👁 Image
ElBruno.LocalEmbeddings.Npu 1.5.0

LocalEmbeddings

👁 NuGet
👁 NuGet Downloads
👁 Build Status
👁 GitHub stars
👁 Twitter Follow

A .NET library for generating text embeddings locally using ONNX Runtime and Microsoft.Extensions.AI abstractions — no external API calls required.

🎥 Quick Overview

New to local embeddings? Watch this 5-minute video explaining the main goal of the library.

Want to build RAG applications? Read this blog post about 3 RAG approaches in .NET with local embeddings and zero cloud calls.

Interested in image embeddings? Check out the YouTube video and blog post about local image embeddings with CLIP and ONNX.

Features

• Local Embedding Generation — Run inference entirely on your machine using ONNX Runtime
• Microsoft.Extensions.AI Integration — Implements IEmbeddingGenerator<string, Embedding<float>>
• 🦅 Harrier model support — Microsoft Harrier-OSS-v1 (270M, 640-dim, 94+ languages, instruction-tuned) via ElBruno.LocalEmbeddings.Harrier
• Kernel Memory Integration — Companion package ElBruno.LocalEmbeddings.KernelMemory provides a native ITextEmbeddingGenerator adapter for Microsoft Kernel Memory
• VectorData Integration — Companion package ElBruno.LocalEmbeddings.VectorData adds DI helpers for Microsoft.Extensions.VectorData (VectorStore and typed collections)
• Built-in In-Memory Vector Store — ElBruno.LocalEmbeddings.VectorData includes InMemoryVectorStore (no Semantic Kernel connector dependency required)
• HuggingFace Model Support — Use popular sentence transformer models from HuggingFace Hub
• Automatic Model Caching — Models are downloaded once and cached locally
• Dependency Injection Support — First-class IServiceCollection integration
• Single-String Convenience API — GenerateAsync("text") and GenerateEmbeddingAsync("text") — no array wrapping needed
• Similarity Helpers — Cosine similarity, all-pairs Similarity(...) matrix, and one-line FindClosestAsync(...) semantic search
• Thread-Safe & Batched — Concurrent generation and efficient multi-text processing

📦 NuGet Packages

Installation

dotnet add package ElBruno.LocalEmbeddings

For Harrier model support, install the companion package:

dotnet add package ElBruno.LocalEmbeddings.Harrier

For Kernel Memory integration, also install:

dotnet add package ElBruno.LocalEmbeddings.KernelMemory

For VectorData integration, install:

dotnet add package ElBruno.LocalEmbeddings.VectorData

Quick Start

1) Generate one embedding

using ElBruno.LocalEmbeddings;

await using var generator = await LocalEmbeddingGenerator.CreateAsync();
var embedding = await generator.GenerateEmbeddingAsync("Hello, world!");
Console.WriteLine(embedding.Vector.Length); // 384

2) Generate embeddings for multiple texts

var inputs = new[] { "first text", "second text", "third text" };
var embeddings = await generator.GenerateAsync(inputs);
Console.WriteLine(embeddings.Count); // 3

3) Compare two texts with cosine similarity

using ElBruno.LocalEmbeddings.Extensions;

var pair = await generator.GenerateAsync(["I love coding", "I enjoy programming"]);
var score = pair[0].CosineSimilarity(pair[1]);
Console.WriteLine(score);

4) Semantic search in one line

var corpus = new[]
{
 "Python for data science",
 "JavaScript for web apps",
 "Swift for iOS development"
};

var corpusEmbeddings = await generator.GenerateAsync(corpus);
var results = await generator.FindClosestAsync(
 "best language for websites",
 corpus,
 corpusEmbeddings,
 topK: 2,
 minScore: 0.2f);

foreach (var result in results)
 Console.WriteLine($"{result.Score:F3} - {result.Text}");

5) Using Harrier model (higher quality, multilingual)

using ElBruno.LocalEmbeddings.Harrier;

var generator = await HarrierEmbeddingGenerator.CreateAsync();
var embedding = await generator.GenerateEmbeddingAsync("Hello, world!");
Console.WriteLine($"Dimensions: {embedding.Vector.Length}"); // 640

For custom models and runtime behavior, use the options-based constructor: new LocalEmbeddingGenerator(new LocalEmbeddingsOptions { ... }).

Note: The synchronous constructor remains available for backward compatibility, but performs blocking initialization when downloads are needed.

Want to go further? Read the and the other docs in this repo for DI, configuration, VectorData, Kernel Memory, and full RAG examples.

Prefer a containerized dev environment? See the Dev Container section in the .

Samples

All sample applications are located in src/Samples/. See the for prerequisites and run instructions.

Configuration

var options = new LocalEmbeddingsOptions
{
 ModelName = "sentence-transformers/all-MiniLM-L6-v2", // HuggingFace model
 MaxSequenceLength = 512, // Max tokens
 CacheDirectory = null, // Auto-detect per platform
 EnsureModelDownloaded = true, // Download if missing
 NormalizeEmbeddings = false // L2 normalize vectors
};

See for supported models, local model paths, and cache locations.

Common model options (with model cards)

Estimated download sizes below are approximate and can vary by ONNX variant (fp32/int8) and tokenizer assets.

• sentence-transformers/all-MiniLM-L6-v2 (default, ~90–100 MB)
• sentence-transformers/all-MiniLM-L12-v2 (~130–140 MB)
• sentence-transformers/paraphrase-MiniLM-L6-v2 (~90–100 MB)
• BAAI/bge-large-en-v1.5 (large, ~1.3 GB)
• intfloat/e5-large-v2 (large, ~1.3 GB)

Documentation

Have an idea for a new feature or sample? Please open an issue and share your suggestion.

Building from Source

git clone https://github.com/elbruno/elbruno.localembeddings.git
cd elbruno.localembeddings
dotnet build
dotnet test

Requirements

• .NET 10.0 SDK or later
• ONNX Runtime compatible platform (Windows, Linux, macOS)

👋 About the Author

Hi! I'm ElBruno 🧡, a passionate developer and content creator exploring AI, .NET, and modern development practices.

Made with ❤️ by ElBruno

If you like this project, consider following my work across platforms:

• 📻 Podcast: No Tienen Nombre — Spanish-language episodes on AI, development, and tech culture
• 💻 Blog: ElBruno.com — Deep dives on embeddings, RAG, .NET, and local AI
• 📺 YouTube: youtube.com/elbruno — Demos, tutorials, and live coding
• 🔗 LinkedIn: @elbruno — Professional updates and insights
• 𝕏 Twitter: @elbruno — Quick tips, releases, and tech news

License

This project is licensed under the MIT License — see the file for details.