ClusterRAG: Cluster-Based Collaborative Filtering for Personalized Retrieval-Augmented Generation

Gibson Nkhata; Uttamasha Anjally Oyshi; Quan Mai; Susan Gauch

cs.IRApr 14, 2026

ClusterRAG: Cluster-Based Collaborative Filtering for Personalized Retrieval-Augmented Generation

Authors: Gibson Nkhata, Uttamasha Anjally Oyshi, Quan Mai, Susan Gauch

Organizations: University of Arkansas, Fayetteville, AR 72701, USA · Walmart Inc., Bentonville, AR 72716, USA

Abstract

Personalized Retrieval-Augmented Generation (RAG) relies on accurately selecting user-relevant documents. In practice, existing RAG approaches often suffer from high retrieval costs and overlook that collaborative signals from similar users can enhance personalized generation for the current user. We propose ClusterRAG, a Cluster-Based Collaborative Filtering for Personalized Retrieval-Augmented Generation. ClusterRAG represents users through their profile documents, organizes users into semantically coherent clusters using density-based clustering, and performs retrieval at both the cluster and document levels via cluster-level similarity and fine-grained ranking. Extensive experiments on the LaMP benchmark demonstrate that jointly leveraging the target user's profile and profiles from top similar users consistently yields the best performance across diverse tasks. Further analysis shows that ClusterRAG integrates seamlessly with different dense retrievers and rankers, and remains effective when paired with both fine-tuned and zero-shot language models.

Figures & tables

Figure 1: Overview of the ClusterRAG framework.

Task	#users	#train	#dev	#test
LaMP-1	6542	6542	1500	1500
LaMP-2	929	5073	1410	1557
LaMP-3	20000	20000	2500	2500
LaMP-4	1643	12500	1500	1800
LaMP-5	14682	14682	1500	1500
LaMP-7	13437	13437	1498	1500

Table 1: Statistics of the LaMP benchmark with time-based data split.

Models	LaMP-1		LaMP-2		LaMP-3		LaMP-4		LaMP-5		LaMP-7
	Acc. $\uparrow$	F1 $\uparrow$	Acc. $\uparrow$	F1 $\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	R-1 $\uparrow$	R-L $\uparrow$	R-1 $\uparrow$	R-L $\uparrow$	R-1 $\uparrow$	R-L $\uparrow$
vanillaRAG	0.630	0.630	0.520	0.440	0.371	0.709	0.171	0.154	0.462	0.413	0.310	0.273
LaMP-IPA	0.674	0.664	0.570	0.522	0.289	0.608	0.175	0.169	0.472	0.423	0.508	0.457
ROPG	0.644	0.322	0.468	0.031	0.346	0.692	0.184	0.163	0.464	0.396	0.353	0.288
CFRAG	0.633	0.327	0.534	0.036	0.354	0.707	0.162	0.141	0.473	0.425	0.375	0.306
ClusterRAG-C	0.674 ∗	0.673 ∗	0.644	0.607	0.284	0.624	0.179	0.157	0.480 ∗	0.430 ∗	0.507	0.454

Table 2: Comparison of the performance of ClusterRAG with baselines on the LaMP benchmark. Best results are shown in bold and second-best in underlined ; boldface indicates statistically significant improvements over the second-best (

p<0.05

). The symbol ∗ denotes the second-best ClusterRAG variant.

Retrievers	LaMP-1		LaMP-2		LaMP-7
	Acc. $\uparrow$	F1 $\uparrow$	Acc. $\uparrow$	F1 $\uparrow$	R-1 $\uparrow$	R-L $\uparrow$
Random	0.640	0.639	0.609	0.608	0.500	0.449
Recency	0.659	0.650	0.618	0.610	0.507	0.456
BM25	0.662	0.658	0.629	0.621	0.510	0.460
Contriever	0.681	0.681	0.649	0.623	0.511	0.459
BGE	0.684	0.682	0.658	0.613	0.509	0.461

Table 3: Comparison of the performance of ClusterRAG under different retrievers on LaMP-(1,2,7).

Panel of Figure 2 — Figure 2: Retrivers’ ROGUE-1 scores on LaMP-5.

LLMs	LaMP-1		LaMP-2		LaMP-5
	Acc. $\uparrow$	F1 $\uparrow$	Acc. $\uparrow$	F1 $\uparrow$	R-1 $\uparrow$	R-L $\uparrow$
nFlan	0.546	0.540	0.451	0.448	0.431	0.398
pFlan	0.648	0.647	0.601	0.601	0.484	0.432
nQwen2	0.602	0.600	0.521	0.521	0.457	0.419
pQwen2	0.639	0.635	0.610	0.606	0.488	0.447

Table 4: Performance comparison of ClusterRAG using different LLMs on LaMP-(1,2,5).

