# Multi-Field RAG Enhances Maritime Accident Root Cause Analysis > Source: > Published: 2026-06-12 04:59:50.256566+00:00 # Multi-Field RAG Enhances Maritime Accident Root Cause Analysis According to the arXiv submission (arXiv:2606.13249), Seongjin Kim and one other author present a **multi-field hybrid retrieval-augmented generation (RAG)** framework for automated maritime root cause analysis. The paper builds a structured knowledge base of **13,329** Korea Maritime Safety Tribunal (KMST) adjudication reports spanning **1971-2025**, creating indexed "incident cards" with three fields: **Summary**, **Causes**, and **Disposition**. The authors report a field-aware hybrid retrieval that fuses sparse and dense rankings via RRF (Reciprocal Rank Fusion), improving **NormRecall@100** from **0.18** to **0.55**, and raising an LLM-as-a-judge quality score from **3.34** to **3.72** over an LLM-only baseline, per the arXiv abstract. The paper suggests that field-aware RAG can speed precedent search and improve consistency in RCA drafting, according to the submission. Editorial analysis: For practitioners, the results indicate that domain-structured indexing plus hybrid retrieval can materially raise retrieval recall and downstream generation quality in regulated, document-heavy verticals such as maritime safety. ### What happened According to the arXiv submission (arXiv:2606.13249), Seongjin Kim and one other author propose a **multi-field hybrid retrieval-augmented generation (RAG)** pipeline aimed at automating maritime accident root cause analysis (RCA). The paper constructs a structured knowledge base from **13,329** Korea Maritime Safety Tribunal (KMST) reports covering **1971-2025**, converting adjudications into indexed "incident cards" with three explicit fields: **Summary**, **Causes**, and **Disposition**, and pairing entries with a hierarchical L1/L2 cause taxonomy, per the submission. The authors evaluate a field-aware hybrid retrieval strategy that fuses sparse and dense rankings using RRF (Reciprocal Rank Fusion) and report improvements in retrieval and generation metrics: **NormRecall@100** increases from **0.18** to **0.55**, and an LLM-as-a-judge score rises from **3.34** to **3.72** versus an LLM-only baseline, according to the abstract. ### Technical details Editorial analysis - technical context: The approach combines three practical elements commonly used in applied RAG systems: 1) structured, multi-field indexing to preserve document semantics across distinct report components; 2) hybrid retrieval that merges sparse (e.g., BM25) and dense (embedding) ranks; and 3) fusion via RRF to produce consolidated candidate lists. The paper measures retrieval using ceiling-normalized recall and nDCG based on a metadata-derived proxy relevance score, a pragmatic choice given the absence of large-scale expert relevance annotations reported in the submission. ### Context and significance Editorial analysis: For practitioners working on vertical RAG, this paper provides an empirical case that domain-specific document structuring plus hybrid ranking can substantially lift recall and improve downstream LLM outputs. The magnitude of the reported retrieval improvement (**0.18** to **0.55** NormRecall@100) is notable for workflows where precedent discovery is the bottleneck. The use of a multi-field index mirrors common legal and regulatory IR patterns where different document segments carry distinct evidentiary weight. ### What to watch Editorial analysis: Observers should look for follow-up artifacts from the authors-released code, index schemas, embedding model choices, and evaluation scripts-that would enable reproducibility and transfer to other regulated domains. Additional signals of practical impact would include human-in-the-loop evaluations with investigators, error analyses showing failure modes across cause taxonomy levels, and comparisons using expert relevance labels rather than metadata proxies. ## Scoring Rationale The paper reports substantive, domain-specific retrieval and generation gains using a large, real-world KMST dataset, which is notable for practitioners building vertical RAG systems, but it is not a frontier-model or broadly generalizable release. Practice interview problems based on real data 1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with. [Try 250 free problems](/problems)