03 · Tool / Function Calling

What you’ll learn

A model alone can only produce text. An agent wraps the model in a loop that can act: the model decides whether a tool is needed, the tool runs, and the result is fed back so the model can finish the answer.

In the demo, try What is 23 * 19? (the model picks the calculator) versus Look up vector database (it picks dataset search) versus a plain question (no tool needed).

The tool-calling loop

1. Decide — inspect the prompt for intent (here, simple keyword/pattern matching stands in for the model’s function-calling decision).
2. Execute — run the chosen tool with structured arguments.
3. Incorporate — pass the tool’s output back as context for the final answer.

import { decideToolCall, runTool, mockLLM } from '@lib/js';

const prompt = 'What is 23 * 19?';
const call = decideToolCall(prompt);
// → { tool: 'calculator', args: { expression: '23 * 19' } }

const result = runTool(call.tool, call.args, corpus);
// → { tool: 'calculator', output: '437' }

const answer = mockLLM(prompt, {
  context: [`Tool ${result.tool} returned: ${result.output}`],
});

from _shared.tools import decide_tool_call, run_tool
from _shared.mock_llm import mock_llm

prompt = "What is 23 * 19?"
call = decide_tool_call(prompt)
# → {"tool": "calculator", "args": {"expression": "23 * 19"}}

result = run_tool(call["tool"], call["args"], corpus)
# → {"tool": "calculator", "output": "437"}

answer = mock_llm(prompt, context=[f"Tool {result['tool']} returned: {result['output']}"])

Why tools matter

• Fresh & exact. Tools do math, hit APIs, or query data the model can’t recall.
• Verifiable. A calculator’s 437 is checkable; a guessed number is not.
• Composable. Stack several tools and a loop, and you have a capable agent.

Every project from here on is built on this loop. Master it and the rest is orchestration.