Mscc.GenerativeAI 1.2.0

Gemini AI Client for .NET and ASP.NET Core

Access and integrate the Gemini API into your .NET applications. The packages support both Google AI Studio and Google Cloud Vertex AI.

Name	Package	Status
Client for .NET	Mscc.GenerativeAI
Client for ASP.NET (Core)	Mscc.GenerativeAI.Web
Client for .NET using Google API Client Library	Mscc.GenerativeAI.Google

Read more about Mscc.GenerativeAI.Web and how to add it to your ASP.NET (Core) web applications. Read more about Mscc.GenerativeAI.Google.

Install the package 🖥️

Install the package Mscc.GenerativeAI from NuGet. You can install the package from the command line using either the command line or the NuGet Package Manager Console. Or you add it directly to your .NET project.

Add the package using the dotnet command line tool in your .NET project folder.

> dotnet add package Mscc.GenerativeAI

Working with Visual Studio use the NuGet Package Manager to install the package Mscc.GenerativeAI.

PM> Install-Package Mscc.GenerativeAI

Alternatively, add the following line to your .csproj file.

  <ItemGroup>
    <PackageReference Include="Mscc.GenerativeAI" Version="1.2.0" />
  </ItemGroup>

You can then add this code to your sources whenever you need to access any Gemini API provided by Google. This package works for Google AI (Google AI Studio) and Google Cloud Vertex AI.

Features (as per Gemini analysis) ✦

The provided code defines a C# library for interacting with Google's Generative AI models, specifically the Gemini models. It provides functionalities to:

• List available models: This allows users to see which models are available for use.
• Get information about a specific model: This provides details about a specific model, such as its capabilities and limitations.
• Generate content: This allows users to send prompts to a model and receive generated text in response.
• Generate content stream: This allows users to receive a stream of generated text from a model, which can be useful for real-time applications.
• Generate a grounded answer: This allows users to ask questions and receive answers that are grounded in provided context.
• Generate embeddings: This allows users to convert text into numerical representations that can be used for tasks like similarity search.
• Count tokens: This allows users to estimate the cost of using a model by counting the number of tokens in a prompt or response.
• Start a chat session: This allows users to have a back-and-forth conversation with a model.
• Create tuned models: This allows users to provide samples for tuning an existing model. Currently, only the text-bison-001 and gemini-1.0-pro-001 models are supported for tuning
• File API: This allows users to upload large files and use them with Gemini 1.5.

The package also defines various helper classes and enums to represent different aspects of the Gemini API, such as model names, request parameters, and response data.

Authentication use cases 👥

The package supports the following use cases to authenticate.

• Google AI: Authentication with an API key
• Google AI: Authentication with OAuth
• Vertex AI: Authentication with Application Default Credentials (ADC)
• Vertex AI: Authentication with OAuth (using Mscc.GenerativeAI.Google)
• Vertex AI: Authentication with Service Account (using Mscc.GenerativeAI.Google)

This applies mainly to the instantiation procedure.

Getting Started 🚀

Use of Gemini API in either Google AI or Vertex AI is almost identical. The major difference is the way to instantiate the model handling your prompt.

Using Environment variables

In the cloud most settings are configured via environment variables (EnvVars). The ease of configuration, their wide spread support and the simplicity of environment variables makes them a very interesting option.

Using any environment variable provides a simplified access to a model.

using Mscc.GenerativeAI;

var model = new GenerativeModel();

Choose an API and authentication mode

Google AI with an API key

using Mscc.GenerativeAI;
// Google AI with an API key
var googleAI = new GoogleAI(apiKey: "your API key");
var model = googleAI.GenerativeModel(model: Model.GeminiPro);

// Original approach, still valid.
// var model = new GenerativeModel(apiKey: "your API key", model: Model.GeminiPro);

Google AI with OAuth. Use gcloud auth application-default print-access-token to get the access token.

using Mscc.GenerativeAI;
// Google AI with OAuth. Use `gcloud auth application-default print-access-token` to get the access token.
var model = new GenerativeModel(model: Model.GeminiPro);
model.AccessToken = accessToken;

Vertex AI with OAuth. Use gcloud auth application-default print-access-token to get the access token.

using Mscc.GenerativeAI;
// Vertex AI with OAuth. Use `gcloud auth application-default print-access-token` to get the access token.
var vertex = new VertexAI(projectId: projectId, region: region);
var model = vertex.GenerativeModel(model: Model.Gemini10Pro);
model.AccessToken = accessToken;

The ConfigurationFixture type in the test project implements multiple options to retrieve sensitive information, ie. API key or access token.

