Structured Output

A mode where an LLM returns its answer in a predefined format (usually JSON), guaranteed to be valid and ready for downstream processing.

What is structured output?

Structured output is an LLM feature in which the model returns its answer in a strict format instead of free text - typically JSON conforming to a supplied schema. The model is guaranteed to produce syntactically valid output, so it can be fed directly into the next step of a workflow without risking parsing errors.

How to enable structured output

JSON mode: A simple toggle in the API, the model returns valid JSON (but no guarantee about specific fields)
JSON schema: You pass a schema with required fields and types - the model follows it exactly
Tool/function schema: With function calling, structured output is effectively mandatory

Typical use cases

Data extraction from emails or documents (name, ID, amount, date)
Ticket classification: {"category": "billing", "priority": "high"}
Chatbot outputs for machine processing
Steps of AI agents - each next action is a structured record

Recommendations

For structured output set temperature close to 0 for consistency
Define the schema as tightly as possible - optional fields increase variability
Always validate the schema on your side too, not all models comply 100%