EmbeddingGemma.SemanticKernel 1.0.1

EmbeddingGemma.NET

EmbeddingGemma.NET provides .NET bindings for Google DeepMind's EmbeddingGemma-300m model, enabling fully local, offline text embedding with no API key, no cloud dependency, and no data egress.

Two NuGet packages are available: one for general .NET applications and one for Microsoft Semantic Kernel.

Why EmbeddingGemma?

Installation

Model Setup

The ONNX model and tokenizer files must be present on disk before the service can be used. They are hosted at onnx-community/embeddinggemma-300m-ONNX on Hugging Face.

Option A — PowerShell script (recommended)

Run the included script once from the repository root. It downloads all required files into a .embedding_resources folder by default; pass -OutputPath to use a different location.

.\Initialize-Embedding-Resources.ps1

Option B — Manual download

Manually download the following files and place them in the same directory:

The resulting directory must have the following structure:

<model-directory>/
├── model.onnx
├── model.onnx_data
├── tokenizer.json
├── tokenizer.model
└── tokenizer_config.json

Usage

Dependency Registration

EmbeddingGemma.Core via IServiceCollection

using EmbeddingGemma.Core;

// Registers IEmbeddingGenerator<string, Embedding<float>> as a singleton.
builder.Services.AddGemmaTextEmbeddingGenerator(options => options.ModelDirectory = @"C:\path\to\model-directory");

EmbeddingGemma.SemanticKernel via IKernelBuilder

using EmbeddingGemma.SemanticKernel;
using Microsoft.SemanticKernel;

var builder = Kernel.CreateBuilder();

builder.AddGemmaTextEmbeddingGenerator(options => options.ModelDirectory = @"C:\path\to\model-directory");

var kernel = builder.Build();

Generating Embeddings

Resolve IEmbeddingGenerator<string, Embedding<float>> from the DI container and call GenerateAsync.

using Microsoft.Extensions.AI;

var generator = serviceProvider.GetRequiredService<IEmbeddingGenerator<string, Embedding<float>>>();

// Without a task type — no prompt prefix is added.
var embeddings = await generator.GenerateAsync(["Hello, world!"]);

// With a task type — the appropriate prompt prefix is applied automatically.
var options = new EmbeddingGemmaGenerationOptions 
{ 
    TaskType = EmbeddingGemmaTaskType.RetrievalQuery 
};
var queryEmbeddings = await generator.GenerateAsync(["What is semantic search?"], options);

For document embeddings, supply an optional DocumentTitle to improve retrieval quality:

var docOptions = new EmbeddingGemmaGenerationOptions
{
    TaskType = EmbeddingGemmaTaskType.RetrievalDocument,
    DocumentTitle = "Introduction to Semantic Search"
};

var docEmbeddings = await generator.GenerateAsync(["Semantic search ranks results by meaning..."], docOptions);

Task Types

Set EmbeddingGemmaGenerationOptions.TaskType to have the service automatically prepend the correct prompt prefix for your scenario. When TaskType is null, no prefix is added.

For detailed guidance on prompt formatting, refer to:

• EmbeddingGemma Model Card — Prompt Instructions
• Using Prompts with EmbeddingGemma