Does the model actually run my function?

No. The model only decides which function to call and produces the arguments as JSON. Your code parses those arguments, runs the real function, and sends the result back. The model never executes code itself in the Chat Completions API — you stay in control of every action.

How do I make sure the arguments are valid?

Define a strict JSON schema with types, required fields, and enums where possible, and enable strict mode so the model is constrained to it. Still validate the parsed arguments in your code before acting, because a clear schema reduces but does not eliminate the need for guardrails.

What is the tool loop?

It is the cycle of call the model, the model requests a tool, you execute it and append the result, then call the model again. The loop continues until the model returns a normal message instead of a tool call. Cap the number of iterations to avoid runaway loops.

Can the model call several functions at once?

Yes. A single response can contain multiple tool_calls, and you should execute each one and append a separate tool message keyed by its tool_call_id. The model then sees all results together and produces a combined answer.

When should I use function calling versus the Assistants API?

Use function calling with Chat Completions when you want full control over execution, context, and the loop. Use the Assistants API when you prefer managed threads and hosted tools. Function calling is the lower-level, more portable building block underneath most agent frameworks.

How to Use Function Calling with OpenAI

What function calling is

Function calling (also called tool use) lets a language model take actions in the real world without ever running code itself. You describe the functions your application can perform as JSON schemas; the model, given a user request, decides which function to call and produces the arguments as structured JSON. Your code then executes the real function and feeds the result back. This is how an LLM goes from a chatbot to an assistant that can check the weather, query a database, book a meeting, or hit any API you expose.

How the loop works

You send the conversation plus a tools array describing your functions. The model responds in one of two ways: a normal message (it is done), or a tool_calls entry containing a function name and JSON arguments (it wants you to act). When you get a tool call, you parse the arguments, run your real function, and append a tool message containing the result and the matching tool_call_id. Then you call the model again with the updated conversation. It now sees the result and either calls another tool or produces the final answer. That cycle — call, execute, return, call again — is the tool loop, and you cap its iterations so it can never run forever.

The make-or-break detail is the schema. A clear function name, a precise description of when to use it, and typed parameters (with enums and required fields) dramatically improve how reliably the model picks the right tool and fills in valid arguments. The builder below lets you define a function and generates both the tool schema and the Python loop that executes it.

Tips and guardrails

Write descriptions for the model, not for humans — state exactly when the function should and should not be used. Use enums and required fields to constrain arguments, and enable strict mode for schema-guaranteed JSON. Always validate parsed arguments in your own code before acting; the schema guides the model but your function is the last line of defence. Handle multiple parallel tool calls in one response. And bound the loop with a maximum iteration count so a confused model cannot spin forever and burn tokens.