Ecommerce ops · Production

DoorDash LLM-Assisted Personalization Framework for Multi-Vertical Retail Discovery

The problem

As DoorDash expanded beyond restaurants into grocery, convenience, alcohol, retail, flowers, and gifting verticals, it faced the challenge of personalizing discovery across a catalog of hundreds of thousands of SKUs at a scale that makes naive prompting or brute-force generation impractical.

Workflow diagram · grounded in source
1
Five-step pipeline initiation
trigger
“Our end-to-end pipeline organizes decisions into five repeatable steps—attribute blending, collection prospecting, item retrieval and ranking, collection targeting, and presentation with LLMs assisting throughout”
2
Two-tower embedding retrieval
ai_action
“We power this with a two-tower embedding model that learns both customer and item representations from sparse order histories, engagement sequences, numerical/context features, and pre-trained embeddings. At serving time, we score via do…”
3
Multi-task ranker scoring
ai_action
“we apply multi-task rankers with a mixture-of-experts design to optimize for multiple outcomes simultaneously — click-through, add-to-cart, in-session conversion, and delayed conversion”
4
Deals Generation Engine targeting
ai_action
“Our Deals Generation Engine actively pairs the right discounts with the right customers, within budget, efficiency, and marketplace constraints.”
5
Cross-vertical novelty via knowledge graphs
ai_action
“translating restaurant history into retail discovery by combining consumer clusters with food and retail knowledge graphs”
6
LLM pipeline assistance
ai_action
“with LLMs assisting throughout: generating topical collections, summarizing past orders into vector context, rewriting queries, explaining recommendations, and augmenting the product knowledge graph”
7
Hierarchical RAG for scalable LLM context
ai_action
“Hierarchical Retrieval-Augmented Generation (RAG) – Rather than dumping the entire catalog into a prompt, we narrow context using category trees and structured retrieval before calling the LLM. This keeps prompts compact, inference fast,…”
8
Personalized surfaces delivered
output
“because these deals are surfaced across discovery carousels, search results, and notifications, they're visible at the moments that matter most”
Reported outcome

DoorDash built a production LLM-assisted personalization framework described as a paradigm shift, with scalable, cost-effective techniques reusable across search, recommendations, and downstream tasks, providing a shared semantic layer for future agentic workflows.

Reported metrics
Catalog scalehundreds of thousands of SKUs
System scalabilityscalable, cost-effective, and reusable across surfaces
Reported stack
large language modelstwo-tower embedding modelmulti-task rankersDeals Generation EngineHierarchical Retrieval-Augmented Generation (RAG)Semantic IDs
Source
https://careersatdoordash.com/blog/doordash-kdd-llm-assisted-personalization-framework/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

DoorDash built a production LLM-assisted personalization framework described as a paradigm shift, with scalable, cost-effective techniques reusable across search, recommendations, and downstream tasks, providing a sha…

What tools did this team use?

large language models, two-tower embedding model, multi-task rankers, Deals Generation Engine, Hierarchical Retrieval-Augmented Generation (RAG), Semantic IDs.

What results were reported?

Catalog scale: hundreds of thousands of SKUs; System scalability: scalable, cost-effective, and reusable across surfaces (source-reported, not independently verified).

How is this ecommerce ops AI workflow structured?

Five-step pipeline initiation → Two-tower embedding retrieval → Multi-task ranker scoring → Deals Generation Engine targeting → Cross-vertical novelty via knowledge graphs → LLM pipeline assistance → Hierarchical RAG for scalable LLM context → Personalized surfaces delivered.

Related ecommerce ops cases