Combining schema discovery and data catalog seems like it might be a hard problem requiring a lot of LLM prompt engineering gymnastics but maybe I underestimate the state of the art.