I'm not using namespaces right now. Each chunk type has a 'level' property set to either "verse", "section", or "chapter". For querying the index, I kick off a pool of three concurrent requests; one for each 'level'. I'm getting 5 verses, 5 sections, and 2 chapters. That can be a lot, or not as much, it's just what I'm doing atm. What I need to implement this week is a way to assess if the user message even needs to run through the whole retrieval pipeline. I was thinking I might add a request at the start for an LLM to assess that. Although it might add a little latency to requests where the RAG is needed, on average, responses might seem quicker or the same since not every message will need the extra context.