Derivatives	LaMP-3		LaMP-7
	MAE $\downarrow$	RMSE $\downarrow$	R-1 $\uparrow$	R-L $\uparrow$
w/o user clustering	0.320	0.637	0.458	0.371
w/o intra-cluster sim	0.329	0.639	0.501	0.442
w/o doc ranking	0.331	0.642	0.462	0.413
Centroids only	0.400	0.643	0.472	0.438
$k$ -means	0.291	0.610	0.502	0.453

Table 5: Ablation study of ClusterRAG on LaMP-(3,7).

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Task	#users	#train	#dev	#test	Input Length	Output Length	#Profile Size	#classes
LaMP-1	6542	6542	1500	1500	51.43 $\pm$ 5.70	–	84.15 $\pm$ 47.54	2
LaMP-2	929	5073	1410	1557	92.39 $\pm$ 21.95	–	86.76 $\pm$ 189.52	15
LaMP-3	20000	20000	2500	2500	128.18 $\pm$ 146.25	–	185.40 $\pm$ 129.30	5
LaMP-4	1643	12500	1500	1800	29.97 $\pm$ 12.09	10.07 $\pm$ 3.10	204.59 $\pm$ 250.75	–
LaMP-5	14682	14682	1500	1500	162.34 $\pm$ 65.63	9.71 $\pm$ 3.21	87.88 $\pm$ 53.63	–
LaMP-7	13437	13437	1498	1500	29.72 $\pm$ 7.01	16.96 $\pm$ 5.67	15.71 $\pm$ 14.86	–

Appendix

Table 6: Detailed statistics of the LaMP benchmark with time-based data split.

Hyperparameter	Tested values
Learning rate	5 $\times 10^{-5}$ , $3\times 10^{-3}$ , $10^{-3}$ , $10^{-4}$
Weight decay	5 $\times$ $10^{-6}$ , 10 -4 , $10^{-3}$
Warm-up ratio	0.05 to 0.10
Batch size	8, 16, 32, 64
Epochs	10, 20 30, 50, 70, 100
Max seq length	64, 128, 256, 512

Appendix

Table 7: Hyperparameter tuning grid used for training and optimizing ClusterRAG.

$k$ value	LaMP-1		LaMP-2		LaMP-3		LaMP-4		LaMP-5		LaMP-7
	Acc. $\uparrow$	F1 $\uparrow$	Acc. $\uparrow$	F1 $\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	R-1 $\uparrow$	R-L $\uparrow$	R-1 $\uparrow$	R-L $\uparrow$	R-1 $\uparrow$	R-L $\uparrow$
1	0.690	0.690	0.661	0.620	0.270	0.594	0.190	0.176	0.490	0.440	0.521	0.470
2	0.692	0.697	0.665	0.631	0.269	0.582	0.193	0.179	0.496	0.444	0.524	0.469
3	0.700	0.700	0.668	0.642	0.270	0.595	0.196	0.180	0.495	0.445	0.528	0.468
4	0.690	0.688	0.660	0.621	0.272	0.595	0.192	0.178	0.497	0.441	0.523	0.469
5	0.689	0.690	0.658	0.619	0.273	0.596	0.191	0.177	0.492	0.438	0.521	0.467

Appendix

Table 8: Effect of the number of similar users (

k

) on ClusterRAG performance in hybrid mode across LaMP tasks.

$m$ value	LaMP-1		LaMP-2		LaMP-3		LaMP-4		LaMP-5		LaMP-7
	Acc. $\uparrow$	F1 $\uparrow$	Acc. $\uparrow$	F1 $\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	R-1 $\uparrow$	R-L $\uparrow$	R-1 $\uparrow$	R-L $\uparrow$	R-1 $\uparrow$	R-L $\uparrow$
1	0.674	0.673	0.648	0.612	0.289	0.608	0.182	0.166	0.480	0.431	0.515	0.464
2	0.690	0.690	0.661	0.620	0.270	0.594	0.190	0.176	0.490	0.440	0.521	0.470
3	0.691	0.691	0.665	0.632	0.269	0.591	0.191	0.178	0.501	0.465	0.528	0.474
4	0.693	0.692	0.669	0.639	0.268	0.590	0.196	0.179	0.510	0.471	0.530	0.473
5	0.695	0.695	0.675	0.650	0.267	0.586	0.198	0.180	0.512	0.473	0.533	0.475

Appendix

Table 9: Effect of the number of retrieved profile documents (

m

) on ClusterRAG performance in hybrid mode across LaMP tasks.

