rag-powered chatbots for maritime classification rules

Large language models (llms) such as gpt-5, grok-4, gemini 2.5, and deepseek perform well on open-domain tasks but struggle with proprietary, domain-specific knowledge. in maritime engineering—where regulatory accuracy is essential—this limitation increases risks of omissions and hallucinations in classification-rule queries. we present a retrieval-augmented generation (rag) system grounded in the asia classification society (acs) 2022 rules and regulations for the classification of ships (3,229 pages). using 500 expert-annotated q–a pairs, we benchmark our rag pipeline against leading llms in default settings. rag consistently outperforms all baselines across correctness (ragas), semantic similarity (sbert), and bertscore, achieving nearly double the accuracy of the strongest model. by anchoring responses in verified regulatory text, the system minimizes hallucinations, enhances factual precision, and delivers traceable answers. these findings establish rag as a reliable framework for domain-specific chatbots in maritime regulation and, more broadly, in safety-critical, rule-based industries.