Skip to content

ttumiel/chat2func

Repository files navigation

chat2func

CI License PyPI

chat2func automatically generates JSON schemas from Python, allowing ChatGPT to talk to your code.

Installation

pip install chat2func

# Or install with developer dependencies
pip install chat2func[develop]

Quick Start

Annotate your function with @json_schema

from chat2func import json_schema

@json_schema
def my_function(x: float, y: float) -> bool:
    """This is a sample function.

    Args:
        x: The first float.
        y: Another float.
    """
    return x > y

After this, my_function will have an additional .__schema__ attribute containing its JSON schema.

print(my_function.__schema__)

{'description': 'This is a sample function.',
 'name': 'my_function',
 'parameters': {'properties': {'x': {'description': 'The first float.',
                                     'type': 'number'},
                               'y': {'description': 'Another float.',
                                     'type': 'number'}},
                'required': ['x', 'y'],
                'type': 'object'},
 'responses': {'200': {'content': {'application/json': {'schema': {'type': 'boolean'}}},
                       'description': 'OK'}}}

JSON Schema

You can generate a JSON schema from any type of callable: classes, functions, or dataclasses.

Using Custom Classes

json_schema works with classes or dataclasses too. Set descriptions=False to not generate object descriptions from docstrings.

@json_schema(descriptions=False)
@dataclass
class Data:
    a: int = 0

print(Data.__schema__)
{'name': 'Data',
 'parameters': {'type': 'object', 'properties': {'a': {'type': 'integer', "default": 0}}}}

Creating a ChatGPT Plugin

You can easily create and demo a ChatGPT plugin using the included Flask server. First, define the functions you want ChatGPT to be able to use, expose them with the server and connect to them from ChatGPT.

from chat2func.server import FunctionServer

def addition(x: float, y: float) -> float:
    "Add two floats."
    return x + y

server = FunctionServer({"addition": addition})
server.run() # Visit http://localhost:3333/ to see the available functions.

Deploy a Plugin

The PluginServer is a simple Flask app, so you can deploy easily using any method that works for Flask. The server provides an additional export function that generates the necessary files to deploy a plugin using firebase.

# ...
server.export("./my-plugin")

After export, double check that all necessary code and imports are inside main.py. Also add a logo to /public.

cd my-plugin
python3.11 -m venv functions/venv && source functions/venv/bin/activate
pip install -r functions/requirements.txt
firebase emulators:start # test the function locally
firebase deploy # deploy to firebase

Don't forget to update the urls in ai-plugin.json once you've deployed.

Calling Functions with JSON Arguments

function_call provides additional functionality for calling and validating functions with JSON arguments. It also validates the function call, raising FunctionCallError if something is unexpected.

import json
from chat2func import function_call

def addition(x: float, y: float) -> float:
    "Add two floats."
    return x + y

# Arguments are passed as a JSON string.
arguments = json.dumps({"x": 1.0, "y": 2.0})
result = function_call(addition, arguments)
print(result) # 3.0

# We can optionally validate the function arguments too. Defaults to on.
arguments = json.dumps({"x": "a", "y": 2.0})
result = function_call(addition, arguments)
# FunctionCallError: Arguments do not match the schema. 'a' is not of type 'number'

Using OpenAI's Function Calling API

experimental

We can use the function calling API directly too. Here we demonstrate using ChatGPT to generate structured data (in the form of dataclasses) from unstructured knowledge about the book Dune.

from chat2func import function_calls
from chat2func.api import FunctionCallingAPI, Role

@json_schema
@dataclass
class Character:
    name: str
    age: Optional[int] = None
    house: Optional[str] = None

# Setup our function calling API
api = FunctionCallingAPI({"Character": Character})
api.chat.add_message(Role.USER.value, "List the heads of houses in the Dune series. Give best estimates of age's where appropriate.")

# Generate 5 different Characters (remember the samples are stochastic!)
for _ in range(5):
    # Force the API to call a function
    message = api.reply(force_function="Character")
    api.chat.messages.append(message)

    # Call the function (and validate the inputs!)
    fn_name = message.function_call["name"]
    args = message.function_call["arguments"]
    result = function_calls(fn_name, functions, args, return_json=False)

    # Add the result to the chat
    api.chat.add_message(Role.FUNCTION.value, str(result), name="character")
    print(result)

# Character(name='Duke Leto Atreides', age=50, house='Atreides')
# Character(name='Baron Vladimir Harkonnen', age=80, house='Harkonnen')
# Character(name='Emperor Shaddam Corrino IV', age=70, house='Corrino')
# Character(name='Lady Jessica', age=36, house='Atreides')
# Character(name='Paul Atreides', age=15, house='Atreides')

Contributing

I'm happy to accept contributions! Please open an issue or pull request if you have any ideas or find any bugs.