A language model does not know your documents. RAG first finds the relevant fragments and then passes them to the model, so that the answer relies on them.
Quality depends first of all on data preparation and search: an outdated manual will give an outdated answer.
A good system shows the source so that the answer can be checked, and says “I don't know” when there is no data.
Before launch, quality should be measured on a set of real questions.