If using RAG, one thing that comes to mind is how you are searching. A straight cosine similarity search will get you the results you are getting. Sounds to me like you need more of a hybrid search, with more emphasis on keyword than semantic.
In fact, I’m wondering if you couldn’t handle the database calls with function calling as opposed to embeddings? See this discussion about using data from CSV file: Private Chat with CSV data
Essentially, you can create a nice shiny high-tech LLM front end but use good old fashioned dbms with SQL for data retrieval. You should be able to do this with either the Assistants or Chat Completion APIs.