I see, I heard of such trickery first here https://www.youtube.com/watch?v=c9QE9oIvlEU

He was not using any kind of DB there at all :D

Rather he had collection of documents

Was using LM to create keyowrds and descriptions and had index file that is small enough to fit in context.

So when you add article to this system

LLM is asked to produce description/keywords

they are added to index file

whole document is saved

when you search it first opens index file, finds relevant descriptions/keywords there, based on the opens full articles :D

LLM as a DB :D

But yeah that should be combined with RAG for cost/performance reasons.