A lightweight, cross-platform .NET library for building RAG (Retrieval-Augmented Generation) pipelines with local embedding models and SQLite vector storage.

🚀 Features

• Local Embedding Models: Use ONNX models for offline, privacy-focused embeddings
• SQLite Vector Storage: Lightweight, file-based vector database with no external dependencies
• Multi-Format Support: Process PDFs and text files with intelligent chunking
• Flexible Training Strategies: Retrain from scratch, incremental updates, or add-only modes
• Real-time Ingestion: Stream text content directly into your vector store
• Metadata Preservation: Maintain document context and metadata throughout the pipeline
• Cross-Platform: Works on Windows, macOS, and Linux with .NET 8.0+

🎯 Quick Start

Installation

dotnet add package RAGamuffin

Basic Usage

using RAGamuffin.Builders;
using RAGamuffin.Core;
using RAGamuffin.Embedding;
using RAGamuffin.Enums;

// 1. Set up your embedding model (download from HuggingFace)
var embedder = new OnnxEmbedder("path/to/model.onnx", "path/to/tokenizer.json");

// 2. Configure your vector database
var vectorDb = new SqliteDatabaseModel("documents.db", "my_collection");

// 3. Build and train your pipeline
var pipeline = new IngestionTrainingBuilder()
    .WithEmbeddingModel(embedder)
    .WithVectorDatabase(vectorDb)
    .WithTrainingStrategy(TrainingStrategy.RetrainFromScratch)
    .WithTrainingFiles(new[] { "document.pdf" })
    .Build();

var ingestedItems = await pipeline.Train();

// 4. Search your documents
string[] results = await pipeline.SearchAndReturnTexts("What is the company policy?", 5);

Real-time Text Ingestion

// Stream text content directly into your vector store
var textItems = new[]
{
    new TextItem("Meeting notes from Q1", "Q1 was successful with 15% growth..."),
    new TextItem("Product roadmap", "Next quarter we'll launch feature X...")
};

var (ingestedItems, model) = await pipeline.TrainWithText(textItems);

Search Existing Vector Store

// Search without retraining
var vectorStore = new SqliteVectorStoreProvider("documents.db", "my_collection");
var searchResults = await vectorStore.SearchAsync("your query", embedder, 5);

// Get metadata
var metadata = await vectorStore.GetAllDocumentsMetadataAsync();

📚 Examples

Check out the comprehensive examples in the Examples/ directory:

• TrainAndSearch: Complete RAG pipeline with training and search
• SearchExistingVectorStore: Query existing vector stores with metadata
• IncrementalTraining: Add new documents to existing collections
• RealTimeIngestion: Stream text content in real-time
• MetadataRetrieval: Work with document metadata and statistics

🔧 Configuration

Embedding Models

RAGamuffin supports ONNX models for cross-platform compatibility. Recommended starter model:

• Model: all-mpnet-base-v2 from HuggingFace
• Download: Model | Tokenizer

Training Strategies

• RetrainFromScratch: Drop all existing data and retrain
• IncrementalAdd: Add new documents (skip if exists)
• IncrementalUpdate: Add new documents and update existing ones
• ProcessOnly: Only process documents, no vector operations

Chunking Options

// PDF processing options
.WithPdfOptions(new PdfHybridParagraphIngestionOptions
{
    MinSize = 0,        // Minimum chunk size
    MaxSize = 800,      // Maximum chunk size
    Overlap = 400,      // Overlap between chunks
    UseMetadata = true  // Include document metadata
})

// Text processing options
.WithTextOptions(new TextHybridParagraphIngestionOptions
{
    MinSize = 500,      // Minimum chunk size
    MaxSize = 800,      // Maximum chunk size
    Overlap = 400,      // Overlap between chunks
    UseMetadata = true  // Include document metadata
})

🏗️ Architecture

RAGamuffin is built with a modular architecture:

• Abstractions: Clean interfaces for embedding, ingestion, and vector storage
• Core: Main pipeline logic and data models
• Embedding: ONNX-based embedding providers
• Ingestion: PDF and text processing engines
• VectorStores: SQLite vector database implementation
• Builders: Fluent API for pipeline configuration

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

📄 License

This project is licensed under the MIT License - see the LICENSE.txt file for details.

🔗 Related Projects

• InstructSharp: LLM client library for .NET
• PdfPig: PDF processing library
• Microsoft.ML.OnnxRuntime: ONNX model inference

---

RAGamuffin - Making RAG pipelines simple and accessible for .NET developers.