Skip to content

Structured outputs & tools

response_format

Azure AI Foundry’s chat completion reference and OpenAI’s chat completion reference document response_format with:

  • { "type": "text" } (default)
  • { "type": "json_object" } (JSON mode) — valid JSON, but schema adherence is not enforced
  • { "type": "json_schema", "json_schema": {...} } (Structured Outputs) — enforces a JSON Schema

Important note from Azure docs: in JSON mode (json_object), you should also instruct the model to output JSON; otherwise the model may stream whitespace until it reaches the token limit.

Microsoft Learn describes Structured Outputs as a way to force the model to follow a supplied JSON Schema, and contrasts it with older JSON mode that only guarantees valid JSON.

tools and tool_choice

Azure OpenAI REST reference documents:

  • tools: list of function tools the model may call
  • tool_choice: control whether tools are used (none/auto/required) or force a specific tool.

This playbook uses tools sparingly, but you can incorporate them into experiments to measure impacts on accuracy and cost.

Next: Experimentation.