If you mean in the sense of dealing with documents that are very similar but not binary identical, a locality sensitive hash would do the job.

A Rabin finger printing algorithm and a hash of the data it generates.

Reference count the hashes.