Agent-Based Multimodal Retrieval and Regulation Analysis for Automated Ship Design Compliance

Abstract

Ship design faces major challenges due to multidisciplinary complexity and manual regulatory verification, making the process labour-intensive and difficult to master. However, AI assistants can improve engineers’ efficiency and efficacy, especially during the stage of ship design. In this paper, we developed a two-agent framework that develops a knowledge database and lets users search it via a LLM (Large Language Model). These two agents are the parser agent and the retrieval agent. The parser agent uses YOLOv11 to analyse the document structure and then divides it into various elements. All text and figures in these elements are extracted by Qwen3-VL and imported into our knowledge database. The retrieval agent uses RAG (Retrieval-Augmented Generation) and/or Graph-based RAG to retrieve answers from our knowledge database. The experiment aims to compare the efficacy of our framework, which uses Qwen3-VL, with ChatGPT and Gemini. The results show that our framework using Qwen3-VL achieves an accuracy rate of 0.772, correctly answering over 50% of the questions. Although the performance of our Qwen3-VL-based framework is lower than that of ChatGPT under the same knowledge base, it demonstrates the potential for further development.