Field notes from teams that ship to production.

Designing evaluation harnesses for production RAG systems

How we structure eval pipelines that catch retrieval regressions before they hit production and keep model behavior measurable as datasets evolve.