Multi-Field Hybrid Retrieval-Augmented Generation for Maritime Accident Root Cause Analysis

This paper introduces a multi-field hybrid retrieval-augmented generation (RAG) framework specifically designed for automating maritime accident root cause analysis (RCA) using a dataset of 13,329 Korea Maritime Safety Tribunal reports. The framework employs a field-aware retrieval strategy that combines sparse and dense rankings through Reciprocal Rank Fusion (RRF), achieving a significant improvement in retrieval performance with NormRecall@100 increasing from 0.18 to 0.55. The results indicate that this approach not only enhances the retrieval process but also improves the quality of RCA generation when grounding the generator on retrieved precedents, thus streamlining maritime safety investigations.