LLM FOO: Tool annotation to help using tools in python

Community
, ,
1

Facing complexities with OpenAI’s APIs? I made a small project called LLM FOO to help with tool / function calling. The project aims to reduce verbosity and ease documentation efforts with the @tool decorator. It’s a work in progress, and I’m eager for your insights and contributions. I would love to get some feedback to make it really useful for everybody.

Basic idea is to automatically generate Function json schema with GPT-4-turbo and have convenience methods for function calling.

Seems I can’t make links but pip install llmfoo and you can find the project from github robocorp llmfoo.

  • have a nice day!
1 Like
2

That sounds really interesting, but you kind of lost me with the description. Why not just describe functions in structured prompts and let ChatGPT (likely using gpt-4-turbo or similar itself) crank out the functions?

1 Like
3

Thanks for the interest!

The tool does the following (assume you want to expose adder as a tool / function calling):

from llmfoo.functions import tool


@tool
def adder(x: int, y: int) -> int:
    return x + y

This will call on the first time gpt-4-turbo to crank out the function definition JSON. This is stored to a file in the same directory as the original function is ($FILE.tool.json).
Notice that you do not have to do any additional documentation if you do not want to.

After first run the annotation will use the definition generated into $FILE.tool.json and not call gpt-4-turbo. You may also modify the definition there.

It looks something like this for the adder in $FILE.tool.json

"adder": {
        "type": "function",
        "function": {
            "name": "adder",
            "description": "Adds two integer numbers together and returns the sum.",
            "parameters": {
                "type": "object",
                "properties": {
                    "x": {
                        "type": "integer",
                        "description": "The first integer to be added."
                    },
                    "y": {
                        "type": "integer",
                        "description": "The second integer to be added."
                    }
                },
                "required": [
                    "x",
                    "y"
                ]
            }
        }
    },

The annotation adds few convenience fields to the function being annotated:

adder.openai_schema  # The JSON for tool / function calling definition
...
messages.append(adder.openai_tool_call(response.choices[0].message.tool_calls[0]))  # Message handler that returns directly response as valid a function call response message.
...
adder.openai_tool_output(tool_call)  # Handler that returns directly response in the Assistant API tool output.

Hope this helps define what the tool does. I think there is still a lot room to improve but the basic idea is there:
Let the ChatGPT define the docs for ChatGPT and re-use the definition. Allow humans to modify the definitions.