What Is Function Calling in LLM APIs?

Function calling is an LLM API capability that lets developers define functions with typed parameter schemas, have the model decide when to invoke them, and receive the invocation as structured JSON rather than free text. OpenAI introduced it to its API in June 2023, and every major provider — Anthropic (tool use blocks), Google (Gemini function declarations) — now ships an equivalent.

The mechanism in one loop

The developer sends a prompt plus function definitions ("get_weather(city: string)"). The model, instead of answering directly, may return a call: {"name": "get_weather", "arguments": {"city": "Berlin"}}. The developer's code executes the real function and returns the result to the model, which incorporates it into a final answer. The model never runs code itself — it emits intentions; the application supplies the execution. The Model Context Protocol (Anthropic, November 2024) standardized how such tools are exposed across vendors.

Why a marketing glossary includes an API feature

Because AI visibility measurement is built on it. A tracking pipeline runs prompts against engines, then must analyze each answer at scale. Function calling (or its sibling, structured output modes) is what makes that analysis reliable: instead of asking an analysis model "was the brand mentioned?" and parsing prose, the pipeline defines a schema — record_analysis(mentioned: boolean, position: integer, sentiment: enum, competitors: string[]) — and gets machine-validated data back on every call. That structural guarantee is the difference between a dashboard you can trust and one that silently miscounts, and it underpins the data behind visibility scoring.

Reliability notes from production use

Schema compliance is enforced by modern APIs, but semantic accuracy is not: a model can return perfectly valid JSON containing a wrong judgment. Production pipelines therefore pair function calling with few-shot-anchored rubrics and spot-check audits. Costs also compound — an analysis call per tracked answer means the measurement layer can rival the sampling layer in token spend, a real consideration when tracking hundreds of prompts weekly across engines. Adjacent concepts — tool use, LLM-as-judge, eval harness — are defined in the glossary.