Function Calling (Tool Use)

Developers register a set of tools by providing the model with their names, descriptions, and parameter schemas (typically as JSON Schema). When the model decides a tool is needed, it halts generation and outputs a structured function call object. The host application executes the real function, then returns the output as a new message. The model resumes generation with the grounded result in context.

• • Tool schemas describe what each function does and what parameters it accepts.
• • The model decides when (and if) to call a tool based on the conversation.
• • Results are fed back into the context for the model to use in its final response.

Modern models can issue multiple function calls in a single turn (parallel tool use) when different pieces of information are independent. For dependent queries, the model issues calls sequentially, using each result to determine the next action. This design pattern is what transforms a chat model into a capable AI agent.