Skip to main content
Turning natural language into API calls is one of the most direct applications of structured output. The user says “refund $25 for order ORD-9842,” and the model produces a valid POST /refunds request with the right body fields. The schema defines every endpoint as a oneOf branch with const values for method and path, so the model can’t invent endpoints that don’t exist or mix parameters from different calls. This pattern is especially valuable for internal tools where users interact with backend APIs through a natural language interface. The schema acts as a whitelist of allowed operations.

​
Goal

Map natural language intent into exactly one valid API call definition, with validated parameters, ready for your backend to execute.

​
Schema contract

{
  "oneOf": [
    {
      "type": "object",
      "properties": {
        "endpoint": { "const": "get_order" },
        "method": { "const": "GET" },
        "path": { "const": "/orders/{order_id}" },
        "path_params": {
          "type": "object",
          "properties": {
            "order_id": { "type": "string", "pattern": "^ORD-[0-9]{4,10}$" }
          },
          "required": ["order_id"],
          "additionalProperties": false
        }
      },
      "required": ["endpoint", "method", "path", "path_params"],
      "additionalProperties": false
    },
    {
      "type": "object",
      "properties": {
        "endpoint": { "const": "create_refund" },
        "method": { "const": "POST" },
        "path": { "const": "/refunds" },
        "body": {
          "type": "object",
          "properties": {
            "order_id": { "type": "string", "pattern": "^ORD-[0-9]{4,10}$" },
            "amount": { "type": "number", "minimum": 0.01 },
            "reason": {
              "type": "string",
              "enum": ["duplicate", "fraud", "requested_by_customer", "other"]
            }
          },
          "required": ["order_id", "amount", "reason"],
          "additionalProperties": false
        }
      },
      "required": ["endpoint", "method", "path", "body"],
      "additionalProperties": false
    }
  ]
}

​
Example input

Issue a $25 refund for order ORD-9842 because the customer was charged twice.

​
Example output

{
  "endpoint": "create_refund",
  "method": "POST",
  "path": "/refunds",
  "body": {
    "order_id": "ORD-9842",
    "amount": 25,
    "reason": "duplicate"
  }
}

​
Implementation tips

  • const on method and path locks each branch. The model can’t produce "method": "DELETE" for the refund endpoint or change the path. This makes the schema a whitelist of allowed operations.
  • oneOf ensures a single action. The model commits to one endpoint per request. If you need multi-step operations, run the model in a loop; don’t try to batch multiple API calls into one schema.
  • Parameter-level constraints catch bad input early. The order_id pattern (^ORD-[0-9]{4,10}$) and amount minimum (0.01) reject invalid values at generation time, before your API sees them.
  • Validate before executing. Even with a strict schema, run the output through your API’s own validation layer. The schema catches structural issues; your API validates business rules (e.g., “this order is not eligible for refund”).