OneSearch - Kuaishou's End-to-End Generative Framework for E-commerce Search

What is OneSearch

OneSearch is an end-to-end generative framework for e-commerce search launched by Kuaishou. It optimizes the cascaded architecture of traditional e-commerce search, improving search accuracy and user experience. Its three major innovations include: the Keyword-enhanced Hierarchical Quantization Encoding (KHQE) module, which extracts core product attributes and generates hierarchical codes (SIDs) to strengthen Query-product relevance constraints; the multi-perspective user behavior sequence injection strategy, which constructs behavior-driven user identifiers (UIDs), integrating explicit short-term behaviors with implicit long-term sequences to precisely model user preferences; and the Preference-Aware Reward System (PARS), which combines multi-stage supervised fine-tuning with adaptive reward reinforcement learning to capture fine-grained user preference signals. OneSearch significantly outperforms traditional systems in offline experiments, and in online experiments increases order volume by 3.22% and buyer count by 2.4%, with outstanding performance in long-tail queries and cold-start scenarios for new products.

Main Features of OneSearch

• Precise Matching: Through the Keyword-enhanced Hierarchical Quantization Encoding (KHQE) module, it precisely extracts core product attributes and generates hierarchical codes, significantly improving the discriminative ability and accuracy of generative retrieval. It can understand colloquial, vague, or even incomplete expressions and transform them into efficient shopping instructions.
• Efficient Ranking: It adopts a multi-perspective user behavior sequence injection strategy to construct behavior-driven user identifiers (UIDs), integrating explicit short-term behaviors with implicit long-term sequences to comprehensively and precisely model user preferences and achieve smarter result ranking.
• Cost Optimization: After launch, online inference costs were reduced by 75.4%, and machine computing efficiency increased by 8 times, significantly lowering operational costs.
• Improved User Experience: In manual evaluations, OneSearch significantly outperforms traditional systems in overall page satisfaction, product quality, and query-item relevance. It can more comprehensively understand user intent and significantly improve the accuracy and user experience of personalized search.
• Empowering Merchants: It performs particularly well in cold-start scenarios, with results significantly better than in regular scenarios, indicating that generative retrieval models can more effectively address the ranking challenges of long-tail users and newly listed products.

Technical Principles of OneSearch

• Keyword-enhanced Hierarchical Quantization Encoding (KHQE): By extracting the core attributes of products, such as brand, category, color, material, etc., it generates a hierarchical "smart ID" (SID) for each product, thereby significantly improving the discriminative ability and accuracy of generative retrieval.
• Multi-perspective User Behavior Sequence Injection: It constructs behavior-driven user identifiers (UIDs), integrating explicit short-term behaviors with implicit long-term sequences to comprehensively and precisely model user preferences and achieve smarter result ranking.
• Preference-Aware Reward System (PARS): It combines multi-stage supervised fine-tuning (SFT) with an adaptive reward reinforcement learning mechanism to capture fine-grained user preference signals and enhance the model's personalized ranking ability.

OneSearch Project Address

• arXiv Technical Paper：https://arxiv.org/pdf/2509.03236

Application Scenarios of OneSearch

• E-commerce Search: Through precise matching and efficient ranking, OneSearch significantly improves users' search experience on e-commerce platforms, helping users find the products they need faster.
• Long-tail Product Recommendation: In cold-start scenarios, OneSearch can more effectively handle the ranking problems of long-tail users and newly listed products, improving the exposure and sales opportunities of long-tail products.
• Personalized search: Through a multi-perspective user behavior sequence injection strategy, OneSearch can accurately model user preferences, provide personalized search results, and meet the needs of different users.
• Improve merchant operational efficiency: By optimizing search results, OneSearch helps merchants increase product exposure and conversion rates, improving overall operational efficiency.