A proof of concept for regulatory support for navigation with LLM and RAG

The interpretation of the International Regulations to Prevent Collisions at Sea (RIPEAM) in operational environments presents challenges associated with the regulatory complexity and the high amount of information involved in the naval context. In this scenario, this work discusses the design of the tool called Cognitive Agent for Analysis of Interrelated Problems in Maritime Navigation (CAPTAIMN), an intelligent agent based on the integration between Large Language Models (LLM) and the Retrieval-Augmented Generation (RAG) architecture. The article focuses on the methodological modeling of the proposed architecture, contemplating processing steps, contextual retrieval of information and assisted generation of responses applied to the RIPEAM domain. For a preliminary analysis of the approach, an exploratory study was conducted involving normative consultations evaluated by navigation experts. The results obtained offer initial indications of the applicability of the architecture in assisted interpretation activities of the regulation, in addition to demonstrating the potential of using LLMs and RAG in support systems aimed at the maritime context.