One Open Source Project a Day (Part 61): knowledge_graph - Turn Any Text into a Knowledge Graph

Introduction

"Convert any text to a graph of knowledge. Graph Retrieval Augmented Generation (GRAG) — a new and improved version of RAG."

This is Part 61 of the "One Open Source Project a Day" series. Today's project is knowledge_graph (GitHub).

Want to turn documents and PDFs into queryable, visualizable knowledge graphs? knowledge_graph uses an LLM to extract concepts (not entities), build nodes and edges, and supports GRAG (Graph Retrieval Augmented Generation), centrality analysis, and community detection. No-GPT approach: local Mistral 7B + Ollama, zero API cost, Docker one-click run.

What You'll Learn

Knowledge graph construction pipeline
Concepts vs entities
Dual weights: W1 (semantic) and W2 (contextual proximity)
GRAG vs RAG
Local Ollama + Docker quick start

Prerequisites

Basic understanding of RAG and knowledge graphs
Docker or local Ollama (optional)

Project Background

Project Overview

knowledge_graph is rahulnyk's open-source text-to-knowledge-graph project. From PDFs and other text corpora, it uses an LLM to extract concepts (not entities): e.g., "Bangalore" is an entity, "Pleasant weather in Bangalore" is a concept. Concepts co-occurring in the same text chunk are treated as related. The resulting graph supports GRAG, centrality, community clustering, and more.

Author

Author: rahulnyk
Philosophy: Concepts are more meaningful than entities; local LLM = zero cost; extensible to graph databases

Project Data

⭐ GitHub Stars: 3.1k+
🍴 Forks: 491+
📄 License: MIT
🌐 Sample graph: rahulnyk.github.io/knowledge_graph
📝 Method details: Medium article

Core Features

Pipeline

Chunking: Split text into chunks, assign chunk_id
Extract concepts & relations: LLM extracts concepts and semantic relations per chunk, weight W1
Contextual proximity: Concepts in the same chunk are related, weight W2
Merge edges: Merge weights and concatenate relations for same concept pairs
Degree & communities: Node degree (sizing), communities (coloring)

Concepts vs Entities

Type	Example	Description
Entity	Bangalore, doctor	Concrete objects, names, places
Concept	Pleasant weather in Bangalore, doctor–patient relationship	Situations, relations, abstractions
This project	Extracts concepts	Author: concepts yield more meaningful KG

Use Cases

GRAG: Use graph as retriever for deeper document Q&A
Centrality: Identify most important concepts in text
Community detection: Cluster concepts, analyze topic structure
Visualization: Pyvis for web-hostable interactive graphs

Quick Start

Docker (recommended):

git clone https://github.com/rahulnyk/knowledge_graph.git
cd knowledge_graph
docker build -t knowledge-graph .
docker run -p 8888:8888 knowledge-graph

Access Jupyter on port 8888.

Local Ollama:

Install Ollama
Run ollama run zephyr (or Mistral 7B OpenOrca)
Edit extract_graph.ipynb with your text/PDF path and run

Core notebook: extract_graph.ipynb

Tech Stack

Component	Description
Mistral 7B OpenOrca	LLM for concept extraction, via Ollama
Ollama	Local model hosting, zero API cost
Pandas	Graph schema dataframes (can switch to graph DB later)
NetworkX	Graph structure and algorithms
Pyvis	Web-hostable JS graph visualization

Project Advantages

Comparison	knowledge_graph	Traditional NER + RE	Pure vector RAG
Extraction	Concepts	Entities	No explicit structure
Relations	Semantic + contextual	Predefined types	None
Cost	Local LLM, zero API	Varies	Embedding cost
Retrieval	Graph (GRAG)	Graph query	Vector similarity
Interpretability	High, traceable edges	Medium	Low

Deep Dive

Dual Weight Design

W1: Semantic relations from LLM; multiple relations per concept pair
W2: Contextual proximity (same chunk)
Merge: Same concept pair → sum weights, concatenate relations

Suggested Improvements (community)

Backend:

Embeddings to deduplicate similar concepts ("doctor" vs "doctors")
Filter redundant/outlier concepts
Better contextual proximity to avoid overweighting frequent concepts

Frontend:

Expand by interest/topic
More useful graph browsing and navigation

Directory Structure

knowledge_graph/
├── extract_graph.ipynb   # Core extraction notebook
├── ner.ipynb            # NER related
├── data_input/           # Input
├── data_output/          # Output
├── helpers/              # Helpers
├── ollama/               # Ollama
├── dockerfile
└── pyproject.toml

Project Links and Resources

Official Resources

🌟 GitHub: https://github.com/rahulnyk/knowledge_graph
🌐 Sample graph: rahulnyk.github.io/knowledge_graph
📝 Medium: How to convert any text into a graph of concepts
🐛 Issues: GitHub Issues

Target Audience

Developers building knowledge graphs from documents
Teams exploring GRAG and graph retrieval
Users wanting zero-API-cost, local RAG/GRAG
Researchers interested in concept extraction and graph visualization

Visit my homepage for more useful knowledge and interesting products