No comments yet.
These models don’t see application code. They can’t ask teammates. They just generate SQL from whatever context the database gives them.
Most of the time, that context is not enough. Schema names don’t explain business logic or domain-specific meaning.
So we built a semantic catalog for Postgres: a layer of natural language descriptions for tables, columns, functions, and key facts.
We tested it by comparing SQL generation with and without the catalog. Same model. Same prompts. The only difference was the context.
Accuracy improved by 27 percent.
The results were strong, so we open sourced it. Try it, break it, or tell us what’s wrong.
