GenFacet: End-to-End Generative Faceted Search via Multi-Task Preference Alignment in E-Commerce

Researchers from JD.com have introduced GenFacet, an industrial-grade generative AI framework that reimagines faceted search for massive e-commerce catalogs. Deployed on JD.com's platform, it tackles the limitations of traditional rule-based systems, which struggle with new vocabulary and semantic gaps. GenFacet reframes the problem as two coupled generative tasks within a single Large Language Model: Context-Aware Facet Generation, which dynamically creates relevant navigation options, and Intent-Driven Query Rewriting, which translates user clicks into precise search queries to improve retrieval.

To align the model's generative capabilities with real-world search utility, the team developed a novel multi-task training pipeline. This combines teacher-student distillation with GRPO (Group Relative Policy Optimization), directly optimizing the model for downstream user satisfaction metrics. The system was rigorously validated through offline evaluations and large-scale online A/B tests on China's largest self-operated e-commerce platform, JD.com.

The results were substantial. Online deployment showed a relative increase of 42.0% in facet Click-Through Rate (CTR) and a 2.0% lift in User Conversion Rate (UCVR). These metrics provide strong evidence that an end-to-end generative approach can significantly enhance query understanding and user engagement in production information retrieval systems, moving beyond static taxonomies to a dynamic, AI-powered navigation experience.