PDF Knowledge Extraction & Q&A System

A simplified PDF knowledge extraction and Q&A chatbot system for college demo. Extracts knowledge from pre-ingested company PDFs, builds a Neo4j knowledge graph with vector embeddings, and provides a Streamlit chat interface with grounded citations.

Key Features

• Offline PDF Processing: Pre-ingest PDFs (24 US federal documents included)
• Layout-Aware Extraction: Uses PyMuPDF for text extraction with bounding boxes
• Knowledge Graph: Neo4j with vector embeddings (local sentence-transformers)
• Hybrid Retrieval: Vector similarity + graph traversal + full-text search
• Local Embeddings: No API calls for embeddings - instant, free, and private
• Grounded Answers: Top 5 citations with PDF screenshots and bbox highlights
• Simple UI: Streamlit chat interface with example questions

Project Structure

pdf-knowledge-system/
├── data/
│   ├── company_1/
│   │   ├── pdfs/                # 24 sample PDFs (US federal documents)
│   │   └── images/              # 61 page screenshots (300 DPI)
│   ├── processed/company_1/     # 24 extracted JSON files
│   └── companies.yaml           # Company metadata
├── scripts/
│   ├── extract_simple.py        # PDF extraction (PyMuPDF)
│   ├── build_knowledge_graph.py # Neo4j population
│   └── test_query.py           # Test RAG system
├── backend/                     # FastAPI REST API
├── frontend/                    # Streamlit chat UI
├── neo4j_data/                  # Neo4j database (529MB, persisted)
└── docker-compose.yml           # Neo4j container setup

Quick Start

1. Prerequisites

• Python 3.12+
• Docker & Docker Compose
• Google Gemini API key (get one here)
• CUDA GPU (optional, for faster local embeddings)

2. Installation

# Clone repository
cd pdf-knowledge-system

# Create virtual environment (recommended)
python3 -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r backend/requirements.txt
pip install streamlit requests pillow

# Setup environment
cp .env.example .env
nano .env  # Add your GEMINI_API_KEY

# Start Neo4j
docker compose up -d

# Wait 15 seconds for Neo4j to initialize
sleep 15

3. Download Pre-built Database (Recommended)

Option A: Use Pre-populated Database

1. Download neo4j_database_backup.tar.gz (20MB) from shared location
2. Extract: tar -xzf neo4j_database_backup.tar.gz
3. Restart Neo4j: docker compose restart neo4j
4. Skip to step 5 (testing)

Option B: Build From Scratch (30-40 minutes)

# PDFs already included in data/company_1/pdfs/ (24 files)

# Step 1: Extract PDFs (2-3 minutes)
python3 scripts/extract_simple.py

# Step 2: Build knowledge graph (30-35 minutes due to Gemini API limits)
python3 scripts/build_knowledge_graph.py

Database stats:

• 3,239 nodes (2,095 Entities, 1,061 Chunks, 58 Pages, 24 Documents)
• 4,941 relationships
• 1,061 chunks with 768-dim embeddings (local, no API)

4. Test the System

# Quick test without starting servers
python3 scripts/test_query.py "What happened to NSF's Total Liabilities in FY 2011?"

5. Start the System

# Terminal 1: Start FastAPI backend
cd backend
uvicorn app.main:app --reload --host 0.0.0.0 --port 8000 > ../backend.log 2>&1 &

# Terminal 2: Start Streamlit frontend  
cd frontend
streamlit run streamlit_app.py --server.port 8501 > ../frontend.log 2>&1 &

Access:

• Streamlit UI: http://localhost:8501 ← Main interface
• FastAPI docs: http://localhost:8000/docs
• Neo4j Browser: http://localhost:7474

Stop all services:

pkill -f uvicorn && pkill -f streamlit && docker compose down

Restart:

docker compose up -d && sleep 15
# Then start backend and frontend as above

Testing the System

Example Questions

Try these in the Streamlit UI (based on actual documents):

NSF Financial Information:

• "What happened to NSF's Total Liabilities in FY 2011?"
• "Summarize NSF's financial position"

Nanotechnology Research:

• "What is the NNI's approach to addressing ethical questions?"
• "What are nanotechnology's potential environmental impacts?"

