Comparing AI Search Architectures: DoorDash vs. Instacart vs. Uber Eats

Pavan_Belagatti · x · 2026-08-01

This post breaks down the fundamentally different search architectures used by three major delivery platforms when integrating LLMs:

DoorDash: Keeps retrieval mostly classical. They use LLMs offline to enrich an existing knowledge graph and employ an inverted RAG approach as a guardrail to safely parse queries at runtime.
Instacart: Splits traffic. Common queries hit a deeply context-engineered offline RAG cache, while cold-start "tail" queries are handled in real-time by a fine-tuned Llama-3-8B model.
Uber Eats: Goes deep with vector search. They fine-tuned a Qwen LLM as the embedding backbone in a two-tower setup, embedding queries in real-time to match against pre-computed document vectors.

More from Research