BlueHomer
Text-to-SQL
Text-to-SQL converts a plain-language question into a database query. It fails when the model must guess business meaning it was never given — producing confident, wrong results.
Definition
Text-to-SQL is the translation of a plain-language question into a database query by a language model. It works well when the question maps onto a well-modelled dataset, and fails when the model has to infer business meaning — grain, join paths, filters, definitions — that nobody gave it.
Why it matters
- The failure is silent: a query can be syntactically valid, run successfully, and return the wrong number in a confident tone.
- Users cannot easily detect the error, because the interface offers an answer rather than a derivation.
- It is the main reason "AI analytics" pilots stall on trust rather than on capability.
How BlueHomer implements it
BlueHomer removes the guesswork rather than tuning the guess: metrics and measures are defined once in a governed semantic layer with grain and meaning declared, assets carry certification the agent can read, queries execute under the acting user’s permissions, and every answer carries attribute-level lineage so it can be checked instead of trusted.
Frequently asked questions
Can text-to-SQL be made accurate?
Accuracy improves dramatically when the model is given governed definitions instead of raw schema: declared grain, one legitimate join path, certified assets and permission-scoped execution. Accuracy is mostly a property of the foundation, not of the model.