Multi-Mode Document Intelligence

A Flask web application for comparing five distinct retrieval strategies using a shared OpenAI-compatible LLM backend. Features a minimal white & coffee design with serif typography.

Features

• Mode 1: Traditional RAG - Vector similarity search using ChromaDB with local embeddings
• Mode 2: Agentic RAG - LLM-driven search with reasoning and expansion
• Mode 3: Vectorless RAG - PageIndex tree structure with human-like navigation
• Mode 4: Graph RAG - Knowledge graph with entity extraction and graph traversal
• Mode 5: Hybrid Keyword RAG - BM25 + semantic search with cross-encoder re-ranking

UI Features

• Minimal Design - Clean white background with coffee (#6F4E37) accents
• Serif Typography - Georgia/Times New Roman throughout
• Fixed Layout - Viewport-locked design with independent scrolling
• Markdown Rendering - Assistant responses formatted as rich Markdown
• Processing Indicators - Animated status messages during AI processing
• Real-time Metrics - Token usage, retrieval time, and latency tracking

Quick Start

Prerequisites

• Python 3.10+
• pip

Installation

1. Clone the repository

   git clone <repository-url>
   cd RAGPro

2. Install Dependencies

   pip install -r requirements.txt

   For local embeddings, this will also download the sentence-transformer model (~80MB) on first use.

3. Configure Environment

   cp .env.example .env
   # Edit .env with your API credentials

4. Run the Application

   python app.py

5. Open Browser Navigate to http://localhost:5000

Configuration

Create a .env file in the project root:

# OpenAI-compatible API Configuration
OPENAI_API_BASE=https://api.openai.com/v1
OPENAI_API_KEY=your_openai_api_key_here

# Model Configuration
MODEL_ID=gpt-4o-mini

Supported Backends

• OpenAI - Official OpenAI API
• Ollama - Local LLMs (http://localhost:11434/v1)
• Groq - Fast inference API
• Any OpenAI-compatible API

Project Structure

RAGPro/
├── app.py                      # Flask application entry point
├── requirements.txt            # Python dependencies
├── .env.example               # Environment template
├── .gitignore                 # Git ignore rules
├── README.md                  # This file
│
├── engine/                    # RAG engine implementations
│   ├── traditional.py         # Mode 1: ChromaDB with local embeddings
│   ├── agentic.py             # Mode 2: Agentic RAG with LLM-driven search
│   ├── vectorless.py          # Mode 3: PageIndex tree structure
│   ├── graph.py               # Mode 4: Knowledge graph with entity extraction
│   └── hybrid_keyword.py      # Mode 5: BM25 + semantic with cross-encoder
│
├── utils/                     # Utility modules
│   ├── config.py              # Environment configuration loader
│   ├── metrics.py             # Query metrics logging
│   └── text_utils.py          # Text processing utilities
│
├── templates/                 # Flask HTML templates
│   └── index.html             # Main application page
│
├── static/                    # Static assets
│   ├── css/
│   │   └── style.css          # White & coffee styling
│   └── js/
│       └── app.js             # Frontend JavaScript
│
└── uploads/                   # Temporary upload directory (auto-created)

Usage

1. Select a Mode from the sidebar

   • Traditional RAG: Fast vector similarity search
   • Agentic RAG: Adaptive multi-step reasoning
   • Vectorless RAG: Hierarchical tree navigation
   • Graph RAG: Knowledge graph exploration with entity linking
   • Hybrid Keyword RAG: Combined BM25 and semantic with re-ranking

2. Upload a Document (PDF, Markdown, or TXT)

   • Drag & drop or click to select
   • Click "Index Document" to process

3. Chat with Your Document

   • Type questions in the input field
   • View processing status indicators
   • Expand metrics and reasoning traces for details

4. Manage Documents

   • See indexed documents in the sidebar
   • Clear all documents with one click
   • Documents persist across sessions in ChromaDB

Architecture

Traditional RAG

• Uses ChromaDB vector store with cosine similarity
• Local sentence-transformer embeddings (all-MiniLM-L6-v2)
• No API calls for embeddings - fully local
• Simple and fast retrieval

Agentic RAG

• LLM decides search strategy dynamically
• Can expand search with neighboring chunks
• Multiple reasoning steps with trace
• More adaptive but requires more tokens

Vectorless RAG (PageIndex)

• Documents converted to hierarchical tree
• LLM navigates like a human (ToC → Sections → Content)
• No embeddings or vector DB needed
• Fully explainable retrieval paths

Graph RAG

• LLM extracts entities and relationships from documents
• Builds knowledge graph using NetworkX
• Entity-aware search with graph traversal
• Explains reasoning through entity connections

Hybrid Keyword RAG

• Combines semantic embeddings with BM25 keyword search
• Uses spaCy for NLP feature extraction (NER, keywords)
• Reciprocal Rank Fusion (RRF) merges search results
• Cross-Encoder re-ranking for improved relevance

API Endpoints

Endpoint	Method	Description
/	GET	Main application page
/api/status	GET	Get document status for all modes
/api/upload	POST	Upload and index a document
/api/chat	POST	Send chat query
/api/clear	POST	Clear indexed documents
/api/metrics	GET	Get query metrics history

Metrics Tracked

• Total Tokens - Prompt + Completion token count
• Retrieval Time - Time to find relevant context
• Total Latency - End-to-end response time
• Reasoning Trace - Step-by-step agent decisions (Modes 2 & 3)

Development

Running in Debug Mode

FLASK_DEBUG=1 python app.py

Clearing Vector Store

To clear all indexed documents:

1. Click "Clear All Documents" button in sidebar, or
2. Delete the chroma_db/ directory and restart

Troubleshooting

Embeddings Model Not Loading

The sentence-transformer model downloads automatically on first use. If you have issues:

python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('all-MiniLM-L6-v2')"

API Connection Errors

• Verify your .env file has correct OPENAI_API_BASE and OPENAI_API_KEY
• For Ollama, ensure the server is running (ollama serve)
• Check that the model name in MODEL_ID exists on your API

Port Already in Use

Change the port in app.py:

app.run(debug=True, host='0.0.0.0', port=5001)  # Change to 5001 or any available port

License

MIT License

Credits

• PageIndex inspiration from VectifyAI
• Sentence Transformers for local embeddings
• ChromaDB for vector storage
• Flask for web framework