OpenAI Functions

pip install funcmodels

from funcmodels import openai_function

@openai_function

The most intuitive, robust and "pure" way to implement functions for OpenAI function calling.

Designed as a more straightforward and ergonomic alternative to jxnl/instructor. Rather than defining your function as a BaseModel, you define your function as a function, and the BaseModel is created for you.

With @openai_function, the ...

• Pydantic model for data validation,
• definition of your OpenAI Function JSON schema,
• parsing and validation of raw json string arguments,
• logic for handling the arguments and producing a result

... are all explicitly defined and encapsulated in one place: your decorated function.

@openai_function will turn your function into a Pydantic BaseModel class, using your function parameters as attributes. It parses your docstring for a description and parameter descriptions, and combines those with the BaseModel's JSON schema to produce a complete OpenAI Function definition. You'll receive back an extended version of this BaseModel class, equipped with a class attribute, .schema with the function definition, a .from_json() class method for creation and validation directly from raw function call arguments from OpenAI, and an execute() method that passes the instance's validated attributes to the handler defined by your decorated function.

from typing import Literal

@openai_function
def get_stock_price(ticker: str, currency: Literal["USD", "EUR"] = "USD"):
    """
    Get the stock price of a company, by ticker symbol

    Parameters
    ----------
    ticker
        The ticker symbol of the company
    currency
        The currency to use
    """
    return f"182.41 {currency}, -0.48 (0.26%) today"


get_stock_price

OpenaiFunction({
    "name": "get_stock_price",
    "description": "Get the stock price of a company, by ticker symbol",
    "parameters": {
        "properties": {
            "ticker": {
                "type": "string",
                "description": "The ticker symbol of the company"
            },
            "currency": {
                "default": "USD",
                "enum": [
                    "USD",
                    "EUR"
                ],
                "type": "string",
                "description": "The currency to use"
            }
        },
        "required": [
            "ticker"
        ],
        "type": "object"
    }
})

Get our OpenAI function definition dictionary

get_stock_price.schema

{'name': 'get_stock_price', 'description': 'Get the stock price of a company, by ticker symbol', 'parameters': {'properties': {'ticker': {'type': 'string', 'description': 'The ticker symbol of the company'}, 'currency': {'default': 'USD', 'enum': ['USD', 'EUR'], 'type': 'string', 'description': 'The currency to use'}}, 'required': ['ticker'], 'type': 'object'}}

Instantiate our pydantic model, validating arguments

validated_function_call = get_stock_price(ticker="AAPL")

Or, go directly from raw json arguments from OpenAI

raw_arguments_from_openai = '{"ticker": "AAPL"}'
validated_function_call = get_stock_price.from_json(raw_arguments_from_openai)
validated_function_call.currency

'USD'

Call our function, with already-validated arguments

validated_function_call.execute()

'182.41 USD, -0.48 (0.26%) today'

If you prefer Pydantic syntax, we can achieve the same thing using Fields

from pydantic import Field

@openai_function
def get_stock_price(
    ticker: str = Field(description="The ticker symbol of the company"),
    currency: Literal["USD", "EUR"] = Field("USD", description="The currency to use."),
):
    "Get the stock price of a company, by ticker symbol"
    return f"182.41 {currency}, -0.48 (0.26%) today"

Here, the field descriptions are defined in the parameters themselves, rather than the docstring.

The result is the exact same function definition as before:

get_stock_price

OpenaiFunction({
    "name": "get_stock_price",
    "description": "Get the stock price of a company, by ticker symbol",
    "parameters": {
        "properties": {
            "ticker": {
                "type": "string",
                "description": "The ticker symbol of the company"
            },
            "currency": {
                "default": "USD",
                "enum": [
                    "USD",
                    "EUR"
                ],
                "type": "string",
                "description": "The currency to use"
            }
        },
        "required": [
            "ticker"
        ],
        "type": "object"
    }
})

Function Groups

def get_stock_price(ticker: str):
    return '182.41 USD, -0.48 (0.26%) today'

def get_weather(city: str):
    return "Sunny, 72 degrees, 0% chance of rain"

group = openai_function_group(functions=[get_stock_price, get_weather])
group

OpenaiFunctionGroup([
    {
        "name": "get_stock_price",
        "parameters": {
            "properties": {
                "ticker": {
                    "type": "string"
                }
            },
            "required": [
                "ticker"
            ],
            "type": "object"
        }
    },
    {
        "name": "get_weather",
        "parameters": {
            "properties": {
                "city": {
                    "type": "string"
                }
            },
            "required": [
                "city"
            ],
            "type": "object"
        }
    }
])

