Source: Gemini API Structured Outputs Docs
Google Gemini API documentation on JSON Schema structured outputs: schema compiled to a state machine, unenforced keywords and complexity limits.
Contents
- What: official API documentation page
- Author/publisher: Google
- Date: living document; read 2026-11-14
- Original: Structured outputs
Summary
Gemini constrains responses to a JSON Schema supplied as response_json_schema (generateContent) with response_mime_type: application/json, or via the Interactions API’s response_format object (mime_type: application/json plus schema); SDKs accept Pydantic and Zod. The API compiles the schema into a state machine for constrained decoding and rejects schemas that exceed state, branching or depth limits.
Key claims
- Use cases: data extraction, structured classification, generating structured inputs for agentic workflows.
- Structured output formats the final answer; function calling is for taking actions mid-conversation.
minLength,maxLength,minProperties,maxProperties,uniqueItems,containsare silently stripped (not enforced);minItems/maxItemsare enforced.- Over-complex schemas get
400 INVALID_ARGUMENT; fix by moving rules into descriptions, flattening nesting, bounding recursion, and reducingenum/anyOfbranching. - Output is syntactically valid but values must still be validated in the application.
- Streamed structured output arrives as valid partial JSON strings that concatenate to the full object.
- Gemini 2.0 needed an explicit
propertyOrderinglist.