What is Function calling?

Function calling, also called tool calling in OpenAI's documentation, lets a model reach systems and data that sit outside its training. The developer describes the functions the model may request. The model never runs them itself. It returns a tool call, and the application is responsible for executing the function.

OpenAI describes a loop of five steps. The application sends a request that lists the available tools. The model returns a tool call. The application runs the function with the supplied arguments. The application sends the output back to the model. The model then produces a final answer, or issues more tool calls, and the loop can repeat.

A function definition has a type of function, a name, a description of when and how to use it, a parameters field written as a JSON schema, and an optional strict flag. In strict mode the call is meant to follow the schema reliably, which requires additionalProperties set to false on every object and every property listed as required. Optional values are expressed by allowing null.

The page lists some limits. Function definitions count against the context window and are billed as input tokens. Some JSON schema features are not supported in strict mode. As a soft guideline, it suggests keeping fewer than about 20 functions available at the start of a turn.