RAG.io - Retrieval-Augmented Generation

Enterprise-grade RAG platform with multi-provider LLM support, document intelligence, and real-time streaming

📋 Table of Contents

• Overview
• Demo
• Features
• Architecture
• Tech Stack
• Installation
• Configuration
• API Documentation
• Usage Guide
• Development
• Contributing
• License

---

Overview

RAG Multi-Expert is a production-ready Retrieval-Augmented Generation platform that combines the power of vector databases, intelligent document processing, and multi-provider LLM integration. Built for enterprises and developers who need:

• Intelligent Document Search: ChromaDB-powered semantic search with adjustable retrieval parameters
• Multi-Provider LLM Support: Seamless integration with OpenAI, Claude, Gemini, Ollama, and 10+ providers
• Project-Based Organization: Isolate document collections and conversations by project
• Real-Time Streaming: Server-Sent Events for progressive responses
• Fine-Grained Control: Per-conversation temperature, top-k, and context window management
• Enterprise Security: JWT authentication, AES-256 encryption, GDPR compliance

Key Differentiators

Feature	Traditional RAG	This Platform
Context Window	Fixed 4K-8K	Adaptive 8K-2M (auto-detects model)
Chunking	Static 512 tokens	Smart chunking with overlap (100-2000 tokens)
History Management	No truncation	Intelligent truncation with recency bias
Provider Support	Single provider	13+ providers with fallback
Real-Time UX	Blocking requests	SSE streaming with auto-reconnect

---

Demo

1. RAG Chat Interface

*Real-time streaming with source attribution and intelligent context management*

2. Document Upload & Processing

*Repo/Manual upload with automatic chunking and vectorization*

3. Project Dashboard

*Multi-project organization with statistics and document management*

4. Provider Configuration

*Graphical configuration for 13+ LLM providers*

---

Features

Core RAG Capabilities

Document Processing

• Supported Formats: PDF, DOCX, TXT, MD, HTML, CSV, JSON (50+ file types)
• Smart Chunking: Adaptive chunk size (100-2000 tokens) with configurable overlap
• Metadata Extraction: Automatic filename, page number, and document type tagging
• Token Tracking: Real-time token counting for cost estimation
• Batch Processing: Background async processing with progress tracking

Semantic Search

• Vector Database: ChromaDB with HNSW indexing
• Embedding Models: sentence-transformers/all-MiniLM-L6-v2 (default), OpenAI embeddings
• Adjustable top-k: Dynamic retrieval (1-20 chunks) based on model context
• Distance Scoring: Cosine similarity with configurable threshold
• Metadata Filtering: Filter by document type, date, or custom tags

Multi-Provider LLM Support

Provider	Models	Context Window	Streaming	Temperature
OpenAI	GPT-4, GPT-4-turbo, o1-preview	8K-128K	✅	0.0-2.0
Anthropic	Claude 3.5 Sonnet, Opus, Haiku	200K	✅	0.0-1.0
Google	Gemini 1.5 Pro/Flash, 2.0	2M	✅	0.0-2.0
Ollama	Llama 3.1, Mistral, Phi-3	8K-128K	✅	0.0-2.0
Groq	Llama 3, Mixtral	32K	✅	0.0-2.0
OpenRouter	200+ models	Varies	✅	0.0-2.0
HuggingFace	Custom models	Varies	✅	0.0-2.0

Features:

• ✅ Auto-detect model context limits
• ✅ Dynamic top-k adjustment (more chunks for large context models)
• ✅ Intelligent history truncation (keeps recent messages + summary)
• ✅ RAG context capped at 50% of total context
• ✅ Real-time token counting with overflow protection

💬 Conversation Management

• Multi-Conversation: Unlimited conversations per project
• History Storage: PostgreSQL with full conversation replay
• Smart Truncation: Keeps recent messages, summarizes older ones
• Export: JSON, Markdown, HTML formats
• Search: Full-text search across all conversations

Advanced Features

🎛️ Fine-Grained Control

• Temperature: Per-conversation creativity control (0.0-2.0)
• Top-k: Adjustable chunk retrieval (1-20)
• Chunk Size: Configurable per project (100-2000 tokens)
• Overlap: Smart overlap to preserve context (10-500 tokens)

Analytics & Monitoring

• Real-time token usage tracking
• Per-provider cost estimation
• Document processing statistics
• Conversation metrics (latency, token count)

Security & Privacy

• JWT-based authentication
• AES-256 API key encryption
• CORS protection
• Rate limiting (60 req/min)
• GDPR-compliant data storage

