I can match the performance in their paper with lz4, and then by just explicitly comparing character-level n-grams between a query document and all documents belonging to the same topic.