DoorDash Moves Beyond Static Models to Agent-Driven Recommendations Powered by Semantic IDs

DoorDash is shifting its recommendation infrastructure from static, batch-driven models toward a more dynamic agentic approach that keeps track of consumer behavior and preferences in real time. The transition aims to significantly improve conversion rates and recommendation quality across the company's expanding product categories—grocery, convenience, alcohol, and retail.

Moving Beyond One-Shot Predictions#

The traditional approach of issuing a single prediction for each user lookup has given way to a more sophisticated system. By treating recommendations as part of an ongoing dialogue with the consumer rather than a one-time event, DoorDash can adapt to context changes and refine suggestions as new signals arrive.

Semantic Representation at Scale#

A key technical innovation is the use of RQ-VAE semantic IDs to represent items in DoorDash's catalog. Rather than storing high-dimensional embeddings, this approach produces compact, meaningful identifiers that capture product relationships and attributes, making it possible to reason about catalog similarities and perform grounded search across millions of items without computational overhead.

Language-Native Consumer Memory#

The new system embeds consumer interaction history directly into language models, allowing agents to reference past behavior, preferences, and context without explicit feature engineering. This "language-native" memory layer enables more nuanced personalization—the model can weigh seasonal preferences, recent search patterns, and category affinity in a unified way.

Why It Matters#

For a marketplace operating across diverse verticals, better recommendations directly drive higher conversion rates. By moving toward agentic systems with better context awareness, DoorDash can serve increasingly personalized experiences as it expands into new categories, reducing the cold-start problem and improving the relevance of suggestions for both new and returning consumers.

For more on the engineering practices behind scaling AI workloads safely, see QCon AI.