Folder-based functions for ChatGPT's function calling with Pydantic support 🚀

SageAI is a framework for GPT 3.5/4 function calling for creating folder-based functions that is easy to organize and scale.

With a built-in vector database used to store and retrieve functions, the number of tokens sent to the model is significantly reduced, making it faster and cheaper to call your functions.

Read the blog post for a more in-depth explanation of the motivation behind SageAI here.

Key Features

Function organization through folder-centric functions.
Strong typing for functions using Pydantic.
Built-in Qdrant vector database with in-memory support for function storage and retrieval, with the option to integrate your own.
Easily test each function with an associated test.json file, supporting both unit and integration tests.
Built with CI/CD in mind, ensuring synchronicity between your vector db and the functions directory across all environments using the index method.
Lightweight implementation with only three dependencies:
- openai
- pydantic
- qdrant-client

Requirements

python >=3.9, <3.12
pydantic >=1.6, <=1.10.12
openai >=0.27.0
qdrant-client >=1.4.0

Installation

# pip
$ pip install sageai

# poetry
$ poetry add sageai

Design

SageAI is built around the concept of a functions directory, which contains all of your functions. Each function is defined in a Python file function.py, and is associated with an optional test.json file for testing.

The format of the function.py file must contain two things in order for SageAI to work:

The function itself
The Function object

Input and output types may be defined using Pydantic models, and are automatically validated by SageAI. They can also be defined outside the function.py file, and imported into the file.

Here is a simplified example of how SageAI might handle a function that fetches the current weather for a given location.

# functions/get_current_weather/function.py
from enum import Enum
from typing import Optional

from pydantic import BaseModel, Field

from sageai.types.function import Function


class UnitTypes(str, Enum):
    CELSIUS = "Celsius"
    FAHRENHEIT = "Fahrenheit"


class FunctionInput(BaseModel):
    location: str = Field(
        ..., description="The city, e.g. San Francisco"
    )
    unit: Optional[UnitTypes] = Field(
        UnitTypes.CELSIUS, description="The unit of temperature."
    )


class FunctionOutput(BaseModel):
    weather: str

    def __eq__(self, other):
        if not isinstance(other, FunctionOutput):
            return False
        return self.weather == other.weather


def get_current_weather(params: FunctionInput) -> FunctionOutput:
    weather = (
        f"The weather in {params.location} is currently 22 degrees {params.unit.value}."
    )
    return FunctionOutput(weather=weather)


function = Function(
    function=get_current_weather,
    description="Get the current weather in a given location.",
)

We'll break down the above example into its components below.

Setup

Create a functions directory in the root directory, and add your functions as described in Design.

Then initialize SageAI.

from sageai import SageAI

sage = SageAI(openai_key="")

Then index the vector database.

sage.index()

That's it! You're now set up and ready to interact with SageAI through natural language queries. 🚀

message = "What's the weather like in Toronto right now?"
response = sage.chat(
    messages=[dict(role="user", content=message)],
    model="gpt-3.5-turbo-0613",
    top_n=5,
)
# response:
# {
#   'name': 'get_current_weather',
#   'args': {'location': 'Toronto'},
#   'result': {'weather': 'The weather in Toronto is currently 22 degrees Celsius.'}
# }

Convention

SageAI follows a convention over configuration approach to make it easy to define functions.

Ensure that your function.py file contains the following:

A function object that is an instance of Function.
A function that is the actual function that will be called by ChatGPT.
The function must have typed input and output Pydantic models.
Each field in the input model must have a description.

Minimal example:

def my_function(params: PydanticInput) -> PydanticOutput:
    return PydanticOutput(...)


function = Function(
    function=my_function,
    description="My function description.",
)

API

SageAI Initialize

The SageAI constructor accepts the following parameters:

Parameter	Description	Defaults
openai_key	The API key for OpenAI.	Required
functions_directory	Directory containing functions.	`/functions`
vectordb	An implementation of the `AbstractVectorDB` for vector database operations.	`DefaultVectorDBService`
log_level	Desired log level for the operations.	`ERROR`

SageAI Methods

1. `chat`

Initiate a chat using OpenAI's API and the provided parameters.

Parameters:

Parameter	Description	Defaults
-	Accepts the same parameters as OpenAI's chat endpoint	-
top_n	The number of top functions to consider from the vector database.	Required

Returns:

dict(
    name="function_name",
    args={"arg1": "value1", "arg2": "value2"},
    result={"out1": "value1", "out2": "value2"},  # Optional
    error="",  # Optional
)

Either result or error will be present in the response, but not both.

2. `get_top_n_functions`

Get the top n functions from the vector database based on a query.

Parameters:

Parameter	Description	Defaults
query	The query to search against.	Required
top_n	The number of functions to return.	Required

Returns:

A dict of function names to Function definitions.

3. `run_function`

Execute a function based on its name and provided arguments.

