PLAID-PRF: Pseudo-Relevance Feedback with Centroid-like Tokens in PLAID

Multi-vector dense retrieval models like ColBERT provide high accuracy by matching individual query tokens to document tokens, but they are computationally expensive. While PLAID reduces this cost through centroid-based quantization, researchers have sought ways to further improve retrieval quality using Pseudo-Relevance Feedback (PRF). The authors investigate whether they can perform effective PRF within the PLAID framework without the heavy computational overhead of traditional clustering or additional model inference.

PLAID-PRF treats the centroid codes already stored in the PLAID index as proxies for document tokens. The process follows four steps:

1. Initial Retrieval: A standard PLAID search retrieves a small set of top-ranked documents.
2. Usefulness Scoring: The system identifies informative tokens by calculating the frequency of centroid codes within the feedback documents, weighted by their global rarity (similar to TF-IDF).
3. Expansion: A diverse set of high-utility centroid-based vectors is selected and appended to the original query.
4. Final Retrieval: The expanded query is used to perform a second, more precise retrieval pass.

By reusing the precomputed centroid codebook, the method avoids the need for expensive online document-token clustering or extra transformer inference, making it significantly faster than existing PRF techniques for multi-vector models.

Experiments on the MSMARCO and BEIR benchmarks demonstrate that PLAID-PRF consistently outperforms standard PLAID. The method achieves notable gains in both nDCG@10 and MRR@10 while maintaining low query latency. Because it operates directly on the existing index structure, it provides a more efficient Pareto-optimal trade-off between retrieval effectiveness and computational cost compared to previous PRF approaches like ColBERT-PRF or CWPRF.