Structured Output

Structured output refers to the ability of language models to generate responses that conform to a predefined data schema. While JSON mode ensures the output is valid JSON, structured output goes further by enforcing that the JSON contains specific fields with specific types, matching a schema provided by the developer. For example, when extracting information from a customer email, a structured output schema might specify that the response must contain a 'sentiment' field (enum: positive/negative/neutral), a 'topic' field (string), an 'urgency' field (boolean), and a 'key_entities' field (array of strings). The model's output is guaranteed to match this schema. This is achieved through constrained decoding — modifying the model's token selection process to only allow tokens that would produce valid output according to the schema. This eliminates parsing failures and ensures that every API call returns data in exactly the expected format.

Structured output is what makes AI reliable enough for production automation. When an AI model's output feeds directly into a database, triggers a workflow, or populates a user interface, that output must conform to a precise format. A missing field, wrong type, or malformed structure causes failures that degrade user experience and operational reliability. Before structured output, developers spent significant effort on output parsing, validation, retry logic, and error handling. Structured output eliminates these concerns, reducing development time and making AI integrations more robust. It also simplifies testing, as the output format is deterministic. Common business applications include data extraction from documents (extracting structured data from invoices, contracts, or forms), content classification (categorising support tickets with specific labels), and workflow automation (generating structured action plans that feed into project management tools).