Hyper-RAG: Combating LLM Hallucinations

Using Hypergraph-Driven Retrieval-Augmented Generation

Presented byAMAL UB
Reg. No PRN23CS028
GuideDR. AJISH S
Overview

Content Overview

01
Abstract
02
Problem Statement
03
Objectives
04
Introduction (Domain Details)
05
Literature Review
06
Methodology (Parts 1–3)
07
Experimental Setup
08
Result and Discussion
09
Conclusion
10
References & Q&A
Section 01

Abstract

  • Background Study: Large language models (LLMs) have transformed decision-making in medicine but face hurdles due to training-induced data compression that generates hallucinations.
  • Specific Problem: Classical graph-based RAG models only map pairwise (binary) correlations, causing significant data loss regarding complex, multi-entity interactions.
  • Performance: Hyper-RAG achieves an average accuracy improvement of 12.3% over direct LLM usage and outscores GraphRAG and LightRAG by 6.3% and 6.0%, respectively, on medical datasets.
  • Future Opportunities: Promising directions include exploring automated data organization and developing methods to handle cross-chunk correlation extraction.
Section 02

Problem Statement

  • The Diagnostic Risk: LLMs generate highly confident but factually aberrant outputs (hallucinations), presenting severe risks to clinical decision-making.
  • Structural Pairwise Constraints: Traditional graph representations encode knowledge topologies relying strictly on pairwise (binary) correlations.
  • Relational Pruning: Simple graphs fail to model beyond-pairwise (high-order) interactions (such as multi-medication drug interactions or concurrent narrative events), multiplying retrieval errors.
Traditional Graph Representation
Patient COVID-19 Azithromycin Chest CT Fever Pneumonia
Captures Higher-Order Relations
Hypergraph Representation (Hyper-RAG)
CLINICAL DIAGNOSIS TREATMENT PLAN Patient COVID-19 Azithromycin Chest CT Fever Pneumonia
Section 03

Seminar Objectives

  • To Engineer Hyper-RAG: Build a robust, hypergraph-driven Retrieval-Augmented Generation (Hyper-RAG) framework to mitigate hallucinations.
  • To Capture High-Order Relations: Systematically map both low-order (pairwise) and high-order (beyond-pairwise/group) correlations from raw domain-specific corpora.
  • To Limit Error Propagation: Suppress the system-level Error Propagation Coefficient ($\alpha$) to enhance search accuracy on complex, multi-hop queries.
Section 04

Introduction (Domain Details)

  • High-Stakes Environments: Specialized medical diagnostic domains demand strict adherence to factual accuracy, where minor deviations can cause adverse outcomes.
  • Retrieval-Augmented Generation (RAG): Acts as a crucial structural safeguard by constraining LLM generation using vector-based retrieval from verified, domain-specific repositories.
  • Structural Evolution: RAG methodologies have evolved from naive, unstructured keyword chunking to structured graph-based strategies to map hierarchical connections.
  • The Hypergraph Paradigm Shift: Transitioning representation models to hypergraphs ensures complete contextual coverage of complex multi-way data spaces.
Section 05

Literature Review

Year Model Reference Key Contribution Structural Limitation
2023 Standard RAG Gao et al. Partitions raw corpora into chunks; retrieves text via keyword/vector similarity. Lacks explicit relational and structural encoding.
2024 GraphRAG Edge et al. Extracts knowledge graphs; establishes hierarchy using clustering techniques. Strictly limited to mapping pairwise (binary) connections.
2025 LightRAG Guo et al. Indexes dual-layered (local and global) knowledge graphs. Omits clustering, losing critical high-order group relations.
2026 Hyper-RAG Feng et al. Hypergraph topology capturing multi-order (pairwise + group) interactions. Mitigates information loss & error propagation
Section 06 · Methodology (Part 1)

Knowledge Base Construction

  • Semantic Text Chunking: Preprocesses and partitions raw corpora ($D$) into semantic chunks ($D_i$) to preserve local context.
  • LLM Entity Extraction: Leverages LLMs to extract entity set ($K_v$) (diseases, symptoms, drugs, treatments, concepts).
  • Relationship Extraction: Maps both pairwise ($E_{low}$) and multi-entity ($E_{high}$) correlations.
  • Vertex Consolidation: Merges duplicate entity descriptions across chunks using an LLM to unify vertex DB ($V$).
Offline Knowledge Base Construction Architecture
Raw Corpus
Domain Docs ($D$)
Chunking
Semantic $D_i$
LLM Extract
Diseases Drugs Symptoms
Relations
Pairwise + Multi
Consolidate
Merge Vertex $V$
Hypergraph KB
Vertices Hyperedges Embeddings
Section 06 · Methodology (Part 2)

Hypergraph Formation & Topology

  • Clinical Entity Initialization: Identifies core entities (Patient, COVID-19, Fever, Chest CT, Azithromycin, Pneumonia).
  • Pairwise Limitations (GraphRAG): Connects nodes strictly 1-to-1, causing severe relational pruning.
  • Hyperedge Formation (Hyper-RAG): Replaces multiple binary edges with unified multi-entity hyperedges.
  • Textual Narrative Enclosure: Encapsulates multi-entity groups with LLM-synthesized narrative descriptions.
Hypergraph Formation Step-by-Step
Clinical Diagnosis
Treatment Plan
Patient
COVID-19
Azithromycin
Chest CT
Fever
Pneumonia
Section 06 · Methodology (Part 3)

