TechPrufer
Sign inGet started

Blogs›Uber›AI-Powered Cart Generation and Management

AI-Powered Cart Generation and Management

AI-Powered Cart Generation and Management

1
posts
2026

Uber is developing advanced AI capabilities to transform user interactions with its commerce platforms, starting with Uber Eats. This initiative focuses on moving beyond traditional search-based interactions to more intuitive, intent-driven experiences. Key developments include the Cart Assistant, which uses multi-prompt state graphs and LLMs to interpret natural language and image inputs, generate draft grocery carts, and manage complex constraints like pricing and quantity. This evolution aims to improve the user experience by enabling agentic grocery shopping, where the system plans and constructs a cart based on user intent, rather than requiring manual item-by-item search. The architecture involves a multi-prompt state graph with distinct stages for planning, retrieval, relevance judging, quantity selection, and guardrails, all driven by LLMs and deterministic systems. Evaluation-driven development is a core part of the process to ensure reliable measurement and improvement of the system's behavior.

2026

Cart Assistant: Agentic Grocery Shopping on Uber Eats

6/16/2026

This post details the technical architecture and implementation of Uber Eats' Cart Assistant, an AI-powered system for agentic grocery shopping. It describes the shift from a search-first to a cart-first shopping paradigm, enabled by a multi-prompt state graph. Key technical contributions include: the Cart Plan Generation stage using LLMs for structured output of planned items with search terms, context, and constraints; Candidate Retrieval and Enrichment via backend search and catalog APIs; Semantic Relevance Judging using LLM rubrics; Price and Deal Constraint Enforcement through deterministic optimization; Quantity Selection combining LLM reasoning with deterministic arithmetic; Guardrails (deterministic and LLM-based) for safety and boundedness; and optimizations for parallelism and latency. The post also highlights the importance of an evaluation-driven development framework for measuring and improving LLM-based systems.