Sorry, I don't understand what you're asking. The way it works now is: each document is read and summarized into a bunch of word frequencies, once. It looks like each is being read many times, but since I'm memoizing, it only happens once. Is that helpful or are you suggesting something else?
So, let's say we have two documents, and want to calculate their distance. Each document has 2,000 words. We can either scan each word in each document once, counting them in memory and processing a total of 4,000 words, or we can start with a word in one document, scan both documents for all occurrences of that word, move on to the next word, scan both documents for all occurrences of that word, and so on, for a total of ~N! scans. Given a choice between a really efficient function running in factorial time, or a longer, uglier function running in constant time ... I'd kind of prefer the latter.
Am I way off the mark here? Is there something I misunderstood?