WeChat Deploys WeGR Generative Retrieval: Orders Up 2.5%, Compute Down Over 25%
WeChat E-commerce Search deployed the WeGR framework, showing a 2.5% increase in orders and over 25% reduction in CPU/GPU usage in an A/B test.
ImportanceMaterialEvidenceE2 unreplicatedWrite-upQuick
The WeChat E-commerce Search system has deployed a unified generative retrieval framework named WeGR. According to an arXiv preprint submitted on October 12, 2026, a two-month online A/B test demonstrated relative improvements of 2.539% in orders, 1.508% in click-through rate, and 1.381% in conversion rate. Simultaneously, online CPU and GPU usage decreased by 25.5% and 26.1%, respectively.
WeGR addresses challenges in existing generative retrieval methods, specifically the difficulty of incorporating collaborative signals into Semantic ID tokenizers and the sparsity of reinforcement learning rewards. The framework includes the CoCo-VAE tokenizer and Dynamic Horizon Policy Optimization (DHPO). The system is currently deployed across multiple search scenarios in WeChat, serving billions of users with tens of millions of daily page views.