Crate ruvector_matryoshka 

Matryoshka-aware coarse-to-fine vector search for RuVector.

Three ANN search variants that trade distance-computation cost against recall:

Variant	Stages	Distance ops
FullDim	Single HNSW at full dim	All at D dims
TwoStage	Coarse HNSW + full-dim rerank	Traverse at D1, rerank at D
ThreeStage	Coarse → mid → full-dim funnel	Traverse D1, filter D2, rerank D

The primary metric is recall@k: fraction of the true top-k (found by brute-force at full dimension) that each variant recovers.

Modules

dataset

Deterministic synthetic dataset generator simulating Matryoshka embedding structure.

hnsw

Minimal HNSW graph parameterized by a working dimension.

Structs

FullDimIndex

Standard HNSW search at the full embedding dimension. Baseline for both recall and latency.

MatryoshkaConfig

ThreeStageIndex

Three-stage funnel: coarse_dim → mid_dim filter → full_dim rerank.

TwoStageIndex

Coarse HNSW at coarse_dim, then full-dim rerank of the candidate set.

Traits

Searcher

Functions

brute_force_knn

Brute-force exact top-k at full dimension (ground truth).

l2_normalize

L2-normalise a vector in-place.

l2_sq

prefix_project

Return the first dim elements, L2-normalised.

recall_at_k

Recall@k: fraction of ground-truth top-k that appear in the result set.