Tool Calling With Ollama

The same loop, running locally

The langchain course's tool-calling lessons (13-16) teach a loop: the model reads a question, decides it needs a tool, describes which one and with what arguments, your code actually runs it, and the result goes back so the model can write a final answer. Some models on Ollama support the exact same loop, over the exact same OpenAI-compatible tools format, which means code you write against a local model here transfers almost directly to any cloud provider using that convention.

Not every model on Ollama supports tool calling, it depends on how the model was fine-tuned. llama3.2 does; check a model's Capabilities section in ollama show <model> (Lesson 2) if you're unsure about another one.

The code, piece by piece

def get_weather(city: str) -> str:
    fake_weather = {"Paris": "18C, cloudy", "Tokyo": "25C, sunny"}
    return fake_weather.get(city, "unknown city")

A plain Python function. It has no idea it's being offered to an AI, that connection is made entirely by the TOOLS description below.

TOOLS = [
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "Get the current weather for a city",
            "parameters": {"type": "object", "properties": {...}, "required": [...]},
        },
    }
]

A JSON Schema again, same shape as format in Lesson 10, just describing a function's arguments instead of an entire reply.

response = ollama.chat(model="llama3.2", messages=messages, tools=TOOLS)

Passing tools=TOOLS tells the model what's available. The model itself never runs anything, it decides whether a tool is needed and, if so, returns a tool_calls list on response.message describing which function and which arguments, instead of writing a normal reply.

call = response.message.tool_calls[0]
result = get_weather(**call.function.arguments)

Your code is the one that actually calls get_weather, using the arguments the model chose. call.function.arguments is already a plain dict, so **call.function.arguments unpacks it directly into keyword arguments.

messages.append({"role": "tool", "content": result, "tool_name": call.function.name})
final_response = ollama.chat(model="llama3.2", messages=messages, tools=TOOLS)

The function's return value goes back into the conversation as a "tool" role message. A second call to ollama.chat() lets the model read that result and write a normal, natural-language final answer.