Tool Calling

Definition

The mechanism that allows an LLM to interact with external software. The developer provides a list of available tools, and the LLM outputs a structured command to execute one of those tools when needed.

Explain Like I'm New

Giving the AI an API. You tell the AI: 'You have a tool called getWeather(city)'. The user asks 'Is it raining in London?'. The AI realizes it can't know this from memory, so it pauses and says 'Run getWeather(London)'. The system runs it, gives the data back to the AI, and the AI answers the user.

Real World Example

ChatGPT Plugins. When you ask ChatGPT to do complex math, it uses Tool Calling to write a Python script, send it to a hidden Python interpreter tool, execute the math, and read the result.

Common Use Cases

•Live data access
•Mathematical computation
•Executing actions

Interview Questions

basic

Does the LLM actually run the Tool's code itself?

intermediate

How does the LLM know which tool to use?

Flash Cards

Question

Does LLM run the code?

Click to reveal answer

Answer

No. The LLM simply generates a JSON string (e.g., `{"tool": "calculator", "args": "5+5"}`). Your backend server intercepts this JSON, actually runs the calculator script, and feeds the answer back to the LLM.

Question

How does it know?

Click to reveal answer

Answer

The developer writes a 'Tool Description' in the System Prompt (e.g., 'Use this tool when the user asks about the weather'). The LLM's attention mechanism reads the user's prompt, reads the descriptions, and logically deduces which tool matches the intent.