Problem
Whatnot is a marketplace with products and live shows. Its search users often misspell a word ("jewlery" for "jewelry") or type an abbreviation ("lv" for "louis vuitton", "nyfw" for "new york fashion week"). Such queries return almost empty pages or few results, and engagement after them is lower. Users then assume the app has no jewelry content at all.
Example from the article: a search for "jewlery" shows a nearly empty page, while "jewelry" shows a full page of categories, live shows and products that users engage with and buy from.
Solution in brief
Collect frequent query tokens from the logs and send them to GPT offline, outside the request path. GPT returns spelling corrections and abbreviation expansions with confidence scores, which go into a key-value cache. At query time, search looks up each token and adds the cached expansions to the original query.
Questions this case answers
- How to correct misspelled and abbreviated search queries at low serving latency without maintaining a hand-built dictionary or knowledge graph
- How to stop misspelled search queries from returning empty or poor result pages
- How to make abbreviation queries return the same results as their full names
- How to add LLM-based query correction to a search stack with a latency budget of about 250 ms
Tools and methods
- GPT — flags misspelled tokens and abbreviations, returns up to 3 candidates each with a confidence level
- Query expansion cache — key-value store mapping a token to its corrections and expansions with confidence
- Search logs — query, filters and landing results tab, joined in the warehouse at tab, query and search-session level
- Query expansion at serving time — adds cached expansions to the query expression next to the original terms
Google or a one-time email code. No password.