Using Google AI Gemini API

Working with Google AI in your application requires an API key. Get an API key from Google AI Studio.

using Mscc.GenerativeAI;

var apiKey = "your_api_key";
var prompt = "Write a story about a magic backpack.";

var model = new GenerativeModel(apiKey: apiKey, model: Model.GeminiPro);

var response = model.GenerateContent(prompt).Result;
Console.WriteLine(response.Text);

Using Vertex AI Gemini API

Use of Vertex AI requires an account on Google Cloud, a project with billing and Vertex AI API enabled.

using Mscc.GenerativeAI;

var projectId = "your_google_project_id"; // the ID of a project, not its name.
var region = "us-central1";     // see documentation for available regions.
var accessToken = "your_access_token";      // use `gcloud auth application-default print-access-token` to get it.
var prompt = "Write a story about a magic backpack.";

var vertex = new VertexAI(projectId: projectId, region: region);
var model = vertex.GenerativeModel(model: Model.GeminiPro);
model.AccessToken = accessToken;

var response = model.GenerateContent(prompt).Result;
Console.WriteLine(response.Text);

More examples

Supported models are accessible via the Model class. Since release 0.9.0 there is support for the previous PaLM 2 models and their functionalities.

Text-and-image input

using Mscc.GenerativeAI;

var apiKey = "your_api_key";
var prompt = "Parse the time and city from the airport board shown in this image into a list, in Markdown";
var model = new GenerativeModel(apiKey: apiKey, model: Model.GeminiVisionPro);
var request = new GenerateContentRequest(prompt);
await request.AddMedia("https://ai.google.dev/static/docs/images/timetable.png");

var response = await model.GenerateContent(request);
Console.WriteLine(response.Text);

The part of InlineData is supported by both Google AI and Vertex AI. Whereas the part FileData is restricted to Vertex AI only.

Chat conversations

Gemini enables you to have freeform conversations across multiple turns. You can interact with Gemini Pro using a single-turn prompt and response or chat with it in a multi-turn, continuous conversation, even for code understanding and generation.

using Mscc.GenerativeAI;

var apiKey = "your_api_key";
var model = new GenerativeModel(apiKey: apiKey);    // using default model: gemini-pro
var chat = model.StartChat();

// Instead of discarding you could also use the response and access `response.Text`.
_ = await chat.SendMessage("Hello, fancy brainstorming about IT?");
_ = await chat.SendMessage("In one sentence, explain how a computer works to a young child.");
_ = await chat.SendMessage("Okay, how about a more detailed explanation to a high schooler?");
_ = await chat.SendMessage("Lastly, give a thorough definition for a CS graduate.");

// A chat session keeps every response in its history.
chat.History.ForEach(c => Console.WriteLine($"{c.Role}: {c.Text}"));

// Last request/response pair can be removed from the history.
var latest = chat.Rewind();
Console.WriteLine($"{latest.Sent} - {latest.Received}");

Use Gemini 1.5 with large files

With Gemini 1.5 you can create multimodal prompts supporting large files.

The following example uploads one or more files via File API and the created File URIs are used in the GenerateContent call to generate text.

using Mscc.GenerativeAI;

var apiKey = "your_api_key";
var prompt = "Make a short story from the media resources. The media resources are:";
IGenerativeAI genAi = new GoogleAI(apiKey);
var model = genAi.GenerativeModel(Model.Gemini15Pro);

// Upload your large image(s).
// Instead of discarding you could also use the response and access `response.Text`.
var filePath = Path.Combine(Environment.CurrentDirectory, "verylarge.png");
var displayName = "My very large image";
_ = await model.UploadMedia(filePath, displayName);

// Create the prompt with references to File API resources.
var request = new GenerateContentRequest(prompt);
var files = await model.ListFiles();
foreach (var file in files.Where(x => x.MimeType.StartsWith("image/")))
{
    _output.WriteLine($"File: {file.Name}");
    request.AddMedia(file);
}
var response = await model.GenerateContent(request);
Console.WriteLine(response.Text);

Read more about Gemini 1.5: Our next-generation model, now available for Private Preview in Google AI Studio.

Create a tuned model

The Gemini API lets you tune models on your own data. Since it's your data and your tuned models this needs stricter access controls than API-Keys can provide.

