Quick Start

Please note that the code provided in this page is purely for learning purposes and is far from perfect. Remember to null-check all responses!

Breaking Changes Notice

If you've just updated the package, it is recommended to check the changelogs for information on breaking changes.

Setup

This plugin only provides abstraction for the chat completion APIs provided by other third party libraries. The plugin itself does not contain the code to interact with the Gemini or GPT APIs. The Gemini client relies on UGemini, and the OpenAI client relies on com.openai.unity.

Authentication is also handled by the third party libraries. See UGemini: Setup and com.openai.unity: Authentication

Now you can create an instance of the desired API client. For Gemini:

using Uralstech.UAI.Abstraction;
using Uralstech.UAI.Abstraction.Providers.Gemini;

IModelClient client = new GeminiModelClient();

For OpenAI:

using Uralstech.UAI.Abstraction;
using Uralstech.UAI.Abstraction.Providers.OAI;

IModelClient client = new OAIModelClient();

Chat

You can chat with the model using IModelClient.Chat.

using Uralstech.UAI.Abstraction.Chat;

ChatInferenceResult result = await client.Chat(new Message[]
{
    new Message(Role.System, "You are a helpful, friendly assistant."),
    new Message(Role.User, "What is the capital of India?")
});

Debug.Log("Response: " + result.Messages[^1].Content);

The method returns a ChatInferenceResult object with the generated messages, including function calls and responses. The model's plain-text response will likely be the last message in the array.

If you don't provide a model to use, it uses the model defined in the client's DefaultModelId. You can change this by either changing the default model, or by providing the model ID in the method's model parameter.

For clients that support the feature, you can also turn of the safety filters of the model by setting tryRemoveFilters to true.

Function Calling

You can define functions using the Function class. They can then be passed into a variant of the Chat function to be called by the model.

using Newtonsoft.Json.Linq;
using Uralstech.UAI.Abstraction.Chat;
using Uralstech.UAI.Abstraction.Tools;

ChatInferenceResult result = await client.Chat(new Message[]
{
    new Message(Role.System, "You are a helpful, friendly assistant."),
    new Message(Role.User, "Can you print \"Hello, World\" to the console?")
},
new Function[]
{
    new Function(
        nameof(PrintToConsole),
        "Prints text to the console.",
        new Parameter[]
        {
            new Parameter()
            {
                Name = "text",
                Description = "The text to print.",
                Type = ParameterType.String,
            }
        },
        PrintToConsole
    )
});

Debug.Log("Response: " + result.Messages[^1].Content);

async Awaitable<JObject> PrintToConsole(JToken input)
{
    await Awaitable.MainThreadAsync(); // Just to make this a valid Awaitable.

    Debug.Log(input["text"].ToObject<string>());
    return new JObject()
    {
        ["status"] = "Printed text to console."
    };
}