SH-RAG (self-healing retrieval augmented generation): a self-correcting RAG framework for large language models
Ayush Kumawat, Anand Jawdekar, Sanjay Patsariya, Vicky Gupta · IET conference proceedings. · 2025
Large Language Models (LLMs) integrated with Retrieval-Augmented Generation (RAG) have significantly enhanced AI-driven knowledge synthesis by dynamically incorporating external sources. However, these models frequently encounter retrieval inconsistencies, hallucinations, and factual misalignment, leading to unreliable outputs. This paper introduces SH-RAG (Self-Healing Retrieval-Augmented Generation), a novel framework that autonomously detects, corrects, and refines retrieval errors and generated responses. SH-RAG comprises three core components: Self-Verification Module – Implements real-time fact-checking and confidence scoring to assess retrieved content. Retrieval Refinement Engine – Dynamically ranks and filters retrieved documents, ensuring optimal relevance. Self-Feedback Learning Loop – Leverages reinforcement learning to iteratively enhance retrieval strategies and reduce misinformation propagation. We evaluate SH-RAG across multiple domains, including financial analysis, scientific literature retrieval, and real-time AI-driven customer support, benchmarking its performance against conventional RAG approaches. Empirical results demonstrate a significant reduction in hallucination rates, improved retrieval precision, and enhanced response accuracy, establishing SH-RAG as a reliable and adaptive AI framework for knowledge augmentation.