Beyond Hallucination: Adaptive RAG Pipelines with Reinforcement-Guided Retrieval and Knowledge Distillation

Abstract

Retrieval-Augmented Generation (RAG) has recently emerged as a powerful paradigm for extending large language models (LLMs) with access to external knowledge sources, thereby mitigating issues of hallucination, limited corpus coverage, and outdated information. This paper presents an optimized local knowledge-base RAG framework that integrates semantic chunking, vector embeddings, dense passage retrieval, and adaptive prompt engineering. The proposed architecture is organized into six key modules: document ingestion, contextaware segmentation, embedding, similarity search, dynamic prompt construction, and response generation. To further enhance adaptability and efficiency, reinforcement-guided retrieval refinement and knowledge distillation are incorporated. This paper proposes an optimized RAG framework incorporating semantic chunking, adaptive retrieval, context-aware prompt construction, and retrieval optimization techniques for domain-specific knowledge bases. The proposed framework is validated through a campus information service application using the ERNIE-4.0 model and AppBuilder platform, with a specific application to campus information services. Experimental evaluation demonstrates significant improvements in retrieval accuracy, contextual relevance, and factual reliability compared with baseline RAG systems. The findings highlight the potential of advanced RAG pipelines in developing domain-specific, scalable, and trustworthy AI-driven knowledge systems.