semantic-ids

[Pytorch] Generative retrieval model using semantic IDs from "Recommender Systems with Generative Retrieval"

Unofficial implementation of paper "Recommender Systems with Generative Retrieval" (NeurIPS'23)

We introduce the direct document relevance optimization (DDRO) for training a pairwise ranker model. DDRO encourages the model to focus on document-level relevance during generation

[Pytorch] Efficient tokenization for recommendations and generative retrieval. Inspired by STATIC decoding from "Vectorizing the Trie"

The simplest, fastest repository for training/finetuning SemanticIDs for Sequential Recommendation.

Official code for "Generative Retrieval for Cover Song Identification via Audio-Derived Semantic IDs" (accepted at ICECCME 2026, IEEE)

A verified SASRec reproduction and a semantic-ID generative recommender on one backbone, one eval harness, one set of frozen negatives. Full-catalog vs. sampled ranking, a measured seed-noise floor, and every README table CI-asserted against the run that produced it.

SASRec с нуля на PyTorch (causal self-attention вручную) на Yandex Yambda. Аудио-эмбеддинги квантованы в semantic IDs через RQ-VAE: как замена item-эмбеддингов проигрывают, как дополнение дают +39% к NDCG@10 (0.575 vs 0.413). Ablation из шести моделей

Semantic Identifier Generative Index Layer: a trie-constrained generative retrieval (DSI-class) system that maps queries directly to document identifiers by constrained autoregressive decoding over frozen residual-quantization semantic IDs. No ANN index in the serving path.