Multiple Tools

Where we left off

Lesson 14 bound exactly one tool, the calculator, so there was never any real choice involved, if the model wanted a tool at all, it could only mean the calculator. This lesson binds two tools with clearly different purposes, and asks three different kinds of questions to see the model actually choose between them.

A second, unrelated tool

@tool
def word_counter(text: str) -> str:
    """Count how many words are in a piece of text."""
    return str(len(text.split()))

Same @tool pattern from Lesson 13, a completely different job: counting words instead of doing arithmetic. The two tools don't overlap at all, which matters for this lesson, we want it to be obvious which one (if either) a given question should trigger.

Binding both at once

model_with_tools = model.bind_tools([calculator, word_counter])

Same bind_tools from Lesson 14, just handed a list of two tools instead of one. The model now sees descriptions of both, and has to decide, per question, which one (if any) actually applies.

Looking up which tool was actually requested

tools_by_name = {"calculator": calculator, "word_counter": word_counter}
for call in ai_message.tool_calls:
    chosen_tool = tools_by_name[call["name"]]
    result = chosen_tool.invoke(call["args"])

In Lesson 14, there was only one tool, so we always knew which function to run. Now that there are two, each tool_calls entry includes a "name" field telling us which one the model actually picked, and we look up the matching real function by that name before running it. This tools_by_name dictionary pattern is how you'd generalize this to any number of tools, not just two.

Three questions, three outcomes

ask("What is 84 times 17?")
ask("How many words are in the sentence: 'The quick brown fox jumps'?")
ask("What is the capital of France?")

Run the lesson and you'll see three different behaviors:

1. The math question triggers calculator. 2. The word-counting question triggers word_counter. 3. The capital-of-France question triggers neither, the model just answers directly, exactly like every plain question since Lesson 1.

That third case matters as much as the first two. Binding tools to a model doesn't force it to use one, it gives it the option. A well-behaved model only reaches for a tool when the question actually calls for it, answering directly the rest of the time.