Gemini function calling connects models to external tools

Function calling on Google Cloud’s Gemini Enterprise Agent Platform lets AI models request external tools to retrieve current information or perform actions. Developers supply tool definitions, and the model returns structured function names and parameters when needed. The application executes the requested operation and sends the result back, allowing the model to answer with external data or request further operations.

Functions use an OpenAPI-compatible schema. Available controls let developers allow automatic tool selection, enforce schema adherence, require function calls, or disable them without removing tool definitions. Parallel calls require all corresponding results to be returned together, with responses correctly matched to requests. When thinking is enabled, applications must preserve and return thought signatures with the conversation.

For Gemini 3 and later models, preview capabilities include multimodal tool responses containing images or documents and streaming function arguments as they are generated. Argument streaming is not supported in Gemini 3.1 Flash-Lite.

Recommended practices include precise function descriptions, strongly typed parameters, relevant tool selection, and clarifying questions when requests lack necessary information. Developers should validate consequential actions, such as orders or database updates, with users before execution. Potential applications include customer support, database queries, workflow automation, and knowledge-base retrieval.