Accessibility Regulations:

• "Where must accessible parking spaces be located?"
• "What are the accessibility requirements for entrances?"

Verify System Health

# Test RAG system directly
python3 scripts/test_query.py "What happened to NSF's Total Liabilities in FY 2011?"

# Check Neo4j data (in browser at http://localhost:7474)
MATCH (n) RETURN labels(n)[0] AS type, count(n) AS count ORDER BY count DESC;

# Verify embeddings
MATCH (c:Chunk) WHERE c.embedding IS NOT NULL RETURN count(c);

# Check processed files
ls -lh data/processed/company_1/  # Should show 24 JSON files
ls -lh data/company_1/images/     # Should show 61 PNG files

Architecture

Tech Stack

Extraction:

• Docling - IBM's AI PDF text and layout extraction
• PyMuPDF (fitz) - PDF text extraction with bboxes
• pdf2image + poppler - 300 DPI screenshots

Knowledge Graph:

• Neo4j 5.15 - Graph database with vector index
• sentence-transformers (all-mpnet-base-v2) - Local 768-dim embeddings
• Google Gemini (gemini-2.5-flash) - Triplet extraction only

Backend:

• FastAPI - REST API
• Neo4j Python driver - Database queries
• Local embeddings - Zero API calls for retrieval

Frontend:

• Streamlit - Chat UI with citations
• PIL - Image display with bbox overlays

API Endpoints

• POST /query - Ask questions, get grounded answers with citations
• GET /companies - List available companies

Component Documentation

Scripts

• scripts/extract_simple.py - PyMuPDF extraction with bboxes
• scripts/extract_docling.py - extraction using Docling
• scripts/build_knowledge_graph.py - Neo4j population (local embeddings + Gemini triplets)
• scripts/test_query.py - Test RAG system without API
• See scripts/README.md for detailed usage

Backend

• backend/app/main.py - FastAPI application with CORS
• backend/app/rag_engine.py - Hybrid retrieval (vector + text + graph) + answer generation
• backend/app/config.py - Configuration (TOP_K_CHUNKS = 5)
• backend/app/models.py - Pydantic models with citations

Frontend

• frontend/streamlit_app.py - Chat UI with:
  • Company selector
  • Example questions sidebar
  • Citation display with screenshots
  • Bounding box overlays (scaled for 300 DPI)

Configuration

Environment Variables

GEMINI_API_KEY=your-api-key        # Required: Gemini API key
NEO4J_URI=bolt://localhost:7687    # Neo4j connection
NEO4J_USER=<DB_NAME>                   # Neo4j username
NEO4J_PASSWORD=<DB_PASSWORD>         # Neo4j password

Company Metadata

Edit data/companies.yaml to customize company names and descriptions.

System Stats

Current Dataset:

• PDFs: 24 US federal documents (company_1)
• Pages: 58 total pages
• Screenshots: 61 PNG files at 300 DPI
• Neo4j Nodes: 3,239 (2,095 Entities, 1,061 Chunks, 58 Pages, 24 Documents, 1 Company)
• Relationships: 4,941
• Database Size: 529MB (persisted in neo4j_data/)

Configuration:

• Chunk size: 512 tokens with 50-token overlap
• Embeddings: 768-dim (sentence-transformers all-mpnet-base-v2, local)
• Retrieval: Top 5 most relevant chunks per query
• Gemini usage: Only for triplet extraction (graph building) and answer generation
• API calls per query: 1 (answer generation only, embeddings are local)

Additional Documentation

• scripts/README.md - Extraction and graph building details
• backend/README.md - Backend API documentation
• frontend/README.md - Frontend UI details