softrag 0.1.1

Minimal local-first Retrieval-Augmented Generation (RAG) library using SQLite with sqlite-vec

softrag

Minimal local-first Retrieval-Augmented Generation (RAG) library powered by SQLite + sqlite-vec.
Everything—documents, embeddings, cache—lives in a single .db file.

---

🌟 Features

• Local-first – All processing happens locally, no external services.
• SQLite + sqlite-vec – Documents, embeddings, and cache in a single .db file (no separate vector store or account needed).
• No cloud service dependency – Plug in any LLM backend; no forced API keys for the core storage layer.
• Blazing-fast – Designed for minimal overhead and maximum throughput on small- and medium-scale corpora.
• Perfect for small & medium use cases – Ideal when you need a lightweight, self-contained RAG solution.
• Configurable chunking – Default RecursiveCharacterTextSplitter (400/100) or your own strategy.
• Model-agnostic – Works with OpenAI, Hugging Face, Ollama, etc.
• Zero heavy deps – Core pulls only minimal extras (langchain-text-splitters optional).

📋 Requirements

• Python 3.12+
• Dependencies: sqlite-vec, trafilatura, pymupdf (for PDFs)
• Access to embedding models and LLMs (uses OpenAI by default)

🚀 Installation

pip install softrag

🔧 Basic Usage

from softrag import Rag
from langchain_openai import ChatOpenAI, OpenAIEmbeddings

chat  = ChatOpenAI(model="gpt-4o")
embed = OpenAIEmbeddings(model="text-embedding-3-small")

rag = Rag(embed_model=embed, chat_model=chat)      # default splitter (RCTS)

rag.add_file("document.pdf")
rag.add_web("https://example.com/page")

query = rag.query("What is the main information in this content?")

📚 Examples

See the examples/ folder for more detailed examples:

• simple.py: Basic example with OpenAI
• local.py: Example using local Transformers models

🔄 How It Works

SoftRAG uses a hybrid approach for retrieval:

1. Extraction: Content is extracted from documents and web pages
2. Splitting: Text is divided into smaller chunks
3. Indexing: Each chunk is indexed by text (SQLite FTS5) and vector embedding
4. Retrieval: Queries combine keyword search and vector similarity
5. Generation: The most relevant chunks are sent to the LLM along with the question

🤝 Contributing

Contributions are welcome! Please feel free to submit Pull Requests.

📜 License

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

Developed with ❤️ for AI community