I think they might be coming back. Have you seen www.graphiumlabs.com?

This post breaks down why understanding precision and recall is essential when building search and information retrieval systems for high stakes decision making:

www.graphiumlabs.com/blog/precisi...

In high-stakes environments, like medical diagnostics, legal research, and threat detection, the trade off between high recall and high precision isn’t just a theoretical optimization problem. The choice has real-world consequences.

Ideally, users want both at 100% – all (good) signal, and zero noise. But the way search works under the hood often forces a trade off: higher recall requires looser filters to bring in more results, and consequentially, more irrelevant results or noise, which bring down precision.

Precision means: “Of the results that were returned, how many were relevant (correct)?”

And recall says: “Of all the correct results, how many were returned?”

In search and information retrieval systems, precision and recall are more than just evaluation metrics—they reflect how well a system aligns with the user’s needs and expectations of relevance and completeness.

So when nuance is important, semantic search built on vector similarity tends to miss the mark by a really, really wide margin.

I'll start: vector embeddings don't encode semantics, they encode substitutability. It _looks right_ if you squint at it, or if the use case is pretty trivial (e.g. "brown" vs. "chocolate" when describing a sofa).

But opposites also have high substituability (good/bad, dark/light, rich/poor, etc.)

I once failed the "check the checkbox" test by checking it... Wrong? I guess?

Hey Tim let's talk.