Stop Benchmarking Embedding Models. 90% of Your Search Quality Lives Upstream.

Brief intro on context. I'm CTO at Vaultt (formerly StudentVenture), a recruitment marketplace for top 1% non-traditional talent. 10,000+ candidate profiles, semantic matching in production for over a year. We run everything on pgvector inside our main Postgres database, with LLM-generated summaries as the embedding input and hybrid filtering at query time. Every few months a new "best embedding model in 2026" benchmark lands, and founders ask me the same question: should we be using model X ins