Do AI coding assistants actually understand your codebase?

I've been using AI coding assistants for about six months now, and I keep hitting the same wall. They're great at isolated functions and boilerplate, but the moment a task requires understanding how three services talk to each other, suggestions fall apart. Context windows keep growing, yet the models still seem to pattern-match locally rather than reason about architecture. I've started keeping a short architecture doc in the repo just to feed as context, which feels like a workaround. Curious how others handle this. Are you relying on retrieval tools, repo-wide indexing, or just accepting that AI handles the small stuff while you own the design?