Three major food delivery platforms, DoorDash, Instacart, and Uber Eats, have independently integrated Large Language Models (LLMs) into their search functionalities, aiming to better understand user intent. Despite facing similar challenges in interpreting queries like synonyms, typos, and shorthand, their architectural solutions vary significantly. This divergence highlights how existing infrastructure and specific design choices, rather than the LLM itself, dictate the integration approach.

DoorDash leveraged its pre-existing knowledge graph of items and restaurants. They used LLMs offline to enrich this graph with attributes extracted from SKU data. At runtime, LLMs parse queries into segments that link to the graph, while retrieval remains keyword and graph-driven. This method uses retrieval-augmented generation as a guardrail, prompting the LLM to select from a limited list of known concepts, resulting in a 30% lift in carousel trigger rates.

Instacart faced a challenge where a general LLM's understanding of 'protein' differed from user behavior, which favored protein bars and powders. Their strategy involves layered approaches: context engineering to inject Instacart-specific data into LLM prompts, and post-processing guardrails using semantic similarity filters to ensure LLM outputs align with the original query. This addresses the limitations of general LLMs with company-specific user data and catalog context.