I am working on a NL2SQL pipeline where user sends a query and we have to turn it into a SQL query so I have query expansion rules basically the normalization rules and chart generation rules, SQL generation rules. data and schema detailed info of database so I thought it would be easy just to implement RAG and store this rules in a vector store and retrieve using a semantic search I started learning this topics as a beginner I have some doubts like in RAG do the format of the file really matters actually I am using python file where huge 200line items are instructions in a python dictonary and tried to change them to json and tried reranking as well using BM25 and also tried LLM reranking it still does not work that good like if I have ask a question it might miss a rule 1 out of 3 times which is not good, I think I overengineered it some times and is there any professionals working on similar tasks please help me.