Distill the LLM, Don't Serve It: Search & Personalization at DoorDash — Raghav Saboo, DoorDash

Effective marketplace discovery relies on deep semantic understanding rather than simple engagement optimization. DoorDash integrates Large Language Models (LLMs) into its search and recommendation systems through four core primitives: LLM-based supervision, semantic catalog IDs, consumer memory blocks, and steerable content generation. By distilling complex LLM reasoning into offline labels and hierarchical semantic codes, the system improves relevance and retrieval accuracy without the latency of real-time LLM inference. These shared representations enable highly personalized experiences, such as dynamic collection generation, which has demonstrated measurable gains in order rates and user engagement. This architecture shifts the focus from purely engagement-driven models to systems that interpret specific user intent and constraints, ultimately capturing broader, multi-session shopping missions across diverse retail verticals.