Every message you send takes the same real path, whether it's a chat app or something you'd build. Toggle retrieval on, and watch the path grow by one honest step.
Station
Why the follow-up feels instant
Reprocessing your whole conversation on every single turn would be slow and wasteful — a KV cache just keeps what it already worked out, and only does fresh work on what's new.
Turn 2 still has the whole conversation on the desk — it just doesn't redo the part it already processed.
So what
"Sending a message to AI" is never one step — it's a short pipeline, and retrieval is just one more station on it.