Support for functions

[francisjervis]

Describe the feature or improvement you're requesting

Pythonic support for functions as described in the feature announcement.

Additional context

No response

[Ja-sonYun]

Seems like we can use like this

completion = openai.ChatCompletion.create(
    model="gpt-3.5-turbo-0613",
    messages=[{"role": "user", "content": "What is the weather like in Boston?"}],
    functions=[
        {
            "name": "get_current_weather",
            "description": "Get the current weather in a given location",
            "parameters": {
                "type": "object",
                "properties": {
                    "location": {
                        "type": "string",
                        "description": "The city and state, e.g. San Francisco, CA",
                    },
                    "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]},
                },
                "required": ["location"],
            },
        }
    ],
)

And the response is

{
  ...
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": null,
        "function_call": {
          "name": "get_current_weather",
          "arguments": "{\n  \"location\": \"Boston, MA\"\n}"
        }
      },
      "finish_reason": "function_call"
    }
  ],
  "usage": {
    "prompt_tokens": 82,
    "completion_tokens": 18,
    "total_tokens": 100
  }
}

openai version is 0.27.8

[francisjervis]

Right, saw the docs updated too. Leaving this open for Pythonic support tho haha, ie using either built-in or Pydantic types rather than horrors like "string" ;) Ideally, noting the safety issues, it should be taking an actual function as input, rather than JSON kinda-code.

[Ja-sonYun]

I totally agree with that pythonic support. We want some sort of type safe stuff like TypedDict at least.

[francisjervis]

See how tools are implemented in Langchain, for example. You should just be able to pass a function and have it receive the JSON as kwargs.

[francisjervis]

This is a real issue - even GPT-4 is apparently not yet reliably capable of returning consistent null values. I'm able to get more consistent results with carefully engineered "old fashioned" prompts for structured outputs as it stands anyhow, but my first impression is that this is not a "magic bullet" for getting machine-readable responses back...

[janekb04]

I have made a simple converter that converts a Python function into this JSON: https://github.com/janekb04/py2gpt

[bonelli]

I don't think this works properly with the new "function" feature, because in here it is clear that the "function_call" element of the "messages" map is not transmitted to the API backend:

openai-python/openai/cli.py

Lines 119 to 121 in 041bf5a

	messages = [
	{"role": role, "content": content} for role, content in args.message
	]

[Ja-sonYun]

I don't think this works properly with the new "function" feature, because in here it is clear that the "function_call" element of the "messages" map is not transmitted to the API backend:

openai-python/openai/cli.py

Lines 119 to 121 in 041bf5a

	messages = [
	{"role": role, "content": content} for role, content in args.message
	]

That is for cli, and the ChatCompletion class which we actually use in our code is this class. check out __init__.py, you can see the ChatCompletion is imported from api_resources.

openai-python/openai/__init__.py

Line 19 in 041bf5a

from openai.api_resources import (

[rattrayalex]

Would you want something like runFunctions / runTools from the Node SDK?

[rattrayalex]

We're working on something like this! We're tracking separately, so I'm going to close this issue. Keep an eye on our releases!

[jackmpcollins]

I built https://github.com/jackmpcollins/magentic to solve exactly this issue of translating between python objects and openai LLM queries. @francisjervis if you are still looking for a solution, you might be interested in checking out the docs on how function calling works in magentic: https://magentic.dev/function-calling/