Product catalogs are the backbone of search, discovery, and recommendation in e‑commerce, yet they are often attribute‑sparse: the attributes needed by shoppers and downstream systems are hidden in unstructured fields such as titles and images or are missing entirely. Manual enrichment is infeasible given the scale and rapid growth of catalogs. This paper introduces TRACE, an agentic Large Language Model (LLM) framework for automated attribute enrichment.
TRACE consists of two agents. The ScoutAgent triangulates multimodal evidence from merchant catalogs, syndicated feeds, and identity‑matched web search, proposing candidate attribute values together with supporting evidence. The JudgeAgent then verifies each candidate by assessing evidence consistency and confidence, deciding whether to publish the value directly or route it for human review.
On an offline human‑evaluation dataset, TRACE achieved 98.2% accuracy with 74.7% attribute coverage. Deployed on an industry‑scale catalog, it increased impression‑weighted enrichment coverage across four business verticals by 90.4%. An online experiment further showed that surfacing the enriched attributes on product detail pages raised checkout conversion by 0.48%.
Blogger's Review: TRACE demonstrates the practical power of LLMs in e‑commerce metadata management. Its multi‑source evidence grounding and agentic verification provide a scalable, low‑cost solution for large‑scale catalog maintenance.