Provides isolated, reusable vector operations including L2 normalization and cosine similarity. These functions are dependency-light (only NumPy), well-tested, and form the foundation for all similarity computations in the library. The module exemplifies clean, focused utility code that could be extracted into any project requiring vector processing.

Defines the core data structures: Strategy enum, Metric enum, and DiversificationResult namedtuple. These types are completely self-contained, have zero internal dependencies, and provide a clear, extensible contract between the library's API and its strategies. The use of typed namedtuples ensures immutability and clarity, making this module ideal for reuse in other algorithmic libraries requiring structured results.

Contains the greedy_select function, a reusable implementation of the greedy selection algorithm used by both MMR and MSD strategies. This function abstracts the core iterative selection logic, accepting a scoring function and selection size, demonstrating how to encapsulate a common algorithmic pattern. Its isolation (depends only on NumPy) and clear interface allow it to be repurposed for other selection-based optimization problems with minimal changes.