The dashboard is green. Uptime looks perfect. Latency is low. The API returns answers fast. Everything about the system seems healthy. But healthy metrics don’t mean correct answers. A RAG system can run smoothly while quietly giving users wrong information. In one widely cited case, an airline’s RAG system gave a customer an incorrect refund policy, and the company ended up losing in court.
Retrieval and generation get measured separately from correctness. Retrieval can pull chunks that look plausible but miss the exact evidence needed. Generation can sound fluent even when it’s built on weak context. An answer can seem relevant to the question and still be factually wrong. Standard monitoring often misses this gap between “answered” and “answered correctly.”
Plausible chunks and fluent generation can still add up to a wrong answer. Answered isn’t the same as answered correctly.
Faithfulness is one guardrail against this quiet failure. It checks whether an answer is actually supported by the retrieved context. Low faithfulness signals hallucinated or contradictory claims. Many teams use it as their main alert for hallucination. A high faithfulness score lowers hallucination risk, but it doesn’t guarantee the answer is fully correct. Some evaluation systems treat faithfulness as the key threshold before an answer reaches production.
Retrieval quality has its own blind spots. Context relevance checks whether retrieved passages are useful. Context recall checks whether the context holds the information needed for a correct answer. Context precision checks whether irrelevant chunks got filtered out. Recall@K and Precision@K test similar things at the top of search results. None of these metrics alone catch every failure.
Answer quality metrics fill some of those gaps. Answer correctness compares a response to a ground-truth answer. Answer relevancy checks if the answer matches the question’s meaning. A response can sound right and still miss key facts. It can be short and clear while contradicting its own source material.
Hallucination metrics try to catch what’s unsupported by context. Many RAG evaluation tools compute faithfulness, hallucination, precision, and recall together. Some define hallucination as facts not backed by the retrieved context. Scoring focused on only one of these failure types can miss the others entirely, which raises production risk. Researchers still argue over the best way to measure it.
Production guides suggest real thresholds. Some aim for recall@20 near 95%. Others set faithfulness at 90% or higher. Correctness thresholds tend to run lower, since reference answers are harder to perfect. Average scores can still hide bad edge cases. AI systems that analyze large data sets can help surface these hidden failure patterns before they reach end users.
References
- https://cleanlab.ai/blog/rag-tlm-hallucination-benchmarking/
- https://www.braintrust.dev/articles/ai-hallucination-evaluations-metrics-methods-2026
- https://docs.raga.ai/ragaai-catalyst/ragaai-metric-library/rag-metrics/hallucination
- https://futureagi.com/blog/evaluating-rag-systems-ensuring-your-llm-remembers-what-it-reads/
- https://getdevstudio.com/blog/rag-evaluation-and-monitoring-guide/
- https://ragaboutit.com/5-rag-compliance-metrics-that-catch-72-of-hallucinations/
- https://arxiv.org/abs/2503.21157
- https://scadea.com/evaluating-rag-quality-hallucination-detection-and-answer-accuracy-metrics/
- https://www.vectara.com/blog/measuring-hallucinations-in-rag-systems
- https://atlas.latticeflow.ai/evaluation/rag_hallucination/
