#Tool / Function Calling Patterns¶

The production form of ReAct (Module 3): the model requests, your code executes.

The Contract¶

You declare tools with JSON schemas. The model decides whether and how to call them and returns a structured call. You run it, return the result, the model continues.

json

11 lines

1{
2  "name": "get_order_status",
3  "description": "Look up the current status of a customer's order. Use when the user asks where their order is or about delivery.",
4  "parameters": {
5    "type": "object",
6    "properties": {
7      "order_id": { "type": "string", "description": "The order ID, e.g. 'A1234'" }
8    },
9    "required": ["order_id"]
10  }
11}

The Description IS a Prompt¶

The model selects tools and fills arguments using the name + description + parameter descriptions. These are prompt engineering, not afterthoughts:

Weak	Strong
`"name": "func1"`	`"name": "get_order_status"`
`"description": "gets stuff"`	`"description": "Look up order status. Use when the user asks about delivery or where their order is."`
`order_id: string`	`order_id: "The order ID, e.g. 'A1234'"` (with example)

The Full Loop¶

1. Send user message + tool schemas
2. Model returns: tool_call(get_order_status, {order_id:"A1234"})
3. Your code executes the real function
4. Send the tool result back to the model
5. Model produces the final natural-language answer

Patterns & Guardrails¶

Force / restrict tool use when needed (e.g. "must call search before answering factual questions").
Parallel tool calls: modern models can request several at once — execute concurrently, return all results.
Validate every argument before executing. The model can hallucinate IDs or out-of-range values. Never pass model output straight into a DB query or shell.
Keep observations small. Summarise large tool outputs before returning them or you'll exhaust context.
Handle tool errors gracefully: return a structured error the model can reason about ("ORDER_NOT_FOUND") rather than crashing.

Mini Worked Example¶

User: "Where's order A1234?" → model calls get_order_status({order_id:"A1234"}) → your DB returns {status:"shipped", eta:"2026-05-20"} → model: "Your order A1234 has shipped and should arrive around May 20, 2026."

Security note: Tool calling is an injection surface. Treat all model-produced arguments as untrusted input — validate and sanitise (deep dive: Module 5).