Source: Santhanam et al. 2022: PLAID

Paper describing PLAID, the ColBERTv2 search engine that prunes candidates using centroids, cutting latency up to 7x on GPU and 45x on CPU.

Contents

Summary

Exploits the centroids already present in ColBERTv2’s compressed index: each passage is first treated as a cheap “bag of centroids”, scored against the query, and pruned, so the expensive residual decompression and full MaxSim run only on a small survivor set.

Key claims

  • Centroid interaction quickly discards low-scoring passages; centroid pruning sparsifies each passage’s bag of centroids.
  • Up to 7x lower latency on GPU and 45x on CPU versus vanilla ColBERTv2, without quality loss.
  • Tens of milliseconds on GPU and tens to a few hundred milliseconds on CPU at 140M passages.

Feeds

Late-interaction retrieval

Suggest changes →