Late interaction
Late interaction: A retrieval architecture comparing multiple token representations after separate query and document encoding. It retains fine detail without full cross-encoding of every pair. Multi-vector indices use more memory than a single embedding. This perspective connects the definition with practical behaviour in artificial intelligence systems.
In italiano: Late InteractionLegge I · Embedding, feature e spazi vettorialiIt retains fine detail without full cross-encoding of every pair. Multi-vector indices use more memory than a single embedding.