Structured Output

Structured Output is a language model capability that forces the response to conform to a predefined schema, such as JSON matching a specific set of fields and types, instead of free-form prose. It exists because software cannot reliably parse a paragraph. Without it, developers wrote fragile regular expressions to pull values out of an answer and retried whenever the model added a preamble, a code fence, or an extra field. With schema enforcement, the model is constrained during generation so the output parses on the first attempt. OpenAI ships this as Structured Outputs backed by JSON Schema, Anthropic and Google reach the same result through tool and function definitions, and libraries such as Zod, Pydantic, and Instructor let a developer declare the shape once and validate the response against it. Typical uses are extracting fields from an invoice, classifying a support ticket into fixed categories, and returning arguments for a function the application will call. The pitfall is confusing valid with correct: the schema guarantees the shape, never the facts, so a confidently wrong value still arrives well-formed. Overly deep or ambiguous schemas also degrade quality, and flat, clearly named fields work better.