What is Function Calling?

Function calling is a structured AI model capability that allows large language models to generate well-formed function invocations — complete with the correct function name and parameters — based on natural language input. Rather than producing free-form text that needs to be parsed, function calling models output structured JSON that directly maps to predefined function signatures, making it reliable enough for production applications that need to trigger real actions based on user requests.

The function calling workflow involves several steps. First, the developer provides the model with a list of available functions, each described with a name, description, and parameter schema (typically JSON Schema). When a user makes a request, the model decides whether to call a function and which one, then generates a structured response containing the function name and arguments. The application executes the function, returns the result, and the model generates a natural language response incorporating the function's output. OpenAI, Anthropic, Google, and other providers have implemented function calling in their APIs with slight variations in syntax but similar concepts.

Function calling enables a broad range of practical applications. Structured data extraction converts unstructured text into typed objects — for example, extracting product names, prices, and quantities from an email. Form-filling assistants collect information through conversation and call a submission function with the gathered data. Smart home controllers map voice commands to device control functions. CRM integrations let users query customer data through natural language. Workflow automation tools use function calling to trigger actions in external systems like Slack, Jira, or Salesforce.

Best practices for function calling include writing detailed, unambiguous function descriptions, using enum types to constrain parameter values, providing clear parameter descriptions with examples, handling cases where the model calls the wrong function or provides invalid parameters, and designing function schemas that are as simple as possible while capturing the necessary information. Parallel function calling — where the model generates multiple function calls in a single turn — is supported by some providers and is particularly useful for gathering information from multiple sources simultaneously.