Graph RAG: From Notes to Knowledge Graphs

This project turns unstructured notes into a queryable knowledge graph using LLMs. Instead of retrieving text chunks based on vector similarity, it leverages graph traversal and entity relationships to support multi-hop reasoning.

The system includes:

• A FastAPI backend for document ingestion and graph-based retrieval
• A React frontend with D3-based graph visualization
• A chat interface grounded entirely in knowledge graph context

Demo

graph-whisperer-react-app-GoogleChrome2025-06-1600-57-52-ezgif.com-video-speed.mp4

Features

• Upload and parse raw text or notes
• Extract (Entity A) —[Relation]→ (Entity B) triples using LLM prompts
• Construct a live, visual knowledge graph
• Query the graph using a chat interface (no vector database required)
• Modular API-first backend designed for reuse in agents or bots

Tech Stack

Layer	Technology
Frontend	React, TailwindCSS
Backend	FastAPI (Python)
RAG Engine	LlamaIndex (Graph RAG)
Visualization	D3.js
Future Plans	Neo4j, spaCy, Hybrid RAG

Quick Start

1. Clone the Repository

git clone https://github.com/Sarita-Joshi/note-to-knowledge.git
cd note-to-knowledge

2. Create and Activate a Virtual Environment

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install Backend Dependencies

pip install -r backend/requirements.txt

4. Run the Backend (FastAPI)

python app.py

5. Start the Frontend

cd frontend
npm install
npm run dev

Environment Variables

Create a .env file in the backend root directory with the important keys like

OPENAI_API_KEY=your_openai_key
GEMINI_API_KEY=your_gemini_key
GROQ_API_KEY=your_groq_key
NEO4J_URI=bolt://localhost:7687
NEO4J_USER=neo4j
NEO4J_PASSWORD=your_password

API Endpoints

See app.py for the following endpoints:

• POST /upload — Ingest text or notes
• GET /graph — Retrieve graph nodes and edges
• GET /chat — Query graph via LLM-backed reasoning

Auto-generated Swagger docs coming soon.

Known Issues

Async Handling in LlamaIndex

Some LlamaIndex methods are asynchronous internally, which may cause issues with FastAPI's sync/async routing.

Common error:

RuntimeWarning: coroutine '...' was never awaited

Workarounds attempted:

• Wrapping calls with asyncio.run()
• Background task queues
• Refactoring into sync-friendly wrappers

Contributions welcome if you’ve solved this more cleanly.

Roadmap

• Neo4j support for persistent graph storage
• Entity resolution via spaCy or sentence embeddings
• Hybrid Graph + Vector RAG
• Feedback loop to validate or revise extracted triples
• Graph memory integration for agent workflows

Contribution

Contributions are welcome. To get started:

git checkout -b feature-name
# make your changes
git commit -m "Add feature"
git push origin feature-name

Then open a pull request.

Source: https://github.com/Sarita-Joshi/note-to-knowledge

License

MIT License

Author

Built as part of a personal Mini Project Series focused on building E2E Applciations with AI, LLMs, and data-driven systems.