---

Tech Stack

Backend

• Framework: FastAPI 0.104+ (async, type-safe)
• Vector DB: ChromaDB 0.4+ (persistent, HNSW indexing)
• Database: PostgreSQL 15+ (production) / SQLite (dev)
• ORM: SQLAlchemy 2.0+ (async support)
• Embeddings: sentence-transformers (all-MiniLM-L6-v2)
• Document Processing: PyPDF2, python-docx, beautifulsoup4
• Auth: JWT (PyJWT), AES-256 (cryptography)
• Streaming: SSE (Server-Sent Events)

Frontend

• Framework: React 18.2 (Vite)
• State Management: Zustand
• Routing: React Router v6
• UI: Tailwind CSS 3.4
• Icons: Lucide React
• HTTP: Axios with interceptors
• Markdown: react-markdown, react-syntax-highlighter

Infrastructure

• Containerization: Docker + Docker Compose
• Reverse Proxy: Nginx (production)
• Deployment: Single docker-compose up

---

Installation

Prerequisites

# Required
- Docker 24.0+
- Docker Compose 2.20+
- 8GB RAM minimum (16GB recommended)
- 10GB disk space

# Optional (for local development)
- Python 3.11+
- Node.js 18+
- PostgreSQL 15+

Quick start with Kubernetes (Production-ready)

Helm Chart + Documentation

📚 Documentation Kubernetes officiel →

https://iwebbo.github.io/RAG.io/

Quick Start (Docker)

# 1. Clone repository
git clone https://github.com/iwebbo/rag.io.git
cd rag.io

# 2. Copy config files
cp backend/.env.example backend/.env
cp frontend/.env.example frontend/.env

# 3. Generate secrets
python3 -c "import secrets; print(secrets.token_hex(32))"  # SECRET_KEY
python3 -c "from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())"  # ENCRYPTION_KEY

# 4. Edit backend/.env with your secrets

# 4.1 Edit backend/.env if (need to be run from VM/PROD Server)
# Change by your hostname.fqdn or IP
CORS_ORIGINS=http://localhost:5173,http://localhost:3000,http://localhost,http://localhost:80
# Change by your hostname.fqdn or IP

# 4.1 Edit frontend/.env if (need to be run from VM/PROD Server)
# Change by your hostname.fqdn or IP
VITE_API_URL=http://localhost:8000 

# Will be solve in 1.1.0
cd backend/
mkdir -p /app/data/chromadb
chmod -R 777 /app/data

# 5. Start application
docker-compose up -d

# 6. Check status
docker-compose ps

# 4. Access application
# Frontend: http://localhost:3000
# Backend: http://localhost:8000
# API Docs: http://localhost:8000/docs

First-time setup:

# Create first user
curl -X POST http://localhost:8000/api/auth/register \
  -H "Content-Type: application/json" \
  -d '{"username":"admin","email":"admin@example.com","password":"SecurePass123!"}'

# Login
curl -X POST http://localhost:8000/api/auth/login \
  -d "username=admin&password=SecurePass123!"

Manual Installation (Development)

Backend

cd backend

# Create virtual environment
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Setup database
python -c "from app.database import init_db; init_db()"

# Run development server
uvicorn app.main:app --reload --host 0.0.0.0 --port 8000

Frontend

cd frontend

# Install dependencies
npm install

# Configure API URL
echo "VITE_API_URL=http://localhost:8000" > .env.local

# Run development server
npm run dev

---

API Documentation

API Docs: http://localhost:8000/docs

Usage Guide

1. Create Your First Project

# Via API
curl -X POST http://localhost:8000/api/projects/ \
  -H "Authorization: Bearer {token}" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Company Docs",
    "description": "Internal documentation",
    "chunk_size": 500,
    "chunk_overlap": 50
  }'

Or via UI: Projects → New Project

Upload Documents

# Via API
curl -X POST http://localhost:8000/api/documents/projects/{project_id}/upload \
  -H "Authorization: Bearer {token}" \
  -F "file=@employee_handbook.pdf"

Or via UI: Project → Documents → Upload

Processing Status:

• processing: Document is being chunked and vectorized
• completed: Ready for RAG queries
• failed: Check error_message field

Monitor Performance

Check Projects → Statistics for:

• Document processing status
• Total chunks and tokens
• Conversation metrics
• Provider usage statistics

---

License

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

Built with ❤️ for Community

⭐ Star us on GitHub if you find this useful!