This foundational paper introduced the term Retrieval-Augmented Generation and the basic idea that still underlies most RAG systems: combine a language model’s parametric memory with an external, searchable document index. A neural retriever selects relevant passages, and the generator uses those passages while composing an answer. The authors describe two formulations—RAG-Sequence and RAG-Token—and show improved performance and factual specificity on knowledge-intensive tasks.
For Process Safety Management, the central lesson is that an LLM need not rely only on knowledge embedded during training. It can retrieve current, organization-controlled evidence from PHAs, MOCs, incident reports, procedures, standards, and equipment records at query time. This supports updating knowledge without retraining, can provide provenance for answers and reduces the occurance of hallucinations. The paper is technical, but it supplies the conceptual architecture needed to understand what the retriever, vector index, and generator each do.
