ColBERTv2: Effective and Efficient Retrieval via Lightweight Late Interaction

Late interaction's multi-vector representations inflate the model's space footprint by an order of magnitude, limiting index size.

Aggressive RESIDUAL COMPRESSION - dimensionality reduction plus sparse outlier retention, discarding the dimensions that carry little relevance signal - combined with a denoised supervision strategy that trains against false negatives. The query is also optionally truncated.

Establishes state-of-the-art quality within and outside the training domain while reducing the SPACE FOOTPRINT of late-interaction models by 6-10x. The baseline is ColBERT and the metric is index size, which is the memory-bound dimension that matters for serving.

Venue NAACL 2022 (crossref: ACL, 'Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics'). This is the paper that made late interaction deployable: 6-10x index reduction is the difference between a ColBERT index that fits and one that does not. Combined with a dimensionality-aware query, ColBERTv2 is the reference configuration for any late-interaction deployment, and it is the checkpoint PLAID is built against. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass): venue_track set to 'main' from the ACL Anthology bib index (anthology.bib, 131650 entries) by normalized title match. Evidence: anthology id 2022.naacl-main.272, booktitle 'Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies'. The Anthology id is the authoritative signal here and it is unambiguous, so this is an additive clarification of a track, not a change of venue - `venue` is untouched and no status change is implied. The reason this matters for inference readers: a Findings paper and a main-track paper at the same venue are not the same kind of evidence, and until this field existed the repo could not say which was which.