🔗 GraphRAG

A Graph-enhanced Retrieval Augmented Generation system for analyzing Tech Business News

---

📖 Overview

GraphRAG combines the power of knowledge graphs with vector embeddings to provide context-rich, accurate responses about tech business news. Unlike traditional RAG systems that rely solely on semantic similarity, GraphRAG leverages entity relationships to understand the broader context of your queries.

✨ Key Features

• Hybrid Retrieval — Combines vector search (Qdrant) with graph traversal (Neo4j) using RRF fusion
• Smart Entity Extraction — Automatically extracts companies, people, products, and events from documents
• 2-Hop Graph Context — Retrieves related entities for comprehensive context
• Intelligent Reranking — Uses Cohere reranker with Gemini fallback for optimal results
• PDF Processing — Native support for PDF document ingestion

---

🏗️ Architecture

flowchart TB
    subgraph Input
        PDF[📄 PDF Documents]
    end
    
    subgraph Processing
        Loader[PDF Loader] --> Chunker[Text Splitter]
        Chunker --> Extractor[Entity Extractor]
        Chunker --> Embedder[Gemini Embeddings]
    end
    
    subgraph Storage
        Extractor --> |Entities & Relations| Neo4j[(Neo4j)]
        Embedder --> |Vectors| Qdrant[(Qdrant)]
    end
    
    subgraph Retrieval
        Query[🔍 User Query] --> VecSearch[Vector Search]
        Query --> GraphSearch[Graph Traversal]
        VecSearch --> Qdrant
        GraphSearch --> Neo4j
        Qdrant --> |Semantic Matches| Fusion[RRF Fusion]
        Neo4j --> |2-Hop Context| Fusion
        Fusion --> Rerank[Reranker]
        Rerank --> Response[📝 Response]
    end
    
    PDF --> Loader

Loading

---

📂 Project Structure

GraphRAG/
├── 📄 docker-compose.yml     # Neo4j & Qdrant infrastructure
├── 📄 requirements.txt       # Python dependencies
├── 📄 pyproject.toml         # Project metadata
├── 📄 .env.example           # Environment template
├── 📁 src/
│   ├── config.py             # Settings management
│   ├── 📁 schema/            # Graph schema definitions
│   ├── 📁 extraction/        # PDF loading & entity extraction
│   ├── 📁 storage/           # Neo4j & Qdrant clients
│   └── 📁 retrieval/         # Hybrid retriever
├── 📁 scripts/
│   └── ingest_documents.py   # Document ingestion script
├── 📁 tests/                 # Test suite
└── 📁 data/sample/           # Sample documents

---

🚀 Quick Start

Prerequisites

• Python 3.11+
• Docker & Docker Compose
• Google Gemini API Key

1. Clone & Setup

git clone <repository-url>
cd GraphRAG

2. Start Infrastructure

docker-compose up -d

This starts:

Service	Port	Purpose
Neo4j Browser	7474	Web UI
Neo4j Bolt	7687	Driver connection
Qdrant HTTP	6333	REST API
Qdrant gRPC	6334	gRPC API

3. Configure Environment

cp .env.example .env

Edit .env with your API keys:

# Required
GOOGLE_API_KEY=your_google_api_key_here

# Optional (for enhanced features)
OPENAI_API_KEY=your_openai_api_key_here
COHERE_API_KEY=your_cohere_api_key_here

4. Install Dependencies

pip install -r requirements.txt

5. Ingest Documents

python scripts/ingest_documents.py data/sample/

6. Query the System

from src.retrieval import HybridRetriever

retriever = HybridRetriever()
results = retriever.retrieve("What companies did Apple acquire?")

for result in results:
    print(result)

---

📊 Graph Schema

Node Types

Type	Properties
Company	name, ticker, industry, headquarters
Person	name, title, role
Product	name, category, launch_date
Event	name, date, type, location
Article	title, published_date, source

Relationships

Person  ──[LEADS|FOUNDED|WORKS_AT]──▶ Company
Company ──[ACQUIRED|INVESTED_IN|SUED_BY|PARTNERS_WITH|COMPETES_WITH]──▶ Company
Company ──[LAUNCHED]──▶ Product
*       ──[MENTIONED_IN]──▶ Article

---

⚙️ Configuration

Variable	Default	Description
GOOGLE_API_KEY	required	Gemini API key
OPENAI_API_KEY	optional	For GPT-4o extraction
NEO4J_URI	bolt://localhost:7687	Neo4j connection
NEO4J_USERNAME	neo4j	Neo4j username
NEO4J_PASSWORD	graphrag_password	Neo4j password
QDRANT_HOST	localhost	Qdrant host
QDRANT_PORT	6333	Qdrant port
COHERE_API_KEY	optional	For reranking
GEMINI_MODEL	models/gemini-2.0-flash	LLM model
CHUNK_SIZE	1024	Text chunk size
CHUNK_OVERLAP	128	Chunk overlap

---

🧪 Testing

# Run all tests
pytest

# Run with verbose output
pytest -v

# Skip integration tests
pytest -m "not integration"

---

📚 Tech Stack

Component	Technology
LLM	Google Gemini 2.0 Flash
Embeddings	Gemini text-embedding-004
Graph Store	Neo4j 5.15
Vector Store	Qdrant 1.7
Framework	LlamaIndex 0.10+
Reranking	Cohere (primary), Gemini (fallback)

---

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

---

📄 License

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

---

Built with ❤️ using LlamaIndex, Neo4j, and Qdrant