A Transferable Retrieval-Augmented Generation Framework for Vertical-Domain Question Answering: From Academic Competitions to Cultural Tourism and Financial Technology

Vertical-domain question answering often relies on domain-specific retrieval pipelines and prompt designs, which limits robustness when transferred across heterogeneous domains. This paper presents a transferable Retrieval-Augmented Generation framework, where Retrieval-Augmented Generation (RAG) integrates external knowledge retrieval with large language model generation for grounded answering. The proposed framework targets cross-domain transfer from academic competition problem solving to cultural tourism services and financial technology applications by unifying query normalization, hybrid retrieval, and citation-consistent generation. Specifically, a domain router predicts an inference policy that adaptively configures sparse retrieval, dense retrieval, and neural re-ranking, while a query rewriting module converts user questions into a structured canonical form to reduce domain shift. Retrieved evidence is further standardized through evidence canonicalization to provide a consistent input schema for downstream generation. To improve reliability, the generation module incorporates evidence alignment and post-generation verification to reduce unsupported statements and enhance citation correctness. A transfer-oriented training strategy is introduced by combining contrastive retrieval learning, lightweight domain adaptation, and domain-invariant regularization, enabling effective adaptation under limited target-domain supervision. Experiments across three representative scenarios demonstrate that the framework improves answer accuracy, evidence recall, and citation consistency under both in-domain evaluation and few-shot transfer settings, indicating strong transferability and practical potential for deployable vertical-domain question answering systems.