Late Interaction
A neural search architecture where queries and documents are encoded independently, and interaction between them happens only at the very final scoring stage.
Think of It Like This
Like a blind taste test where the chef and the critic prepare in total isolation and only meet at the moment of tasting.
Unlike early interaction models (like cross-encoders) that process query and document together at high computational cost, late interaction models (like ColBERT) keep them separate until the end. This allows document embeddings to be pre-computed and cached. It provides a sweet spot combining the speed of bi-encoders with the accuracy of cross-encoders.