Struct CentroidSelector 

pub struct CentroidSelector;

Phase 1: Select top centroids per query token.

For each query token, compute its similarity with all centroids and select the top-nprobe centroids above the score threshold.

Implementations

impl CentroidSelector

pub fn select( query: &MultiVectorEmbedding, centroids: &[f32], dim: usize, config: &WarpSearchConfig, ) -> Vec<Vec<(usize, f32)>>

Select top centroids for each query token.

Arguments

• query - Query multi-vector embedding
• centroids - Flattened centroid vectors [num_centroids × dim]
• dim - Token embedding dimension
• config - Search configuration

Returns

For each query token, a vector of (centroid_id, centroid_score) pairs sorted by score descending.

pub fn batch_scores( query_token: &[f32], centroids: &[f32], dim: usize, ) -> Vec<(usize, f32)>

Batch compute centroid scores for a single query token.

Returns scores for all centroids sorted by score descending.