This project implements a two-stage generative retrieval model for recommendations. First, it trains an RQ-VAE (Residual Quantized Variational Autoencoder) to map items in a corpus to tuples of semantic IDs. Second, it trains a transformer-based decoder on sequences of these semantic IDs to predict the next items in a user interaction sequence. The model supports datasets like Amazon Reviews and MovieLens, with configurable hyperparameters via gin-config.