ElBruno.LocalEmbeddings.Harrier

ElBruno.LocalEmbeddings.Harrier

High-quality multilingual embeddings using Microsoft Harrier-OSS-v1 locally with ONNX Runtime.

Installation

dotnet add package ElBruno.LocalEmbeddings.Harrier

Quick Start

using ElBruno.LocalEmbeddings.Harrier;

// Create with default settings (downloads INT8 quantized model on first run)
await using var generator = await HarrierEmbeddingGenerator.CreateAsync();

// Generate embeddings
var embeddings = await generator.GenerateAsync(["Hello world!", "Hola mundo!"]);
Console.WriteLine($"Dimensions: {embeddings[0].Vector.Length}"); // 640

Model Details

• Publisher: Microsoft (MIT license)
• Architecture: Decoder-only (Gemma 3 based)
• Parameters: 270M
• Embedding dimensions: 640
• Context window: 32,768 tokens
• Languages: 94+
• MTEB-v2 ranking: #1
• Default variant: INT8 Quantized (~270 MB)

Features

• 🦅 Top-ranked embedding model on MTEB-v2 — best semantic quality
• 🌍 Multilingual support — 94+ languages
• 📦 Multiple quantization variants — Choose between quality (FP32/FP16) and speed (INT8/Q4)
• 🎯 Instruction-tuned — Specify task-specific prefixes for better results
• 📚 Long context — Supports up to 32,768 tokens
• ⚡ ONNX Runtime — Runs locally on CPU, zero external API calls
• 🔒 Private by default — No data sent to external services

Configuration

var options = new HarrierEmbeddingsOptions
{
    // Model variant (FP32, FP16, Quantized, Q4)
    ModelVariant = HarrierModelVariant.Quantized,
    
    // Instruction prefix for task-specific embeddings
    InstructionPrefix = "Instruct: Retrieve semantically similar text\nQuery: ",
    
    // Token limit
    MaxSequenceLength = 8192,
    
    // Cache and download
    EnsureModelDownloaded = true,
    CacheDirectory = null  // Auto-detect per platform
};

await using var generator = await HarrierEmbeddingGenerator.CreateAsync(options);

Dependency Injection

using ElBruno.LocalEmbeddings.Harrier.Extensions;
using Microsoft.Extensions.AI;

services.AddHarrierEmbeddings(options =>
{
    options.ModelVariant = HarrierModelVariant.Q4;
});

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

Learn More

📖 Full Harrier Integration Guide — Instruction prefixes, configuration, troubleshooting, and migration from MiniLM.

🎬 HarrierConsoleApp Sample — Complete working example with 6 progressive scenarios.

License

MIT — See LICENSE