Function calling
When the LLM picks a tool from a list and emits the arguments to call it — text in, structured action out. This allows Roost to actually perform tasks like adding a meeting to your calendar.
Function calling allows an LLM to bridge the gap between text generation and software execution. By outputting structured data that matches a specific API schema, the model can trigger external tools to perform actions like scheduling meetings or querying databases.
Function calling is the capability of an LLM to map natural language input to a pre-defined function signature, resulting in the emission of structured data (typically JSON) that conforms to a specific schema. This mechanism enables the model to act as an orchestrator, invoking external software interfaces to execute deterministic tasks based on the generated arguments.
evolution
- 2020-06 · historyGPT-3 API Launch
OpenAI introduced the ability to prompt models for structured outputs, laying the groundwork for tool-use patterns.
- 2022-11 · historyLangChain Framework
The release of LangChain popularized 'Agents' that could chain LLM reasoning with external API execution.
- 2023-06 · historyOpenAI Function Calling API
OpenAI formally introduced a dedicated API endpoint that fine-tuned models to reliably output JSON arguments for specific functions.
- 2023-11 · historyGPT-4 Turbo Tool Use
The introduction of 'Parallel Function Calling' allowed models to trigger multiple tools in a single turn, significantly increasing agentic efficiency.