HNSW Explorer

Concentric spheres: inner core = entry point (sparse) → outer shell = base layer (dense)

INDEX

Build Params

Set at construction time. Requires rebuild to change.

M (connections/node)Maximum bidirectional links per node. Higher M means better recall but more memory and slower builds. Layer 0 uses 2xM connections, upper layers use M. Typical range: 4-48.4

ef_constructionBeam width used when building the index. Higher values produce a better-connected graph at the cost of slower index construction. Typical range: 64-512.100

Vector count100K

RecallThe percentage of true nearest neighbors found by the approximate search compared to an exact brute-force search. 100% recall means the approximate search found all the same results as exact search.

~28%

Build

Fast

MRLMatryoshka Representation Learning allows truncating embeddings to fewer dimensions while preserving search quality. Lower dimensions reduce storage and speed up queries.

Dimensions

Dimension truncation. Lower dims = less storage, faster search, slight accuracy loss.

Full dimensions3072

MRL truncation1536

Used: 50%Savings: 50%

Quality

~100%

Drag to orbit · Scroll to zoom · Right-drag to pan