MultiVectorEncoder supports ColBERT-style late interaction retrieval, loading any PyLate checkpoint, any Stanford-NLP ColBERT checkpoint, and colpali-engine models through a single API. Unlike dense embedding models that compress text into one vector, multi-vector models retain per-token vectors and score using the MaxSim operator, preserving token-level matching for stronger retrieval. The tradeoff is larger index size, though compression methods like PLAID bring storage costs in line with standard dense indexes.
No score is assigned. Sources and their independence are shown in the citation chain below.