Evaluating search?
Don’t jump to complex labeling systems, just do simple side-by-sides.
When I worked at Reddit, I would socialize a spreadsheet of search results. On each side were search results for a test query. One side was “control” - autogenerated from prod. The other “test” - my new fancy algorithm.
But the reviewers didn’t know which was which. They were blind.
What they saw: one side labeled “Pepsi” the other “Coke” (referencing the old commercial). They’d give me a preference over several dozen queries - isolating to types of queries my change impacted. This gave me a good feeling whether continuing with deeper evaluation (ie A/B test) was worth it.
Your search evals need not involve a PhD. Start grug-brained. Don’t get ahead of your skis!
https://softwaredoug.com/blog/2025/06/22/grug-brained-search-eval
-Doug
This is part of Doug’s Daily Search tips - subscribe here
Upcoming course: Build your own vector database
Want to understand what makes embedding retrieval fast, relevant, and useful in real AI systems? Join Build your own vector database and build the core pieces yourself, from embeddings and indexing to search and retrieval.