Module matryoshka 

Matryoshka adaptive-dim querying.

For embedding models trained with Matryoshka Representation Learning (MRL), prefix-truncated vectors retain semantic structure. We exploit this for memory-bandwidth-bound HNSW: coarse pass at low dim → exact rerank at full dim.

Cache-efficiency

A 256-dim coarse pass reads 256 × 4 = 1 KiB per vector vs 1536 × 4 = 6 KiB for full-dim traversal — a ≥6× reduction in L1/L2 memory bandwidth during the dominant HNSW graph traversal phase.

Supported models

• text-embedding-3 (OpenAI): [256, 512, 1024, 1536]
• Gemini Embedding: [256, 512, 768, 3072]
• Nomic Embed: [64, 128, 256, 512, 768]

Structs

MatryoshkaSearchOptions

Options for a two-stage Matryoshka search.

MatryoshkaSpec

Per-collection Matryoshka configuration.

Functions

matryoshka_search

Two-stage Matryoshka search.

truncate

Stride-truncate a vector to its first dim components.