This is the same route I followed for https://zenquery.app .... It uses LLM to generate SQL rather than working directly on data files. Saves a ton of costs as well since you don't need to send entire file(s) to LLM, just the schema.