Module plaid 

PLAID-style centroid-based candidate pruning for multi-vector search.

Builds K-means centroids over all document vectors. Each document is encoded as a sorted bag of centroid IDs. At query time the query’s centroid bag is computed and only documents whose centroid bag overlaps the query bag are returned as candidates.

Reference: Santhanam et al., “PLAID: An Efficient Engine for Late Interaction Retrieval”, CIKM 2022.

Structs

PlaidPruner

PLAID centroid-based candidate pruner.