Uber Eats Tries Agentic Shopping

Uber Eats is rolling out Cart Assistant, a new feature designed to streamline grocery shopping by transforming natural language or image prompts into draft shopping carts. This initiative represents a significant shift from traditional search-first e-commerce to a more intuitive, intent-driven approach. The system aims to bridge the gap between a user's vague plan, like "healthy breakfasts for the week," and a finalized order, as detailed on Uber Engineering.

Cart Assistant moves beyond basic keyword searches. Users can now simply describe their needs or upload an image, and the system generates a preliminary cart for review. This process fundamentally alters the shopping flow from a manual item-by-item selection to an intent-to-draft-cart model.

The Agentic Architecture

At its core, Cart Assistant employs a multi-prompt state graph. This architecture orchestrates a series of specialized tasks, each handled by either a Large Language Model (LLM) or a deterministic system. LLMs tackle ambiguity and interpret user intent, while deterministic systems manage data retrieval, pricing, and cart construction.

Each stage in the graph has a distinct responsibility:

• Cart Plan Generation: Interprets input to create planned items with search terms and constraints.
• Candidate Retrieval and Enrichment: Queries store inventory for relevant products.
• Semantic Relevance Judging: Determines the best product matches within context.
• Price and Deal Constraint Enforcement: Applies user-defined price limits and preferences.
• Quantity Selection: Translates colloquial quantities into purchasable units.
• Guardrails: Validates inputs, outputs, and system behavior for safety and accuracy.
• Cart Assembly: Consolidates all items into a final draft cart.
• Content Refinement: Generates user-facing text for recipes or meal plans.

This modular design allows for parallel processing of items, significantly reducing latency. Instead of a slow, sequential build, the system processes multiple items concurrently.

Handling Complexity and Ensuring Safety

Translating natural language requests into precise shopping actions is complex. For instance, a prompt like "pasta for two, plus paper towels and vegan protein powder under $20" requires the system to capture multiple distinct intents and constraints.

Quantity selection is particularly challenging, moving from terms like "a dozen eggs" or "snacks for a party" to store-specific units. Cart Assistant uses a combination of LLM reasoning and deterministic validation to resolve these discrepancies, considering product packaging and common grocery knowledge.

Robust guardrails are essential for agentic grocery shopping. These include deterministic checks for schema validity, price limits, and availability, alongside LLM-based assessments for in-domain requests and prompt injection attempts. This layered safety approach ensures the system operates within defined boundaries.

Evaluation-Driven Development

The stochastic nature of LLMs necessitates a rigorous evaluation framework. Uber Eats employs a system that combines strict deterministic verification with an LLM-as-a-judge approach. This allows for precise measurement of changes and continuous refinement of the system's performance against a wide range of test cases.

This focus on evaluation ensures that improvements in one area, like recipe generation, do not negatively impact other functionalities, such as simple shopping lists. The system is continuously monitored using both synthetic and real-world data.

Cart Assistant represents a significant step towards more intelligent, personalized e-commerce experiences, turning complex user intent into actionable shopping carts with user control maintained throughout the process.