An MCP-Style Modular Multi-View RAG-Fusion Method for Question Answering over HAZOP Reports

Hazard and operability (HAZOP) reports store field-linked evidence that single-query retrieval-augmented generation (RAG) may inadequately cover. This paper proposes an MCP-style modular multi-view RAG-Fusion method inspired by the Model Context Protocol (MCP). It identifies question intent and fields, expands queries through engineering perspectives, combines reciprocal rank fusion (RRF), vector similarity, and field-consistency correction, and organizes retrieved evidence into structured answers. Experiments use 191 records and 60 questions. In a controlled seven-method comparison, the complete method exceeds BM25, single-query dense retrieval, and hybrid retrieval on all four retrieval metrics, achieving recall at rank 10 (R@10) of 0.9167 and normalized discounted cumulative gain at rank 10 (nDCG@10) of 0.8623. Its nDCG@10 gains meet the question-level Holm-adjusted significance threshold; parameter-group sensitivity analysis also supports the BM25 and dense comparisons. Answer-quality evaluation favors the method in context precision, context recall, and answer correctness. For complex engineering questions, the separate question-type evaluation reports nDCG@10 improvement of approximately 22.6% over RAG-only, and nDCG@10 and answer correctness improvements of approximately 13.1% and 10.4% over MCP-only. These findings support the integrated framework for evidence retrieval and structured answering in non-time-critical HAZOP report consultation within the evaluated setting.