Self hosted AI to train on an existing database

Vector embeddings with ChromaDB. Basically you pre compute the word embeddings of every row / table / whatever granularity you want and then stick that into a vector DB. Then you do an embedding computation of your query and compare similarity. You can either return the table / row / whatever you want that’s most similar (“semantic search”) or you use that as context for an LLM (“RAG”)

You could use something like https://huggingface.co/defog/sqlcoder-34b-alpha / https://github.com/defog-ai/sqlcoder however I haven’t used this one myself but it builds up on CodeLlama model so the quality should be good. I haven’t seen other models specifically for SQL queries yet.

Edit: it includes all the hardware requirements and some demo examples in the links