Rethinking E-commerce Search: A Large Language Model and Database Powered Information Retrieval System

The two biggest challenges for e-commerce search are understanding queries and leveraging world knowledge to match queries and products. Large language models are not only powerful in query understanding, but they also possess a vast amount of world knowledge. Still, databases are a critical component of the e-commerce infrastructure as they manage accurate and sometimes time-changing facts about the products. This position paper examines how the strengths of database systems and large language models can be synthesized to create information retrieval systems that better support e-commerce search. We believe the solution is to convert structured and semi-structured data such as the product catalog, taxonomies, and ontology to natural language text and train a language model, which is used as an end-to-end solution for e-commerce search. This is a clear departure from previous e-commerce search approaches that focus on converting unstructured data such as product descriptions, customer reviews, web pages, etc. to structured data through a costly information extraction process and using a myriad of algorithms and machine learning models to support a complex system with separate modules for indexing, retrieval, and ranking.