Note: Use group.function_definitions to get a list of the raw function schemas to use in the functions argument to OpenAI

Now, when we get a function call from OpenAI, we can let the group handle it.

function_call_from_openai = {
    "name": "get_weather",
    "arguments": '{"city": "Denver"}',
}

validated_function_call = group.evaluate_function_call(function_call_from_openai)
validated_function_call

get_weather(city='Denver')

result = validated_function_call.execute()
result

'Sunny, 72 degrees, 0% chance of rain'

Create your own ChatGPT

Because each @openai_function encapsulates all the information needed to facilitate OpenAI function calling, it's very easy for us to build our own ChatGPT, with automated function call handling.

Step 1. Define conversation handler

import json
from dataclasses import dataclass
from openai import OpenAI
from openai.types.chat.chat_completion import Choice
from funcmodels import OpenaiFunctionGroup, openai_function_group


@dataclass
class ChatGPTConversation:
    model: str
    client: OpenAI
    functions: OpenaiFunctionGroup
    messages: list[dict]

    def get_openai_response(self) -> Choice:
        response = self.client.chat.completions.create(
            messages=self.messages,
            model=self.model,
            functions=self.functions.function_definitions,
            function_call="auto",
        )
        return response.choices[0]

    def handle_function_call(function_call) -> dict:
        # match the function call to the right function, and validate arguments
        validated_call = self.functions.evaluate_function_call(function_call)
        # Call the function with the validated arguments
        result = validated_call.execute()
        return dict(role="function", name=function_call.name, content=str(result))

    def chat(self):
        result = self.get_openai_response()
        self.add_message(result.message.model_dump(exclude_unset=True))

        if (function_call := result.message.function_call) is not None:
            self.add_message(self.handle_function_call(function_call))
            if result.finish_reason == 'function_call':
                self.chat()

    def add_message(self, message: dict):
        print(json.dumps(message, indent=4))
        self.messages.append(message)

    def send_message(self, prompt: str):
        self.add_message({"role": "user", "content": prompt})
        self.chat()

Here, chat() is a recursive function that continues sending API requests for as long as the response's finish_reason='function_call'.

Step 2. Define openai functions

def get_stock_price(ticker: str):
    "Get the stock price of a company, by ticker symbol."
    return "182.41 USD, −0.48 (0.26%) today"

def get_weather(city: str):
    "Get the current weather in a city."
    return "Sunny and 75 degrees"

def get_current_datetime(city: str):
    "Get the current date and time in a city."
    return "Friday, Nov. 10, 2023, 10:00 AM"

group = openai_function_group(functions=[get_stock_price, get_weather, get_current_datetime])
group

OpenaiFunctionGroup([
    {
        "name": "get_stock_price",
        "description": "Get the stock price of a company, by ticker symbol.",
        "parameters": {
            "properties": {
                "ticker": {
                    "type": "string"
                }
            },
            "required": [
                "ticker"
            ],
            "type": "object"
        }
    },
    {
        "name": "get_weather",
        "description": "Get the current weather in a city.",
        "parameters": {
            "properties": {
                "city": {
                    "type": "string"
                }
            },
            "required": [
                "city"
            ],
            "type": "object"
        }
    },
    {
        "name": "get_current_datetime",
        "description": "Get the current date and time in a city.",
        "parameters": {
            "properties": {
                "city": {
                    "type": "string"
                }
            },
            "required": [
                "city"
            ],
            "type": "object"
        }
    }
])

Step 3. Create a new conversation

chatgpt = ChatGPTConversation(
    model="gpt-4-1106-preview",
    client=OpenAI(api_key=os.environ["OPENAI_API_KEY"]),
    functions=group,
    messages=[
        dict(role="system", content="You are a helpful AI assistant."),
    ]
)

Step 4. Exchange messages

You'll need to use Jupyter notebooks (or interactive terminal) for this

chatgpt.send_message("Hello, how are you?")

