use a k-means based cluster approach to speed up similarity searches