Structured Output

Structured output is the practice of requiring a model's response to conform to a declared schema — typically expressed as JSON Schema — so the result can be consumed by code without parsing prose. It underlies both Tool Use and any pipeline that needs machine-readable results.

The important subtlety is that not every part of a schema is enforced the same way. Providers commonly compile the schema into a grammar that constrains generation, but only for the parts a grammar can express: object shape, property names, types, enumerations, required-versus-optional. Constraints like minimum length, maximum length, item counts, and regular-expression patterns are typically stripped out of the grammar and appended to the field's description as prose. The model is then merely asked to respect them, while the client library still validates the reply against the full schema — so a bound set close to the length actually wanted produces intermittent hard failures.

Two consequences follow. Put the intended target in the description and keep any hard limit far away as a runaway backstop, per Anchoring Effect. And treat anything the grammar cannot express — custom refinements, cross-field rules — as runtime validation only, invisible to the model, which means the failure message is the only channel that teaches it what is legal.