RAG passes evaluations but fails in prod. How to detect and prevent regressions before users notice?

Imagine you have deployed a Retrieval-Augmented Generation (RAG) platform that successfully passed all offline evaluations. 

The system achieved excellent scores for:

1. Faithfulness
2. Answer Relevance
3. Context Precision
4. Context Recall
5. Hallucination Rate

using frameworks such as RAGAS and DeepEval. 

However, after deployment, user begin reporting:

A. Incorrect answers
B. Missing context
C. Increased hallucinations
D. Irrelevant document retrieval
E. Higher response latency

The engineering team discovers that production behavior differs significantly from offline evaluation results. 

How would you design a production-grade monitoring and evaluation framework to:

1. Detect regressions early
2. Monitor retrieval quality continuously
3. Track hallucination trends
4. Identify data drift and embedding drift
5. Measure real-world user satisfaction
6. Prevent bad model or prompt deployments

What strategies would you use for:

A. Online vs offline Evaluation
B. Golden Test Sets
C. RAGAS/DeepEval Metrics
D. Canary Deployments
E. LangSmith Observability
F. User Feedback Loops
G. Data Drift Detection
H. Prompt Versioning 

Explain your production approach and trade-offs.