This is neat. I bet if you use a hypercircle technique then you could project wikipedia articles into a GiST or BallTree or similar n dimensional tree structure. Then search can be done by projecting a query into the hypercircle and returning results close by.
You would have a representative dataset but you'd still be stuck with the curse of dimensionality that plagues nearest neighbors: in high dimensions, everything is far away from each other and nothing is close to the center.
This sounds like something that could be done with a materialized view, or am I missing something?
The point of this algorithm is sort of to construct a materialized view, but how it's done is the cool part. It's not a simple database query -- it reduces the number of dimensions into a manageable number.
Indeed, the core contribution (section 1.2) is the cool part.