I assume the same, feels like their product is just summarizing the top n results? I wouldn't need the whole vector db thing, if local models (or hardware) would be able to run with a context of this size.