fix: prevent VertexAiRagRetrieval from blocking the event loop

VertexAiRagRetrieval.run_async() called rag.retrieval_query() directly. That
call issues a synchronous, blocking gRPC request which freezes the asyncio
event loop for ~1-2s on every invocation, even though run_async is an async
method. This is especially harmful for real-time/streaming use cases
(StreamingMode.BIDI with gemini-live models), causing audio/video desync and
dropped frames.

Defer the blocking call to a worker thread via asyncio.to_thread, matching the
established pattern already used across ADK (e.g. the GCS/file artifact
services and the Spanner/BigQuery tools).

Closes #5033

Co-authored-by: Haran Rajkumar <haranrk@google.com>
PiperOrigin-RevId: 934014525

Author: 2 people

src/google/adk/tools/retrieval/vertex_ai_rag_retrieval.py

@@ -16,6 +16,7 @@
 
 from __future__ import annotations
 
+import asyncio
 import logging
 from typing import Any
 from typing import TYPE_CHECKING
@@ -94,7 +95,8 @@ async def run_async(
   ) -> Any:
     from ...dependencies.vertexai import rag
 
-    response = rag.retrieval_query(
+    response = await asyncio.to_thread(
+        rag.retrieval_query,
         text=args['query'],
         rag_resources=self.vertex_rag_store.rag_resources,
         rag_corpora=self.vertex_rag_store.rag_corpora,