- https://arxiv.org/pdf/2205.09707
- Source: arxiv
TL;DR
PLAID is a serving engine for ColBERT that uses centroid interaction as a cheap pre-filter to prune irrelevant passages, achieving 4–7x latency reduction with no measurable loss in retrieval quality.
Summary
PLAID (Performance-optimized Late Interaction Driver) is a serving engine for ColBERT that eliminates the bottleneck of full MaxSim computation over entire corpora. It introduces centroid interaction as a cheap pre-filter: query tokens are scored against cluster centroids first, pruning passages unlikely to be relevant before the expensive per-token scoring step. A subsequent candidate generation stage further narrows the set. The result is a 4–7x speedup over vanilla ColBERT with no measurable loss in retrieval quality, enabling late interaction at billion-scale.
Key contributions
- Introduces centroid interaction as a two-stage pre-filter that scores query tokens against cluster centroids before full MaxSim.
- Achieves 4–7x latency reduction versus vanilla ColBERT at identical recall levels on MS-MARCO.
- Integrates directly with ColBERTv2’s residual-compressed index without requiring reindexing.
- Demonstrates that late interaction retrieval is viable at billion-scale corpora in production settings.
When to cite
- When discussing how to make late interaction retrieval practical at production scale and billion-document corpora.
- When applying centroid-based pruning as a transferable pattern for speeding up multi-vector retrieval systems.
- When arguing that indexing and serving co-design can unlock deployment of more expressive retrieval models.
- When evaluating ColBERT-based RAG pipelines where retrieval latency is a bottleneck.
LLM AI VDB RAG ColBERTv2 - Effective and Efficient Retrieval via Lightweight Late Interaction