← Builder galleryReference pattern
ColBERT late-interaction retrieval
Per-token MaxSim scoring instead of one pooled vector — retrieval that survives catalog drift.
per-token
query/document interaction (MaxSim)
no re-embed
on catalog copy changes
edge-sized
runs beside the model, not a vector DB SaaS
The recipe
Embed every token of the query and document, score with MaxSim (each query token takes its best-matching document token), sum. Late interaction keeps token-level signal that single-vector embeddings average away — misspellings, rare attributes, and exact identifiers keep their weight.
Honest rough edge: index size grows with tokens, not documents. For catalog-scale corpora, quantize the token embeddings and prune stopword positions before comparing footprint with single-vector systems.
When to use it
RAG over catalogs or documents that change weekly, where re-embedding the corpus on every drift is the hidden cost of the pooled-vector approach.