ClusterRAG: Collaborative Filtering for Personalized RAG
ClusterRAG is a new approach for personalized retrieval-augmented generation that uses cluster-based collaborative filtering. It represents users via profile documents, groups them into semantic clusters with density-based clustering, and retrieves documents at both cluster and document levels. Experiments on the LaMP benchmark show that combining the target user's profile with those of similar users yields the best performance across diverse tasks. The method integrates seamlessly with existing RAG systems.
Key facts
- ClusterRAG is a Cluster-Based Collaborative Filtering for Personalized Retrieval-Augmented Generation.
- It uses density-based clustering to organize users into semantically coherent clusters.
- Retrieval occurs at both cluster and document levels via cluster-level similarity and fine-grained ranking.
- Experiments on the LaMP benchmark show best performance when leveraging profiles from top similar users.
- ClusterRAG integrates seamlessly with different RAG systems.
- The approach addresses high retrieval costs and leverages collaborative signals from similar users.
- The paper is available on arXiv with ID 2605.18769.
- The method represents users through their profile documents.
Entities
Institutions
- arXiv