tool use / function use

Tool use enables LLMs to invoke external APIs, functions, and systems to extend their capabilities. **Capabilities extended**: Real-time information (web search, APIs), computation (calculators, code execution), actions (send emails, database operations), specialized tools (image generation, retrieval). **Implementation patterns**: Function calling APIs (structured JSON output), ReAct (reasoning + action in text), tool tokens (special vocabulary for tool invocation). **Tool definition**: Name, description, parameters with types, return format - clear descriptions improve selection accuracy. **Execution loop**: User query → model reasoning → tool selection → argument generation → execution → result injection → continued generation. **Popular frameworks**: LangChain, LlamaIndex, Semantic Kernel, Haystack. **Multi-tool scenarios**: Model chains multiple tools, routes between options, handles failures. **Security**: Sandboxed execution, argument validation, permission controls, audit logging. **Best practices**: Minimal tool set (reduce confusion), clear descriptions, error handling, rate limiting. Tool use transforms LLMs from knowledge sources into capable agents.