FaithGate: A Proposed Architecture for Real-Time Hallucination Detection and Correction in Retrieval-Augmented Generation Systems

Retrieval-Augmented Generation (RAG) was introduced as a remedy for the tendency of large language models to fabricate information, grounding generation in externally retrieved documents rather than relying solely on parametric memory. In practice, however, RAG systems continue to produce ungrounded, contradictory, or subtly fabricated claims, because retrieval and generation remain functionally disconnected: nothing in the standard pipeline verifies that what the model writes is actually entailed by what it retrieved. This paper proposes FaithGate, a structured, multi-layered framework for real-time hallucination detection and correction in RAG systems.

The architecture integrates four interdependent components: a Claim Decomposer that segments generated responses into atomic, independently verifiable claims; a Groundedness Estimator that scores each claim against retrieved evidence using entailment-based methods; a Contradiction Detector that flags claims which conflict with source material; and a Correction Loop that regenerates, hedges, or removes claims that fail verification, governed throughout by a Transparency Logger that produces an auditable record of every correction. The framework is grounded in a review of foundational and recent literature spanning retrieval-augmented generation, hallucination detection, natural language inference, and faithfulness evaluation. An architectural diagram and a pseudoalgorithmic specification of the core verificationcorrection loop are provided to bridge conceptual design and practical implementability.

The paper concludes by identifying limitations around latency overhead, evaluationmetric disagreement, and dependence on retrieval quality, along with a forward-looking research agenda for empirical validation across enterprise, legal, and healthcare RAG deployments.