TSGR: Taobao Search Generative Retrieval

In industrial e-commerce search, generative retrieval (GR) models often struggle to balance semantic relevance with business objectives like conversion rates and GMV. Traditional GR systems treat all items as semantically equal and lack access to rich item side-info during the generation process, often leading to the retrieval of semantically relevant but commercially inferior products. This paper asks: how can we integrate business value awareness directly into a generative retrieval framework to improve downstream business metrics?

To address these limitations, the authors propose Taobao Search Generative Retrieval (TSGR), which introduces two primary innovations:

1. Query-aware Parallel SID (QP-SID): Instead of a single static SID, TSGR constructs multiple parallel SID orderings for each item. These orderings are derived from query-item click statistics, ensuring that items with higher business value for a specific query are assigned to earlier, more frequently visited token indices.
2. Value-aware Ranking Module (VRM): The authors integrate a ranking module directly into the generative backbone. This module uses the model's hidden states and external item side-info (such as behavioral statistics and category features) to score candidates generated during beam search. Because the VRM is jointly optimized with the generative model, it functions as both a retriever and a pre-ranker without requiring a separate, decoupled ranking stage.

These components are supported by a progressive training pipeline that includes multi-task pretraining and supervised fine-tuning with weighted multi-positive supervision to align the model with user preferences and business goals.

TSGR demonstrates that embedding business value into the retrieval process significantly improves performance. Offline experiments on Taobao search data showed a 9.16% improvement in HR@1000. Online A/B testing confirmed the practical impact of the framework, yielding a 0.43% increase in item page views (IPV), a 1.12% increase in transaction counts, and a 1.64% increase in GMV compared to existing baselines.