Complete the claim: the two-axis decomposition of RAG eval that diagnoses which layer failed.

RAG has two parts that can each fail: the retriever that fetches chunks, and the generator that writes the answer. Good evaluation measures them separately. **Faithfulness** is the answer side metric. It asks whether every claim in the generated answer is supported by the retrieved chunks, and that nothing contradicts them. A faithful answer never invents facts beyond what the context says. **Context recall** is the context-side metric. It asks whether the chunks the retriever pulled actually contain the information needed to answer the question. High recall means the right evidence made it into the prompt. Why split them? Because a single end to end accuracy score hides the cause of a failure. If you only see a wrong answer, you don't know whether retrieval missed the page or the model ignored a page it had. Measuring both axes turns a vague failure into a diagnosis: high faith with low recall is a retrieval failure, and low faith with high recall is a generation failure. That tells you exactly which layer to fix.

RAG evaluation decomposes pipeline quality into two orthogonal axes. The split lets a failure point at the layer that caused it. **Faithfulness** is the answer side axis. It scores whether the generated answer's claims are entailed by the retrieved context and not contradicted by it. It is the headline metric RAGAS popularized. It is typically computed by extracting claims from the answer, then checking each one against the chunks. Context recall is the context-side axis: given a ground-truth answer or reference, it measures whether the retrieved chunks contain the information needed to produce it. It is a property of the retriever, independent of what the LLM then does. The diagnostic power comes from the cross-product. High faith with low recall means the system was honest but ignorant: it stayed true to its context, but the context lacked the answer, a retrieval failure. Low faith with high recall means the LLM ignored or contradicted evidence it actually had, a generation failure. Keeping the axes orthogonal is the whole point. Collapsing them into one accuracy number tells you something is broken but not where. You then can't decide whether to tune chunking and the retriever or to fix the prompt and the model.