"Make the search 'one-stop'! Kuaishou (01024) proposes an end-to-end generative search solution OneSearch."

Currently, e-commerce platforms generally adopt a cascade-style search architecture of "recall, coarse ranking, fine ranking." Although this architecture is mature and stable, it still faces many challenges: confusing product descriptions, prominent relevancy issues, bottlenecks in the cascade structure, and cold start problems, leading to search results that are often unsatisfactory. To address these challenges, Kuaishou (01024) has proposed the industry's first industrial-grade deployed e-commerce search end-to-end generative framework - OneSearch. Currently, the system has been successfully deployed in multiple e-commerce search scenarios at Kuaishou, serving millions of users daily and generating tens of millions of page views. Breaking the traditional architecture and proposing innovative solutions The OneSearch framework integrates three major innovations: the Keyphrase Hierarchical Quantitative Encoding (KHQE) module, the Multi-Perspective User Behavior Sequence Injection Strategy, and the Preference-aware Reward System (PARS). In the Keyphrase Hierarchical Quantitative Encoding (KHQE) module, the RQ-OPQ encoding scheme is utilized to model product characteristics in both vertical and horizontal dimensions, generating an intelligent "ID card" with rich semantic levels for each product, greatly enhancing the discriminative ability and accuracy of generative retrieval. The Multi-Perspective User Behavior Sequence Injection Strategy allows OneSearch to effectively capture users' recent preferences and long-term interests, building distinctive user identities based on users' short-term and long-term behavior sequences. This enables the system to achieve a more comprehensive and deep understanding of user intent, 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 diversity, effectively avoiding the "reward cracking" problem. Significant improvements in metrics such as order volume and number of buyers Offline experiments show that OneSearch significantly outperforms existing cascade systems in various metrics. The online deployment results are even more prominent: order volume increased by 3.22%, and the number of buyers grew by 2.4%. This marks the first time a generative model has completely replaced traditional search links in a large-scale industrial setting, with significant implications. In manual evaluations, OneSearch not only performs excellently in CVR and CTR but also significantly outperforms traditional cascade systems in overall page satisfaction, product quality, and query-item relevance. Additionally, the system's online performance is impressive: machine compute efficiency (MFU) has improved by a factor of 8, online inference costs (OPEX) have decreased by 75.40%, and resource utilization has been greatly optimized. Of particular note, OneSearch performs exceptionally well in cold start scenarios, outperforming conventional scenarios, demonstrating that generative retrieval models can effectively handle sorting challenges posed by long-tail users and newly listed products. The successful deployment of OneSearch marks a major breakthrough in replacing traditional links with generative models in large-scale industrial search scenarios, pointing the way for the future development of e-commerce search technology. Multiple technical breakthroughs by the relevant team have been published in top international conferences such as RecSys, CIKM, and KDD, attracting widespread industry attention. Moving forward, Kuaishou will continue to explore online real-time encoding schemes to narrow the gap between predefined encoding and streaming training. Additionally, stronger reinforcement learning mechanisms will be introduced to more accurately match user preferences, and multimodal product features such as images and videos will be incorporated to further enhance model inference and user experience. With ongoing technological iterations, the future of e-commerce search will become more intelligent, accurate, and personalized, allowing users to truly achieve the ideal search experience of "getting it right in one step."