MEGA-RAG: a retrieval-augmented generation framework with multi-evidence guided answer refinement for mitigating hallucinations of LLMs in public health

MEGA-RAG is a retrieval-augmented generation framework that fuses dense (FAISS), sparse (BM25), and biomedical knowledge-graph retrieval with cross-encoder reranking and a discrepancy-aware refinement module to reduce LLM hallucinations in public health question answering.

Problem Statement

LLMs are increasingly used for health communication and policy support, but their tendency to hallucinate factually incorrect content poses serious risks in clinical and public-health contexts where accuracy is critical. Standard RAG pipelines typically rely on a single retrieval modality and lack an explicit mechanism to detect and correct discrepancies between generated text and retrieved evidence, leaving residual hallucinations unaddressed.

Key Novelty

Multi-source evidence retrieval that fuses dense vector search (FAISS), keyword-based retrieval (BM25), and structured biomedical knowledge graphs for richer, complementary evidence coverage
Cross-encoder reranking stage that filters retrieved evidence for deep semantic relevance rather than relying on retrieval scores alone
A discrepancy-aware answer refinement module that explicitly checks generated outputs against retrieved evidence and corrects inconsistencies, directly targeting residual hallucinations after generation

Evaluation Highlights

Reduces hallucination rate by over 40% compared to four baselines (PubMedBERT, PubMedGPT, standalone LLM, standard RAG-augmented LLM)
Achieves best-in-class accuracy (0.7913), precision (0.7541), recall (0.8304), and F1 score (0.7904) across all compared models

Signal Assessment

5/10 The work thoughtfully engineers an integrated pipeline from established components (FAISS, BM25, knowledge graphs, cross-encoders) plus a novel discrepancy-aware refinement step, yielding strong domain-specific gains, but it is an applied systems contribution rather than a fundamentally new algorithm or paradigm.

Methodology

Retrieve candidate evidence for a query using three complementary methods: dense FAISS embedding search, BM25 keyword retrieval, and biomedical knowledge graph lookup
Rerank pooled evidence with a cross-encoder to select passages with the highest semantic relevance to the query
Generate an initial answer with the LLM conditioned on the reranked evidence
Apply the discrepancy-aware refinement module to compare the generated answer against evidence and revise any unsupported or conflicting statements
Evaluate the refined output against baselines using accuracy, precision, recall, F1, and hallucination rate metrics

System Components

Dense Retriever (FAISS)

Performs vector-similarity search over embedded biomedical text to retrieve semantically relevant passages

Sparse Retriever (BM25)

Retrieves passages via keyword/lexical matching to complement dense retrieval, especially for rare terms and specific entities

Biomedical Knowledge Graph Retrieval

Queries structured medical knowledge graphs to supply factual, relational evidence not easily captured by text retrieval alone

Cross-Encoder Reranker

Jointly encodes query-evidence pairs to score and rank retrieved candidates by fine-grained semantic relevance

Discrepancy-Aware Refinement Module

Detects inconsistencies between the LLM's generated answer and the retrieved evidence, then refines the answer to improve factual alignment and reduce hallucinations

Results

Metric	Best Baseline (PubMedBERT/PubMedGPT/LLM/RAG-LLM)	MEGA-RAG	Delta
Hallucination Rate	Higher baseline rate (exact value not reported)	Reduced hallucination rate	>40% relative reduction
Accuracy	Lower than 0.7913 (exact value not reported)	0.7913	Highest among all compared models
Precision	Lower than 0.7541 (exact value not reported)	0.7541	Highest among all compared models
Recall	Lower than 0.8304 (exact value not reported)	0.8304	Highest among all compared models
F1 Score	Lower than 0.7904 (exact value not reported)	0.7904	Highest among all compared models

Key Takeaways

Blending dense, sparse, and knowledge-graph retrieval gives more comprehensive evidence coverage than single-modality RAG, which matters for specialized, high-stakes domains like public health
Adding an explicit post-generation discrepancy-check/refinement step can catch hallucinations that survive standard retrieval-augmented generation, suggesting refinement loops are a valuable addition to RAG pipelines
Domain-grounded RAG (with biomedical knowledge graphs) can outperform domain-pretrained encoder/generative models (PubMedBERT, PubMedGPT) on factuality-sensitive tasks
This multi-retrieval + rerank + refine architecture pattern is likely transferable to other high-stakes domains (e.g., legal, financial, regulatory) where hallucination mitigation is a priority

Abstract

Introduction The increasing adoption of large language models (LLMs) in public health has raised significant concerns about hallucinations-factually inaccurate or misleading outputs that can compromise clinical communication and policy decisions. Methods We propose a retrieval-augmented generation framework with multi-evidence guided answer refinement (MEGA-RAG), specifically designed to mitigate hallucinations in public health applications. The framework integrates multi-source evidence retrieval (dense retrieval via FAISS, keyword-based retrieval via BM25, and biomedical knowledge graphs), employs a cross-encoder reranker to ensure semantic relevance, and incorporates a discrepancy-aware refinement module to further enhance factual accuracy. Results Experimental evaluation demonstrates that MEGA-RAG outperforms four baseline models [PubMedBERT, PubMedGPT, standalone LLM, and LLM with standard retrieval-augmented generation (RAG)], achieving a reduction in hallucination rates by over 40%. It also achieves the highest accuracy (0.7913), precision (0.7541), recall (0.8304), and F1 score (0.7904). Discussion These findings confirm that MEGA-RAG is highly effective in generating factually reliable and medically accurate responses, thereby enhancing the credibility of AI-generated health information for applications in health education, clinical communication, and evidence-based policy development.