In my system, RAG is the second-brain layer. The agents pull domain context from a wiki I maintain rather than relying on training data alone. It's the difference between an assistant that knows things in general and one that knows things about me.
The job of the RAG layer is to surface the right slice of the wiki for the question at hand — the relevant documents, the relevant past decisions, the relevant memos — and put them in the model's context before it answers. The model doesn't "know" me from training. It knows me from what the retrieval surfaces.
Most discussions of RAG focus on the retrieval algorithm: which embedding model, which vector DB, which similarity threshold. Those matter, but they're not the load-bearing decision. The load-bearing decision is what's in the wiki. A RAG layer that retrieves brilliantly from a thin or unstructured corpus surfaces noise. A RAG layer that retrieves crudely from a well-curated corpus still surfaces signal. The corpus is the asset; the retrieval is the index.
The architecture is what's permanent. The specific embedding model, the specific vector DB, the specific similarity threshold all get upgraded over time. The wiki gets edited every week. What stays the same is the pattern: agents read from a structured store of my context before they answer.
Without the corpus, no amount of retrieval skill helps. With the corpus, the retrieval can be primitive and still useful. That's the relationship.