Confidence-Calibrated Hallucination Reduction in RAG-Augmented LLM Systems

Large Language Models (LLMs) exhibit impressive generative capability but remain unsafe for high-stakes deployment because they can produce fluent, plausible, and factually incorrect outputs. This hallucination problem is not merely an accuracy issue; it is fundamentally a confidence alignment issue. Raw model confidence is often miscalibrated, and Retrieval-Augmented Generation (RAG), while improving factual grounding, does not eliminate the problem. In noisy retrieval conditions, contradictory or weakly relevant documents can intensify rather than reduce model overconfidence. This paper presents Confidence-Calibrated Hallucination Reduction (CCHR), a model-agnostic post-generation architecture for improving reliability in RAG-augmented LLM systems. CCHR integrates multi-signal confidence estimation, context-aware calibration, utility-based action selection, response control, evaluation, and online learning into a unified framework. The architecture estimates raw confidence from five complementary signals, calibrates that confidence using retrieval quality and domain context, selects response actions through expected-utility maximization rather than fixed thresholds, and continuously improves through explicit, implicit, and system-level feedback. The resulting framework transforms an LLM pipeline from a static generator into a calibrated, selective, and self-improving decision system suitable for enterprise and high-risk AI deployment.