Retrieval Augmented Generation with Collaborative Filtering for Personalized Text Generation

CFRAG adapts collaborative filtering principles from recommender systems into Retrieval-Augmented Generation (RAG), enabling LLMs to leverage histories of similar users—not just the target user—for improved personalized text generation.

Problem Statement

Existing personalized RAG methods only retrieve from a single user's history, ignoring the rich signal available from similar users' histories. This limits personalization quality, especially for users with sparse interaction histories. The work addresses the lack of collaborative inter-user information in RAG-based LLM personalization pipelines.

Key Novelty

First framework to integrate collaborative filtering into personalized RAG, enabling cross-user history retrieval for LLM generation
Contrastive learning-based user embedding training to identify similar users without requiring explicit user similarity labels
LLM-feedback-driven fine-tuning of a personalized retriever and reranker that jointly considers user preferences and generation quality

Evaluation Highlights

Validated on the LaMP (Language Model Personalization) benchmark, showing effectiveness of CFRAG over existing personalized RAG baselines
Ablation analysis confirms that collaborative information (cross-user retrieval) provides measurable gains beyond single-user RAG baselines

Breakthrough Assessment

6/10 CFRAG is a solid, well-motivated contribution that bridges recommender systems and RAG in a novel way, but it is an incremental architectural improvement rather than a paradigm shift; the core techniques (contrastive learning, reranking, LLM feedback) are individually established.

Methodology

Train user embeddings via contrastive learning using interaction histories as supervision signals, enabling retrieval of similar users without explicit similarity labels
Retrieve top-k documents from the combined histories of the current user and identified similar users using a personalized retriever that encodes user preference context
Rerank retrieved documents with a personalized reranker, then fine-tune both retriever and reranker using feedback from the LLM's generation quality as a reward signal

System Components

User Embedding Module

Encodes users into a shared embedding space using contrastive learning on interaction histories to enable similarity-based user retrieval without labeled pairs

Collaborative User Retriever

Identifies top-N similar users based on embedding similarity, expanding the candidate document pool beyond the target user's own history

Personalized Document Retriever

Retrieves top-k documents from the expanded user pool, conditioned on the current user's preference representation and query context

Personalized Reranker

Reorders retrieved documents by relevance to both the query and the user's personal preferences to surface the most generation-supportive documents

LLM Feedback Fine-tuning

Uses the downstream LLM generation quality as a feedback signal to iteratively fine-tune the retriever and reranker, aligning retrieval with generation needs

Results

Benchmark	Best Baseline (Personalized RAG)	CFRAG	Delta
LaMP (overall)	Competitive personalized RAG baseline	Best reported on LaMP tasks	Positive improvement across tasks
Collaborative info ablation	Single-user RAG (no collaborative info)	CFRAG with collaborative filtering	Meaningful gain confirming collaborative signal utility

Key Takeaways

When building personalized LLM systems, incorporating histories from similar users (collaborative filtering) can meaningfully improve generation quality beyond single-user RAG, especially for sparse-history users
Contrastive learning is a practical way to learn user similarity without needing explicit labels, making CFRAG deployable in real-world settings where user similarity ground truth is unavailable
Closing the loop between retrieval and generation via LLM feedback fine-tuning is an effective strategy to align retrieved context with what actually helps downstream generation, and is worth adopting in RAG system design

Abstract

Recently, the personalization of Large Language Models (LLMs) to generate content that aligns with individual user preferences has garnered widespread attention. Personalized Retrieval-Augmented Generation (RAG), which retrieves relevant documents from the user's history to reflect their preferences and enhance LLM generation, is one commonly used approach for personalization. However, existing personalized RAG methods do not consider that the histories of similar users can also assist in personalized generation for the current user, meaning that collaborative information between users can also benefit personalized generation. Inspired by the application of collaborative filtering in recommender systems, we propose a method called CFRAG, which adapts Collaborative Filtering to RAG for personalized text generation. However, this presents two challenges: (1) how to incorporate collaborative information without explicit user similarity labels? (2) how to retrieve documents that support personalized LLM generation? For Challenge 1, we use contrastive learning to train user embeddings to retrieve similar users and introduce collaborative information. For Challenge 2, we design a personalized retriever and reranker to retrieve the top-k documents from these users' histories. We take into account the user's preference during retrieval and reranking. Then we leverage feedback from the LLM to fine-tune the personalized retriever and reranker, enabling them to retrieve documents that meet the personalized generation needs of the LLM. Experimental results on the Language Model Personalization (LaMP) benchmark validate the effectiveness of CFRAG. Further analysis confirms the importance of incorporating collaborative information.