What Is Structured Output? – Definition and How It Works

Learn what structured output is, how schema-constrained generation works, why it beats parsing prose, and how to handle validation failures.

Structured output is a model returning data that conforms to a declared schema, so downstream code can consume typed fields instead of parsing prose.

1. What It Is

A contract on the shape of the response.

• Key idea: the caller declares the shape; the model fills it.
• Mechanism: a JSON Schema supplied with the request, enforced during generation or validated after.
• Goal: make model output a dependable input to ordinary software.

Example Schema

{
  "type": "object",
  "properties": {
    "title":  { "type": "string" },
    "price":  { "type": "number" },
    "in_stock": { "type": "boolean" },
    "currency": { "type": "string", "enum": ["USD", "EUR", "GBP"] }
  },
  "required": ["title", "in_stock"]
}

2. How Enforcement Works

Two mechanisms exist, and they differ in strength.

• Constrained decoding. The schema restricts which tokens may be generated, so output is valid by construction. Strongest guarantee.
• Post-generation validation. The model is asked for JSON and the result is validated, with a re-prompt on failure. Weaker, and works with any provider.

Constrained decoding guarantees shape, not correctness: a valid object can still contain a wrong value.

3. Why It Beats Parsing Prose

That last row is the real gain: a schema turns a silent wrong value into a visible failure.

4. Designing a Good Schema

• Use enums for closed sets; they eliminate a whole class of invalid value.
• Mark required fields deliberately, and allow null where absence is meaningful.
• Describe each field. Descriptions are read by the model and improve accuracy.
• Keep it flat where possible; deep nesting increases error rates.
• Avoid free-form catch-alls, which reintroduce parsing.

5. Handling Failures

Even with validation, values can be wrong or implausible. Layer checks:

1. Schema validation for shape.
2. Range and plausibility checks for values.
3. Cross-field consistency — a discount larger than the price is invalid whatever the schema says.
4. A bounded re-prompt path that includes the specific validation error, since a model can correct a fault it is told about.

6. Common Uses

• Extracting fields from unstructured text or scraped pages.
• Classification with a fixed label set.
• Tool arguments, which are structured output by another name.
• Multi-step pipelines where one step's output is the next step's typed input.

7. Limits

• Shape is not truth. A confidently wrong number passes validation.
• Constraints can distort. Forcing a value where none exists produces invention; allow null.
• Complex schemas degrade quality. Split large extractions into several calls.
• Support varies. Not every provider offers constrained decoding.

8. Summary

Structured output makes a model's response conform to a declared schema, turning generation into a typed interface rather than text to be parsed.

Design schemas with enums, descriptions, and honest nullability; validate shape and then plausibility; and remember that a schema guarantees the container, never the contents.

Potential Related Glossary

• Data Structure
• Parsing
• Unstructured Data
• Data Wrangling