AIGQ: An End-to-End Hybrid Generative Architecture for E-commerce Query Recommendation

Alibaba researchers have unveiled AIGQ (AI-Generated Query), a novel end-to-end hybrid generative architecture designed to revolutionize search query recommendations on Taobao, China's largest e-commerce platform. The system specifically targets the platform's HintQ feature, which provides search suggestions on the homepage. Traditional methods relied heavily on ID-based matching and co-click heuristics, leading to shallow semantics and poor performance for new users or items. AIGQ represents the first fully generative framework built to overcome these limitations through three core technical innovations.

First, the team developed Interest-Aware List Supervised Fine-Tuning (IL-SFT), a list-level training approach that aggregates user session behavior and uses an interest-guided re-ranking strategy to create more nuanced training samples. Second, they designed Interest-aware List Group Relative Policy Optimization (IL-GRPO), a novel policy gradient algorithm. This component uses a dual-reward mechanism to optimize both individual query relevance and the overall quality of the suggestion list, with rewards informed by the platform's live click-through rate (CTR) model.

For deployment, the architecture is split into two pathways to meet Taobao's strict latency requirements. AIGQ-Direct handles nearline, personalized user-to-query generation, while AIGQ-Think acts as a reasoning-enhanced variant that creates trigger-to-query mappings to boost the diversity of suggestions. According to the paper, extensive offline evaluations and large-scale online A/B tests on Taobao have demonstrated that AIGQ delivers substantial improvements in key business metrics related to both platform effectiveness and user engagement, marking a significant step beyond traditional retrieval-based recommendation systems.