Google_Gemini 0.10.2-dev.44

Google.Gemini

Features 🔥

• Fully generated C# SDK based on official Google.Gemini OpenAPI specification using AutoSDK
• Same day update to support new features
• Updated and supported automatically if there are no breaking changes
• All modern .NET features - nullability, trimming, NativeAOT, etc.
• Support .Net Framework/.Net Standard 2.0
• Microsoft.Extensions.AI IChatClient and IEmbeddingGenerator support

Usage

using Google.Gemini;

using var client = new GeminiClient(apiKey);

Microsoft.Extensions.AI

The SDK implements IChatClient and IEmbeddingGenerator:

using Google.Gemini;
using Microsoft.Extensions.AI;

// IChatClient
IChatClient chatClient = new GeminiClient(apiKey);
var response = await chatClient.GetResponseAsync(
    [new ChatMessage(ChatRole.User, "Hello!")],
    new ChatOptions { ModelId = "gemini-2.0-flash" });

// IEmbeddingGenerator
IEmbeddingGenerator<string, Embedding<float>> generator = new GeminiClient(apiKey);
var embeddings = await generator.GenerateAsync(
    ["Hello, world!"],
    new EmbeddingGenerationOptions { ModelId = "gemini-embedding-001" });

Live API (Real-time Voice/Video)

The SDK supports the Gemini Live API for real-time bidirectional voice and video interactions over WebSocket:

using Google.Gemini;

using var client = new GeminiClient(apiKey);

// Connect to the Live API
await using var session = await client.ConnectLiveAsync(new LiveSetupConfig
{
    Model = "models/gemini-2.5-flash-native-audio-latest",
    GenerationConfig = new GenerationConfig
    {
        ResponseModalities = [GenerationConfigResponseModalitie.Audio],
    },
});

// Send text and receive audio responses
await session.SendTextAsync("Hello, how are you?");

await foreach (var message in session.ReadEventsAsync())
{
    // Audio data in message.ServerContent.ModelTurn.Parts[].InlineData
    if (message.ServerContent?.TurnComplete == true)
        break;
}

Voice selection and speech config:

await using var session = await client.ConnectLiveAsync(new LiveSetupConfig
{
    Model = "models/gemini-2.5-flash-native-audio-latest",
    GenerationConfig = new GenerationConfig
    {
        ResponseModalities = [GenerationConfigResponseModalitie.Audio],
        SpeechConfig = new SpeechConfig
        {
            VoiceConfig = new VoiceConfig
            {
                PrebuiltVoiceConfig = new PrebuiltVoiceConfig
                {
                    VoiceName = "Kore", // Aoede, Charon, Fenrir, Kore, Puck, etc.
                },
            },
        },
    },
});

Multi-turn conversation:

// Send conversation history before triggering a response
await session.SendClientContentAsync(
    turns:
    [
        new Content
        {
            Role = "user",
            Parts = [new Part { Text = "My name is Alice" }],
        },
        new Content
        {
            Role = "model",
            Parts = [new Part { Text = "Nice to meet you, Alice!" }],
        },
        new Content
        {
            Role = "user",
            Parts = [new Part { Text = "What's my name?" }],
        },
    ],
    turnComplete: true);

System instruction (customize model behavior):

await using var session = await client.ConnectLiveAsync(new LiveSetupConfig
{
    Model = "models/gemini-2.5-flash-native-audio-latest",
    GenerationConfig = new GenerationConfig
    {
        ResponseModalities = [GenerationConfigResponseModalitie.Audio],
    },
    SystemInstruction = new Content
    {
        Parts = [new Part { Text = "You are a friendly pirate. Always respond in pirate speak." }],
    },
});

Tool calling:

var config = new LiveSetupConfig
{
    Model = "models/gemini-2.5-flash-native-audio-latest",
    GenerationConfig = new GenerationConfig
    {
        ResponseModalities = [GenerationConfigResponseModalitie.Audio],
    },
    Tools = [new Tool { FunctionDeclarations = [myFunction] }],
};

await using var session = await client.ConnectLiveAsync(config);
await session.SendTextAsync("What's the weather in London?");

await foreach (var message in session.ReadEventsAsync())
{
    if (message.ToolCall is { } toolCall)
    {
        // Handle function call and send response
        await session.SendToolResponseAsync([new FunctionResponse
        {
            Name = toolCall.FunctionCalls![0].Name,
            Id = toolCall.FunctionCalls[0].Id,
            Response = new { temperature = "15C" },
        }]);
    }

    // Tool calls cancelled due to user interruption
    if (message.ToolCallCancellation is { } cancellation)
    {
        Console.WriteLine($"Tool calls cancelled: {string.Join(", ", cancellation.Ids!)}");
    }

    if (message.ServerContent?.TurnComplete == true)
        break;
}

Session resumption (reconnect without losing context):

var config = new LiveSetupConfig
{
    Model = "models/gemini-2.5-flash-native-audio-latest",
    GenerationConfig = new GenerationConfig
    {
        ResponseModalities = [GenerationConfigResponseModalitie.Audio],
    },
    SessionResumption = new LiveSessionResumptionConfig(),
};

await using var session1 = await client.ConnectLiveAsync(config);
// ... interact ...
var handle = session1.LastSessionResumptionHandle;

