This AIChE paper provides a direct process-safety example of RAG in operation. Air Products developed a tool that converts free-text incident narratives into structured classifications for incident cause and product loss. Its multi-RAG design retrieves both similar historical incidents and expert guidance notes. Those materials are inserted into a structured prompt, after which an LLM produces controlled labels that can be stored, trended, and analyzed. Validation against human-generated classifications achieved greater than 90% agreement, and the reported hose-burst example reduced an analysis that had required about 160 hours to less than an hour.
The paper is highly relevant to Process Safety Management because it shows how RAG can combine institutional experience, classification rules, and current incident text without model fine-tuning. It also demonstrates practical controls: predefined labels, low model temperature, structured outputs, expert-developed guidance, validation against subject-matter experts, and traceable contextual support.
