← Retour aux guides

Tool Use and Function Calling: A Beginner's Guide

Modern AI applications do more than generate text. They can look up an order, query a database, calculate a price, create a calendar event, or call an external API. This capability is usually called tool use or function calling. Understanding the difference and the basic execution flow makes it much easier to build reliable AI features.

What Is Tool Use?

Tool use means allowing a language model to request an action from your application. The model does not directly access your database or execute arbitrary code. Instead, you describe available tools to the model, and it decides whether one is useful for the user request.

For example, a weather assistant might have a tool called get_weather. Its definition could require a city name and a temperature unit. When a user asks for the weather in Paris, the model returns a structured request such as the tool name and arguments. Your server then calls the weather service and sends the result back to the model, which writes a normal response.

What Is Function Calling?

Function calling is the structured form of tool use popularized by OpenAI-compatible APIs. A function is described with a name, purpose, and JSON Schema for its arguments. The model can then produce a function call instead of guessing an answer or writing unstructured instructions.

The terms are often used interchangeably, but tool use is the broader concept. A tool might be a function, a code interpreter, a search capability, or another structured operation. Claude APIs commonly refer to tools, while GPT APIs often use function calling or tools. The underlying pattern is similar.

The Tool-Calling Loop

A dependable integration normally follows these steps:

• Define each tool with a clear name, description, required fields, types, and validation rules.
• Send the user message and tool definitions to the model through the API.
• Inspect the response. If it contains a tool call, do not display it as the final answer yet.
• Validate the arguments, check permissions, and execute the matching function on your server.
• Return the tool result to the model using the provider's expected tool-result message format.
• Ask the model to continue. It can then call another tool or produce the final response.

This loop may run more than once. For example, an assistant could retrieve a customer record, check inventory, and then calculate shipping before answering.

Why JSON Schema Matters

A tool description is an interface contract. If a tool expects a date, define it as a string with a documented format such as YYYY-MM-DD. If a field accepts only a small set of values, use an enum. Mark essential fields as required and explain ambiguous terms in the description.

Good schemas reduce invalid calls, but they do not replace server-side validation. Always verify types, access rights, account ownership, and business rules before executing a requested action. Treat model-generated arguments as untrusted input, just as you would treat form data from a browser.

Important Safety Practices

• Use allowlists so the model can call only explicitly registered tools.
• Require confirmation before irreversible actions such as refunds, deletions, purchases, or email delivery.
• Set timeouts and rate limits for external services.
• Log tool names, sanitized arguments, results, errors, and execution time for debugging.
• Return concise, useful errors to the model instead of exposing secrets or internal stack traces.

Choosing an API for Experiments

Tool calling can become expensive while you iterate because every loop adds input and output tokens. 59API is a practical low-cost relay for developers who want pay-as-you-go access to native Claude models, including Opus, Sonnet, Haiku, and Fable, as well as GPT models. It uses the API base URL https://api.59api.com and is compatible with Claude Code, Codex, and OpenAI SDK-based applications.

Because the relay provides official-quality models without a downgrade and is among the cheapest relay options, it can be useful for testing multi-step agents without committing to a large fixed plan. Developers can also receive a referral rebate. If you are ready to test a tool-calling workflow, sign up for 59API and start with a small budget, then monitor latency, token usage, and tool error rates before expanding.

Start With One Safe Tool

The easiest first project is a read-only function such as looking up a product, checking weather, or searching internal documentation. Define a narrow schema, implement strict validation, test malformed arguments, and verify that the model returns a helpful answer when the tool fails. Once that path is reliable, add more tools one at a time. This approach keeps the execution loop understandable and gives you clear evidence about accuracy, cost, and safety.

Tarifs · Docs