hallucination detection

Turn RAG from guesswork to observability with drop-in Bash: define the right metrics (retrieval overlap/groundedness, latency p50/p95, error and cache rates, token cost, freshness), log every turn as JSONL, compute rolling stats with rag-metrics.sh, export Prometheus textfile gauges, trigger Slack/email alerts via rag-alert.sh, and gnuplot quick charts—catch drift, cut latency, and rein in spend.

A practical guide to Linux- and CI-friendly RAG evaluation: measure retrieval recall@k, context precision@k, answer similarity, and latency to separate retriever vs generator issues, prevent regressions, and speed iteration. Includes deps install, a small JSONL dataset, a 100-line Python evaluator, Make-based CI gates, tuning k/embeddings/chunking, and optional faithfulness checks.