cluster-quantized-knn

What problem does it solve?

This Skill reduces expensive pairwise distance computations for KNN scoring by approximating distances using precomputed cluster centroid relationships plus per-point intra-cluster offsets.

Core Features & Use Cases

• Cluster-Quantized O(1) Distance Scoring: Approximates dist(i, j) with centroid-to-centroid distance plus intra-cluster distances for i and j.
• Precomputation Protocol: Builds centroids, intra distances, and a centroid distance matrix once to make runtime lookups fast.
• Two-Path Routing for Correctness: Uses an approximate distance for unfiltered walk-mode traversal and switches to exact distance when filtering breaks cluster validity.
• Walk Mode Scorers: Supports MMR-style diversity reranking and stochastic density-walk selection for exploration.

Quick Start

Use the cluster-quantized-knn skill to precompute centroid and intra-cluster distance tables from your labeled embedding corpus, then route unfiltered queries through the O(1) approximate distance while routing filter-then-rank queries through exact distances over the filtered survivor set.