Open-Source Local LLM-Based Knowledge Graph (KG) Generation Released

This is an open-source project that extracts Concepts, not just simple entities, from text corpora (such as PDFs) to build a knowledge graph composed of nodes and edges. The project aims to go beyond traditional vector-based RAG toward graph-based Retrieval-Augmented Generation (Graph RAG).

Key features are as follows.

• No-GPT approach: To reduce costs, it uses local models such as Mistral 7B OpenOrca instead of GPT, run via Ollama.
• Concept-centric extraction: Going beyond simple named entity recognition (NER), it extracts meaningful 'concepts' from context to form a deeper knowledge structure.
• Tech stack: NetworkX handles graph algorithms, Pyvis provides interactive visualization, and Pandas manages the data.
• Implementation method: Text is split into chunks, and the LLM extracts concept pairs and relationships, followed by a 4-step process that calculates contextual proximity to assign weights.

It can currently be easily installed and run via Docker, and future improvements will include concept deduplication using embeddings and outlier filtering.