Function Calling (LLM)

Function calling, also called tool use, is a deployment pattern in which a language model is given a registry of named tools described by a JSON schema, and is expected to produce a structured invocation when it decides one is needed. The model itself does not run the function: it returns a payload such as a tool_use content block or a tool_calls field, and the host application parses the call, runs the real code, and feeds the result back as a follow-up message. Commercial providers expose this as a first-class API feature. OpenAI's interface defines a parameters schema per tool and returns tool_calls entries; Anthropic's API uses an input_schema field and returns tool_use blocks. Both support a strict mode that constrains generations to validate against the supplied schema, which makes malformed or hallucinated arguments much rarer in practice. The Model Context Protocol, introduced by Anthropic in late 2024, generalises the pattern by letting external servers advertise tools to any compatible client. Function calling underpins most modern agent frameworks and is a core mitigation for capability hallucination: because only registered tools can be invoked and arguments are schema-checked, a fabricated call to a nonexistent send_email tool is rejected at the runtime rather than silently succeeding in prose. It composes naturally with ReAct Prompting, whose Action steps map directly onto tool calls.