Currently, e-commerce platforms commonly adopt a cascaded search architecture of "recall, coarse ranking, and fine ranking." While this architecture is mature and stable, it still faces numerous pain points: chaotic product descriptions, prominent relevance issues, bottlenecks in cascaded structures, and cold start challenges, leading to often unsatisfactory search results. To address these challenges, KUAISHOU-W (01024) has proposed the industry's first industrial-grade deployed e-commerce search end-to-end generative framework—OneSearch. Currently, this system has been successfully deployed across multiple e-commerce search scenarios on Kuaishou, serving millions of users daily and generating tens of millions of page views.
**Breaking Traditional Architecture with Innovative Solutions**
The OneSearch framework integrates three major innovations: the Keyword-enhanced Hierarchical Quantized Encoding (KHQE) module, multi-perspective user behavior sequence injection strategy, and Preference-Aware Reward System (PARS).
In the Keyword-enhanced Hierarchical Quantized Encoding (KHQE) module, the RQ-OPQ encoding scheme is adopted to model product features from both vertical and horizontal dimensions, generating an "intelligent ID card" with rich semantic layers for each product, significantly improving the discrimination capability and accuracy of generative retrieval.
The multi-perspective user behavior sequence injection strategy enables OneSearch to effectively capture users' recent preferences and long-term interests, constructing distinctive user identities based on users' short-term and long-term behavior sequences. This allows the system to achieve more comprehensive and deep user intent understanding, significantly improving personalized search accuracy and user experience.
The Preference-Aware Reward System (PARS) combines multi-stage supervised fine-tuning with adaptive reward reinforcement learning mechanisms to capture fine-grained user preference signals. This mechanism improves ranking performance while ensuring generation diversity and effectively avoiding "reward hacking" problems.
**Significant Improvements in Orders, Buyers, and Multiple Metrics**
Offline experiments show that OneSearch achieves significant improvements across all metrics compared to existing cascaded systems. Online deployment results are even more outstanding: order volume increased by 3.22% and buyer count grew by 2.4%. This marks the first time a generative model has completely replaced traditional search pipelines in large-scale industrial scenarios, demonstrating significant practical importance.
In human evaluation, OneSearch not only excels in CVR and CTR but also significantly outperforms traditional cascaded systems in overall page satisfaction, product quality, and query-item relevance. Additionally, the system's online performance is equally impressive: machine efficiency utilization (MFU) improved by 8 times relatively, online inference cost (OPEX) decreased by 75.40%, and resource utilization efficiency was greatly optimized.
Particularly noteworthy is OneSearch's exceptional performance in cold start scenarios, showing significantly better results than regular scenarios, demonstrating that generative retrieval models can effectively address ranking challenges for long-tail users and newly listed products.
The successful deployment of OneSearch marks a major breakthrough for generative models replacing traditional pipelines in large-scale industrial search scenarios, pointing the direction for future development of e-commerce search technology. Multiple technical breakthroughs from the related team have been published at top international conferences including RecSys, CIKM, and KDD, attracting widespread industry attention.
Moving forward, Kuaishou will continue exploring online real-time encoding solutions to reduce the gap between predefined encoding and streaming training. The company will also introduce more powerful reinforcement learning mechanisms to match user preferences more precisely and incorporate multimodal product features such as images and videos to further enhance model inference effects and user experience.
With continuous technological iteration, future e-commerce search will become more intelligent, precise, and "understanding," enabling users to truly achieve the ideal "one-step complete" search experience.