Retrieval & Prompt Augmentation

  • Query Keyword Partitioning: Splits query $q$ into Entity keywords ($X_{ent}$) and Correlation keywords ($X_{cor}$).
  • Dual Vector Search & Expansion: Maps keywords to target vertices/edges and performs 1-step network diffusion.
  • Context Ranking & Prompt Building: Formats graph paths into natural-language context and appends it to prompt.
  • Grounded LLM Response: Passes augmented prompt to LLM, eliminating training-induced compression loss.
Online Retrieval & Prompt Augmentation Workflow
User Query
Query $q$
Keywords
$X_{ent}$ $X_{cor}$
Vector Search
Hypergraph DB
Expansion
1-Step Paths
Prompt
Q: Medical Query
Entity: Vertices
Hyper: Narrative
LLM Engine
Grounded Response
Key Presenter Insight (Seminar Defense): "The hypergraph itself is never directly fed into the LLM. Instead, Hyper-RAG retrieves the most relevant nodes and hyperedges, converts them into natural-language context, appends that context to the user's query, and only then sends the augmented prompt to the LLM."
Section 07

Experimental Setup

Diverse Benchmarks

Evaluated across 9 domain-specific datasets (prominently featuring clinical NeurologyCrop and LegalCrop).

Baseline LLMs
  • Validated across 6 major models (GPT-4o mini, Qwen-Plus, LLaMa-3.3-70B, GLM-4-Air, DeepSeek-V3, Doubao-1.5-Pro)
  • 1200-token global chunks & 319-token semantic chunks
Evaluation Frameworks
  • Multi-dimensional Scoring-Based Assessments (0–100 scale)
  • Comparative Selection-Based Assessments
Mechanistic Hallucination Metrics

Key Knowledge Missing Rate (KMR), Hallucination Error Rate (HER), and Error Propagation Coefficient ($\alpha$).

Section 08

Result and Discussion

+12.3%
Accuracy Improvement
(vs direct LLM usage)
-50%
Hallucination Error Rate
(HER reduction)
0.14
Error Propagation (α)
Suppressed from 0.59 (-76%)
Error Propagation Coefficient (α) Suppression Comparison
Standard
0.68
GraphRAG
0.59
Hyper-RAG
0.14
  • Performance Breakthroughs: Hyper-RAG boosts baseline accuracy by an average of 12.3%, outperforming GraphRAG and LightRAG by 6.3% and 6.0%, respectively.
  • Cross-Domain Generalisation: Achieves a 35.5% average performance leap over LightRAG across 9 datasets (peaking at +55.3% on LegalCrop).
Section 09

Conclusion

  • Effective Hallucination Defense: Hyper-RAG successfully addresses LLM factual divergence by grounding responses in hypergraph-driven structures.
  • Context Preservation: Modeling beyond-pairwise relationships via hyperedges prevents severe information loss during knowledge graph construction.
  • Stable under Complexity: Retains stable performance on nested, multi-stage queries, unlike traditional RAG models that degrade.
  • Scalable Pathway: Hyper-RAG-Lite provides a rapid, lightweight retrieval framework ideal for real-time diagnostic systems.
Section 10 · References (Part 1 / 3)

Core & RAG Architectures

  • Feng, Y. et al. (2026). Hyper-RAG: Combating LLM hallucinations using hypergraph-driven retrieval-augmented generation. Nature Communications.
  • Edge, D. et al. (2024). From local to global: A graph RAG approach to query-focused summarization. arXiv preprint arXiv:2404.16130.
  • Guo, Z. et al. (2025). LightRAG: Simple and fast retrieval-augmented generation. Findings of the ACL: EMNLP 2025.
  • Gao, Y. et al. (2023). Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997.
  • Gao, Y., Feng, Y., Ji, S. & Ji, R. (2022). HGNN+: General hypergraph neural networks. IEEE Trans. Pattern Anal. Mach. Intell.
  • Qian, H. et al. (2025). MemoRAG: Moving towards Next-Gen RAG Via Memory-Inspired Knowledge Discovery. Proc. of the ACM on Web Conference 2025.
  • Yu, Y. et al. (2024). Rankrag: Unifying context ranking with retrieval-augmented generation in LLMs. Adv. Neural Inf. Process. Syst.
Section 10 · References (Part 2 / 3)

Graph Learning & Document Structuring

  • Zhu, X. et al. (2024). Structugraphrag: Structured document-informed knowledge graphs for retrieval-augmented generation. Proc. AAAI Symp. Series.
  • Jiang, X. et al. (2024). Ragraph: A general retrieval-augmented graph learning framework. Adv. Neural Inf. Process. Syst.
  • He, X. et al. (2024). G-retriever: Retrieval-augmented generation for textual graph understanding and question answering. Adv. Neural Inf. Process. Syst.
  • He, J. et al. (2025). Clara: Bridging retrieval and generation with continuous latent reasoning. arXiv preprint arXiv:2511.18659.
  • Es, S. et al. (2024). Ragas: Automated evaluation of retrieval augmented generation. Proc. of 18th Conf. of European Chapter of ACL.
  • Yu, H. et al. (2024). Evaluation of retrieval-augmented generation: A survey. CCF Conf. on Big Data.
Section 10 · References (Part 3 / 3)

Clinical Domain & Hallucination Mitigation

  • Xiong, G. et al. (2024). Benchmarking retrieval-augmented generation for medicine. Findings of the ACL 2024.
  • Singhal, K. et al. (2023). Large language models encode clinical knowledge. Nature.
  • Hager, P. et al. (2024). Evaluation and mitigation of the limitations of large language models in clinical decision-making. Nat. Med.
  • Huang, L. et al. (2025). A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Trans. Inf. Syst.
  • Chuang, Y.-S. et al. (2024). DoLa: Decoding by contrasting layers improves factuality in large language models. Proc. Twelfth ICLR.
  • Dhuliawala, S. et al. (2024). Chain-of-verification reduces hallucination in large language models. Findings of ACL 2024.
  • Kuhn, L. et al. (2023). Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. Proc. Eleventh ICLR.

Thank You!