Skip to main content
ZICQ

Wiki AI Concepts

Structured Output

AI Concepts
Aliases: Structured Output Structured Outputs Schema-constrained Generation 结构化输出 ·2026-10-07

Structured Output

Structured output returns data in an agreed shape for programmatic consumption. It addresses format contracts, not factual accuracy or authorization.

Levels of constraint

Method Intended constraint Not established
Prompt asks for JSON Instruction-based formatting Always parseable JSON
JSON mode Valid JSON, implementation dependent A particular schema
Schema-constrained output Supported fields and types Truth, evidence, authorization
Strict tool arguments Parameter structure Successful execution

JSON Schema describes structure; complex semantic relationships may require application validation.

Check completion status

Consult documentation for the model and supported schema subset. Claude structured-output documentation describes refusals and token-limit truncation that may break schema compliance. Check completion state before parsing and validating; HTTP success alone is insufficient.

Engineering recommendations

For invoice extraction, define invoice_number, currency, total, and source_page. Permit null for missing evidence and represent incomplete extraction explicitly.

Validate amounts, currencies, relationships, and source locations. Check authorization before database writes. Version the schema and input, log validation failures, and bound retries.

See JSON mode, function calling, hallucination, and prompt injection.