RQ-VAE-Recommender

Why we included this project

Most recommender systems still rely on embedding-based two-tower or matrix-factorization designs. This project implements a different idea from the paper "Recommender Systems with Generative Retrieval": an RQ-VAE first compresses every item in the corpus into a short tuple of semantic IDs, then a decoder-only transformer learns to predict the next ID in a user's sequence, which is effectively the next item they are likely to engage with. The repo includes training scripts for both stages, gin-config files for tuning, and KMeans initialization for the quantizer, and it runs on Amazon Reviews and MovieLens datasets with no manual download step. Pretrained RQ-VAE checkpoints for Amazon Beauty are published on Hugging Face, so you can test the tokenizer before committing to a full training run. If you are researching or prototyping generative retrieval, this is a good place to start reading and adapting the approach.