Design of a Multi-Modal Retrieval-Augmented Framework with Human-in-the-Loop Validation for Maritime Regulatory Compliance

Abstract

Maritime laws, specifically focusing on Vessel Inspection Questionnaire (SIRE 2.0), are known for their level of detail and heavy use of visual diagrams. Standard Retrieval-Augmented Generation (RAG) architectures tend to perform poorly in such a scenario, due to context pollution by irrelevant images and a lack of safety guarantees. This paper proposes a Maritime AI Assistant, which is an Agentic RAG model tailored to handle complex maritime safety data. This is achieved through two major components: (1) a “Ruthless” Image Injection Algorithm that uses aggressive scoring to counter context window pollution, and (2) a Human-in-the-Loop (HITL) Gatekeeper module that manually verifies “Risk” queries with high stakes. Experimental results show that the proposed setup significantly improves technical diagram retrieval accuracy and removes errors in safety-critical responses compared to standard RAG models.