OneSearch: A Preliminary Exploration of the Unified End-to-End Generative Framework for E-commerce Search

The evolution of e-commerce search systems has progressed from simple keyword-based matching to complex multi-stage cascading architectures. Traditional methods like EBR, DCN, etc., have made significant advances in recall and ranking stages but still face issues of fragmented computation and objective collision. Recently, Generative Retrieval (GR) has emerged as a promising paradigm, treating large-scale retrieval as sequence generation tasks, showing potential superior to traditional methods.

Traditional e-commerce search systems employ multi-stage cascading architectures, balancing computational efficiency and business conversion but limited by fragmented computation and optimization objective collisions. Additionally, these systems perform poorly in handling cold-start queries and long-tail sessions, struggling to accurately capture user intent.

OneSearch addresses traditional issues through three core innovations: 1) Keyword-enhanced Hierarchical Quantization Encoding (KHQE) module preserves hierarchical semantics and unique attributes, ensuring strong query-item relevance; 2) Multi-view User Behavior Sequence Injection strategy constructs behavior-driven user IDs, integrating short-term and long-term sequences to model user preferences comprehensively; 3) Preference-Aware Reward System (PARS) captures fine-grained user preferences through multi-stage supervised fine-tuning and adaptive reward-weighted ranking.

• �� Keyword-enhanced Hierarchical Quantization Encoding (KHQE): Highlights core attributes of items through keyword-enhanced semantic collaborative encoding. Uses RQ-Kmeans for hierarchical feature encoding and OPQ for unique feature quantization of each item. • Multi-view User Behavior Sequence Injection strategy: Introduces weighted decay click behavior sequence into user ID to construct a distinctive user representation, explicitly incorporating short behavior sequences in prompts to learn recent user preferences and implicitly including long behavior sequences to model the user profile. • Preference-Aware Reward System (PARS): Designs a multi-stage supervised fine-tuning process for semantic alignment and personalization, followed by an adaptive reward system leveraging hierarchical user behavior signals and list-wise preference optimization.

Experiments were conducted on large-scale industry datasets, using baselines including traditional multi-stage cascading architectures. Key metrics included item CTR, buyer volume, and order volume. Experiments also involved ablation studies to assess the contribution of each component. OneSearch demonstrated excellent performance in offline evaluations, with online A/B tests further validating its effectiveness in practical applications.

OneSearch demonstrated superior performance in offline evaluations, with online A/B tests showing significant improvements at the same exposure position: +1.67% item CTR, +2.40% buyer, and +3.22% order volume. Furthermore, OneSearch reduces operational expenditure by 75.40% and improves Model FLOPs Utilization from 3.26% to 27.32%. These results indicate OneSearch's superior performance in high-quality recall and ranking.

OneSearch has been successfully deployed across multiple search scenarios in Kuaishou, serving millions of users and generating tens of millions of PVs daily. Its application in e-commerce search demonstrates the potential of generative retrieval, paving the way for the development of smarter search systems.

OneSearch still faces performance limitations in handling cold-start queries and long-tail sessions, as user intent and item information are harder to capture in these scenarios. Additionally, the system may be misled by lengthy, noisy item information, affecting the accuracy of representation models. Future research will focus on optimizing performance in these scenarios.