This repository contains a resource of an open resource for vision and text-enhanced Retrieval-Augmented Generation (RAG) in the recommendation domain. RAG-VisualRec provides a new, reproducible test‑bed for multimodal RAG research. It is designed as a transparent, modular, and extensible resource for rigorous multimodal recommendation research, with a primary goal of bridging the gap between theoretical advances (e.g., fusion techniques, textual data augmentation, multi-modal retrieval, augmented generation) and practical, reproducible workflows that any researcher can adapt or extend.
The overall pipeline of RAG-VisualRec contains below steps:
- Data preparation and Ingestion, starting with MovieLens (
latest-smalland1M) dataset - Multimodal embedding extraction (including textual, visual, and audio)
- Applying fusion strategies (e.g., concatenation, PCA, CCA)
- Applying user embedding construction (random, averag, or temporal)
- Candidate retrieval
- Profile augmentation and LLM prompting (manual or LLM-based)
- Evaluation and logging (accuracy, beyond-accuracy, fairness/robustness)
The whole framework is fully configurable through a centralized parameter block in the Google Colab notebook. Check the codes directory for more information.
This research is submitted to ACM Transactions on Recommender Systems.
@article{tourani2025rag,
title={RAG-VisualRec: An Open Resource for Vision-and Text-Enhanced Retrieval-Augmented Generation in Recommendation},
author={Tourani, Ali and Nazary, Fatemeh and Deldjoo, Yashar},
journal={arXiv preprint arXiv:2506.20817},
year={2025}
doi={https://doi.org/10.48550/arXiv.2506.20817}
}This project is licensed under the GPL-3.0 license - see the LICENSE for more details.
If you have any questions or collaboration opportunities, please open an issue or contact the authors.