OpenAI Function Calling Schema Builder

Function calling (also called 'tool use') lets LLMs decide when to invoke predefined functions and extract the right arguments from natural language. Instead of parsing free-form text, the model returns structured JSON that maps exactly to your function's parameters. This enables reliable structured output, API orchestration, and multi-step agent workflows.

Both use JSON Schema for parameter definitions but have slightly different wrapper formats. OpenAI uses type: 'function' with a nested 'function' object containing name, description, and parameters. Anthropic uses a flatter structure with name, description, and input_schema directly in the tool object. This builder generates both formats.

Both APIs support the core JSON Schema types: string, number, integer, boolean, array, object. Strings can have enum constraints (picklist), format hints (date, email, uri), and minLength/maxLength. Numbers support minimum/maximum. Arrays support items type. Objects support nested properties recursively.

OpenAI's strict mode (strict: true) enforces that the model can only output parameters defined in your schema — no extra fields. This guarantees structured output but requires all required fields to be explicitly listed and does not allow additionalProperties. Use strict mode when you need guaranteed JSON compliance in production.

Keep it focused. 1–3 functions works best for most use cases. More than 10 functions can confuse the model and increase latency. If you have many operations, group related ones into a single function with an action parameter (e.g. action: 'create'|'update'|'delete') rather than three separate functions. Always test which function(s) actually get called with your specific prompts.