OpenAI introduced Function Calling in the Chat Completions API (gpt-4-0613 and gpt-3.5-turbo-0613), fine-tuning models to detect when an external function is needed and emit a typed JSON argument object matching developer-supplied schemas.
Tool Use & Function Calling
Also known as: Function Calling · Tool Use · Tool Calling · BFCL
Plain English
Teaching an AI model to ask external tools—like databases, web search, code runners, or APIs—for live information or actions instead of guessing everything from memory.
Technical Overview
Function calling pairs model fine-tuning on tool-invocation schemas with constrained or schema-validated JSON argument emission, returning tool results back into the model's context loop. Standardized across model APIs and MCP, reliable multi-turn and parallel tool use underpins modern agent harnesses and coding assistants.
Timeline (3)
2023
- Release
2024
- Release
Fireworks AI released Firefunction-v2 on Llama 3, showing that targeted open-weight Post-Training (SFT, RLHF, DPO & GRPO) could match frontier commercial models on parallel and multi-turn function calling at much lower serving latency.
- Ecosystem
Anthropic open-sourced the Model Context Protocol (MCP), decoupling tool discovery and execution from individual model-vendor schemas so Claude, Cursor, and other agent harnesses can share reusable tool servers.
See also (bkz.) (12)
Open full graphTool Use & Function Calling