arXiv · Jul 29, 2026
Improving Item Discoverability in e-Commerce Search via Related Intent Generation
A paper on arXiv proposes a discovery-augmented search system for e-commerce that uses intent-conditioned recall expansion. It employs a two-stage hybrid architecture: closed-weight LLMs for head queries and a finetuned small language model (SLM) with LoRA adapters and teacher-student distillation for tail queries. Evaluation uses LLM-as-a-judge metrics validated against human preferences and end-to-end session-level purchase analysis.
What happened
Traditional e-commerce search prioritizes precision over recall, limiting discovery of substitute, complementary, and related items. This paper introduces a scalable system that generates implicit user intents to expand candidate recall while maintaining relevance. A two-stage hybrid architecture uses closed-weight LLMs for head queries and a distilled SLM for tail queries, balancing cost and quality. Evaluation combines LLM-as-a-judge metrics with session-level purchase analysis, showing improved discoverability.
Technical significance
The system addresses the cost-quality tradeoff of generative retrieval by combining closed-weight LLMs for high-frequency queries with a LoRA-finetuned SLM distilled via teacher-student training for tail queries. This hybrid approach enables scalable intent generation across the query distribution. LLM-as-a-judge metrics are validated against human preferences, providing a reliable proxy for semantic quality.
Industry impact
E-commerce platforms, especially in grocery, can benefit from discovery-augmented search to increase user satisfaction and commercial outcomes. The hybrid architecture offers a practical path to deploy generative retrieval at scale, balancing performance and cost. The use of distilled small models for tail queries suggests a trend toward efficient, specialized models in production search systems.
What to watch
Next signals include potential adoption by major e-commerce platforms, further refinement of distillation techniques for retrieval, and extension to other domains like video or music recommendation. The approach may also influence how LLM-as-a-judge metrics are standardized for search evaluation.
Decision value
Improved item discoverability can lead to higher conversion rates, larger basket sizes, and better user retention in e-commerce. The hybrid architecture reduces reliance on expensive closed-weight LLMs for all queries, lowering operational costs while maintaining relevance.