This survey maps the RAG landscape as of late 2023, tracing an evolution through three paradigms: Naive RAG (chunk, embed, retrieve, generate), Advanced RAG (query rewriting, re-ranking, iterative retrieval, context compression), and Modular RAG (pluggable retrieval, memory, and routing components). It dissects each pipeline stage — document chunking, embedding model selection, vector indexing, retrieval strategy, post-retrieval enhancement, and generation-time grounding — alongside evaluation methods and characteristic failure modes: irrelevant retrieval, incomplete evidence, and hallucination.
For Process Safety Management, the paper functions as a design and procurement checklist. PSM knowledge is siloed, multi-format, and dense, and the survey explains why the specific engineering choices matter: how to chunk a 200-page PHA report without severing a deviation from its safeguards, why hybrid keyword-plus-semantic search outperforms pure semantic retrieval on regulatory and technical text, why re-ranking improves answer precision, and how metadata filtering preserves document provenance and access permissions across PHA, MOC, incident, and procedure repositories. Its central lesson for PSM leaders is that installing a vector database does not produce a trustworthy system — every stage must be engineered, cited, and independently evaluated for both retrieval quality and answer fidelity before an assistant is trusted in a high-hazard environment.
