[morkalork]

I don't think such tools will be reliable until backed with knowledge of the codebase that populates the data. Just because a column has a certain name and has plausible looking data in it doesn't mean that's what's really in there. Especially if there are many tables and columns with similar names and data that have nuances for what differentiate them. Nuances that live in code and the heads of developers. I would not trust any LLM right now to do any financial reporting on any of the production data warehouses I've seen, it would be a recipe for shooting yourself in the dick.

[mambrogi]

In practice you'd probably try to limit to just a few key tables. Probably limit columns the model is controlling as well and provide a lot of detail about what each captures. I think this would help a lot. But overall I don't disagree.