HNSW Explorer

Real HNSW links; shells separate hierarchy levels, so distances across shells are not vector distances.

INDEX

Build Params

Seeded incremental HNSW construction on the rendered 3D vectors.

M (connections/node)Neighbors selected at insertion. Pruning caps stored outgoing links at 2×M on layer 0 and M above. Higher M uses more memory; links need not stay reciprocal after pruning.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

Sizing vector count100K

Rendered vectors96

M and ef_construction rebuild the real demo graph. Sizing vector count affects the memory estimates below; it does not add millions of scene points.

MRLOnly MRL-trained embedding models support useful dimension truncation. The quality trade-off must be measured on a retrieval dataset.

Storage sizing

Lower dimensions reduce storage. Search quality depends on the embedding model and dataset.

Full dimensions3072

MRL truncation1536

Used: 50%Savings: 50%

Byte estimates only. The scene uses actual 3D vectors. Truncation quality needs an MRL-compatible model and retrieval tests; no quality percentage is inferred here.

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

HNSW Explorer

Real HNSW links; shells separate hierarchy levels, so distances across shells are not vector distances.

INDEX

Build Params

Seeded incremental HNSW construction on the rendered 3D vectors.

M (connections/node)Neighbors selected at insertion. Pruning caps stored outgoing links at 2×M on layer 0 and M above. Higher M uses more memory; links need not stay reciprocal after pruning.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

Sizing vector count100K

Rendered vectors96

M and ef_construction rebuild the real demo graph. Sizing vector count affects the memory estimates below; it does not add millions of scene points.

MRLOnly MRL-trained embedding models support useful dimension truncation. The quality trade-off must be measured on a retrieval dataset.

Storage sizing

Lower dimensions reduce storage. Search quality depends on the embedding model and dataset.

Full dimensions3072

MRL truncation1536

Used: 50%Savings: 50%

Byte estimates only. The scene uses actual 3D vectors. Truncation quality needs an MRL-compatible model and retrieval tests; no quality percentage is inferred here.

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