SimilariPy

Fast: powered by Cython and OpenMP for fast parallel execution
Sparse matrix focus: optimized for Scipy CSR/COO formats
Wide similarity coverage: Dot Product, Cosine, Tversky, P3α, RP3β, and more
Flexible normalization: includes L1, L2, TF-IDF, BM25, BM25+

Welcome to SimilariPy’s documentation!

SimilariPy is a Python package that provides high-performance KNN similarity and normalization functions, optimized for sparse matrices.

SimilariPy achieves high performance through several key optimizations:

Cython compilation: Core algorithms compiled to C++ for native-speed execution
OpenMP parallelization: Multi-threaded execution scales with available CPU cores
GIL-free computation: Releases Python's Global Interpreter Lock during heavy computation
CSR-optimized: Direct memory access patterns optimized for Compressed Sparse Row format
Pre-allocated buffers: Reuses row buffers across iterations, eliminating millions of allocations per computation
Float32 precision: Uses 32-bit floats for matrix values instead of 64-bit for 2× memory bandwidth and better SIMD vectorization
Fused types: Single codebase supports both int32/int64 matrix indices without overhead or code duplication
Top-K filtering: Early termination and efficient heap structures minimize unnecessary computation