Ecommerce ops · Production

DoorDash uses LLM-generated profiles and content embeddings to improve semantic search and recommendations across verticals

The problem

DoorDash faced a persistent bottleneck where personalization depended on embedding quality, which in turn depended on data quality — but sparse metadata flattened catalog richness across all verticals. Behavioral co-visitation approaches tried to bypass this dependency but could not capture identity, context, and intent.

Workflow diagram · grounded in source
1
LLM profile generation
ai_action
“LLMs produce consistent, high-quality narratives for merchants and items such as ingredients, preparation, attributes, or context that reduce reliance on human-labeling efforts”
2
Incremental embedding pipeline
integration
“We use incremental inference via Metaflow, which only requires re-embedding entities when their underlying content has changed”
3
Embedding index publishing
output
“Publishing writes embeddings to persistent storage/index so that downstream experiments can consume them consistently”
4
Semantic query retrieval
ai_action
“Embed the query online and retrieve against offline store/item profile embeddings, so that even rare or novel queries can retrieve semantically aligned candidates”
5
Item-level reranking
ai_action
“a fine-tuned Qwen 3 Rerank model that scores each candidate by consuming the search query, the item profiles of the top-k most relevant items within a store, and the store profile”
6
Generative personalized carousels
ai_action
“An LLM generates a carousel theme from the consumer profile and context, such as time of day, then embeds the theme and retrieves nearest-neighbor stores and representative dishes within the delivery radius. Final ordering uses the exist…”
Reported outcome

Deploying LLM-generated profiles with content embeddings improved semantic search and homepage discovery: null search rate fell 3.65%, core search CVR rose 0.66%, and generative personalized carousels drove a 2.4% relative increase in homepage order rate.

Reported metrics
7D active customer share+0.0724%
Null search rate−3.65%
core search session CVR+0.66%
Dish query ranking improvement7.8%
Show all 14 reported metrics
7D active customer share+0.0724%
null search rate−3.65%
core search session CVR+0.66%
dish query ranking improvement7.8%
cuisine query ranking improvement1.4%
trial merchant visit rate+0.435%
homepage clicks per impression+0.110%
consumer homepage order rate2.4%
consumer 7-day reorder rate+0.164%
variable profit per order0.32%
offline precision@10 on homepage68% to 85%
item-to-item Hit@5 with LLM profiles+31.22%
item-to-item Hit@5 data and model combined+37.55%
store-to-store Hit@5 data and model combined+209%
Reported stack
LLMsgemini-embedding-001Qwen 3 Rerank modelMetaflowMiniLMGoogle Gemini embeddingsQwen embedding modelstext-embedding-005text-embedding-3-largeOpenAICohere
Source
https://careersatdoordash.com/blog/doordash-llms-to-build-content-embeddings-for-search-and-recommendations/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Deploying LLM-generated profiles with content embeddings improved semantic search and homepage discovery: null search rate fell 3.65%, core search CVR rose 0.66%, and generative personalized carousels drove a 2.4% rel…

What tools did this team use?

LLMs, gemini-embedding-001, Qwen 3 Rerank model, Metaflow, MiniLM, Google Gemini embeddings, Qwen embedding models, text-embedding-005, text-embedding-3-large, OpenAI.

What results were reported?

7D active customer share: +0.0724%; Null search rate: −3.65%; core search session CVR: +0.66%; Dish query ranking improvement: 7.8% (source-reported, not independently verified).

How is this ecommerce ops AI workflow structured?

LLM profile generation → Incremental embedding pipeline → Embedding index publishing → Semantic query retrieval → Item-level reranking → Generative personalized carousels.

Related ecommerce ops cases