Get near cross-encoder accuracy at bi-encoder speed with ColBERT late interaction (MUVERA + RAGatouille)

ML Journey·medium signal

ColBERT stores a per-token embedding for every query and document token and scores relevance as the sum of each query token's max similarity (MaxSim) against the document — so phrases match the specific passage where they're relevant instead of competing in one pooled vector. The 2026 production recipe uses RAGatouille for indexing/search and a MUVERA+ approach: fast single-vector candidate generation followed by exact ColBERT MaxSim reranking over the top-K. The tradeoff is storage — multi-vector indexes are an order of magnitude larger — so apply it as a reranker, not the primary index.