The traditional multi-stage search funnel—where products are painstakingly dragged through recall filters, coarse ranking, and fine ranking—has officially reached an architectural dead end. Engineers at Taobao and Tmall Group (Alibaba), led by Jianbo Zhu, have concluded that the classic pipeline is far too fragmented for modern e-commerce. Their latest research details a transition to generative search, which collapses every stage into a single, end-to-end neural network model.
Instead of hunting for unique identifiers (SKUs) for millions of items, the new CQ-SID (Category-and-Query constrained Semantic ID) architecture encodes products into hierarchical semantic clusters using the RQ-VAE method. Effectively, search is transformed into a machine translation task: a natural language query is directly converted into a structured cluster ID.
Implementing this at scale is primarily a battle against hallucinations. As Jing Wang’s team explains, they introduced a four-stage progressive learning strategy, moving from simple product-to-ID mapping to personalized user-query-SID refinement. To keep the model grounded in business logic, Alibaba developed EG-GRPO (Expert-Guided Group Relative Policy Optimization). This reinforcement learning technique stabilizes the process by incorporating ground-truth reference samples into gradient groups.
This approach prevents the ‘click and impression collapse’ typical of standard optimization, forcing the generator to suggest high-probability purchases rather than just popular informational noise. Taobao’s infrastructure is already seeing the benefits: by optimizing beam search width, they have radically slashed latency and computational costs without sacrificing results. This isn't just a cosmetic upgrade; it is proof that generative search is viable in high-load environments.
Alibaba has essentially turned the search engine from a passive filter into a predictive generator. The industry’s future now officially lies in semantic clusters rather than rigid database queries. This architecture will likely become the benchmark for any platform managing millions of daily catalog updates—provided their engineers have the expertise to use Reinforcement Learning to ground generative models in the reality of sales.