Picking an Embedding Model

MTEB rankings measure performance across a benchmark suite — mostly web text, mostly English, mostly generic. Your legal corpus, medical notes, or codebase lives somewhere else in embedding space.

A model that's 3 points higher on MTEB can be worse on your domain because the benchmarks never tested it there. The honest eval: build 50–200 pairs of your queries → your relevant chunks (from logs, from labeling, from synthetic generation verified by hand), then measure recall on those. A weekend of domain eval beats a leaderboard argument.