Task	HDBSCAN Score	$k$ -means Score
LaMP-1	0.601	0.389
LaMP-2	0.535	0.326
LaMP-3	0.551	0.328
LaMP-4	0.570	0.274
LaMP-5	0.562	0.347
LaMP-7	0.537	0.323

Appendix

Table 10: Silhouette scores of user clusters produced by ClusterRAG using HDBSCAN and

k

-means on the LaMP benchmark.

User Query:

Movie description : When the Davison family comes under attack during their wedding anniversary getaway, the gang of mysterious killers soon learns that one of their victims harbors a secret talent for fighting back.

Gold Output: Violence

Target User Profile:

(1) Movie tag : “twist ending”, Movie description : “Soon after his insufferably arrogant father wins the Nobel Prize for chemistry, Barkley Michaelson is kidnapped by Thaddeus James, a young genius who claims to be Barkley’s illegitimate half-brother. Motivated not so much by money as revenge, Thaddeus tries to convince Barkley to help him carry out a multimillion-dollar extortion plot against their patriarch.”

(2) Movie tag : “action”, Movie description : “The Bride unwaveringly continues on her roaring rampage of revenge against the band of assassins who had tried to kill her and her unborn child. She visits each of her former associates one-by-one, checking off the victims on her Death List Five until there’s nothing left to do …but kill Bill.”

(3) Movie tag : “action”, Movie description : “NYPD cop John McClane’s plan to reconcile with his estranged wife is thrown for a serious loop when, minutes after he arrives at her office, the entire building is overtaken by a group of terrorists. With little help from the LAPD, wisecracking McClane sets out to single-handedly rescue the hostages and bring the bad guys down.”

Appendix

Table 11: A case study illustrating how ClusterRAG leverages user profile and similar-user information for personalized RAG.

Density-Based Clustering

See all 15 papers

Rethinking Retrieval-Augmented Generation as a Cooperative Decision-Making Problem
Feb 21, 2026Lichang Song, Ting Long, Yi ChangRetrieval-Augmented Generation RAG Robustness
Jilin University
SproutRAG: Attention-Guided Tree Search with Progressive Embeddings for Long-Document RAG
Jun 16, 2026Amirhossein Abaskohi, Issam H. Laradji, Peter West +1Document Chunking Retrieval-Augmented Generation
University of British Columbia · ServiceNow Research
Uncertainty-Aware Hybrid Retrieval for Long-Document RAG
Jun 11, 2026Hoin Jung, Xiaoqian WangRetrieval-Augmented Generation Information Retrieval
Elmore Family School of Electrical and Computer Engineering Purdue University West Lafayette, IN 47907

Feb 21, 2026J/K move · Enter open · S save

Rethinking Retrieval-Augmented Generation as a Cooperative Decision-Making Problem

Lichang Song, Ting Long, Yi Chang

Jilin University

Retrieval-Augmented Generation RAG Robustness LLM Reranking

Open paper AskPDF arXiv

Retrieval-Augmented Generation (RAG) has demonstrated strong effectiveness in knowledge-intensive tasks by grounding language generation in external evidence. Despite its success, many existing RAG systems are built based on a ranking-centric, asymmetric dependency paradigm, where the generation quality of the generator is highly dependent on reranking results of the reranker. To overcome this limitation, we propose Cooperative Retrieval-Augmented Generation (CoRAG), a framework that treats the reranker and the generator as peer decision-makers rather than being connected through an asymmetric dependency pipeline. By jointly optimizing their behaviors toward a shared task objective, the reranker and generator are encouraged to cooperate, ensuring that document reranking and generation work in concert to improve the final response. Experimental results demonstrate good generalization and improved generation stability of CoRAG, even when the model is trained on only around 10K PopQA samples. Our model released in https://github.com/CoderrrSong/CoRAG.

Abstract

Density-Based Clustering

Explore similar work

Rethinking Retrieval-Augmented Generation as a Cooperative Decision-Making Problem

SproutRAG: Attention-Guided Tree Search with Progressive Embeddings for Long-Document RAG

Uncertainty-Aware Hybrid Retrieval for Long-Document RAG