Right now you create new httpx client per request, which is probably inefficient.
It not only waste resources to initialization/finalization every time, but also establish new connection every time, preventing HTTP connection pooling.
Instead, httpx client should be initialized once [1] [2] at initialization of class which use it and then closed when class instance is no longer used, which is more efficient.
OpenAI doing exactly that for their client:
https://github.com/openai/openai-python/blob/b95be16e7c8a76c3d63335df13ab0d55ba3d5c35/src/openai/_base_client.py#L1449