Is there any way I could store meta data for each assistant response (I'm using custom llm) Maybe I can somehow input that meta data into "source": "meta data here"