Shopify 商业引擎背后的生成式推荐器

A buyer's journey through Shopify isn't a click. It's a sequence—searches, views, add-to-carts, favorites, purchases—spread across storefronts and the Shop app, stretching back months. That sequence carries meaning in its order, timing, and gaps between events.

买家在 Shopify 的购物之旅不是一次点击。它是一个序列——searches、views、add-to-carts、favorites、purchases——分布在店面和 Shop app 中，延伸回数月。这个序列在其顺序、时机和事件之间的间隙中蕴含意义。

A good recommendation reads that sequence and predicts what comes next. At Shopify's scale (millions of products and billions of events), getting this right requires a model that can work with the full sequence, not just a summary of it.

一个好的推荐系统读取该序列并预测下一个项目。在 Shopify 的规模（数百万产品和数十亿事件）下，做到这一点需要一个能够处理完整序列的模型，而不仅仅是其摘要。

This is how we built a foundational generative recommender: a system that learns directly from raw event sequences, operates within real production latency constraints, and delivers measurable impact at scale. We'll cover the architecture decisions, the training strategies that actually moved the needle, and what happened when we shipped it.

这就是我们构建基础生成式推荐系统的方式：一个直接从原始事件序列中学习、在真实生产延迟约束内运行，并在规模上交付可衡量影响的系统。我们将介绍架构决策、真正见效的训练策略，以及我们发布它时发生了什么。

The scale of the opportunity

机会的规模

The Black Friday Cyber Monday weekend is the clearest reminder that recommendations at Shopify are built in an environment defined by scale. During BFCM 2025, Shopify observed 2.2 trillion edge requests, and more than 81 million consumers bought from Shopify-powered brands. That customer count is based on unique buyer emails associated with purchases made from Shopify merchants over the BFCM weekend.

Black Friday Cyber Monday 周末是最清楚的提醒，表明 Shopify 的推荐是在规模定义的环境中构建的。在 2025 BFCM 期间，Shopify 观察到 2.2 万亿 edge requests，超过 8100 万消费者从 Shopify-powered brands 购买。该客户数量基于 BFCM 周末从 Shopify 商家购买关联的唯一买家电子邮件。

Those numbers describe the amount of signal available to learn from, and they also describe the engineering constraints of serving recommendations in real time. If you want recommendations that feel consistently relevant, you need a model that can read that stor...