How search evals lead you astray


I wrote a detailed blog article critiquing the traditional judgment list search evaluation regime.

It's impossible to NOT have some kind of sampling bias. IE your labels all come from BM25 or some legacy search algorithm. So solutions can, at best, measure their effectiveness reranking within that sample.

Indeed, its hard to think of an alternative. Labeling random results will just end up wasting time (99% of the corpus will be irrelevant). In the end, you need a sampling algorithm to mine for useful samples. And that aggregated combination of example mining techniques turns into a sampling bias too.

Or, like we talk about in our AI Powered Search course, turn this into feature discovery. Labeling BM25 results is a way of finding out how for example, BM25, matters. As ranking constantly needs to seek new ways a result can be relevant / irrelevant, you just stay in a constant loop of trying sampling with a new feature, sampling every way a document can be relevant / irrelevant with that feature, discovering how/if it matters, and iterating over and over.

The article:

​https://softwaredoug.com/blog/2026/10/26/search-golden-sets-fail​

Cheers,

-Doug

PS AI Powered Search starts next week​

​Slack Community * Events · Consulting · Training (use code search-tips)

You're subscribed to Doug Turnbull's daily search tips where I share tips, blog articles, events, and more. You can always manage your profile:

Doug Turnbull

I share search tips, blog articles, and free events I'm hosting about the search+retreval industry, vector databases, information retrieval and more.

Read more from Doug Turnbull

In a world where agents easily drive browsers, why spend time on classic ranking metrics like NDCG? Instead, task an agent with a search use case. See how it fairs in the UI and ranking. In this article, I discuss an alternative eval philosophy. Instead of LLM as a Judge, let agents simulate search users. Observe how they interact with the search app to solve real use-cases. We try to answer the question: can we make search lower effort and easier to accomplish common tasks? Build the...

I am a loyal customer of Alamo Drafthouse and Cinema - a boutique movie theater chain in the US. One nice perk is they allow you to order with a QR code ordering system from your table. Sadly the search doesn't work that well, when you search for "burger" you get "nachos" as the first result. Well that's an annoyance purpose-built for me. So I took it upon myself to recreate the search, then fix it. Did I build vector search? No. Did I try to improve query understanding? No. What I did...

This week+next there's some upcoming, free search events I'm cohosting: AI-Search Fundamentals (free 'intro to search' training) Today 12PM ET (w/ Me + Trey Grainger) Sign up -- https://maven.com/p/b9e696/ai-search-fundamentals-vector-spaces-matching-ranking This session covers foundational concepts in information retrieval around matching and ranking. We'll cover sparse vs. dense vector spaces, lexical (BM25) vs. semantic (embedding) search, and matching and relevance ranking fundamentals....