Workflow
生成式搜索
icon
Search documents
让搜索“一步到位”!快手提出端到端生成式搜索方案OneSearch
Zhi Tong Cai Jing· 2025-09-23 11:45
当前,电商平台普遍采用"召回、粗排、精排"的级联式搜索架构。该架构虽然成熟稳定,但仍面临诸多痛点:商品描述混乱、相关性问题突出、级联结构存 在瓶颈以及冷启动难题,导致搜索结果往往不尽如人意。 为了解决这些挑战,快手(01024)提出了业界首个工业级部署的电商搜索端到端生成式框架——OneSearch,目前,该系统已在快手多个电商搜索场景中成功 部署,每日服务数百万用户,产生数千万页面浏览量。 订单量、买家数等多项指标实现大幅提升 离线实验表明,OneSearch相比现有级联式系统各项指标均有显著提升。在线部署结果更为突出:订单量提升3.22%,买家数增长2.4%。这是在大规模工业 场景下,这是生成式模型首次在大规模工业场景中完整替代传统搜索链路,具备重要落地意义。 | Method | Item CTR | PV CTR | PV CVR | Buy | | --- | --- | --- | --- | --- | | MCA w/o ranking | -9.97% | -20.33% | -11.55% | -28.78 | | OneSearch4 | -1.10% | -2.06% | +0.39% ...
让搜索“一步到位”!快手(01024)提出端到端生成式搜索方案OneSearch
智通财经网· 2025-09-23 11:25
智通财经APP获悉,当前,电商平台普遍采用"召回、粗排、精排"的级联式搜索架构。该架构虽然成熟稳定,但仍面临诸多痛点:商品描述混乱、相关性问 题突出、级联结构存在瓶颈以及冷启动难题,导致搜索结果往往不尽如人意。 在关键词增强层次量化编码(KHQE)模块中,采用RQ-OPQ编码方案,从纵向与横向两个维度建模商品特征,为每个商品生成具备丰富语义层次的"智能身份 证",极大提升生成式检索的区分能力和准确性。 多视角用户行为序列注入策略则让OneSearch能有效捕捉用户的近期偏好与长期兴趣,基于用户的长短期行为序列构建具有区分性的用户标识。使系统实现 更全面、深层的用户意图理解,显著提升个性化搜索准确性与用户体验。 偏好感知奖励系统(PARS)是结合多阶段监督微调与自适应奖励强化学习机制,捕捉细粒度用户偏好信号。该机制在提升排序性能的同时,保障生成多样 性,有效避免"奖励破解"问题。 订单量、买家数等多项指标实现大幅提升 离线实验表明,OneSearch相比现有级联式系统各项指标均有显著提升。在线部署结果更为突出:订单量提升3.22%,买家数增长2.4%。这是在大规模工业 场景下,这是生成式模型首次在大规模工业场景 ...
速递|百度高管再现变动,T11级别搜索首席架构师离职,暂未确定下一步去向
Z Finance· 2025-03-24 09:50
Core Viewpoint - The article discusses the recent departure of Baidu's Chief Architect Gu Simiao and the implications of executive changes within the company, highlighting the trend of talent migration from Baidu to ByteDance and the strategic shifts in Baidu's organizational structure [1][2]. Group 1: Departure of Key Personnel - Gu Simiao, Baidu's Chief Architect, has recently left the company, holding the T11 position, which is a significant technical rank within Baidu [1]. - Gu Simiao has been a pivotal figure in Baidu's AI strategy, leading the development of key technologies such as knowledge graphs and recommendation systems, and has contributed to increasing the search satisfaction rate to 58% [1]. - His departure may signal a potential shift towards AI entrepreneurship, as he has been a strong advocate for generative search technologies [1]. Group 2: Executive Restructuring - Baidu has been undergoing significant executive changes, with a focus on high-level personnel rotations to enhance business collaboration and strategic focus [2]. - The former CFO, Luo Rong, has transitioned to lead the Mobile Ecosystem Group (MEG), integrating AI technology with search and health services [2]. - The Health Group has been merged into MEG to leverage B2B data service potential, indicating a strategic pivot towards commercial efficiency [2]. Group 3: Talent Migration to ByteDance - Several high-profile executives from Baidu have moved to ByteDance, including four CEO-level individuals, indicating a trend of talent acquisition that could impact Baidu's competitive position [2][3]. - Notable figures such as Yang Zhenyuan and Hong Dingkun, who held significant roles at Baidu, are now in key positions at ByteDance, showcasing the influence of Baidu's former talent on ByteDance's growth [3][4]. - The article lists various executives who have transitioned from Baidu to ByteDance, emphasizing the depth of talent that has migrated and its potential implications for both companies [3][4].