Production-grade RAG Evaluation System

Why this exists

RAG regressions often appear after prompt edits, model swaps, index rebuilds, or chunking changes. Target: measure those changes before merge

Next checks

• golden-set retrieval recall and faithfulness suites
• per-run latency and cost budget
• score records keyed by commit and dataset hash
• GitHub Actions gate that compares a pull request to a baseline branch

Public line

Public after a repo, reproducible benchmark, and one evaluation run exist