Structured Output

It is having an LLM produce an answer in a specified format — typically JSON matching a schema, or a tool call with typed arguments — rather than free prose. The provider enforces the structure during decoding, so your downstream code can parse the output reliably.

Plain JSON mode just guarantees the output is valid JSON — any shape. Structured outputs (OpenAI) or tool-use schemas (Anthropic) also guarantee the JSON matches a supplied schema, with constrained decoding ensuring field names, types, enums, and required fields are all respected.

At each token step, the decoder masks out tokens that would violate the schema — for example, after '{"name":' only string tokens are legal, after a closing brace only whitespace or end-of-message. Tools like Outlines, llguidance, or XGrammar build a finite-state machine from the schema.

Sometimes slightly. The model can be less fluent when constrained, and overly nested or ambiguous schemas can confuse it. Best practice: keep schemas flat where possible, let the model reason in a 'thoughts' field first, and validate with Pydantic/Zod after. For complex tasks, generate reasoning then a structured answer.