When I implemented one I simply filtered the results that came back from previous requests to see if they matched what the user had typed in the mean time. That way the UI might get relevant results with lower latency than would otherwise be possible with no risk that a non-matching result would show up to confuse the user.