DoorDash Reduces Grocery Reorder Latency to 2 Seconds Using AI Precomputation

Latency Reduction in Grocery Reordering

DoorDash has significantly reduced the latency in its AI-powered grocery reorder system, bringing it down from approximately 16 seconds to just 2 seconds. This improvement was achieved by shifting from rebuilding purchase needs with each request to precomputing them in advance.

The optimization led to a doubling of shopping list views and an increase in the reorder flow's order rate by about 34%. This implementation leverages DoorDash's existing shared agent platform and grounded shopping interface, which provide reusable infrastructure and allow consumers to modify shopping lists directly.

Technical Implementation Details

The system integrates batch large language model (LLM) inference, deterministic validation, and live product catalog resolution within Ask DoorDash, the company’s conversational shopping assistant. For grocery reordering, DoorDash analyzes 90 days of purchase history along with consumer profile signals.

Batch LLM inference identifies likely recurring purchases, and deterministic validation checks the generated results before storing reusable shopping list data. This precomputation reduces the workload during live requests.

Serving Path and Catalog Resolution

The serving path retrieves the prepared data and resolves each item against the current product catalog. It attempts to match the original store item first, then a catalog product identifier, and finally a search term. This method accommodates changes in listings and product unavailability without re-evaluating historical reasoning for every request.

Shared Agent Platform Benefits

The shared platform provides orchestration, memory, model access, tracing, evaluation infrastructure, and rollout controls. Domain teams maintain ownership of agent instructions, skills, tools, and evaluation criteria. DoorDash reported that adding its Reservations agent took one week, which is approximately ten times faster than building its initial domain agents.

Design Philosophy

Twisha Jain, a DoorDash engineer, noted that effective agentic systems prioritize deciding what can be processed ahead of time versus what needs to be part of the live agent loop. Moving recurring purchase analysis offline reduces the work required during a live request while still incorporating current catalog information for product selection. The Grocery agent is still invoked for requests requiring additional reasoning, such as finding cheaper alternatives based on current prices, promotions, or package sizes.