LocalReranker 0.3.0

LocalReranker

⚠️ ARCHIVED: This project has been archived and is no longer actively maintained.

Successor: This project has been superseded by LocalAI.Reranker, which offers improved features and is actively maintained.

Existing users are encouraged to migrate to LocalAI.Reranker for continued support and updates.

A lightweight, zero-configuration semantic reranker for .NET. No API keys, no cloud dependencies—just install and use.

Features

• Zero Configuration - Works out of the box with sensible defaults
• Auto Model Download - Models are downloaded and cached automatically from HuggingFace Hub
• Offline First - Fully local inference, no internet required after initial download
• Minimal Dependencies - Only 2 packages: ONNX Runtime + ML.Tokenizers
• Cross-Platform - Windows, Linux, macOS
• Thread-Safe - Single instance can be shared across multiple threads

Installation

dotnet add package LocalReranker

Quick Start

using LocalReranker;

await using var reranker = new Reranker();

var query = "What is machine learning?";
var documents = new[]
{
    "Machine learning is a subset of artificial intelligence.",
    "The weather today is sunny.",
    "Deep learning uses neural networks with many layers."
};

var results = await reranker.RerankAsync(query, documents, topK: 5);

foreach (var result in results)
{
    Console.WriteLine($"[{result.Score:F4}] {result.Document}");
}

Configuration

var reranker = new Reranker(new RerankerOptions
{
    ModelId = "quality",              // Use alias or full HuggingFace ID
    MaxSequenceLength = 512,
    CacheDirectory = "/path/to/models",
    UseGpu = true,
    BatchSize = 64
});

Supported Models (v0.1.0)

Coming in v0.2.0

Note: BGE models require SentencePiece tokenizer support, which is planned for v0.2.0.

ASP.NET Core Integration

// Program.cs - Register as singleton (recommended)
builder.Services.AddSingleton<IReranker>(sp =>
{
    var reranker = new Reranker(new RerankerOptions
    {
        ModelId = "default",
        BatchSize = 32
    });
    reranker.WarmupAsync().Wait(); // Pre-load model at startup
    return reranker;
});

// SearchService.cs
public class SearchService(IReranker reranker)
{
    public async Task<List<Document>> SearchAsync(string query, List<Document> candidates)
    {
        var texts = candidates.Select(d => d.Content).ToArray();
        var ranked = await reranker.RerankAsync(query, texts, topK: 10);
        return ranked.Select(r => candidates[r.OriginalIndex]).ToList();
    }
}

Performance Tips

1. Reuse instances - Create one Reranker and reuse it (thread-safe)
2. Use WarmupAsync() - Pre-load model during app startup
3. Adjust BatchSize - Larger batches are faster but use more memory
4. Use topK - Return only the results you need

Requirements

• .NET 10.0+
• ~90MB disk space (default model)
• ~200MB RAM (model loaded)

Documentation

• API Reference
• Design Overview
• Changelog

License

MIT - See LICENSE for details.