Parameters:

Parameter	Description	Defaults
name	Name of the function.	Required
args	Arguments to pass to the function.	Required

Returns:

The function result as a dict.

4. `call_openai`

Calls the OpenAI API with the provided parameters.

Parameters:

Parameter	Description	Defaults
openai_args	Accepts the same parameters as OpenAI's chat endpoint	Required
top_functions	List of dicts that is a representation of your functions.	Required

Returns:

A tuple of the function name and the function args.

5. `index`

Index the vector database based on the functions directory. This method is useful to update the vectordb when new functions are added or existing ones are updated.

Want more control?

The chat function uses get_top_n_functions, run_function, and call_openai internally. However, we also expose these methods incase you wish to use them directly to implement your own chat logic.

Vector DB

SageAI comes with a built-in in-memory vector database, Qdrant, which is used to store and retrieve functions.

If you wish to use your own vector database, you can implement the AbstractVectorDB class and pass it into the SageAI constructor.

See the advanced example for an example of how to integrate your own vector database.

Testing

As for the optional test.json file in each function, follow this structure:

[
  {
    "message": "What's the weather like in Toronto right now?",
    "input": {
      "location": "Toronto",
      "unit": "Celsius"
    },
    "output": {
      "weather": "The weather in Toronto is currently 22 degrees Celsius."
    }
  }
]

Each object in the array represents a test case.
The message field is the natural language message that will be sent to ChatGPT, and the input field is the expected input that will be passed to the function.
The output field is the expected output of the function.

SageAI offers unit and integration tests.

Unit Tests

Unit tests do not call the vector database nor ChatGPT, and will not cost you OpenAI credits.

Unit tests are used to ensure your functions directory is valid, and it tests the function in isolation.
It tests whether:
- the functions directory exists.
- each function has a function.py file.
- each function.py file has a Function object.
- and more!
It also tests whether the input and output types are valid, and whether the function returns the expected output based on the input alone by calling func(test_case["input"]) == test_case["output"].

Integration Tests

Integration tests will call the vector database and ChatGPT, and will cost you OpenAI credits.

Integration tests are used to test the function by calling ChatGPT and the vector database.
They test whether the vector database is able to retrieve the function, and whether ChatGPT can call the function with the given input and return the expected output.

Because ChatGPT's responses can vary, integration tests may return different results each time. It's important to use integration tests as a tool to ensure ChatGPT is able to call the right function with the right input, and not as a definitive test to measure the test rate of your functions.

Output Equality

You can customize how to determine equality between the expected and actual output by overriding the __eq__ method in the output model.

class FunctionOutput(BaseModel):
    weather: str
    temperature: int

    def __eq__(self, other):
        if not isinstance(other, FunctionOutput):
            return False
        return self.weather == other.weather

In the case above, we only care about the weather field, and not the temperature field. Therefore, we only compare the weather field in the __eq__ method.

This is especially useful when you are returning an object from a database, for example, and you only care to test against a subset of the fields (for example, the id field).

CLI

# To run unit and integration tests for all functions:
poetry run sageai-tests --apikey=openapikey --directory=path/to/functions

# To run unit tests only for all functions:
poetry run sageai-tests --apikey=openapikey --directory=path/to/functions --unit

# To run integration tests only for all functions:
poetry run sageai-tests --apikey=openapikey --directory=path/to/functions --integration

# To run unit and integration tests for a specific function:
poetry run sageai-tests --apikey=openapikey --directory=path/to/functions/get_current_weather

Parameter	Description	Defaults
--apikey	OpenAI API key.	Required
--directory	Directory of the functions or of the specific function to run	/functions
--unit	Only run unit tests	false
--integration	Only run integration tests	false

Examples

Basic - Get started with a simple SageAI function.
Advanced - Dive deeper with more intricate functionalities and use-cases.

Roadmap

Add tests and code coverage
Support multiple function calls
Support streaming
Support asyncio
Support Pydantic V2
Write Chainlit example
Write fullstack example

Contributing

Interested in contributing to SageAI? Please see our CONTRIBUTING.md for guidelines, coding standards, and other details.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

README.md

README.md

Table of Contents

Key Features

Requirements

Installation

Design

Setup

Convention

API

SageAI Initialize

SageAI Methods

1. `chat`

2. `get_top_n_functions`

3. `run_function`

4. `call_openai`

5. `index`

Vector DB

Testing

Unit Tests

Integration Tests

Output Equality

CLI

Examples

Roadmap

Contributing

Files

README.md

Latest commit

History

README.md

File metadata and controls

Table of Contents

Key Features

Requirements

Installation

Design

Setup

Convention

API

SageAI Initialize

SageAI Methods

1. chat

2. get_top_n_functions

3. run_function

4. call_openai

5. index

Vector DB

Testing

Unit Tests

Integration Tests

Output Equality

CLI

Examples

Roadmap

Contributing

1. `chat`

2. `get_top_n_functions`

3. `run_function`

4. `call_openai`

5. `index`