OneBar boosts e-commerce video searches 17-19% with generative queries

Short-video e-commerce platforms like Douyin or TikTok now show clickable search boxes under videos, letting users immediately search for products they see. But traditional query recommendation systems struggle with latency and misaligned objectives, while generative models often fail due to noisy video metadata and shifting user preferences. To solve this, a team from Alibaba (including Yao Tang, Ying Yang, and others) introduced OneBar, an end-to-end generative framework for real-time query recommendations. OneBar has three key innovations: (1) a collaborative-multimodal intent grounding module that combines video understanding with behavior-derived collaborative anchors; (2) a unified architecture with prompt compression for efficient online serving; and (3) a progressive preference learning strategy that internalizes hierarchical behavior preferences directly into the generative policy, eliminating the need for a separately trained reward model.

Deployed in a live e-commerce video feed, OneBar significantly improved user engagement and conversion. Compared to the online baseline, the framework increased Query Exposure by 16.91% and Query Click by 18.68%, while maintaining a slight CTR gain of 0.19%. Importantly, the additional search traffic translated directly to business outcomes: guided orders rose 20.36% and GMV grew 21.67%. These results demonstrate that grounding generative query recommendations in both video content and user behavior can bridge the gap between passive watching and active shopping, making OneBar a powerful tool for any short-video commerce platform looking to boost search-driven revenue.