CatalogAgent Improves E-commerce Data with Self-Learning AI.

E-commerce platforms rely heavily on product catalogs, but these often suffer from missing structured attribute values like material or color. While LLM-based generator-evaluator frameworks can predict these values, they face challenges when the generator and evaluator disagree, as both can make errors. A new agentic system called CatalogAgent has been developed to address this, focusing on continuous improvement for e-commerce catalog enrichment. CatalogAgent introduces a Supervisor Agent that intervenes when conflicts arise, either internally between the LLM generator and evaluator or from external seller feedback. This supervisor makes final decisions. The system also incorporates a Memory Base and Summarizer to store and aggregate the Supervisor Agent's activities and learnings. These insights are then fed back into the worker Generator and Evaluator LLMs through "context engineering," enabling them to self-improve without human intervention. Experiments show this approach significantly boosts performance, improving generator accuracy by 15.24% and evaluator accuracy by 13.98%.