MedRAGChecker: Biomedical RAG Verification

• MedRAGChecker is a comprehensive framework that decomposes biomedical RAG outputs into atomic claims for precise, evidence-based evaluation.
• It integrates NLI and knowledge-graph consistency to assess entailment and detect contradictions, improving safety and reliability in medical QA.
• The system employs ensemble reliability weighting and modular diagnostics to support scalable evaluation in clinical and regulatory environments.

MedRAGChecker is a comprehensive claim-level verification and evaluation framework specifically designed for biomedical retrieval-augmented generation (RAG) systems. Its primary objective is to disentangle granular answer failures such as ungrounded claims, contradictions, and safety-critical errors, providing actionable diagnostics to improve the reliability of medical QA outputs produced by LLMs using retrieved evidence (Ji et al., 10 Jan 2026). MedRAGChecker leverages a modular architecture that integrates atomic claim extraction, evidence-grounded natural language inference (NLI), biomedical knowledge-graph (KG) consistency, and ensemble reliability assessment. This supports both large-scale evaluation and post-hoc correction workflows in clinical decision support, research, and regulatory settings.

1. Architectural Components and Workflow

MedRAGChecker decomposes the QA process into a multistage pipeline:

• Input Structure: Each instance comprises a user question $q$, a collection of top-$k$ retrieved biomedical passages $D = \{d_j\}$, and an LLM-generated answer $a$.
• Claim Extraction: A high-capacity LLM ("teacher"—GPT-4.1) segments $a$ into atomic claims $\mathcal{C} = \{c_1, ..., c_n\}$ of minimal inference scope (typically short statements or SPO triples). A compact "student" model (e.g., Med42-Llama3-8B) is distilled for efficient, scalable extraction.
• Claim Verification: For each $c_i$, both NLI and KG-based verification are executed:
  • NLI component: The claim $c_i$ is assessed under premises $D$ (retrieved context), producing support probabilities $p_{\rm NLI}(c_i) = P(\mathrm{Entail} \mid c_i, D)$ via biomedical LLMs or ensemble critic.
  • KG consistency: Entity and relation surfaces are linked from DRKG; plausibility scores $p_{\rm KGE}(c_i)$ are computed using TransE embeddings and string alignment.
  • Score Fusion: A calibrated support probability $P^\star(c_i)$ is computed by mixing NLI and KG signals in logit space:

  $$P^\star(c_i) = \sigma\left(\beta \cdot \mathrm{logit}\left(p_{\rm NLI}(c_i)\right) + (1 - \beta) \cdot \mathrm{logit}\left(s_{\rm KG}(c_i)\right)\right)$$

  where $s_{\rm KG}(c_i)$ is a weighted sum of embedding and text scores.

• Ensemble Reliability Weighting: Multiple student checkers are F1-weighted per-class (Entail, Neutral, Contradict) for robust aggregation.
• Diagnostics Computation: Per-claim decisions are aggregated into answer-level faithfulness, hallucination rate, under-evidence (ClaimRec), context precision, self-knowledge, and safety-critical error metrics.

2. Verification Methodologies

MedRAGChecker integrates evidence-grounded NLI and KG-based plausibility, addressing biomedical answer verification beyond pure text entailment:

• NLI Verification: Claims are classified into Entail, Neutral, or Contradict labels by fine-tuned biomedical LLMs. Softmax probabilities are used for probabilistic support assessment. Teacher supervision ensures high-fidelity verification consistency across application domains.
• Knowledge-Graph Consistency: DRKG triples covering treats, causes, contraindicates are used for entity-relation alignment. Each claim is scored for string similarity and TransE plausibility, factoring in incomplete KG coverage and linguistic paraphrasing.
• Class-Specific Ensemble Aggregation: Model-specific reliability weights $w_m^{(y)}$ favor checkers with higher F1 on contradictory labels, reducing systematic under-detection of dangerous errors.

3. Diagnostic Metrics and Error Taxonomy

MedRAGChecker quantifies answer reliability along several axes, providing granular insight into system performance:

