Awesome... for databases containing large number of tables, you can pre-process the tables and generate embeddings for each. Then, when user asks a question in plain english, filter relevant tables using in-built vector search and pass metedata of these only as context to LLM.

Happy Holidays..