Mind the Gap: Using LLMs to bridge behavioral silos in multi-vertical recommendations

DoorDash developed an LLM-powered framework to solve the cold-start problem in multi-vertical recommendations by converting restaurant orders and search queries into cross-vertical affinity features. Using hierarchical RAG with GPT-4o-mini, they map user behavior to a four-level product taxonomy, then integrate these semantic features into their multi-task ranking models. The approach achieved 4.4% AUC-ROC improvement offline and 4.3% online, with particularly strong gains for cold-start users, while keeping inference costs practical through prompt caching and optimization strategies.