Hybrid LLM-SLM Architecture for E-Commerce Intent-Conditioned Search Expansion
July 30, 2026
A two-stage retrieval system improves e-commerce item discoverability by generating implicit user intents to expand recall. The architecture uses closed-weight LLMs for head queries and a LoRA-finetuned small language model (SLM) to maintain cost-effective performance on tail queries.
HOW THIS AFFECTS YOU
●
builderYou can implement this two-stage pattern to balance generative retrieval quality with inference costs.
●
founderThis provides a roadmap for increasing marketplace revenue through improved product discoverability.