Yeah. This seems like an area where a “tiny” (2-4GB) local model would be more than sufficient to generate very high quality queries and schema answers to the vast majority of questions. To the point that it feels outright wasteful to pay a frontier model for it.