Manuals, procedures, contracts, closed tickets. Everything somebody looks up by hand today becomes an answer, with the page it came from next to it.
RAG stands for retrieval-augmented generation, and behind the technical name there is something simple: instead of answering from memory, the model first searches your documents and then composes the answer out of what it found.
The practical difference is that the answer can be checked. Every claim has a source, and when there is no source the system says so instead of inventing one. It is the only way an automated answer can be part of a business process.
PDFs, Word files, Confluence, SharePoint, tickets, website pages. Scans included, read with OCR before they are indexed.
Vectors for meaning, keywords for part numbers and standards references. Neither is enough on its own: semantic search loses the codes.
Document, page, paragraph. Whoever is reading can open the original and check, and that is the part that makes the answer genuinely usable.
If a SharePoint document is visible only to finance, the answer that cites it is visible only to finance. Permissions are inherited from the systems of record, not redesigned.
You constrain the model to answer from what the retrieval step handed it, and you give it an explicit way out when it finds nothing. It is not an absolute guarantee, but with the sources next to the answer a mistake is visible immediately.
No, and it is normal. Working out which documents actually matter and which are stale versions is half the job, and that half is ours.
Three things together: answers anchored to the retrieved sources, the citation always visible, and an “I don't know” the system is allowed to say. The third one is the one that is usually missing.
Fewer than you think. You start with one area — support, engineering, finance — and measure there. Indexing everything at once makes it impossible to tell what is not working.
A free 30-minute call, with no commercial follow-up unless you want it.