Scalable Search System Boosts Item Discovery via LLM-Generated Intents

A novel approach to enhancing search capabilities in e-commerce, especially in the grocery sector, utilizes intent-conditioned recall expansion to boost item visibility. This two-tiered hybrid framework initially leverages closed-weight large language models (LLMs) for primary queries to optimize discoverability. Subsequently, a fine-tuned small language model (SLM), enhanced with LoRA adapters and teacher-student distillation, extends its advantages to less common queries. This method tackles the balance between cost and quality in generative retrieval by generating implicit user intents to broaden candidate recall while ensuring relevance. The research can be accessed on arXiv.