Retrieval-Augmented Generation for Vertical-Domain Knowledge Question Answering

Abstract

Large language models have substantially improved natural language understanding and generation, yet their deployment in vertical-domain knowledge question answering remains limited by stale parametric knowledge, factual hallucination, weak provenance, and insufficient adaptation to specialized terminology. Retrieval-augmented generation (RAG) addresses these limitations by coupling a generative model with an external, updateable knowledge base. Before answering, the system retrieves relevant evidence, reranks it, and conditions generation on the selected context. This paper presents a practical RAG framework for vertical-domain question answering. The framework integrates document cleaning, semantic chunking, dense vector retrieval, evidence reranking, prompt construction, and constrained answer generation. We formulate the key retrieval and generation objectives, design a prototype evaluation, and analyze representative results. The study indicates that RAG improves answer accuracy, evidence coverage, and factual consistency compared with direct generation and keyword-based retrieval augmentation. The proposed framework offers a feasible path for knowledge-intensive applications in law, medicine, finance, education, and scientific research.