Interesting approach! I can see some potential latency challenges arising from the LLM's generation of (and inference of) multiple queries, as well as from RRF. Will definitely try this out.