NumPy is used in rec engines because it has an extremely fast matrix library.
If you're doing non-matrix (or representing a matrix without a 2d array) recommendation engines then NumPy could be completely useless.
I'm doing research on this exact problem at Carnegie Mellon and we are using a graph to do things instead. We aren't using basic techniques like kNN however, so that may have something to do with it. Instead, we have someone who has done heavy research in submodularity and we're using an approximation algorithm to the submodular function optimization problem.
I help Khalid and Carlos with their research (I'm an undergrad). Here's Khalid's thesis proposal. The paper recommendation problem is section 2.2 (Beyond Keyword Search).