Metric	Formula/Definition	Purpose
Faithfulness	$$\mathrm{Faith}(a) = \frac{1}{n} \sum_{i} \mathbf{1}[\hat{y}_i = \mathrm{Entail}]$$	Fraction of claims fully supported
Hallucination	$$\mathrm{Halluc}(a) = \frac{1}{n} \sum_{i} \mathbf{1}[\hat{y}_i = \mathrm{Contradict}]$$	Contradicted claims rate
ClaimRec	$$\mathrm{ClaimRec} = \frac{1}{|\mathcal C^{\rm ref}|} \sum_{c \in \mathcal C^{\rm ref}} \mathbf{1}[P^\star(c) \ge \tau]$$	Coverage of reference claims
Context Precision	$$\mathrm{CtxPrec} = \frac{1}{k} \sum_{j} \mathbf{1}[\exists i: d_j \text{ supports } c_i]$$	Evidence use efficiency
Safety Error	$$\mathrm{SafetyErr} = \frac{1}{|\mathcal C_{\rm safety}|} \sum_{c \in \mathcal C_{\rm safety}} \mathbf{1}[\hat y(c)=\mathrm{Contradict}]$$	Error rate on critical biomedical relations
Self-Knowledge	$$\mathrm{SelfKnow} = \frac{1}{n} \sum_{i} \mathbf{1}[p_{\rm NLI}(c_i|\varnothing) \ge \tau_{\rm NLI}]$$	Claims supported by model prior alone

These metrics enable fine-grained analysis of faithfulness, error localization, retrieval performance, and model self-sufficiency.

4. Experimental Evaluation and Comparative Performance

• Datasets: Evaluation on PubMedQA, MedQuAD, TREC LiveQA Medical, MedRedQA using fixed PubMed (BM25 + hybrid dense retrieval) and multiple biomedical LLMs (Med-Qwen2-7B, Med42-Llama3-8B, Meditron3-8B, PMC-LLaMA-13B, LLaMA-3-8B-Instruct).
• Claims Extraction Fidelity: Span F1 vs GPT-4.1 teacher 20–25%, average 4–5 claims per answer.
• Student NLI Prediction Accuracy: ~82–86% three-way, macro-F1 up to 67%; ensemble yields 87.4% accuracy.
• KG-Fusion Gains: SafetyErr reduced (example: Med-Qwen2-7B improves from 19.6%→12.6% with KG fusion).
• End-to-End Diagnostics: Faithfulness rates (Med-Qwen2-7B: 81.4%, Med42-Llama3-8B: 85.3%), Hallucination rates (6–10%), SafetyErr (6–11%). Human ratings confirm claim-set quality (≈4.7–5.0/5); moderate alignment for answer correctness and completeness.
• Distinct Generator Profiles: Some models maximize faithfulness but under-evidence; others yield higher hallucination rates.

5. Integration into Practical Medical RAG Verification

MedRAGChecker serves as an automated screening tool for verifying and diagnosing LLM-generated biomedical QA answers:

• Automated Post-Hoc Screening: Ingest arbitrary QA output, parse into atomic claims, evaluate each via NLI/KG ensemble, and output supported/contradicted/under-evidenced claims.
• Clinical Decision Support: Flag safety-critical contradictions and hallucinations, report detailed metrics for auditing and regulatory review.
• Human-in-the-Loop Option: Present flagged claims and reliability scores for manual review and correction.
• Scalability and Efficiency: Distilled student models enable low-latency, batch-scale evaluation; ensemble weighting optimizes error detection.

6. Limitations and Future Directions

• Teacher Supervision Bias: Dependence on GPT-4.1 as pseudo-ground truth may import systematic limitations, especially for consumer-health and cross-lingual data.
• Coverage of Contradict Label: Contradictions remain an under-detected minority class.
• KG Alignment Scope: String-based entity/relation linking may fail on negated, implied, or paraphrased claims; DRKG coverage remains incomplete.
• Threshold Sensitivity: Diagnostic metrics depend on per-class thresholds and fusion weights, though these demonstrate stability across reasonable ranges.
• Screening Role: MedRAGChecker is not a substitute for expert judgment; validation by domain specialists is indispensable for clinical deployment.

7. Relation to Other Verification Frameworks

MedRAGChecker shares conceptual lineage with end-to-end judge scoring frameworks (e.g., CCRS (Muhamed, 25 Jun 2025)), citation-aware iterative verification tools (MedTrust-RAG (Ning et al., 16 Oct 2025)), meta-analytic evidence filters (META-RAG (Sun et al., 28 Oct 2025)), agent-driven clinical error pipelines (MedReAct'N'MedReFlex (Corbeil, 2024)), and multi-criteria benchmarks (MedRGB (Ngo et al., 2024)). Its distinctive contributions are:

• Robust atomic claim-level assessment combining evidence-grounded inference and biomedical graph consistency.
• Ensemble modeling with class-weighted reliability for contradict/error detection.
• Safety-critical diagnostics for drug–disease and adverse-effect relations in real QA workflows.

This framework reflects the emerging principle that biomedical RAG evaluation requires granular, explainable, and trustworthy verification at the claim level, especially where ungrounded inferences carry direct safety risk (Ji et al., 10 Jan 2026).