Documents can be pre-encoded for indexing, while query-time scoring retains finer token-to-token matches than a single pooled vector. This improves expressiveness at the cost of larger indexes and more scoring work.
Late interaction retrieval compares token-level query and document representations after they have been encoded separately.
Documents can be pre-encoded for indexing, while query-time scoring retains finer token-to-token matches than a single pooled vector. This improves expressiveness at the cost of larger indexes and more scoring work.