Skip to main content
Function calling bridges the gap between language models and real-world systems. Instead of only generating text, the model can signal that it wants to invoke a function you define — looking up a database, calling an external API, or running a calculation. Your code executes the function and returns the result, and the model incorporates that result into its final response. This is the foundational pattern for building AI agents.

​
How It Works

You include a tools array in your request describing the functions available to the model. If the model determines that a function call is the best way to answer the user’s request, it returns a response with finish_reason: "tool_calls" and a tool_calls array instead of a plain text reply. You then execute the requested function with the provided arguments, send the result back as a tool role message, and make one more API call to get the model’s final answer.

​
Step-by-Step

​
Controlling Tool Use with tool_choice

The tool_choice parameter lets you override the model’s default behavior. Use "none" when you want to ask the model a follow-up question without triggering a tool call. Use a specific function name when the user action unambiguously maps to a single function and you want to skip the model’s decision step.

​
Parallel Function Calls

The model can request multiple function calls in a single turn by returning more than one entry in the tool_calls array. Execute all requested functions in parallel, then send all results back in one follow-up request — one tool message per call, each referencing its corresponding tool_call_id. This significantly reduces round-trip latency for agents that need to gather several pieces of information at once.