DoorDash 以 AI 智能體將雜貨補購延遲由 16 秒降至 2 秒

DoorDash cut latency in its AI-powered grocery reorder flow from about 16 seconds to 2 seconds by precomputing recurring purchase needs instead of rebuilding them with each request. The change doubled shopping list views and increased the flow’s order rate by approximately 34%. The implementation builds on DoorDash’s shared agent platform and grounded shopping interface, which provide reusable agent infrastructure and let consumers modify shopping lists directly.

The implementation combines batch large language model (LLM) inference, deterministic validation, and live product catalog resolution within Ask DoorDash, the company’s conversational shopping assistant. Previous InfoQ coverage described how Ask DoorDash uses specialized agents and tools to support conversational discovery and shopping. For grocery reordering, DoorDash processes 90 days of purchase history alongside consumer profile signals. Batch LLM inference identifies likely recurring purchases, while deterministic validation checks the generated results before storing reusable shopping list data.

DoorDash’s grocery reorder architecture (Source: DoorDash Blog Post)

The serving path retrieves this prepared data and resolves each item against the current catalog. It first tries to match the original store item, then a catalog product identifier, and finally a search term. This allows the system to accommodate changing listings and unavailable products without repeating historical reasoning for every request.

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

Ask DoorDash shared agent platform architecture (Source: DoorDash Blog Post)

Twisha Jain, a DoorDash engineer, described the design rationale in a LinkedIn post:

Better agentic systems aren’t always about more reasoning; they’re about deciding what belongs in the agent loop and what can happen ahead of time.

Moving recurring purchase analysis offline reduces the work required during a live request while retaining current catalog information for product selection. The system still invokes the Grocery agent for requests requiring additional reasoning, such as finding cheaper alternatives based on current prices, promotions, package sizes, inventory, and consumer preferences. Simpler operations, including changing quantities or removing products, update the shopping list directly without another LLM round trip.

A DoorDash engineering post about its shopping interface states:

A consumer’s direct edit is authoritative: the next agent turn starts from the edited list, not from what was said earlier.

The design treats the shopping list as a structured artifact that both the consumer and agent can update, rather than relying solely on conversational history. Other grocery platforms also use AI to personalize shopping. Instacart’s AI shopping assistant supports meal planning, list creation, and product discovery using consumer preferences and local grocery offerings. DoorDash’s implementation focuses specifically on reducing reorder latency by separating historical purchase analysis from live product resolution.

About the Author

Leela Kumili

Show moreShow less