ColBERTv2
Late-interaction retrieval: token-level matching with MaxSim scoring
About
ColBERTv2, from Stanford's Future Data Systems lab (Santhanam, Khattab et al.), introduced the late-interaction retrieval paradigm: instead of compressing a document into one vector, it stores an embedding per token and scores queries with MaxSim — the sum over query tokens of each token's best match in the document. The v2 paper added distillation from a cross-encoder plus residual compression that shrinks the index 6-10x while preserving quality, and it set state-of-the-art retrieval results on MS MARCO and BEIR in 2022. It is MIT licensed and remains the foundation for the whole late-interaction family (PLAID, ColPali, ColQwen). The cost is storage: even compressed, token-level indexes are far larger than single-vector ones, which is why most production stacks use it as a reranking layer rather than a first-stage retriever.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
MS MARCO / BEIR SOTA at release (paper-reported, 2022)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: Stanford authorship, late-interaction/MaxSim mechanism, residual compression, MIT license, and arXiv:2112.01488 (v3 Jul 2022) all verified across the paper page, HF card, and a research survey read this session.; 2022 SOTA claims are paper-reported.; Storage cost of token-level indexes stated as the key limitation..