In a recent arXiv submission (arXiv:2606.13249), Seongjin Kim and a co-author introduce a multi-field hybrid retrieval-augmented generation (RAG) framework aimed at automating maritime root cause analysis. Their research involves constructing a comprehensive knowledge base from 13,329 adjudication reports from the Korea Maritime Safety Tribunal (KMST), covering the years 1971 to 2025. This structured database features indexed “incident cards” that categorize information into three distinct fields: Summary, Causes, and Disposition.
The authors highlight an innovative field-aware hybrid retrieval method that combines sparse and dense rankings through Reciprocal Rank Fusion (RRF). This approach significantly enhances retrieval performance, as evidenced by an increase in NormRecall@100 from 0.18 to 0.55. Additionally, the quality score for an LLM-as-a-judge improved from 3.34 to 3.72 compared to an LLM-only baseline, underscoring the effectiveness of their framework.
The findings suggest that employing field-aware RAG techniques can expedite precedent searches and enhance the consistency of drafting root cause analyses. For industry professionals, this research indicates that a domain-structured indexing approach, combined with hybrid retrieval methods, can substantially improve both retrieval recall and the quality of generated outputs in heavily regulated environments like maritime safety.
