Start with the retrieval problem, then earn the right to add generation.
Retrieval-augmented generation projects usually begin with a demo: paste a question, get a fluent answer with a citation. The demo works because the document set is small and the questions are friendly. The production version fails because neither of those things stays true.
We build in a different order. First, retrieval on its own. Can the system find the right passage for a realistic question, from the real corpus, more often than a person using search? Until the answer is yes, generation only hides the problem behind confident prose.
Second, the corpus. Internal documents are duplicated, outdated, and inconsistently formatted. Deciding what is in scope, how it is chunked, and how freshness is tracked matters more than the choice of embedding model.
Third, evaluation. A few dozen real questions with known-good answers, run on every change, is the single most valuable asset in the project. It turns arguments about quality into measurements.
Only then do we add generation, and even then the first version should quote rather than summarise. Summaries are where hallucinations hide. When the retrieval is strong and the evaluation is honest, the generation step becomes the easy part.