Hyper-RAG
A hypergraph-driven retrieval-augmented generation method that models beyond-pairwise relationships between entities in domain-specific knowledge, to reduce hallucinations in medical LLM answers.
Hyper-RAG is a hypergraph-driven retrieval-augmented generation method, introduced in a paper published in Nature Communications, meant to reduce hallucinations when large language models are used in domains such as medicine.
Unlike graph-based RAG, which only models pairwise relationships between entities, Hyper-RAG's hypergraph also captures relationships among more than two entities at once, giving it a richer representation of domain-specific knowledge to retrieve from before generating an answer. On the NeurologyCorp dataset, across six LLMs, it improved accuracy by an average of 12.3% over direct LLM use and outperformed Graph RAG and Light RAG by 6.3% and 6.0% respectively, staying stable as query complexity increased where other methods declined. Across nine further datasets it showed a 35.5% improvement over Light RAG.
A lightweight variant, Hyper-RAG-Lite, retrieves twice as fast as Light RAG while still improving performance by 3.3%. It ships with an installation guide, a quick start, a Docker setup and an evaluation suite.