Our Early Journey to Transform Instacart’s Discovery Recommendations with LLMs

This post details Instacart's early journey in transforming discovery recommendations using LLMs. It outlines a top-down, cascaded generative content pipeline for the Shopping Hub, moving from static content libraries to an AI-native platform. The pipeline includes phases for page design & theme generation, retrieval keyword generation, quality & diversity filtering, and product & pagewise ranking. Key learnings emphasize the benefits of a cascaded approach over a single-step model for personalization, cohesion, and adaptability, and highlight the use of constrained decoding and retrieval-augmented generation (RAG) techniques.