Problem
A retail catalog for grocery, convenience and liquor merchants needs clean, complete attributes for every SKU (stock keeping unit). Raw merchant data comes in varying formats, with missing or wrong values. Tagging was manual, done by contract operators, and a second human had to audit the first, so turnaround was long, cost high and errors frequent. Training an in-house tagger from scratch needs a lot of labeled data (the cold-start problem).
Example from the article: the item name "Dove Silk Glow Body Wash 500 ml" must be split into brand (Dove), size (500) and unit of measure (ml).
Solution in brief
DoorDash replaces a labeled training set with LLMs, in three projects. A brand pipeline sends only low-confidence SKUs to an LLM and checks duplicates against a knowledge graph. An organic-label pipeline cascades string match, LLM reasoning and a web-searching LLM agent. A generalized extractor takes labels from GPT-4 with retrieved golden examples, then fine-tunes an LLM on them.
Questions this case answers
- How to extract structured attributes from unstructured item descriptions at catalog scale when labeled training data is scarce and human annotation is slow and costly
- How to add newly discovered entities to a taxonomy without creating duplicates when the set of real-world entities is open-ended
- How to extract and standardize product attributes from messy merchant SKU data without a large labeled set?
- How to find new brands in a catalog and avoid adding duplicates to a brand taxonomy?
Tools and methods
- GPT-4 — generates attribute labels from in-context examples
- OpenAI Embeddings + approximate nearest neighbors — retrieve similar golden-annotated SKUs (RAG)
- LLM agent with online search — fills gaps when product data is missing
- Knowledge graph of brands — lookup of similar brands and item names for duplicate checks
- Fine-tuned LLM — scalable inference trained on LLM-generated annotations
Google or a one-time email code. No password.