3 Changes to Expand DoorDash's Product Search Beyond Delivery

Sonic Wang, Xiaochang Miao, Yongshuang Wang10 May 2022

Problem

DoorDash search was built for restaurants: nested menus, a food taxonomy and a heuristic ranker. Expanding into groceries, convenience, alcohol and pet food brought discrete SKUs, new product domains and many more merchants. The old system was tightly coupled to food delivery, so each new vertical needed surgical changes in core code. The heuristic ranker lost relevance on any query that was not a merchant name.

Example from the article: a query like "red pizza" has no full match, so the system relaxes it to "pizza" by walking up the taxonomy to raise recall.

Solution in brief

The team changed three layers. Infrastructure got a materialized-view index and a federate-then-blend pipeline with one retriever per vertical. Query and document understanding got food, retail and query-intent taxonomies, filled from manual work up to ML. Ranking moved from BM25-style heuristics to a gradient-boosted pointwise ranker predicting view-to-conversion, then added personalization.

Questions this case answers
  • How to extend a single-domain search system to several heterogeneous catalogs while keeping each catalog's relevance tunable and failures isolated
  • How to build labeled training data and taxonomies for a new catalog domain when volume far exceeds in-house annotation capacity
  • How to move a search ranker from hand-tuned heuristics to a learned model trained on implicit feedback and measured with an evaluation framework
  • How to extend a food-only search engine to groceries, alcohol and other verticals without rewriting core code for each one?

Tools and methods

  • Federate-then-blend search — each vertical retrieves independently, a blending stage merges results
  • Materialized view of search documents — faster, reproducible index rebuilds
  • Taxonomies (food, retail, query intent) — structured meaning of items and queries
  • Human annotation with vendors — labels at scale, with majority vote and audits
  • Pointwise learning-to-rank (gradient boosted trees) — predicts view-to-conversion per store
  • spaCy part-of-speech tagging — query structure for relaxation and expansion
  • Elasticsearch, BM25 heuristic ranker — the baseline being replaced
Continue reading — it's free

Google or a one-time email code. No password.