Abstract: Nuclear power welding procedure documents are the fundamental technical basis for ensuring weld quality, and the quality of their review directly affects the safety of nuclear power plant construction and operation. Aiming at the problems of manual-dependent review, low efficiency and unsystematic quality risk identification in traditional procedure document review, a method for the review and quality risk identification of nuclear power welding procedure documents based on generative artificial intelligence and Retrieval-Augmented Generation (RAG) is proposed. Targeting typical documents such as welding procedure specifications, welding procedure qualification reports and welding procedure cards, RAG is integrated with a knowledge base of nuclear power welding procedures, nuclear safety regulations and relevant standards, and three review mechanisms, namely compliance review, cross-document consistency review and graded quality risk identification, together with corresponding strategies for knowledge base version management and qualification coverage verification, are constructed. Combined with practical application in a nuclear power engineering project, the feasibility of the method in improving review efficiency and enhancing quality risk prediction capability is preliminarily verified, and a reference for the digital transformation of nuclear power welding quality management is provided.