// Later, reconnect with the handle
var config2 = new LiveSetupConfig
{
    // ... same config ...
    SessionResumption = new LiveSessionResumptionConfig { Handle = handle },
};
await using var session2 = await client.ConnectLiveAsync(config2);

Output transcription (get text alongside audio responses):

var config = new LiveSetupConfig
{
    Model = "models/gemini-2.5-flash-native-audio-latest",
    GenerationConfig = new GenerationConfig
    {
        ResponseModalities = [GenerationConfigResponseModalitie.Audio],
    },
    OutputAudioTranscription = new LiveOutputAudioTranscription(),
};

await using var session = await client.ConnectLiveAsync(config);
await session.SendTextAsync("Tell me a joke");

await foreach (var message in session.ReadEventsAsync())
{
    // Text transcription of the audio response
    if (message.ServerContent?.OutputTranscription?.Text is { } text)
        Console.Write(text);

    if (message.ServerContent?.TurnComplete == true)
        break;
}

Send audio/video:

// Send PCM audio (16-bit, 16kHz, little-endian, mono)
await session.SendAudioAsync(pcmBytes);

// Send audio with custom MIME type
await session.SendAudioAsync(audioBytes, "audio/pcm;rate=24000");

// Send video frame
await session.SendVideoAsync(jpegBytes, "image/jpeg");

// Stream video frames in a loop
foreach (var frame in videoFrames)
{
    await session.SendVideoAsync(frame, "image/jpeg");
    await Task.Delay(100); // ~10 fps
}

<details> <summary><b>Advanced features</b> (compression, interruption, usage, GoAway, audio round-trip)</summary>

Context window compression (for longer sessions):

var config = new LiveSetupConfig
{
    Model = "models/gemini-2.5-flash-native-audio-latest",
    GenerationConfig = new GenerationConfig
    {
        ResponseModalities = [GenerationConfigResponseModalitie.Audio],
    },
    ContextWindowCompression = new LiveContextWindowCompression
    {
        SlidingWindow = new LiveSlidingWindow
        {
            TargetTokens = 1024, // tokens to retain after compression
        },
    },
};

Interruption handling (user speaks during model response):

await foreach (var message in session.ReadEventsAsync())
{
    if (message.ServerContent?.Interrupted == true)
    {
        // Model response was cut short — user started speaking
        Console.WriteLine("Model interrupted by user input");
    }

    if (message.ServerContent?.ModelTurn?.Parts is { } parts)
    {
        foreach (var part in parts)
        {
            // Process audio/text parts (may be partial if interrupted)
            if (part.InlineData?.Data is { } audioData)
                PlayAudio(audioData);
        }
    }

    if (message.ServerContent?.TurnComplete == true)
        break;
}

Usage metadata (track token consumption):

await foreach (var message in session.ReadEventsAsync())
{
    if (message.UsageMetadata is { } usage)
    {
        Console.WriteLine($"Prompt tokens: {usage.PromptTokenCount}");
        Console.WriteLine($"Response tokens: {usage.CandidatesTokenCount}");
        Console.WriteLine($"Total tokens: {usage.TotalTokenCount}");
    }

    if (message.ServerContent?.TurnComplete == true)
        break;
}

GoAway handling (graceful session migration):

await foreach (var message in session.ReadEventsAsync())
{
    if (message.GoAway is { } goAway)
    {
        // Server is closing soon — reconnect using session resumption
        Console.WriteLine($"Server closing in {goAway.TimeLeft}, reconnecting...");
        break; // dispose session and reconnect with resumption handle
    }

    if (message.ServerContent?.TurnComplete == true)
        break;
}

Audio round-trip (send and receive audio):

var config = new LiveSetupConfig
{
    Model = "models/gemini-2.5-flash-native-audio-latest",
    GenerationConfig = new GenerationConfig
    {
        ResponseModalities = [GenerationConfigResponseModalitie.Audio],
    },
};

await using var session = await client.ConnectLiveAsync(config);

// Send PCM audio (16-bit, 16kHz, little-endian, mono)
await session.SendAudioAsync(pcmBytes);

// Signal end of user turn
await session.SendClientContentAsync(turns: [], turnComplete: true);

// Receive audio response
await foreach (var message in session.ReadEventsAsync())
{
    if (message.ServerContent?.ModelTurn?.Parts is { } parts)
    {
        foreach (var part in parts)
        {
            if (part.InlineData?.Data is { } audioData)
            {
                // audioData is base64-decoded PCM audio (24kHz)
                PlayAudio(audioData);
            }
        }
    }

    if (message.ServerContent?.TurnComplete == true)
        break;
}

</details>

Embedding Models

The SDK defaults to gemini-embedding-001. For best retrieval quality, use gemini-embedding-2-preview (note: embedding spaces are incompatible between the two models). See Google's embedding guide for details.

API Version

This SDK targets the v1beta API, which is the full-featured version used by Google's own SDKs (Python, JS, Go). The v1 (stable) API only exposes ~30 of the 70+ available endpoints and lacks critical features like tool calling, file upload, context caching, and grounding.

Support

Priority place for bugs: https://github.com/tryAGI/Google_Generative_AI/issues Priority place for ideas and general questions: https://github.com/tryAGI/Google_Generative_AI/discussions Discord: https://discord.gg/Ca2xhfBf3v

Acknowledgments

This project is supported by JetBrains through the Open Source Support Program.

This project is supported by CodeRabbit through the Open Source Support Program.