Research map / Graph learning / Knowledge and temporal graphs
HyperGraphRAG: Retrieval-Augmented Generation via Hypergraph-Structured Knowledge Representation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HyperGraphRAG, a RAG framework that represents n-ary relational facts as hyperedges in a knowledge hypergraph (stored as a bipartite graph with vector indexes), and shows gains over StandardRAG and binary graph-based RAG baselines across five domains.
Problem
Existing graph-based RAG methods use binary edges, which cannot natively represent n-ary relations (n>=2) among multiple entities, causing knowledge fragmentation and sparse retrieval.
Approach
Extract knowledge as natural-language hyperedges connecting multiple entities, store the hypergraph as a bipartite graph with separate entity and hyperedge vector indexes, retrieve via entity and hyperedge similarity with bidirectional expansion, and fuse with chunk retrieval for generation.
Key results
- HyperGraphRAG reaches overall F1 35.35 on Medicine vs 27.90 for StandardRAG, and improves R-S and G-E over all baselines in every domain (e.g., +7.45 F1 vs StandardRAG on Medicine binary source).
- Ablation on Medicine: removing hyperedge retrieval drops F1 from 35.4 to 26.4; removing entity retrieval drops it to 29.8; removing chunk fusion drops it to 29.2.
- Construction costs 3.084 s and $0.0063 per 1k tokens; generation takes 0.256 s per query, between StandardRAG and binary graph baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HyperGraphRAG instead of StandardRAG | RAG across the five domains · Overall | F1: +7.45 (F1) | Helped |
| Use HyperGraphRAG instead of StandardRAG | RAG across the five domains · Overall | R-S: +7.62 (R-S) | Helped |
| Use HyperGraphRAG instead of StandardRAG | RAG across the five domains · Overall | G-E: +3.69 (G-E) | Helped |
| add entity retrieval (ablation) | HyperGraphRAG · Medicine | F1: 35.4 -> 29.8 | Helped |
| add hyperedge retrieval (ablation) | HyperGraphRAG · Medicine | F1: 35.4 -> 26.4 | Helped |
| add chunk retrieval fusion (ablation) | HyperGraphRAG · Medicine | F1: 35.4 -> 29.2 | Helped |
Surprising findings
- Existing binary graph-based RAG methods (GraphRAG, LightRAG, PathRAG, HippoRAG2) often underperform the simple StandardRAG chunk baseline.
Limitations
- Evaluation relies on questions generated by GPT and manually verified, with a modest 512 questions per domain.
- Uses a proprietary GPT-4o-mini for extraction, generation, and LLM-as-judge, so results depend on that model.
- Proofs of Propositions are information-theoretic arguments with informal assumptions rather than empirical validation of the claims.
Experimental setup
Datasets: UltraDomain (Agriculture, CS, Legal, Mix) latest international hypertension guidelines (Medicine)
Benchmarks: UltraDomain-based QA (Medicine, Agriculture, CS, Legal, Mix)
Metrics: F1 Retrieval Similarity (R-S) Generation Evaluation (G-E) construction time and cost generation time and cost
Baselines: NaiveGeneration StandardRAG GraphRAG LightRAG PathRAG HippoRAG2
Builds on: GraphRAG LightRAG PathRAG HippoRAG2
Closest papers by meaning
- When to use Graphs in RAG: A Comprehensive Analysis for Graph Retrieval-Augmented Generation (ICLR 2026 · similarity 0.91)
- Simple is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation (ICLR 2025 · similarity 0.90)
- Can Knowledge-Graph-based Retrieval Augmented Generation Really Retrieve What You Need? (NeurIPS 2025 · similarity 0.89)
- Think-on-Graph 2.0: Deep and Faithful Large Language Model Reasoning with Knowledge-guided Retrieval Augmented Generation (ICLR 2025 · similarity 0.89)
- A Systematic Exploration of Knowledge Graph Alignment with Large Language Models in Retrieval Augmented Generation (AAAI 2025 · similarity 0.89)
- GFM-RAG: Graph Foundation Model for Retrieval Augmented Generation (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.