ClusterRAG Boosts Personalized RAG with Collaborative Filtering

Personalized Retrieval-Augmented Generation (RAG) typically suffers from high retrieval costs and ignores valuable signals from similar users. A new paper by Gibson Nkhata, Uttamasha Anjally Oyshi, Quan Mai, and Susan Gauch introduces ClusterRAG, which applies collaborative filtering to RAG by clustering users based on their profile documents. It uses density-based clustering to form semantically coherent groups, then performs retrieval at both the cluster and document levels—first identifying relevant clusters, then ranking documents within those clusters. This two-tier approach reduces search space while injecting collaborative signals.

Extensive experiments on the LaMP benchmark show that combining the target user’s profile with profiles from top similar users consistently yields the best performance across diverse personalization tasks. ClusterRAG integrates seamlessly with different dense retrievers and rankers, and works effectively with both fine-tuned and zero-shot language models. The paper is accepted at ACL 2026 and published on arXiv. For AI engineers building personalized search or Q&A systems, ClusterRAG offers a practical, efficient way to improve relevance without massive compute overhead.