OpenAI function calling, Anthropic tool use, structured output.

Tool Use — Function Calling, Schema Design, Error Handling

Tool use (function calling) allows LLMs to request execution of defined functions with structured JSON arguments. The model outputs a tool_call object instead of text; the application executes the function and returns the result. Tool schema design is critical: clear descriptions, typed parameters, and example values guide the model. Parallel tool calls enable concurrent execution. Strict mode (OpenAI) enforces JSON Schema compliance in tool arguments.

Interactive Pipeline Flow

UnderstandImplementOptimise

📐 Define — tool JSON Schema

Define tool schema: {name, description, parameters:{type:object, properties:{...}, required:[...]}}. Description is read by the LLM — be specific about what the tool does and when to use it.

Architecture Diagrams

LLMs can request multiple tool calls in parallel.

Loading diagram...

Key Design Decisions

📐Tool descriptions are model prompts

The description field in your tool schema is read by the LLM to decide when and how to call the tool. Write it like documentation: "Searches the web for current information. Use when the user asks about recent events."

⚡Execute parallel tool calls concurrently

When the LLM returns multiple tool calls, execute them with Promise.all() or Task.WhenAll(). Sequential execution wastes time — parallel calls can reduce latency by 3-5x.

⚠️Return errors to the LLM, not to the user

Tool errors should flow back as tool result messages so the LLM can self-correct. Only crash or escalate to the user if max retries are exceeded.

Discussion