arXiv Paper Chat

A beautiful, interactive web application for browsing AI research papers from arXiv and chatting with them using AI-powered PDF chatbots.

Versions Available

Local Edition (Recommended - Privacy First!)

Run everything on your own machine with complete privacy. Uses Ollama for local LLM inference and RAG for accurate answers.

• File: index.html - Local version with Python backend
• Privacy: All processing happens locally - no external API calls
• Cost: Completely free - no API costs
• Models: Use any Ollama model (llama3, mistral, etc.)

Cloud Edition

Uses HuggingFace Spaces for chatbot functionality (requires internet).

• File: index-cloud.html - Cloud version
• Convenience: No setup required, just open in browser
• Dependency: Requires HuggingFace services to be online

Features

Local Edition

• Complete Privacy: All data stays on your machine
• RAG-Powered: Advanced Retrieval-Augmented Generation for accurate answers
• Vector Embeddings: Uses ChromaDB for semantic search
• Source Citations: Shows relevant paper sections used for each answer
• Conversation History: Maintains context across questions
• PDF Processing: Automatic text extraction and intelligent chunking

Both Versions

• Browse Research Papers: Curated collection of cutting-edge AI research papers organized by category
  • LLM Reasoning & Training
  • AI Safety & Alignment
  • Vision-Language Models
• Select Papers: Choose from various influential papers with detailed descriptions
• Download PDFs: Direct links to arXiv papers
• Integrated Experience: Seamlessly download papers and chat with them

Setup Instructions

Local Edition Setup

Prerequisites:

• Python 3.8 or higher
• Ollama installed on your system

Step 1: Install Ollama

# macOS/Linux
curl -fsSL https://ollama.com/install.sh | sh

# Or download from https://ollama.com

Step 2: Pull an LLM Model

# Recommended: Llama 3 (4.7GB)
ollama pull llama3

# Alternatives:
# ollama pull mistral       # Smaller, faster
# ollama pull llama2        # Older but reliable
# ollama pull codellama     # Good for technical papers

Step 3: Install Python Dependencies

cd backend
pip install -r requirements.txt

Step 4: Start the Backend Server

cd backend
python server.py

The server will start on http://localhost:5000

Step 5: Open the Frontend

# In a new terminal, from the project root
python -m http.server 8000

Then open http://localhost:8000 in your browser.

Cloud Edition Setup

No setup required! Simply open index-cloud.html in your browser.

How to Use (Local Edition)

1. Start Backend: Make sure the Python backend is running (python backend/server.py)
2. Open Frontend: Navigate to http://localhost:8000 in your browser
3. Browse Papers: Explore the curated collection of AI research papers
4. Download PDF: Click "Download PDF" on any paper that interests you
5. Open Chat: Click "Open Chat Interface"
6. Upload PDF: Click "Upload PDF" and select the downloaded paper
7. Ask Questions: Start chatting! The AI will answer based on the paper content
8. View Sources: Expand source sections to see which parts of the paper were used

Technology Stack

Frontend

• React 18: Modern UI framework
• Tailwind CSS: Beautiful, responsive styling
• Babel Standalone: In-browser JSX compilation

Backend (Local Edition)

• Flask: Lightweight Python web framework
• Ollama: Local LLM inference
• LangChain: LLM application framework
• ChromaDB: Vector database for embeddings
• Sentence Transformers: Text embedding generation
• PyPDF2: PDF text extraction

Featured Papers

LLM Reasoning & Training

• LightReasoner - Training large models using small models
• BoostStep - Mathematical reasoning enhancement
• Impact of Reasoning Step Length
• Divide and Conquer strategies
• Reverse Curriculum RL

AI Safety & Alignment

• Comprehensive survey of alignment techniques
• Constitutional AI and DPO methods

Vision-Language Models

• AlignVLM - Bridging vision and language
• DeepSeek-VL2 - Mixture-of-Experts approach
• Vision-Language-Action Models
• Chart Understanding with MMC

API Endpoints (Local Backend)

The Flask backend provides the following REST API endpoints:

• GET /api/health - Health check
• GET /api/status - Get system status (PDF loaded, chunk count)
• POST /api/upload - Upload and process a PDF file
• POST /api/chat - Send a message and get AI response
• POST /api/clear - Clear conversation history
• POST /api/reset - Reset system (clear all data)

Architecture

┌─────────────────┐
│   Frontend      │  React SPA
│   (index.html)  │  - Browse papers
│                 │  - Upload PDF
└────────┬────────┘  - Chat interface
         │
         │ HTTP/REST
         ▼
┌─────────────────┐
│  Flask API      │  Python Backend
│  (server.py)    │  - PDF upload
└────────┬────────┘  - Request routing
         │
    ┌────┴────┐
    ▼         ▼
┌─────────┐ ┌──────────┐
│   PDF   │ │   RAG    │
│Processor│ │  Engine  │
└─────────┘ └────┬─────┘
                 │
        ┌────────┴────────┐
        ▼                 ▼
   ┌─────────┐      ┌─────────┐
   │ChromaDB │      │ Ollama  │
   │Vectors  │      │  LLM    │
   └─────────┘      └─────────┘

Customization

Change the LLM Model

Edit backend/rag_engine.py and change the model name:

rag_engine = RAGEngine(
    model_name="mistral",  # Change this
    embedding_model="all-MiniLM-L6-v2"
)

Adjust Chunk Size

Edit backend/pdf_processor.py:

processor = PDFProcessor(
    chunk_size=1500,  # Increase for more context
    chunk_overlap=300  # Increase for better continuity
)

Change Number of Retrieved Chunks

Edit backend/server.py in the chat endpoint:

response = rag_engine.chat(
    query=message,
    conversation_history=conversations[session_id],
    n_context=5  # Retrieve more chunks
)

Troubleshooting

Backend won't start

• Make sure all dependencies are installed: pip install -r backend/requirements.txt
• Check if port 5000 is available: lsof -i :5000
• Ensure Ollama is running: ollama list

"Ollama connection error"

• Start Ollama service: ollama serve
• Pull the model: ollama pull llama3
• Check Ollama is running: curl http://localhost:11434

Slow responses

• Try a smaller model: ollama pull mistral
• Reduce context chunks in server.py (change n_context to 2)
• Consider using GPU acceleration for Ollama

PDF upload fails

• Check file size (max 50MB)
• Ensure PDF is not password protected
• Try a different PDF

Performance Tips

1. GPU Acceleration: Ollama automatically uses GPU if available (NVIDIA, Apple Silicon)
2. Model Selection: Smaller models (mistral) are faster, larger models (llama3) are more accurate
3. Chunk Tuning: Balance between chunk_size (more context) and response time
4. Embeddings: The all-MiniLM-L6-v2 model is fast and lightweight. For better quality, try all-mpnet-base-v2

Deployment

Cloud Edition

Deploy the static HTML file anywhere:

• GitHub Pages: Just commit and enable GitHub Pages
• Netlify: Drag and drop index-cloud.html
• Vercel: Deploy with zero configuration

Local Edition

For production deployment:

• Use Gunicorn instead of Flask dev server
• Add nginx as reverse proxy
• Consider Docker for easy deployment
• Set up systemd service for auto-start

License

This project is open source and available for educational purposes.

Credits

• Local AI powered by Ollama
• Cloud chatbots powered by HuggingFace Spaces
• Papers sourced from arXiv
• Icons inspired by Lucide Icons
• RAG implementation using LangChain and ChromaDB