LLM Tooling Framework

A simple toy framework to enable tool usage with LLMs (like Ollama's models). It supports automatic discovery and registration of tools defined using decorators.

Features

• Auto-discovery of tool functions (class methods or standalone)
• Decorator-based registration using @tool_method
• Supports structured tool calling with arguments and response parsing
• Easy integration with Ollama's chat interface

Quickstart

1. Define your tools

Create tool functions in the tools/ directory. Use the @tool_method decorator to register them: For example you can define a WeatherTool api like below

from main.registry.tool_registry import ToolRegistry as tools

@tools.tool_method(name="get_weather_for_location", description="Get weather for a given location")
def get_weather_for_location(location: str) -> str:
    return f"The weather in {location} is sunny."

2. Initialize in your main app

Use the LlmTooling interface to simplify initialization and tool usage.

from ollama import chat
from main import LlmTooling

# Initialize or register all tools
LlmTooling.init()

messages = [{"role": "user", "content": "What's the weather in Paris?"}]
response = chat(
    model="llama3.2:3b",
    messages=messages,
    tools=LlmTooling.get_all_tools_as_callable() # Get all available tools
)

followup = LlmTooling.parse_and_call_tool_if_available(response) # Depending on the LLM response, call the requried tool if needed 

if followup:
    # Send followup to LLM for continuation
    final_response = chat(
        model="llama3.2:3b",
        messages=followup,
        tools=LlmTooling.get_all_tools_as_callable()
    )
    print(final_response)

3. Tool Decorator API

@tools.tool_method(name="tool_name", description="What the tool does")
def your_tool_function(arg1: str, arg2: int) -> str:
    return "result"

4. Examples:

• chat_with_tool_calling

Example Tool Calling Flow

• LLM chooses to call a tool (via tool_calls in response)
• parse_and_call_tool_if_available runs the function
• Result is sent back to the LLM using the appropriate message format

Notes

• Only functions marked with @tool_method are registered
• Both class methods and standalone functions are supported

License

MIT