PLAID: An Efficient Engine for Late Interaction Retrieval

Late interaction's query-time MaxSim scan over full-resolution document vectors was too slow for interactive search at scale.

Treats every passage as a lightweight BAG OF CENTROIDS (cluster centroids over its token vectors) rather than its full token vectors, and prunes the centroid bags. Centroid interaction plus centroid pruning inside a highly optimised engine eliminates low-scoring passages early, so the full-resolution MaxSim only runs on survivors.

Reduces late-interaction search latency by up to 7x on a GPU and 45x on a CPU versus vanilla ColBERTv2, while preserving state-of-the-art retrieval quality. Baseline (vanilla ColBERTv2) and hardware class (GPU and CPU separately) are both stated - this is one of the few retrieval papers with a CPU number, which matters because CPU-only retrieval serving is common. At 140M passages it reports tens of ms on GPU.

VENUE: Semantic Scholar records CIKM 2022 ('International Conference on Information and Knowledge Management'); crossref did not surface a proceedings entry, so 'other' is used rather than guessing at 'kdd'. THE MOST IMPORTANT IDEA IN THIS SLICE FOR A SERVING ENGINEER: late interaction is a PRUNING problem, not a compute problem. Approximate candidate scoring first, exact MaxSim on survivors - the same coarse-to-fine pattern as speculative decoding, and the reason a 45x CPU speedup is possible when the tail of the score distribution carries most of the answer. It also shows why quantized or centroid-compressed indexes are a practical requirement rather than a nice-to-have at 140M passages. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. Peer-reviewed venue confirmed as CIKM 2022 (same). This is an additive track annotation; `venue` is unchanged at 'other' and no status change is implied by the track alone. VENUE RECLASSIFIED 2026-10-10 (crawler/w11-venue-enum): venue 'other' -> 'cikm', venue_kind 'conference'. Confirmed by record notes and agreeing publisher/index deposit in data/sources/.