Stefan Sarbu - 18 Jun 2026
How do you build RAG systems that work in production?
Most RAG implementations are built to produce answers. The ones that work in high-stakes environments are built to produce answers with calibrated confidence, source attribution the user can verify, and behavior that changes when confidence is low. The hard engineering work sits in four places most teams underestimate: chunking strategy, embedding quality, confidence calibration, and how the system surfaces its own uncertainty. The teams that get RAG into production treat confidence as a product feature rather than a model property, and they treat "I do not know" as a legitimate answer the system needs to be trusted to give.
