Problem
DoorDash search and recommendations span restaurants, groceries, retail and gifting, where catalog metadata is sparse: tags and short descriptions do not tell a spicy Sichuan soup from a mild Cantonese broth. Embeddings learned from clicks inherit popularity bias, and users order about once a week, so most cohorts and tail queries have too little data. New items and small merchants stay invisible until they collect engagement. A retrieval miss cannot be repaired by the ranker.
Example from the article: for the query "Szechuan" the production retrieval returned a single Sichuan restaurant, while retrieval over LLM-profile embeddings returned a diverse set of Chinese stores.
Solution in brief
An LLM first writes a standardized text profile for every store and item (ingredients, preparation, cuisine, dietary attributes). An off-the-shelf embedding model then encodes the profile, and the same vectors serve entity similarity, search retrieval and homepage carousels. Offline tests show that improving the input text moves quality at least as much as swapping the encoder.
Questions this case answers
- How to build entity embeddings that capture meaning when structured metadata is sparse and behavioral data is too thin to train on
- How to compare candidate embedding models offline and split the quality gain between input data and model choice when no human relevance labels exist
- How to get usable embeddings for stores and items when catalog metadata is sparse and clicks are too few to learn from?
- How to choose an embedding model without a human-labeled golden set?
Tools and methods
- LLM-generated merchant and item profiles — standardized text narratives that replace sparse metadata as embedding input
- gemini-embedding-001 (256 dimensions, Matryoshka) — chosen encoder; task types SEMANTIC_SIMILARITY, RETRIEVAL_QUERY, RETRIEVAL_DOCUMENT
- LLM-as-a-judge — builds reference rankings for offline evaluation without human labels
- Hit@K and nDCG@K — offline metrics for similarity and query-to-entity retrieval
- Fine-tuned Qwen3 reranker — rescoring of item-level candidates in search
Google or a one-time email code. No password.