[Bug]: AzureAISearchVectorStore leaks aiohttp sessions created via get_search_client()

[mangekyousharingan]

Bug Description

When AzureAISearchVectorStore is constructed with BOTH a SearchIndexClient and an AsyncSearchIndexClient that are managed with context managers, the store internally calls:

self._async_search_client = self._async_index_client.get_search_client(index_name)

The AsyncSearchIndexClient context manager cleans up its own aiohttp session on exit, but the AsyncSearchClient returned from get_search_client spins up a new aiohttp.ClientSession/TCPConnector that is never closed.

As a result, the asyncio loop logs:

Asyncio unhandled exception in task 'unknown': Unclosed client session
and
Unclosed connector

To avoid this you have to ad finally block which closes the sessions:

    with SyncSearchIndexClient(endpoint=endpoint, index_name=index_name, credential=creds) as sync_client:
        async with AsyncSearchIndexClient(endpoint=endpoint, index_name=index_name, credential=creds) as async_client:
            vector_store = AzureAISearchVectorStore(
                search_or_index_client=sync_client,
                async_search_or_index_client=async_client,
                index_name=index_name,
                index_management=IndexManagement.CREATE_IF_NOT_EXISTS,
                id_field_key="id",
                chunk_field_key="chunk",
                embedding_field_key="embedding",
                metadata_string_field_key="metadata",
                doc_id_field_key="doc_id",
            )
          
            try:
                yield vector_store
            finally:
                async_search_client = getattr(vector_store, "_async_search_client", None)
                if async_search_client:
                    await async_search_client.close()
  
                search_client = getattr(vector_store, "_search_client", None)
                if search_client:
                    search_client.close()

Version

0.12.52

Steps to Reproduce

from azure.search.documents import SearchIndexClient
from azure.search.documents.indexes import SearchIndexClient as SyncSearchIndexClient
from azure.search.documents.indexes.aio import SearchIndexClient as AsyncSearchIndexClient
from azure.core.credentials import AzureKeyCredential
from llama_index.vector_stores.azureaisearch import AzureAISearchVectorStore, IndexManagement

async def main():
    creds = AzureKeyCredential("<api-key>")
    index_name = "llama-test"
    endpoint = "https://<service-name>.search.windows.net"

    with SyncSearchIndexClient(endpoint=endpoint, index_name=index_name, credential=creds) as sync_client:
        async with AsyncSearchIndexClient(endpoint=endpoint, index_name=index_name, credential=creds) as async_client:
            vector_store = AzureAISearchVectorStore(
                search_or_index_client=sync_client,
                async_search_or_index_client=async_client,
                index_name=index_name,
                index_management=IndexManagement.CREATE_IF_NOT_EXISTS,
                id_field_key="id",
                chunk_field_key="chunk",
                embedding_field_key="embedding",
                metadata_string_field_key="metadata",
                doc_id_field_key="doc_id",
            )
            # use vector_store, e.g. insert nodes

    # Exiting the context emits "Unclosed client session" warnings

Relevant Logs/Tracbacks

Asyncio unhandled exception in task 'unknown' (coro: unknown): Unclosed client session
Asyncio context dump: {'client_session': '<aiohttp.client.ClientSession object at 0xffff674efef0>', 'message': "'Unclosed client session'"}
Asyncio context without exception: {
  'client_session': <aiohttp.client.ClientSession object at 0xffff674efef0>,
  'message': 'Unclosed client session',
  'source_traceback': [
    ...,
    <FrameSummary file /Users/sebix/Repositories/Newcode-Nova-backend/src/app/utils/llama/index.py, line 309 in create_embeddings_from_nodes>,
    <FrameSummary file /usr/local/lib/python3.12/site-packages/llama_index/vector_stores/azureaisearch/base.py, line 965 in async_add>,
    <FrameSummary file /usr/local/lib/python3.12/site-packages/azure/search/documents/aio/_search_client_async.py, line 645 in index_documents>,
    <FrameSummary file /usr/local/lib/python3.12/site-packages/azure/search/documents/aio/_search_client_async.py, line 653 in _index_documents_actions>,
    <FrameSummary file /usr/local/lib/python3.12/site-packages/azure/core/pipeline/_base_async.py, line 229 in run>,
    <FrameSummary file /usr/local/lib/python3.12/site-packages/azure/core/pipeline/_base_async.py, line 77 in send>,
    ...,
    <FrameSummary file /usr/local/lib/python3.12/site-packages/azure/core/pipeline/transport/_aiohttp.py, line 149 in open>
  ]
}

[dosubot]

Hi @mangekyousharingan! I'm Dosu and I’m helping the llama_index team.

You’re correct—AzureAISearchVectorStore does not manage the lifecycle of the aiohttp sessions created by the AsyncSearchClient returned from get_search_client. This leads to unclosed session warnings unless you manually close those clients after use. There’s no context manager or cleanup logic for these clients in the current implementation, and no documentation warning about this requirement. Your workaround using a finally block to close both _async_search_client and _search_client is necessary to avoid resource leaks. This is a design gap in the vector store’s resource management and hasn’t been addressed in recent pull requests or discussions.
Source: AzureAISearchVectorStore implementation

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[Prozac614]

Hi @mangekyousharingan and @dosu,

I'd be happy to work on this issue.

From what I understand, the unclosed session warnings come from the AsyncSearchClient created internally by the vector store when using get_search_client(), and since the store doesn't manage its lifecycle, those sessions never get closed.

My plan is to add proper cleanup for internally-created clients and make sure the vector store handles its resources more safely, while keeping everything backward-compatible.

Please let me know if this direction sounds good!