Structured output makes the model return JSON or a fixed schema instead of free-form prose.
DefinitionWhat it means
Structured output constrains a model's response to a predefined schema, typically JSON with specific fields and types, either through prompt instructions, a function or tool-call interface, or constrained decoding that guarantees the grammar is valid. Instead of parsing prose with regular expressions, downstream code reads a predictable object.
Why it mattersWhy you should care
Any system that pipes model output into another function, a database write, an API call, a UI component, needs structured output to be reliable; a missed field or malformed bracket breaks the pipeline. Modern APIs increasingly offer schema-enforced or guaranteed JSON modes precisely because free-text parsing was the single biggest source of silent production failures in early LLM apps.
At a glanceSee it
A taxonomy of how you actually get structured output — from politely asking in the prompt up to grammar-constrained decoding, ordered by how hard each one guarantees the schema.
The runtime validate-and-repair loop — parse, check against the schema, and feed any error back for another attempt, while semantically wrong but well-typed values still slip through as the quiet failure.
Where you see itIn the wild
- Function calling or tool-use APIs that return arguments as typed JSON.
- Data extraction pipelines pulling fields like name, date, and amount from documents.
- Form-filling or UI-generation agents that must emit a schema the frontend can render.