iMoonLab/Hyper-RAG

"Hyper-RAG: Combating LLM Hallucinations using Hypergraph-Driven Retrieval-Augmented Generation" by Yifan Feng, Hao Hu, Shihui Ying, Xingliang Hou, Shiquan Liu, Mingyuan Yang, Junchang Li, Shaoyi Du, Nanning Zheng, Han Hu, and Yue Gao.

What it solves

Hyper-RAG addresses the issue of hallucinations in Large Language Models (LLMs), particularly in high-stakes domains like medical diagnostics. It aims to improve factual accuracy by capturing complex, beyond-pairwise correlations in domain-specific knowledge that traditional graph-based RAG methods often miss.

How it works

Hyper-RAG uses a hypergraph-driven approach to organize and retrieve knowledge. Unlike standard graphs that connect two nodes (pairwise), hypergraphs allow an edge to connect multiple nodes simultaneously, preventing information loss during the modeling of entity relationships. The system integrates with Hypergraph-DB for rapid retrieval of these higher-order associations. It extracts entities and correlations from a raw corpus, builds a knowledge hypergraph, and retrieves relevant prior knowledge to augment the LLM's prompt, thereby reducing hallucinations.

Who it’s for

Developers and researchers working with LLMs in specialized fields where factual precision is critical, such as medicine, finance, or education, and those looking for an alternative to Graph RAG or Light RAG.

Highlights

  • Hypergraph Modeling: Captures complex, multi-node relationships to provide a more comprehensive understanding of the corpus.
  • Native DB Integration: Built on Hypergraph-DB for efficient retrieval of high-order associations.
  • Performance Gains: Outperforms Graph RAG and Light RAG in accuracy on specific datasets (e.g., NeurologyCorp).
  • Hyper-RAG-Lite: A lightweight variant that offers twice the retrieval speed of Light RAG with a slight performance boost.
  • Visualization Tool: Includes a web-based UI for visualizing hypergraphs and interacting with the QA system.