EasyOpenAiTools.Library 0.0.3

EasyOpenAiTools

<h4 align="center">Easily build chatbots with custom functionality on the .NET platform</h4>

<p align="center"> <a href="#about">About</a> � <a href="#goals">Goals</a> � <a href="#how-does-it-work">How does it work</a> � <a href="#future-developement">Future developement</a> </p>

About

EasyOpenAiTools is a lightweight wrapper allowing you to easily build chatbots based on OpenAi. It is built around the new openai/openai-dotnet library.

Goals

Integrating OpenAi LLMs with custom tools and function calling was my main goal and the reason why I created this library.

Registering a tool and running the OpenAi library in a way that makes sure that tools are called correctly takes some boilerplate that would have to be repeated in every project. I created this library to allow me to easily integrate and reuse the code in different projects in the future. And maybe it will help you too 😃

How does it work

Make sure to check the EasyOpenAiTools.Samples project for a running example.

Creating a Model

To create a model you need to supply an OpenAiModelSettings Object containing the following information:

• Your Api Key
• The Type of Model you want to use (currently Gpt3.5 or Gpt4)
• The initial prompt telling the model how to behave check here for inspiration

var openAiApiKey = Environment.GetEnvironmentVariable("OPENAI_API_KEY");
var openAiModelType = OpenAiModelType.Gpt4;
var initialPrompt = "Answer all questions by users in a brief and concise way.";

var openAiModelSettings = new OpenAiModelSettings(openAiApiKey, openAiModelType, initialPrompt);

var openAiModel = new OpenAiModel(openAiModelSettings);

After creating the model, using it is as easy as this:

var thread = openAiModel.CreateThread("Why is the moon flying away further from earth?").Result;
Console.WriteLine($"[Assistant]: {thread.Last().Content[0].Text}");

// [Assistant]: The Moon is gradually moving away from the Earth due to tidal forces. 
// The Earth's rotation causes tidal bulges in the oceans, which transfer angular momentum 
// to the Moon, causing it to slowly accelerate in its orbit. As the Moon gains angular momentum, 
// its orbit becomes slightly larger, leading to an increase in the distance between the Earth and 
// the Moon. This process occurs at a rate of about 3.8 centimeters per year.

The returning object is a thread that contains a list of all previous messages and all function calls the model has made to answer your question. The final response to the latest question can be found in the last message. Currently the Thread is a List that is hard to work with. I am unhappy with the current state and will improve this in future updates.

You can continue the conversation of a thread as follows:

thread = openAiModel.AskInThread("When will the moon be to small to fully cover the sun?", thread).Result;


// [Assistant]: The phenomenon where the Moon covers the Sun completely during a solar eclipse is known 
// as a total solar eclipse. Due to the Moon's gradual increase in distance from the Earth, its apparent 
// size in the sky is shrinking. Scientists estimate that in about 600 million years, the Moon will no
// longer be able to fully cover the Sun, and total solar eclipses will no longer occur. Instead, we will
// only experience annular eclipses, where the Moon appears smaller than the Sun and leaves a ring of the
// Sun visible around its edges.

Creating Tools

Chatting with the LLM is nice and all, but the real fun lies in providing the LLM with custom Tools. These tools can integrate with APIs or Databases, anything you can imagine.

The following class is automatically detected by EasyOpenAiTools and added to the list of tools available to the LLM. For this auto discovery to work you need to provide 3 types of attributes:

• Tool: The tool Attribute is set at the class level, marking it as a tool. You can use it to set the name and the description of the tool. This description will be used by the LLM to identify and use the tool so make sure to explain it correctly and experiment.
• ToolProperty: These are Arguments you want the LLM to set when calling your custom Tool. Again each Argument has a name and Description.
• ToolMethod: This is the method that will be called when the LLM calls your tool

[Tool("GetWeatherAtLocation", 
      "Returns the current temperature of a given city or location")]
internal class WeatherToolNew
{
    [ToolProperty("Location", "The name of a city. This only takes the city name, not numbers or other inputs.")]
    public string Location { get; init; }

    [ToolMethod]
    public async Task<string> Execute()
    {
        var weatherClient = new OpenMeteoClient();
        var weatherData = await weatherClient.QueryAsync(Location);

        return
            $"Weather {Location} now: " +
            weatherData.Current.Temperature +
            weatherData.CurrentUnits.Temperature;
    }
}

No further action is required. When combining this with the model creation from the Creating a Model section you can easily start using the tool.

[User]: What is the current Temperature in Berlin?
[Assistant]: The current temperature in Berlin is 16.9�C.
[User]: How did you find this out?
[Assistant]: I used a weather API to retrieve the current temperature in Berlin.
[User]: Which tool did you use for that?
[Assistant]: I used the `GetWeatherAtLocation` function to find out the current temperature in Berlin.

Multiple calls in parallel are also possible:

[User]: Where is it currently 5�C?
[Assistant]: Currently, none of the places checked (Reykjavik, Helsinki, and Oslo) 
have a temperature of 5�C.

A few things to note:

• The last example only worked with GPT 4. GPT 3.5 always thought it could use the 5�C as a place name and supplied it straight to the API. Overall GPT 4 gives a lot better results when using tools but costs noticeably more as well

• A big limitation though is the context size of the LLM. You should not flat out return a complete dataset through a tool. Instead a tool should provide small functions that enhance the experience.

• Also note the 128 Tool limit

Future developement

I plan to rework a bunch of things. Namely the way exceptions are handled and the way the results are communicated at the moment.

Integrating image support or any other openai/openai-dotnet functions is not planned at the moment but might be added to the todo if I need them in any project.

Since this project is open source you are very welcome to contribute and implement any improvements or features you can imagine.