Combining Hashing and Abstraction in Sparse High Dimensional Feature Spaces: Searching Inside