Taobao's TSGR boosts e-commerce search with value-aware generative retrieval

Taobao (Alibaba) researchers have published TSGR (Taobao Search Generative Retrieval), a novel framework that addresses a critical gap in industrial e-commerce search: existing generative retrieval systems optimize for semantic matching but ignore item business value. TSGR introduces two key innovations: Query-aware Parallel SID (QP-SID) constructs parallel codebooks from query-item statistics to assign better token indices to higher-value, query-relevant items, and a Value-aware Ranking Module (VRM) that combines retrieval and pre-ranking into a single model, eliminating a dedicated pre-ranking stage. A progressive training pipeline aligns the model with semantic relevance, user preferences, and business objectives.

Offline experiments demonstrate a 9.16% improvement in HR@1000, and online A/B tests on Taobao's live traffic validate real-world impact: IPV (item page views) increased by 0.43%, transaction count grew by 1.12%, and GMV (gross merchandise value) rose by 1.64%. These results show that incorporating value awareness directly into the retrieval process can significantly boost business metrics without sacrificing relevance. The paper, authored by Tianyu Zhan and 11 others, was submitted to arXiv on July 21, 2026, and is currently under review for publication.