From 232b9e8e78ff5b7d14376bf7d72173cb469a05f6 Mon Sep 17 00:00:00 2001
From: milan <milan@berri.ai>
Date: Thu, 23 Jul 2026 21:04:42 +0000
Subject: [PATCH] fix(rag): consume top-level filters kwarg to avoid duplicate
 keyword in search

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 litellm/rag/main.py                 |  3 ++-
 tests/test_litellm/rag/test_main.py | 35 +++++++++++++++++++++++++++++
 2 files changed, 37 insertions(+), 1 deletion(-)

diff --git a/litellm/rag/main.py b/litellm/rag/main.py
index a1a5bbd38e9..628680e8a62 100644
--- a/litellm/rag/main.py
+++ b/litellm/rag/main.py
@@ -233,7 +233,8 @@ async def _execute_query_pipeline(
         raise ValueError("No query found in messages for RAG query")
 
     # 2. Search vector store
-    filters = retrieval_config.get("retrieval_filter") or retrieval_config.get("filters")
+    kwargs_filters = kwargs.pop("filters", None)
+    filters = retrieval_config.get("retrieval_filter") or retrieval_config.get("filters") or kwargs_filters
     with _suppressed_sub_call_billing():
         search_response = await litellm.vector_stores.asearch(
             vector_store_id=retrieval_config["vector_store_id"],
diff --git a/tests/test_litellm/rag/test_main.py b/tests/test_litellm/rag/test_main.py
index 30ff0b76cd8..796f8c38750 100644
--- a/tests/test_litellm/rag/test_main.py
+++ b/tests/test_litellm/rag/test_main.py
@@ -301,6 +301,41 @@ async def test_aquery_forwards_retrieval_filter_to_vector_store_search(filter_ke
     assert fake_search.await_args.kwargs["max_num_results"] == 50
 
 
+@pytest.mark.asyncio
+async def test_aquery_top_level_filters_kwarg_does_not_collide():
+    """
+    An SDK caller may pass a top-level `filters` kwarg (it used to flow to the
+    search via **kwargs). Now that the pipeline passes `filters` explicitly, the
+    top-level kwarg must be consumed rather than forwarded twice, otherwise
+    asearch raises TypeError for a duplicate keyword before any search runs.
+    """
+    from litellm.types.vector_stores import VectorStoreSearchResponse
+
+    top_level_filter = {"equals": {"key": "tenant", "value": "a"}}
+
+    fake_search = AsyncMock(
+        return_value=VectorStoreSearchResponse(
+            object="vector_store.search_results.page",
+            search_query="q",
+            data=[],
+        )
+    )
+
+    with patch("litellm.vector_stores.asearch", new=fake_search):
+        response = await litellm.aquery(
+            model="gpt-4o-mini",
+            messages=[{"role": "user", "content": "hello"}],
+            retrieval_config={"vector_store_id": "vs_test_123", "custom_llm_provider": "openai"},
+            filters=top_level_filter,
+            mock_response="hi",
+        )
+
+    assert isinstance(response, ModelResponse)
+    fake_search.assert_awaited_once()
+    assert fake_search.await_args.kwargs["filters"] == top_level_filter
+    assert "filters" not in fake_search.await_args.kwargs.get("kwargs", {})
+
+
 @pytest.mark.asyncio
 async def test_aquery_without_filter_forwards_none():
     """