plaid
Module for indexing and searching using PLAID.
This module provides classes and configurations for indexing and searching using PLAID in the Lightning IR framework.
PLAID (Performance-optimized Late Interaction Driver) is a optimized search engine designed to drastically accelerate multi-vector models like ColBERT without sacrificing their state-of-the-art accuracy. While late interaction architectures provide precision by comparing every query word to every document word, doing this exhaustively across massive databases is computationally slow. PLAID solves this by grouping document token vectors into a smaller set of representative “centroids” during the indexing phase. When a search occurs, the system first performs a rapid filter by matching the query against these lightweight centroid IDs, pruning away irrelevant documents. It then only executes the expensive, exact token-level mathematical comparisons on the small handful of highly relevant candidate passages that remain.
PLAID: PLAID: An Efficient Engine for Late Interaction Retrieval
Modules
Plaid Indexer using fast-plaid library for Lightning IR Framework |
|
Plaid Searcher using fast-plaid library for Lightning IR Framework |