Tool Use / Function Calling

Pain Signals — You Need This Pattern When:

• The LLM needs live data it was not trained on (current prices, user account status, real-time metrics)
• The LLM needs to perform calculations it is unreliable at (maths, date arithmetic, aggregations)
• The task requires side effects — creating records, sending messages, modifying state
• The LLM needs to interact with external APIs or databases to answer questions
• Hard-coded logic would be faster but the tool selection itself requires natural language understanding

Core Principle

Tool Use turns the LLM into a controller that decides which function to call and with what arguments, while the actual execution happens in deterministic, testable code. The LLM is the Strategy selector; the tools are the strategies.

User query → LLM decides which tool(s) → Tool executes → Result returned to LLM → LLM synthesises answer