K-Means Clustering

K-Means clustering for vector embeddings (not database clustering).

Documentation

• K-Means Algorithm - Algorithm details and how clustering speeds up vector search
• Metal Optimizations - Apple Silicon-specific kernel notes

What is K-Means Clustering?

K-Means clustering groups similar vectors together. NornicDB uses it as an inverted-file approximate nearest-neighbor (IVF) candidate generator: queries first identify the most relevant clusters, then perform exact similarity scoring within those clusters. This is faster than full brute-force search on large datasets.

Quick Start

K-means clustering is off by default. Enable it through the feature flag:

export NORNICDB_KMEANS_CLUSTERING_ENABLED=true
./nornicdb serve

The cluster index is built automatically once the dataset reaches NORNICDB_KMEANS_MIN_EMBEDDINGS (default 1000). Below that threshold, brute-force search is used because it is faster on small data. The number of clusters is chosen automatically as K ≈ √(N/2) (capped at 8192) unless you override it via NORNICDB_KMEANS_NUM_CLUSTERS.

For the user-facing tuning surface, see Vector Search → Performance Tuning.

Learn More

• K-Means Algorithm — How K-Means works in NornicDB's IVF candidate generator
• Metal Optimizations — Atomic-float workaround for the Metal kernel
• Vector Search — User-facing vector search guide that consumes the cluster index
• GPU Acceleration — Backend selection (Metal/CUDA/Vulkan) used by the kernel

---

Get started → K-Means Algorithm