Generative retrieval replaces embeddings with semantic barcodes

What it does

This repo implements the two-stage generative retrieval approach from Recommender Systems with Generative Retrieval. First, an RQ-VAE tokenizes every item in the corpus into a short tuple of discrete semantic IDs. Then a decoder-only transformer is trained on user interaction histories—represented as sequences of those frozen semantic IDs—to predict the next item directly, bypassing conventional embedding-based nearest-neighbor retrieval.

The interesting bit

The model sidesteps vector databases entirely by treating items as structured barcodes rather than dense vectors. Retrieval becomes pure autoregressive generation over a compact vocabulary of semantic tokens, collapsing the usual search problem into next-token prediction.

Key highlights

• Supports Amazon Reviews (Beauty, Sports, Toys) and MovieLens (1M and 32M) out of the box with no manual dataset downloads.
• Two separate training pipelines: train_rqvae.py learns the item tokenizer, while train_decoder.py fits the retrieval model using a frozen RQ-VAE.
• Hyperparameters are managed via gin-config, with sample .gin files provided for each supported dataset.
• A pretrained RQ-VAE checkpoint for Amazon Beauty is hosted on Hugging Face.

Caveats

• Only one pretrained checkpoint (Amazon Beauty) is currently available on Hugging Face.
• The README does not report retrieval metrics or benchmark comparisons against standard baselines.
• The tokenizer and retrieval model are trained separately; the README does not describe a joint end-to-end training mode.

Verdict

Worth cloning if you are researching generative retrieval or semantic-ID alternatives to ANN-based recommenders. Pass if you need a production-hardened stack with published offline metrics.