Google Gemini Functions and Structured Output

Choose the correct control

Use response_format with JSON MIME type and a schema when the interaction should return constrained JSON. Use a function tool when the model should request an application capability. Gemini returns a function_call step; the application decides whether it is known, valid, authorized, and approved before execution, then returns a function_result linked by call_id.

Do not invent parity with another provider's strict mode. Close schemas where supported, but validate both structure and domain meaning in application code. Authenticate and authorize with trusted runtime context, require human approval for consequential actions, use idempotency for mutations, and treat tool descriptions as routing hints rather than enforcement.

Define explicit Interactions contracts

The following current request fragments show Gemini Interactions API shapes. Insert one into a reviewed request and replace bracketed values with non-secret domain fields.

const structuredOutput = {
  response_format: {
    type: "text",
    mime_type: "application/json",
    schema: {
      type: "object",
      properties: { decision: { type: "string", enum: ["approve", "review", "reject"] }, reason: { type: "string" } },
      required: ["decision", "reason"],
      additionalProperties: false,
    },
  },
};

const functionTool = {
  tools: [{
    type: "function",
    name: "record_review_decision",
    description: "Record an already authorized review decision for [SYSTEM_OF_RECORD]",
    parameters: {
      type: "object",
      properties: { item_id: { type: "string" }, decision: { type: "string", enum: ["approve", "reject"] } },
      required: ["item_id", "decision"],
      additionalProperties: false,
    },
  }],
};

Expected evidence and verification

For structured output, expect completed status and model_output text containing JSON that passes the same application schema. For function calling, expect requires_action with a function_call step containing a known name, ID, and arguments; after authorized execution, return a function_result with the original call_id and verify the final interaction. Test valid, missing, extra, malformed, unauthorized, duplicate, blocked, and truncated cases.

Do not execute when parsing, validation, approval, or authorization fails. Record a redacted failure and request correction or human review within a retry budget. Rollback means discarding untrusted output; for an executed function, use its idempotency key and tested compensation or reconciliation path. Schema compliance does not guarantee factual correctness or safety.