RAGSharp 1.0.0

RAGSharp: Lightweight RAG Pipeline for .NET

RAGSharp is a lightweight, extensible Retrieval-Augmented Generation (RAG) library built entirely in C#. It provides clean, minimal, and predictable components for building RAG pipelines: document loading, token-aware text splitting, vector storage, and semantic search.

🚀 Why RAGSharp?

Choose RAGSharp if you want:

• Just the RAG essentials - load, chunk, embed, search
• Local-first - Works with any OpenAI-compatible API (OpenAI, LM Studio, Ollama, vLLM, etc.) out of the box.
• No database required - File-based storage out of the box, extensible to Postgres/Qdrant/etc.
• Simple and readable - Straightforward interfaces you can understand quickly

RAGSharp doesn’t aim to replace frameworks like Semantic Kernel or Kernel Memory — instead, it is a minimal, local-friendly pipeline you can understand in one sitting and drop into an existing app.

📦 Installation

dotnet add package RAGSharp

Dependencies:

• OpenAI - Official SDK for embeddings
• SharpToken - GPT tokenization for accurate chunking
• HtmlAgilityPack - HTML parsing for web loaders
• Microsoft.Extensions.Logging - Logging abstractions

✨ Quick Start

Load documents, index them, and search - in under 15 lines:

using RAGSharp.Embeddings.Providers;
using RAGSharp.IO;
using RAGSharp.RAG;
using RAGSharp.Stores;

//load doc
var docs = await new FileLoader().LoadAsync("sample.txt");

var retriever = new RagRetriever(
    embeddings: new OpenAIEmbeddingClient(
        baseUrl: "http://127.0.0.1:1234/v1",
        apiKey: "lmstudio",
        defaultModel: "text-embedding-3-small"
    ),
    store: new InMemoryVectorStore()
);

//ingest
await retriever.AddDocumentsAsync(docs);

//search
var results = await retriever.Search("quantum mechanics",  topK: 3);
foreach (var r in results)
    Console.WriteLine($"{r.Score:F2}: {r.Content}");

This example uses a local embedding model hosted by (LM Studio) and an in-memory vector store.

⚙️ Architecture.

Every RAG pipeline in RAGSharp follows this flow. Each part is pluggable:

   [ DocumentLoader ] → [ TextSplitter ] → [ Embeddings ] → [ VectorStore ] → [ Retriever ]

That’s it — the essential building blocks for RAG, without the noise.

🔌 Extensibility

Built on simple interfaces.

For Embeddings:

public interface IEmbeddingClient
{
    Task<float[]> GetEmbeddingAsync(string text);
    Task<IReadOnlyList<float[]>> GetEmbeddingsAsync(IEnumerable<string> texts);
}

For Vector Store:

public interface IVectorStore
{
    Task AddAsync(VectorRecord item);
    Task AddBatchAsync(IEnumerable<VectorRecord> items);
    Task<IReadOnlyList<SearchResult>> SearchAsync(float[] query, int topK);
    bool Contains(string id);
}

No vendor lock-in. Want Claude or Gemini embeddings? Cohere? A Postgres or Qdrant backend? Just implement the interface and drop it in.

📚 Core Components

1. Document Loading (RAGSharp.IO)

Loaders fetch raw data and convert it into a list of Document objects.

Custom loaders: Implement IDocumentLoader for PDFs, Word docs, databases, etc.

2. Embeddings & Tokenization (RAGSharp.Embeddings)

3. Text Splitting (RAGSharp.Text)

RecursiveTextSplitter - Token-aware recursive splitter that respects semantic boundaries:

var tokenizer = new SharpTokenTokenizer("gpt-4");
var splitter = new RecursiveTextSplitter(
    tokenizer, 
    chunkSize: 512,    // tokens
    chunkOverlap: 50
);

Here’s how it works:

- Input Text
    ↓
Split by paragraphs (\n\n)
    ↓
For each paragraph:
    ├─ Fits in chunk size? → Yield whole paragraph
    └─ Too large? → Split by sentences
                      ↓
                For each sentence:
                    ├─ Buffer + sentence fits? → Add to buffer
                    └─ Too large?
                         ├─ Yield buffer
                         └─ Single sentence too large? → Token window split
                                                       └─ Sliding window with overlap

• Splits by paragraphs first
• Falls back to sentences if chunks too large
• Final fallback: sliding token window with overlap
• Uses SharpTokenTokenizer for token counting

Implement ITextSplitter for custom splitting strategies.

4. Vector Stores (RAGSharp.Stores)

Why no built-in database support?

• Many use cases don't need a database (small knowledge bases, temporary searches)
• Database choices vary widely (Postgres, Qdrant, Redis, SQLite etc)
• Implement IVectorStore in ~50 lines for your preferred database
• Keeps core library lightweight

5. Retriever (RAGSharp.RAG)

The RagRetriever orchestrates the entire pipeline:

• Accepts new documents.
• Chunks them using ITextSplitter.
• Sends chunks to the IEmbeddingClient.
• Stores the resulting vectors in the IVectorStore.
• Performs semantic search (dot product/cosine similarity) against the store based on a query.
• Optional: Pass an ILogger to track ingestion progress, by defualt it uses ConsoleLogger

📚 Examples & Documentation

Clone the repo and run the sample app.

git clone https://github.com/mrrazor22/ragsharp
cd ragsharp/SampleApp
dotnet run

📂 Project Structure

RAGSharp/
├── Embeddings/        (IEmbeddingClient, ITokenizer, providers)
├── IO/                (FileLoader, DirectoryLoader, UrlLoader, WebSearchLoader)
├── RAG/               (RagRetriever, IRagRetriever)
├── Stores/            (InMemoryVectorStore, FileVectorStore)
├── Text/              (RecursiveTextSplitter, ITextSplitter)
├── Logging/           (ConsoleLogger)
└── Utils/             (HashingHelper, VectorExtensions)

SampleApp/
├── data/              (sample.txt, documents/)
└── Examples/          (Example1..Example5)

📜 License

MIT License.

💡 Status

Early stage, suitable for prototypes and production apps where simplicity matters.