It's a nifty technique, but I'm not sure it refutes my statement. You still have agents crawling the book to build an index of metadata by (in your experiment's case) subjects and years mentioned. You throw away the entries for subjects and years you're not searching for, because they're not needed for your specific problem, but that's an optimization. If you didn't do that, your agents would assemble what amounts to a fine-grained (and correspondingly large) card catalog of sentences, indexed by subject and year, which you could then search.