Parameter-Efficient Contextual Calibration for Hallucination Mitigation in Domain-Specific Large Language Model Retrieval-Augmented Generation

Retrieval-Augmented Generation (RAG) has become the gold standard paradigm for deploying Large Language Models (LLMs) in knowledge-intensive and high-stakes domains such as biomedical inquiry, financial compliance, and legal reasoning. Despite providing external grounding documents, LLMs continue to exhibit insidious factual hallucinations—either by fabricating plausible-sounding unsupported assertions or by ignoring conflicting retrieved evidence in favor of memorized parametric training biases. Existing mitigation approaches, such as full-parameter fine-tuning or iterative self-reflection prompting, incur prohibitive computational costs and excessive inference latency. In this paper, we propose CAL-RAG (Context-Aware Low-Rank Calibration for RAG), a parameter-efficient fine-tuning and decoding calibration framework designed to enforce strict contextual faithfulness without compromising generative fluency. CAL-RAG introduces a dual-channel Low-Rank Adaptation (LoRA) mechanism: a Context-Grounded Adapter that measures token-level semantic consistency against retrieved evidence chunks, and an Entropy-Gated Decoding Controller that dynamically modulates vocabulary probability distributions during autoregressive generation based on cross-attention dispersion. We conduct extensive empirical evaluations across three challenging domain benchmarks: BioASQ (biomedical), FinQA (financial reasoning), and LegalBench (legal clause interpretation), utilizing open-source LLM backbones (Llama-3-8B, Mistral-7B-Instruct, and Gemma-7B). CAL-RAG reduces factual hallucination rates by 43.7% relative to standard RAG baselines while improving Faithfulness Score from 0.642 to 0.891 and answer accuracy by +11.8% F1. Remarkably, CAL-RAG adds only 0.4% trainable parameters and introduces less than 6.5 ms token latency overhead, making it highly suitable for enterprise production deployment.