{
    "role": "user",
    "content": "Hello, how are you?"
}
{
    "content": "Hello! I'm just a machine, so I don't have feelings, but I'm functioning optimally. How can I assist you today?",
    "role": "assistant"
}

chatgpt.send_message(
    "I'm enjoying my breakfast here in Denver. Can you list 3 fun things to do here, "
    "then give me a quick morning update with the information available?"
)

{
    "role": "user",
    "content": "I'm enjoying my breakfast here in Denver. Can you list 3 fun things to do here, then give me a quick morning update with the information available?"
}
{
    "content": "Of course! Here are three fun activities to do in Denver:\n\n1. **Visit the Denver Botanic Gardens**: Enjoy a peaceful morning walking through the various plant collections and exhibits. It's a good way to appreciate nature and the diverse plant life.\n\n2. **Explore the Denver Art Museum**: With its wide range of art from across the world, including contemporary, indigenous, and classic pieces, this museum offers a rich cultural experience.\n\n3. **Take a Stroll Around Larimer Square**: This historic district is perfect for a leisurely walk, as you can explore boutique shops, enjoy local eateries, and take in the unique architecture and vibrant atmosphere.\n\nNow, let's get you that morning update for Denver:\n\nI can provide you with the current weather, the current date and time, and check on the stock price of a company if you're following any in particular. Let's start with the weather and the date and time:\n\n- **Weather**: I'll check the current weather conditions for you.\n- **Current Date and Time**: I\u2019ll provide you with the current date and time in Denver.\n\nLet me fetch the latest information for you, just a moment.",
    "role": "assistant",
    "function_call": {
        "arguments": "{\"city\":\"Denver\"}",
        "name": "get_weather"
    }
}
{
    "role": "function",
    "name": "get_weather",
    "content": "Sunny and 75 degrees"
}
{
    "content": null,
    "role": "assistant",
    "function_call": {
        "arguments": "{\"city\":\"Denver\"}",
        "name": "get_current_datetime"
    }
}
{
    "role": "function",
    "name": "get_current_datetime",
    "content": "Friday, Nov. 10, 2023, 10:00 AM"
}
{
    "content": "Here's your morning update for Denver:\n\n- **Weather:** It's currently sunny and 75 degrees in Denver. It seems like a beautiful day to enjoy any of the outdoor activities I mentioned!\n- **Current Date and Time:** It's Friday, November 10, 2023, and the time is 10:00 AM.\n\nIf there's a particular stock you'd like to check on, just let me know the ticker symbol, and I'll get the latest price for you. Have a fantastic day enjoying Denver!",
    "role": "assistant"
}

If we stitch together the content of each assistant message (2nd and last messages), we get a continuous block of response text:

Denver is a vibrant city with plenty of activities to enjoy, ranging from cultural experiences to outdoor adventures. Here are three fun things you might consider doing in Denver:

1. Visit Denver Botanic Gardens: These gardens offer a peaceful and beautiful urban retreat. You can stroll through various plant displays, including a Japanese garden and a conservatory with exotic tropical and subtropical species.

2. Explore the Denver Art Museum: Known for its collection of American Indian Art, the Denver Art Museum also boasts a wide range of other collections, from pre-Columbian artifacts to contemporary art. The museum's architecture is also a work of art in itself.

3. Discover Red Rocks Park and Amphitheatre: Just outside Denver, this world-famous outdoor venue is renowned for its amazing acoustics and stunning red sandstone formations. During the day, you can hike or bike the trails, and in the evening, possibly catch a concert under the stars.

Now, let's get you a morning update on Denver:

• Weather: I'll check the current weather conditions for you.
• Local Time: I'll provide the current date and time.
• Stock Market: If you're interested in a particular stock, I can provide the latest price.

Let me gather the weather and local time information for you. One moment, please. Here's your morning update for Denver:

• Weather: It's a beautiful sunny day with a pleasant temperature of 75 degrees Fahrenheit. A great day to plan an outdoor activity!
• Local Time: The current date and time in Denver is Friday, November 10th, 2023, at 10:00 AM.

If you have any stock in mind or need more information, feel free to let me know! Enjoy your breakfast and have a fantastic day in Denver.

This single response was made up of multiple API calls/responses:

1. Sent:
   • User prompt
2. Received:
   • Content response (PART 1)
   • Function call to: get_weather
3. Sent:
   • Function response from: get_weather
4. Received:
   • Function call to: get_current_datetime
5. Sent:
   • Function response from: get_current_datetime
6. Received:
   • Content response (PART 2)

The response text above combines the content from API responses 2 and 6.