fix(vector_stores): block caller-supplied embedding selection params on query surfaces

litellm/proxy/rag_endpoints/endpoints.py

@@ -42,6 +42,7 @@ from litellm.proxy.rag_endpoints.upload_security import (
 )
 from litellm.proxy.vector_store_endpoints.endpoints import (
     build_request_data_from_managed_vector_store,
+    reject_caller_embedding_selection_params,
 )
 from litellm.proxy.vector_store_endpoints.utils import (
     assert_user_can_access_vector_store_id,
@@ -716,6 +717,7 @@ async def rag_query(
                 status_code=400,
                 detail={"error": "retrieval_config must contain 'vector_store_id'"},
             )
+        reject_caller_embedding_selection_params(payload=retrieval_config, source="retrieval_config")
         resolved_stores: Final = await _authorize_nested_vector_store_ids(
             payload=retrieval_config,
             user_api_key_dict=user_api_key_dict,

litellm/proxy/vector_store_endpoints/endpoints.py

@@ -29,6 +29,29 @@ from litellm.types.vector_stores import IndexCreateRequest, IndexListResponse
 from litellm.vector_stores.vector_store_registry import VectorStoreIndexRegistry
 
 router: Final = APIRouter()
+
+BLOCKED_QUERY_EMBEDDING_SELECTION_PARAMS: Final = frozenset(
+    {
+        "embedding_model",
+        "litellm_embedding_model",
+        "litellm_embedding_config",
+        "litellm_credential_name",
+    }
+)
+
+
+def reject_caller_embedding_selection_params(payload: Mapping[str, object], source: str) -> None:
+    blocked: Final = sorted(BLOCKED_QUERY_EMBEDDING_SELECTION_PARAMS & payload.keys())
+    if blocked:
+        raise HTTPException(
+            status_code=400,
+            detail={
+                "error": f"'{blocked[0]}' cannot be set in {source}. "
+                "Embedding configuration comes from the vector store's server-side registration."
+            },
+        )
+
+
 ########################################################
 # OpenAI Compatible Endpoints
 ########################################################
@@ -134,6 +157,7 @@ async def vector_store_search(
     )
 
     data = await _read_request_body(request=request)
+    reject_caller_embedding_selection_params(payload=data, source="the search request body")
     data["vector_store_id"] = vector_store_id
 
     # Check for legacy vector store registry (non-managed vector stores)

tests/test_litellm/proxy/rag_endpoints/test_rag_endpoints.py

@@ -421,6 +421,30 @@ def test_rag_query_store_params_win_over_user_retrieval_config(client_internal_u
     assert forwarded_config["aws_region_name"] == "eu-west-1"
 
 
+@pytest.mark.parametrize(
+    "blocked_key",
+    ["embedding_model", "litellm_embedding_model", "litellm_embedding_config", "litellm_credential_name"],
+)
+def test_rag_query_rejects_caller_embedding_selection_params(client_internal_user, blocked_key):
+    """
+    Regression: a caller must not pick the embedding model or credential used at
+    search time. Those resolve through the Router with the proxy's credentials,
+    bypassing the key's model permissions, so they may only come from the
+    managed store's server-side registration.
+    """
+    response = client_internal_user.post(
+        "/v1/rag/query",
+        json={
+            "model": "gpt-4o-mini",
+            "messages": [{"role": "user", "content": "hello"}],
+            "retrieval_config": {"vector_store_id": "s3-store", blocked_key: "attacker-choice"},
+        },
+    )
+
+    assert response.status_code == 400, response.json()
+    assert blocked_key in str(response.json())
+
+
 EICAR = r"X5O!P%@AP[4\PZX54(P^)7CC)7}$EICAR-STANDARD-ANTIVIRUS-TEST-FILE!$H+H*"
 INGEST_REQUEST = '{"ingest_options":{"vector_store":{"custom_llm_provider":"openai"}}}'
 

tests/test_litellm/proxy/vector_store_endpoints/test_vector_store_endpoints.py

@@ -3158,3 +3158,35 @@ class TestAzureAIAnalyzeNamedIndexClassification:
             user_api_key_dict=self._team_member("analyze", ["read"]),
         )
         assert result is True
+
+
+@pytest.mark.parametrize(
+    "blocked_key",
+    ["embedding_model", "litellm_embedding_model", "litellm_embedding_config", "litellm_credential_name"],
+)
+def test_vector_store_search_rejects_caller_embedding_selection_params(blocked_key):
+    """
+    Regression: the search request body must not pick the embedding model or
+    credential used to embed the query. Those resolve through the Router with
+    the proxy's credentials, bypassing the key's model permissions, so they may
+    only come from the managed store's server-side registration.
+    """
+    from fastapi.testclient import TestClient
+
+    from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
+    from litellm.proxy.proxy_server import app
+
+    mock_auth = UserAPIKeyAuth(user_id="test_internal_user", user_role=LitellmUserRoles.INTERNAL_USER.value)
+    original_overrides = app.dependency_overrides.copy()
+    app.dependency_overrides[user_api_key_auth] = lambda: mock_auth
+    try:
+        client = TestClient(app)
+        response = client.post(
+            "/v1/vector_stores/s3-store/search",
+            json={"query": "hello", blocked_key: "attacker-choice"},
+        )
+    finally:
+        app.dependency_overrides = original_overrides
+
+    assert response.status_code == 400, response.json()
+    assert blocked_key in str(response.json())