VG-RAG: verification-gated retrieval-augmented generation

Retrieval Augmented Generation (RAG) improves the factual grounding of large language models by conditioning the outputs based on the retrieved passages. The main drawback of the present Retrieval Augmented Generation model is its implicit nature, where the outputs are based on the retrieved passages. If the passages are not relevant, outdated, or contradictory, the generator will propagate hallucinations before the activation of the correction mechanism. We propose Verification-Gated Retrieval-Augmented Generation (VG-RAG) , where the information retrieved by the system is first sanitized by a deterministic pre-generation verification gate. VG-RAG checks the epistemic qualities of the passage on three fronts: semantic relevance, constraint-aware internal consistency, and corroborative agreement. These are combined using a non-compensatory multiplicative gate that requires all reliability thresholds to be met simultaneously. This prevents any single high-quality signal from overcoming the lack of a signal in other dimensions. Any passage that does not meet these criteria is rejected outright; those meeting the middle confidence criteria are conditionally promoted using an external web-based promotion-only audit. If no passage meets these criteria, the system abstains entirely. We perform experiments with the 150-query HaluEval QA and TruthfulQA benchmarks. The Context Leakage Rate of VG-RAG is 83 % lower than the baseline RAG (0.50 \(\rightarrow\) 0.09 on HaluEval; 0.44 \(\rightarrow\) 0.08 on TruthfulQA), and the Verified Purity of VG-RAG is between 0.70 and 0.85 for the queries for which it returns a response. A falsification test, inspired by the Popperian method, verifies the effectiveness of active adversarial web verification by demonstrating its capacity to falsify 89.3 % of counterfactual claims (95 % CI: 0.83–0.96), while leaving factual statements intact. A controlled ablation study verifies the qualitative distinctness of the failure modes targeted by each of the gate components, as well as the non-compensatory scoring, which is a design choice but also a safety imperative for the retrieval-augmented generation task. Repository : https://github.com/Gopikrish-30/VG-RAG