A model calls a create_refund function through the Chat Completions API.

How it works

Function calling lets the model return arguments for your functions. To guarantee those arguments always match the function's JSON Schema, you use structured outputs for function calling, which takes two settings:

1. strict: true in the function definition. That single parameter turns structured outputs on for the function.
2. parallel_tool_calls: false on the request. Structured outputs aren't supported with parallel function calls, so the structured outputs page tells you to set it to false.

Structured outputs constrain the model to your schema. Microsoft contrasts this with the older JSON mode, which guaranteed valid JSON but couldn't ensure strict adherence to the supplied schema.

The other options, and when you'd use them

• response_format json_object: That's JSON mode. It guarantees valid JSON but not that the output matches your schema.
• tool_choice none: It forces a user-facing message, so the model wouldn't call create_refund at all. The default, auto, lets the model decide whether and which function to call.
• parallel_tool_calls true: Structured outputs don't support parallel function calls, so this works against the guarantee instead of helping it.

Exam tip

Schema-exact function arguments = strict: true + parallel_tool_calls: false. JSON mode = valid JSON only.

References

• Structured outputs
• How to use function calling with Microsoft Foundry Models
• Learn how to use JSON mode