Answers grounded in your documents.

A chat tool holds what fits in one conversation. Retrieval searches everything you have, every time, and returns only the passages that matter for the question asked. It also respects who is allowed to see what, keeps working as the corpus grows past any context window, and cites the source so an answer can be checked rather than trusted.

It should say so, and that behaviour has to be designed in. A system that always produces an answer will invent one when the material is thin. We set the threshold for when retrieval has found too little, and the response becomes a refusal with a route to a person.

Measured as whether the right passage was retrieved, on an evaluation set built from your real questions. That number matters more than how the answer reads, because a fluent response over the wrong passage is the failure mode that gets missed. 91% is what we have measured on delivered work.

Yes, and it has to be built in from the start. Permissions are applied at retrieval, so a user's question only ever searches the material they could already open. Filtering after the model has already seen restricted content fails, because that content has influenced the answer.

RAG, in almost every case where the question is about your own documents. Retrieval changes the moment you update a file; a tuned model has to be retrained. Tuning suits a consistent format or house style that prompting cannot reach, and the two combine well, but reaching for tuning first is the common and expensive mistake.

Re-indexing runs on a schedule or on a publish action, so an edit reaches the system in minutes. The harder problem is the superseded document that nobody deleted, which retrieval will happily find. We build the index around a source of truth somebody owns, which is a different thing from a shared folder.