[Bug]: QueryFusionRetriever._aretrieve blocks the event loop during query generation #21159
Description
Bug Description
_aretrieve() calls the synchronous _get_queries(), which blocks on self._llm.complete() instead of awaiting an async equivalent. When num_queries > 1 (the default), this blocks the current event loop during query generation and prevents other coroutines on that same loop from making progress until query expansion finishes.
Version
0.14.19
Steps to Reproduce
import asyncio
from llama_index.core.base.base_retriever import BaseRetriever
from llama_index.core.retrievers import QueryFusionRetriever
from llama_index.core.schema import NodeWithScore, QueryBundle, TextNode
from llama_index.llms.openai import OpenAI
class MockRetriever(BaseRetriever):
def _retrieve(self, query_bundle):
return [NodeWithScore(node=TextNode(text="Hi I am Gautam!"), score=1.0)]
retriever = QueryFusionRetriever(
retrievers=[MockRetriever()],
llm=OpenAI(model="gpt-5"),
num_queries=4,
)
async def other_work():
await asyncio.sleep(0.1)
print("other work ran")
task = asyncio.ensure_future(other_work())
await retriever.aretrieve("What is the capital of France?")
print(f"other_work done: {task.done()}") Relevant Logs/Tracbacks
other_work done: False
So basically the task had the entire duration of a real OpenAI HTTP call to complete its
0.1s sleep and still never got scheduled.