A tiny four kilobyte rulebook added to a database makes even average AI models perform as well as the best in the world.

AI performance on data analytics is usually attributed to the intelligence of the model itself. This benchmark shows that providing a small semantic layer of business definitions matters far more than model size. With this context, different frontier models perform almost identically on text-to-SQL tasks. Most errors are caused by the database schema being too confusing for a machine to understand without help. This means companies don't need smarter models to get accurate data insights, they just need to document their data better. A little bit of context goes a very long way in AI accuracy.