Module maxsim 

High-performance MaxSim scoring for late-interaction (ColBERT) workflows.

This module provides optimized CPU implementations of MaxSim scoring using:

• SIMD instructions (AVX2 on x86_64, NEON on ARM) for fast max reduction
• BLAS-accelerated matrix multiplication via ndarray (when accelerate or openblas features enabled)

Credits

The SIMD optimization techniques in this module are adapted from the maxsim-cpu which provides high-performance MaxSim computation for ColBERT-style late interaction models.

Platform Support

• macOS ARM: Uses NEON SIMD + Apple Accelerate (with accelerate feature)
• Linux x86_64: Uses AVX2 SIMD + OpenBLAS (with openblas feature)
• Other platforms: Falls back to scalar operations

Functions

assign_to_centroids

Assign embeddings to their nearest centroids using batched GEMM.

maxsim_score

Compute MaxSim score for a single query-document pair.