What Is Function Calling?

Function calling lets AI models request the execution of external functions, bridging the gap between language understanding and real-world actions.

The Function Calling Loop

1. Define — Describe available functions with JSON Schema
2. Generate — The model decides which function(s) to call and with what arguments
3. Execute — Your code runs the actual function
4. Return — Send the result back to the model
5. Respond — The model incorporates the result into its response

Schema Design

Good schemas are the foundation of reliable function calling:

{
  "name": "search_products",
  "description": "Search the product catalog by name, category, or price range. Returns matching products with details.",
  "parameters": {
    "type": "object",
    "properties": {
      "query": { "type": "string", "description": "Search term or product name" },
      "category": { "type": "string", "enum": ["electronics", "clothing", "books", "home"] },
      "min_price": { "type": "number", "description": "Minimum price in USD" },
      "max_price": { "type": "number", "description": "Maximum price in USD" },
      "in_stock": { "type": "boolean", "default": true }
    },
    "required": ["query"]
  }
}

Description Best Practices

• Be specific about what the function does and returns
• Include example values in parameter descriptions
• Document edge cases and limitations
• Use enums to constrain values wherever possible

Common Pitfall: Writing Vague Function Descriptions

A vague description like "Search function" or "Gets data" leaves the AI guessing when to call a tool. Write descriptions that specify the exact trigger, the expected inputs, and what the function returns:

{
  "name": "search_products",
  "description": "Search the product catalog by keyword, category, or price range. Use this when the user is looking for specific products or browsing a category. Returns up to 20 matching products with name, price, and stock status.",
  ...
}

Best Practice: Test Tool Calling with Adversarial Prompts

Verify your schema handles edge cases by testing with unusual inputs designed to trigger incorrect behavior:

const tests = [
  "Show me everything",              // Should search, not dump all products
  "I want a laptop under $500",      // Should call with max_price: 500
  "What is the cheapest thing?",     // Should search + sort, not guess
];
for (const prompt of tests) {
  const result = await callModelWithTools(prompt, tools);
  console.log(prompt, "→", JSON.stringify(result.tool_calls));
}

Named Product Reference: Anthropic Claude Tool Use

Claude's tool use is less prone to hallucinating tool calls compared to other providers — it prefers asking for clarification over calling a tool with guessed parameters. For high-stakes use cases (payments, bookings, sends), Claude's conservative tool selection behavior reduces error rates.