Yo!
I am trying to build a RAG system for a database.
The purpose of this would be to give to the RAG system the documentation(txt) of my DB and ask to the LLM(llama3) questions about the DB in real time.
The first step is to get correct SQL queries.
When I give to the RAG ultra-small documentation it work perfectly fine.

HOWEVER after tests when I load large document, I don’t have complete answer.
From my understanding of the RAG & LLMs, question about large DB can’t be perfectly answered because the LLM would have to understand the doc as a whole, and not only parts of it.

For exemple, if the DB contains 100 tables, with 5 fields by tables. How can I be sure that the LLM look every table to make the good queries? RAG system works with embeddings, which mean that they “look for the closest sense” between what you said and the doc you loaded.
Is my understanding wrong? Are LLM not “smart” enough yet? Is my RAG system not good enough?

I have no idea