DoorDash replaces one-shot predictions with agentic recommendations using consumer memory and semantic IDs

The presentation described a transition from traditional one-shot prediction models to an agentic recommendation architecture at DoorDash. This change introduces language-native consumer memory to capture user intent over time. It also adopts RQ-VAE semantic IDs for a unified catalog representation. Finally, grounded search is applied to tie recommendations to real-world item attributes.

Engineers can see how these components together aim to increase relevance and conversion metrics in consumer-facing services. The talk highlighted that the new system supports rapid experimentation across grocery, convenience, alcohol, and retail verticals. It emphasized the role of semantic IDs in reducing catalog sparsity. The approach is presented as a way to scale personalization without relying on isolated point predictions.

The material did not disclose specific implementation costs, resource requirements, or operational overhead for the agentic system. It also did not describe failure modes, edge cases, or scenarios where the approach might underperform. Consequently, readers must look elsewhere for details on adoption trade-offs and limits. The focus remained on the architectural shift and reported metric improvements.