What about some generation-augmented retrieval augmented generation set-up where all your conversations are indexed for regular text search, and then you use the LLMs language knowledge to generate relevant search phrases the results of which are included in the current prompt?
