Concentric mapping
Project embeddings to polar shell + angle coordinates and map order-preserving keys onto a Chord ring.
Sapana Micro Software
Concentric Vector Chord
A kernelized vector distributed hash table that preserves semantic locality across Euclidean and kernel-defined embedding spaces—without a centralized vector index.
LSH / KLSH buckets
Similarity-preserving signatures for cosine, L2, and RBF kernel spaces—including implicit RKHS embeddings.
Chord routing
Finger-table overlay routes queries to responsible nodes with successor overflow for hot buckets.
Key findings
KLSH tightens neighborhoods
Config B (RBF + KLSH) clusters top-8 neighbors above 0.94 similarity, versus a sparser cosine LSH decay in Config A.
Concentric clustering needs overflow
Without Chord successor overflow, bucket saturation failed near 300 vectors on 24 nodes. Overflow stabilized insertion to 512 vectors.
Sub-100 ms simulation queries
In-process init and k-NN query for 256–512 vectors completed in under 100 ms on Apple Silicon (seed 42).
Vulkan validates placement
Interactive visualization confirmed shell guides, ring topology, vector projections, and routing hop paths during development.
Prototype research
This site presents the published technical summary and benchmark observations. Implementation source is proprietary and not distributed here.
View architecture overview →