In the product I work on, there's a Metadata layer that maps user-supplied definitions, notes and commentary onto the raw table descriptions

These are optional but it allows manual override when schema design is poor

This is the key issue. I call it a semantic gap.

Example is a transaction application with a backing database: The users conduct information interactions with the application. The application stores relatively sparse data into the transactional database. The database has just enough metadata to route data from transactions but doesn't have the full meaning.

Example database column "updated_date" with type timestamp: Without looking at the data, how do you know if this column is populated when the application first stores data in a row? How do you know the timezone of the timestamp as a user would see it?

Real world example: Patient height in app and api is SI and both note the units. Patient height in the data warehouse is US Customary with no unit notation. The app and api make elaborate conversions for the underlying transaction database's US customary storage.

I think the promise of an LLM ontology is to read application code to understand the context for any particular database column. The linked software is not that.