Distill the LLM, Don't Serve It: Search & Personalization at DoorDash — Raghav Saboo, DoorDash
DoorDash distills LLM capabilities into traditional search systems using four semantic primitives
“the real bottleneck is semantic understanding. That is, it's knowing what elements actually mean to users, given a certain context, and what the shopper actually intends to do”
DoorDash ML engineer Raghav Saboo describes a four-primitive framework (supervision, catalog semantics, semantic personalization, managed content generation) for integrating LLMs into traditional search and ranking pipelines rather than replacing them. The approach addresses the semantic understanding bottleneck in marketplace discovery across DoorDash's expanded verticals including grocery, retail, and pet. The 'distill don't serve' framing is a practical engineering pattern for production LLM integration at scale.