Find exact nearest neighbors in relatively high dimensional spaces. Supports in-memory and out-of-core data sets (via bcolz and bvec).

Gives realtime performance in 20-100 dimensional feature spaces, over hundreds of thousands of items.

Includes the following similarity measures

• cosine
• jaccard
• generalized

The generalized similarity measure is based on an alternate normalization of cosine similarity, and includes both cosine similarity and lift as special cases.

• efficient calculation of a relevant subset of Bregman Divergences
• subsetting of feature vectors for inclusion in results with boolean vectors (carrays)