Decentralized Retrieval-Augmented Generation (DRAG)

This repository presents a decentralized extension of Retrieval-Augmented Generation (RAG), addressing privacy, scalability, and security issues of traditional RAG systems using IPFS, MQTT, and blockchain. DRAG enables users to interact with local knowledge bases and contribute to a global, shared database, promoting knowledge democratization and incentivizing participation.

DRAG Overview

DRAG improves RAG by decentralizing storage, communication, and processing, ensuring:

• Privacy: Secure, decentralized data storage with IPFS.
• Scalability: Global knowledge contributions without central authority.
• Security: Blockchain for transparency and safe record-keeping, with rewards for contributors.
• Collaborative Learning: Continuous improvement from distributed nodes.

Key Technologies:

• IPFS for decentralized storage.
• MQTT for low-latency communication.
• Blockchain for secure record-keeping.
• Qdrant for vector database storage and retrieval.

Traditional RAG vs. DRAG

The main difference between traditional RAG and DRAG is the decentralization. In DRAG, multiple nodes contribute knowledge, while traditional RAG relies on a centralized approach. Below is a visual comparison:

Traditional RAG: Centralized architecture with a single knowledge base.

DRAG: Decentralized architecture with multiple nodes contributing to a global knowledge base.

Architecture

The system consists of two types of nodes:

1. Data Nodes: Provide domain-specific knowledge to the global vector database.
2. Evaluator Nodes: Verify contributions and ensure the integrity of the global database.

The framework incorporates a blockchain-based reward system that incentivizes nodes for their contributions, improving both system accuracy and knowledge sharing.

Setup and Installation

Clone and Build

1. Clone the repository:

   git clone https://github.com/bayesianinstitute/Decentralized-RAG
   cd Decentralized-RAG 

2. Build the package:

   python setup.py sdist bdist_wheel
   pip install .

Running in Docker Container

Run all services:

docker compose up -d

Download the model and start the application:

bash run.sh

Setting up Qdrant (Vector Database)

1. Download Qdrant Image:

   docker pull qdrant/qdrant

2. Run Qdrant:

   docker run -d -p 6333:6333 -p 6334:6334 \
       -v ./qdrant_data:/qdrant/storage \
       qdrant/qdrant

   For Windows:

   docker run -d --name qdrant_container -p 6333:6333 -p 6334:6334 \
       -v C:/path/to/qdrant_data:/qdrant/storage \
       qdrant/qdrant:latest

Setting up Dependencies

1. Ollama: Install Ollama by following the instructions on Ollama's website.

2. Language Model: Pull a language model from the Ollama library:

   ollama pull llama3:8b

3. Text Embedding Model:

   ollama pull nomic-embed-text:latest

4. Install required Python libraries (e.g., qdrant-client).

Running the Application

1. Configure Node Type:
   Edit main.py to specify the node type:

   • admin: Institute Node (manages the global embedding)
   • data: Data Node (contributes specialized knowledge)

2. Start the Application:

   python main.py --data-dir data --nodetype admin

   Replace data with the desired data directory. Set --nodetype to either admin or data.

IPFS Installation

To set up IPFS for decentralized storage, follow the installation instructions from IPFS Documentation or use the following commands for Windows:

wget https://dist.ipfs.tech/kubo/v0.23.0/kubo_v0.23.0_windows-amd64.zip -Outfile kubo_v0.23.0.zip
Expand-Archive -Path kubo_v0.23.0.zip -DestinationPath .\kubo
cd .\kubo
.\install.bat

Conclusion

This framework showcases the potential of decentralized systems in enhancing RAG applications by reducing retrieval errors, preserving privacy, and fostering continuous, collaborative learning. By leveraging blockchain technology, it ensures secure and transparent record-keeping, while rewarding participants for their contributions. For a complete implementation, visit our GitHub repository: Decentralized RAG GitHub.

References

• Ollama Docker Hub: https://hub.docker.com/r/ollama/ollama
• IPFS Documentation: https://docs.ipfs.tech