[논문 리뷰] SOPRAG: Multi-view Graph Experts Retrieval for Industrial Standard Operating Procedures

Liangtao Lin, Zhaomeng Zhu외 2명|arXiv (Cornell University)|2026. 02. 02.

Advanced Graph Neural Networks인용 수 0

한 줄 요약

tldr: SOPRAG는 세 가지 특화 그래프 전문가(Entity, Causal, Flow), Procedure Card 게이팅 층, 그리고 실행 가능한 SOP를 검색 및 생성하기 위한 LLM-guided Router를 포함하는 Mixture-of-Experts 검색 프레임워크를 도입하고, 산업용 SOP를 위한 자동화 벤치마크를 제공합니다.

ABSTRACT

Standard Operating Procedures (SOPs) are essential for ensuring operational safety and consistency in industrial environments. However, retrieving and following these procedures presents unique challenges, such as rigid proprietary structures, condition-dependent relevance, and actionable execution requirement, which standard semantic-driven Retrieval-Augmented Generation (RAG) paradigms fail to address. Inspired by the Mixture-of-Experts (MoE) paradigm, we propose SOPRAG, a novel framework specifically designed to address the above pain points in SOP retrieval. SOPRAG replaces flat chunking with specialized Entity, Causal, and Flow graph experts to resolve industrial structural and logical complexities. To optimize and coordinate these experts, we propose a Procedure Card layer that prunes the search space to eliminate computational noise, and an LLM-Guided gating mechanism that dynamically weights these experts to align retrieval with operator intent. To address the scarcity of domain-specific data, we also introduce an automated, multi-agent workflow for benchmark construction. Extensive experiments across four industrial domains demonstrate that SOPRAG significantly outperforms strong lexical, dense, and graph-based RAG baselines in both retrieval accuracy and response utility, achieving perfect execution scores in real-world critical tasks.

연구 동기 및 목표

• Identify challenges in industrial SOP retrieval: proprietary structure, condition-dependent relevance, and executable output requirements.
• Propose a structure-aware MoE framework (Entity, Causal, Flow graphs) with a Procedure Card layer to prune search space.
• Develop an intention-aware LLM Router to gate and weight experts per query intent.
• Create an automated, multi-agent benchmark construction pipeline for domain-specific SOP datasets.

제안 방법

• Represent each SOP as a sparse-activation Procedure Card and three specialized graph experts (Entity, Causal, Flow).
• Construct an Entity Graph to link entities to their governing Procedure Cards, addressing proprietary structure.
• Build a Causal Graph to model state transitions from symptoms/causes to the relevant Procedure Card.
• Develop a Flow Graph for each Procedure Card to capture the procedural steps and enable executable prompts.
• Use a coarse-to-fine retrieval: Top-K Procedure Cards, then an LLM-guided router to weight experts (E, C, F) and compute a final relevance score.
• Generate structure-aware responses by linearizing the retrieved Flow Graph into a verified step-by-step prompt.

실험 결과

연구 질문

• RQ1 How can SOP retrieval be improved to respect proprietary equipment context and parameter constraints?
• RQ2 Can a multi-view graph-structured approach better capture causal and sequential dependencies in SOPs than flat chunking or generic graphs?
• RQ3 Does an intent-aware routing mechanism improve retrieval accuracy and generation fidelity for industrial SOPs?

주요 결과

• SOPRAG consistently outperforms lexical, dense, and GraphRAG baselines in retrieval metrics (MRR and Acc@K) across four domains.
• SOPRAG achieves peak MRR 0.76 and Acc@5 0.93 in Airline Services, demonstrating strong domain adaptability.
• Generation quality (faithfulness, relevance, context precision) is highest for SOPRAG, with a perfect SOP Quality Score of 1.0 in the Data Center task.
• Ablation shows the Procedure Card layer and each graph expert contribute meaningfully, with routing-based gating providing additional gains over static averaging.