processor = ProxyBaseLLMRequestProcessing(data=data)
     try:

tests/test_litellm/proxy/auth/test_auth_utils.py

@@ -466,6 +466,25 @@ def test_get_model_from_request_authorizes_all_file_routing_model_sources():
     }
 
 
+def test_get_model_from_request_includes_vector_store_query_and_header_model():
+    assert (
+        get_model_from_request(
+            request_data={},
+            route="/v1/vector_stores",
+            request_query_params={"model": "restricted-model"},
+        )
+        == "restricted-model"
+    )
+    assert (
+        get_model_from_request(
+            request_data={},
+            route="/v1/vector_stores/vs_123",
+            request_headers={"x-litellm-model": "restricted-model"},
+        )
+        == "restricted-model"
+    )
+
+
 def test_get_model_from_request_extracts_simple_encoded_file_id_model():
     from litellm.proxy.openai_files_endpoints.common_utils import (
         encode_file_id_with_model,

tests/test_litellm/proxy/vector_store_endpoints/test_vector_store_endpoints.py

@@ -18,6 +18,7 @@ from litellm.integrations.vector_store_integrations.vector_store_pre_call_hook i
 )
 from litellm.proxy._types import LitellmUserRoles, UserAPIKeyAuth
 from litellm.proxy.vector_store_endpoints.endpoints import (
+    _apply_model_routing_hint,
     _update_request_data_with_litellm_managed_vector_store_registry,
     index_create,
 )
@@ -592,6 +593,116 @@ async def test_update_request_data_passes_through_legacy_embedding_config():
     resolve_mock.assert_not_awaited()
 
 
+def _model_hint_request(query_params=None, headers=None):
+    request = MagicMock(spec=Request)
+    request.query_params = query_params or {}
+    request.headers = headers or {}
+    return request
+
+
+def test_apply_model_routing_hint_uses_query_param():
+    result = _apply_model_routing_hint(
+        data={"vector_store_id": "vs_123"},
+        request=_model_hint_request(query_params={"model": "azure-vector-store"}),
+    )
+
+    assert result["model"] == "azure-vector-store"
+
+
+def test_apply_model_routing_hint_uses_header():
+    result = _apply_model_routing_hint(
+        data={"vector_store_id": "vs_123"},
+        request=_model_hint_request(headers={"x-litellm-model": "azure-vector-store"}),
+    )
+
+    assert result["model"] == "azure-vector-store"
+
+
+def test_apply_model_routing_hint_prefers_body_model():
+    result = _apply_model_routing_hint(
+        data={"vector_store_id": "vs_123", "model": "body-model"},
+        request=_model_hint_request(query_params={"model": "query-model"}),
+    )
+
+    assert result["model"] == "body-model"
+
+
+def test_apply_model_routing_hint_preserves_registry_routing():
+    """A managed vector store row already selected the provider/credentials, so a
+    caller-supplied hint must not repoint the request at another deployment."""
+    data = {
+        "vector_store_id": "vs_123",
+        "custom_llm_provider": "azure_ai",
+        "api_key": "registry-key",
+    }
+
+    result = _apply_model_routing_hint(
+        data=data,
+        request=_model_hint_request(query_params={"model": "other-deployment"}),
+    )
+
+    assert "model" not in result
+    assert result == data
+
+
+def test_apply_model_routing_hint_without_hint_is_noop():
+    data = {"vector_store_id": "vs_123"}
+
+    assert _apply_model_routing_hint(data=data, request=_model_hint_request()) == data
+
+
+@pytest.mark.asyncio
+async def test_vector_store_retrieve_routes_by_model_query_param():
+    from litellm.proxy.vector_store_endpoints import endpoints as vector_store_endpoints
+
+    processor = MagicMock()
+    processor.base_process_llm_request = AsyncMock(return_value={"id": "vs_123"})
+
+    with (
+        patch.object(
+            vector_store_endpoints,
+            "get_litellm_managed_vector_store",
+            AsyncMock(return_value=None),
+        ),
+        patch.object(
+            vector_store_endpoints,
+            "ProxyBaseLLMRequestProcessing",
+            MagicMock(return_value=processor),
+        ) as mock_processor_cls,
+    ):
+        response = await vector_store_endpoints.vector_store_retrieve(
+            request=_model_hint_request(query_params={"model": "azure-vector-store"}),
+            vector_store_id="vs_123",
+            fastapi_response=MagicMock(),
+            user_api_key_dict=UserAPIKeyAuth(),
+        )
+
+    assert response == {"id": "vs_123"}
+    assert mock_processor_cls.call_args.kwargs["data"]["model"] == "azure-vector-store"
+
+
+@pytest.mark.asyncio
+async def test_vector_store_list_routes_by_model_header():
+    from litellm.proxy.vector_store_endpoints import endpoints as vector_store_endpoints
+
+    processor = MagicMock()
+    processor.base_process_llm_request = AsyncMock(return_value={"data": []})
+
+    with patch.object(
+        vector_store_endpoints,
+        "ProxyBaseLLMRequestProcessing",
+        MagicMock(return_value=processor),
+    ) as mock_processor_cls:
+        response = await vector_store_endpoints.vector_store_list(
+            request=_model_hint_request(headers={"x-litellm-model": "azure-vector-store"}),
+            fastapi_response=MagicMock(),
+            user_api_key_dict=UserAPIKeyAuth(),
+        )
+
+    assert response == {"data": []}
+    assert mock_processor_cls.call_args.kwargs["data"]["model"] == "azure-vector-store"
+
+
 class TestCheckVectorStorePermission:
     """Test suite for check_vector_store_permission function."""