Zhi Tong Cai Jing·2025-09-23 11:45Core Insights - Kuaishou has introduced OneSearch, an end-to-end generative framework for e-commerce search, addressing challenges in traditional search architectures [1][2][4] Group 1: Innovation and Technology - OneSearch integrates three innovations: Keyword Enhanced Hierarchical Quantization Encoding (KHQE), Multi-Perspective User Behavior Sequence Injection Strategy, and Preference-Aware Reward System (PARS) [2] - KHQE uses RQ-OPQ encoding to model product features, creating a "smart identity" for each product, enhancing retrieval accuracy [2] - The Multi-Perspective User Behavior Sequence Injection Strategy captures both recent preferences and long-term interests, improving user intent understanding [2] - PARS combines multi-stage supervised fine-tuning with adaptive reinforcement learning to capture fine-grained user preference signals, enhancing ranking performance while ensuring diversity [2] Group 2: Performance Metrics - OneSearch has shown significant improvements in various metrics compared to traditional systems, with a 3.22% increase in order volume and a 2.4% growth in the number of buyers [4] - In offline experiments, OneSearch outperformed existing systems in terms of Click-Through Rate (CTR) and Conversion Rate (CVR), with notable improvements in user satisfaction and item quality [5][6] - The system achieved an 8-fold increase in machine computation efficiency and a 75.40% reduction in online inference costs, optimizing resource utilization [5] Group 3: Market Impact and Future Directions - OneSearch's deployment marks a significant breakthrough in replacing traditional search links with generative models in large-scale industrial scenarios [4][6] - The system excels in cold start scenarios, effectively addressing challenges related to long-tail users and newly listed products [6] - Kuaishou plans to explore online real-time encoding solutions and enhance reinforcement learning mechanisms to better match user preferences, aiming for a more intelligent and precise e-commerce search experience [6]