Isn't it more like "compressing" interactions with and patterns of 500MB of data?

Like, it's not only text search but also has a degree of generalization (e.g. in labelling new documents)