Before you can create a tuned model, you'll need to setup OAuth for your project.

using Mscc.GenerativeAI;

var projectId = "your_google_project_id"; // the ID of a project, not its name.
var accessToken = "your_access_token";      // use `gcloud auth application-default print-access-token` to get it.
var model = new GenerativeModel(apiKey: null, model: Model.Gemini10Pro001)
{
    AccessToken = accessToken, ProjectId = projectId
};
var parameters = new HyperParameters() { BatchSize = 2, LearningRate = 0.001f, EpochCount = 3 };
var dataset = new List<TuningExample>
{    
    new() { TextInput = "1", Output = "2" },
    new() { TextInput = "3", Output = "4" },
    new() { TextInput = "-3", Output = "-2" },
    new() { TextInput = "twenty two", Output = "twenty three" },
    new() { TextInput = "two hundred", Output = "two hundred one" },
    new() { TextInput = "ninety nine", Output = "one hundred" },
    new() { TextInput = "8", Output = "9" },
    new() { TextInput = "-98", Output = "-97" },
    new() { TextInput = "1,000", Output = "1,001" },
    new() { TextInput = "thirteen", Output = "fourteen" },
    new() { TextInput = "seven", Output = "eight" },
};
var request = new CreateTunedModelRequest(Model.Gemini10Pro001, 
    "Simply autogenerated Test model",
    dataset,
    parameters);

var response = await model.CreateTunedModel(request);
Console.WriteLine($"Name: {response.Name}");
Console.WriteLine($"Model: {response.Metadata.TunedModel} (Steps: {response.Metadata.TotalSteps})");

(This is still work in progress but operational. Future release will provide types to simplify the create request.)

Tuned models appear in your Google AI Studio library.

Read more about Tune Gemini Pro in Google AI Studio or with the Gemini API.

More samples

The folders samples and tests contain more examples.

• Simple console application
• ASP.NET Core Minimal web application
• ASP.NET Core MVP web application (work in progress!)

Troubleshooting ⚡

Sometimes you might have authentication warnings HTTP 403 (Forbidden). Especially while working with OAuth-based authentication. You can fix it by re-authenticating through ADC.

gcloud config set project "$PROJECT_ID"

gcloud auth application-default set-quota-project "$PROJECT_ID"
gcloud auth application-default login

Make sure that the required API have been enabled.

# ENABLE APIs
gcloud services enable aiplatform.googleapis.com

In case of long-running streaming requests it can happen that you get a HttpIOException: The response ended prematurely while waiting for the next frame from the server. (ResponseEnded). The root cause is the .NET runtime and the solution is to upgrade to the latest version of the .NET runtime. In case that you cannot upgrade you might disable dynamic window sizing as a workaround: Either using the environment variable DOTNET_SYSTEM_NET_HTTP_SOCKETSHTTPHANDLER_HTTP2FLOWCONTROL_DISABLEDYNAMICWINDOWSIZING

DOTNET_SYSTEM_NET_HTTP_SOCKETSHTTPHANDLER_HTTP2FLOWCONTROL_DISABLEDYNAMICWINDOWSIZING=true

or setting an AppContext switch:

AppContext.SetSwitch("System.Net.SocketsHttpHandler.Http2FlowControl.DisableDynamicWindowSizing", true);

Several issues regarding this problem have been reported on GitHub:

• https://github.com/dotnet/runtime/pull/97881
• https://github.com/grpc/grpc-dotnet/issues/2361
• https://github.com/grpc/grpc-dotnet/issues/2358

Using the tests 🧩

The repository contains a number of test cases for Google AI and Vertex AI. You will find them in the tests folder. They are part of the [GenerativeAI solution]. To run the tests, either enter the relevant information into the appsettings.json, create a new appsettings.user.json file with the same JSON structure in the tests folder, or define the following environment variables

• GOOGLE_API_KEY
• GOOGLE_PROJECT_ID
• GOOGLE_REGION
• GOOGLE_ACCESS_TOKEN (optional: if absent, gcloud auth application-default print-access-token is executed)

The test cases should provide more insights and use cases on how to use the Mscc.GenerativeAI package in your .NET projects.

Feedback ✨

For support and feedback kindly create issues at the https://github.com/mscraftsman/generative-ai repository.

License 📜

This project is licensed under the Apache-2.0 License - see the LICENSE file for details.

Citation 📚

If you use Mscc.GenerativeAI in your research project, kindly cite as follows

@misc{Mscc.GenerativeAI,
  author = {Kirstätter, J and MSCraftsman},
  title = {Mscc.GenerativeAI - Gemini AI Client for .NET and ASP.NET Core},
  year = {2024},
  publisher = {GitHub},
  journal = {GitHub repository},
  note = {https://github.com/mscraftsman/generative-ai}
}

Created by Jochen Kirstätter.