Pick the Phoenix evaluator that flags 'the answer contradicts the retrieved context'

Phoenix from Arize ships a small set of pre-built evaluators that each target one specific failure mode in a RAG pipeline. Knowing which one to reach for is the load-bearing skill. **Hallucination** (sometimes called groundedness or faithfulness depending on the framing) checks one specific thing: given the answer and the retrieved context, is every claim in the answer supported by the context? This is the evaluator that catches the failure 'the model invented something the context did not actually say'. **Relevance** checks the other side of the pipeline: given the question and the retrieved chunks, are those chunks actually about the question? If retrieval drags in irrelevant chunks, relevance flags it. Relevance does not look at the generated answer at all. **QA Correctness** compares the generated answer against a known ground-truth answer, so it requires a labeled dataset. Useful for offline benchmark runs; less useful for online traffic where ground truth does not exist. **Toxicity** is unrelated to grounding; it checks for harmful content. In the question, the failure described is 'the answer contradicts the retrieved context'. That is the hallucination / groundedness / faithfulness evaluator.

**The right pick is Hallucination (or Groundedness / Faithfulness).** Phoenix's hallucination evaluator takes the generated answer and the retrieved context as inputs and asks an LLM judge whether every claim in the answer is supported by the context. It is the only one of the four evaluators in the question that targets the answer vs context contradiction failure mode. The other three each target a different failure mode: - **Relevance**: question vs retrieved chunks. Did retrieval pull the right context? It cannot detect a hallucinated answer drawn from good context, because it never looks at the answer. - **QA Correctness**: answer vs ground-truth answer. Requires a labeled dataset; right tool for offline regression suites where you have a golden set. - **Toxicity**: orthogonal to grounding entirely; it checks safety, not factuality. The naming varies across the literature. Phoenix uses 'hallucination' as the primary name in its evaluator catalog; TruLens uses 'groundedness'; RAGAS uses 'faithfulness'; the OpenAI Evals library uses 'context_relevance' for the related but distinct retrieval-side check. Conceptually the answer vs context check is the same; the implementation differences are in the judge prompt template and how multiple claims are decomposed. The interview signal here is whether the candidate maps failure modes to evaluators rather than picking by name similarity.