Walmart's INSPIRE uses LLMs to decode shopper intent for better product matches

Walmart, which dominates the U.S. e-commerce grocery market, faces a persistent challenge: search queries are short, ambiguous, and often miss explicit dietary or size preferences. This mismatch between user intent and retrieved products can cost engagement and ad revenue at scale. In a new paper, researchers introduce INSPIRE (Intent-aware Neural Sponsored Product Retrieval for E-commerce), a retrieval framework that explicitly models intent as a set of structured, multi-dimensional attributes—covering explicit signals like brand and flavor, and implicit ones like dietary constraints or cuisine type.

INSPIRE first uses a large language model (LLM) as a teacher to generate structured intent labels from product titles and descriptions via weak supervision. These labels then fine-tune a compact student LLM using LoRA-based supervised finetuning, making it lightweight enough for production. The predicted intents are incorporated into a biencoder dense retrieval system, enriching both query and product representations. The result: sponsored product retrieval that better understands what shoppers actually want, even when the query doesn't say it. Accepted to ACM SIGIR E-commerce Workshop 2026.