Retrieval-augmented generation reduces guesswork when contracts, SOPs and SKUs can be retrieved. Garbage documents still produce garbage answers.
RAG is a search problem with a language model on top. If you cannot find the current SOP, the model will invent one. Chunking, access control and freshness matter more than the brand of the embedding.
Documents need owners, versions and a place they are allowed to live. SKU text that contradicts the ERP will be retrieved with authority. Clean the corpus or the answers will look professional and be wrong.
RAG will not tell a picker which bin to walk. It will not post a goods receipt. Those objects belong in WMS or finance software. RAG sits beside them for policies, manuals and contract clauses.
Production RAG is index jobs, permissions, evaluation questions and a refresh schedule. A weekend prototype on a shared drive is fine as a test. It is not a knowledge system.
Request a scope