PLAID-PRF boosts ColBERT retrieval by 7.3% with lightweight feedback

Multi-vector dense retrieval models like ColBERT achieve strong effectiveness by modeling fine-grained token-level interactions between queries and documents. Methods such as PLAID use centroid-based quantization to reduce index size and speed up retrieval. Now, researchers from the University of Glasgow and elsewhere introduce PLAID-PRF, a pseudo-relevance feedback technique that reformulates ColBERT's query vectors using the top-retrieved results. The key innovation: it reuses PLAID's internal centroid vectors as expansion tokens, avoiding costly document-token clustering. This allows the system to select a small, diverse set of high-utility vectors and append them to the original query, then rerun PLAID for refined candidate generation and scoring.

Extensive experiments on in-domain MSMARCO and four out-of-domain BEIR datasets show consistent improvements. PLAID-PRF improves retrieval effectiveness by up to 4.3% in nDCG@10 and 7.3% in MRR@10 over standard PLAID, while introducing substantially less computation than prior PRF methods. The work, accepted at SIGIR 2026, demonstrates that centroid-aware pseudo-relevance feedback offers an effective and lightweight mechanism to improve top-ranked results without sacrificing speed. This makes feedback-aware late-interaction retrieval practical for real-world search systems.