You plan to filter your traces to identify issues while observing how the application is...

A multi-turn chatbot application intermittently produces responses that are grammatically correct and on-topic but contradict earlier turns in the conversation, creating a confusing user experience. CoherenceEvaluator measures exactly this: whether the flow of ideas across a multi-turn conversation is logically consistent and non-contradictory without requiring an external knowledge base. RelevanceEvaluator (option A) measures whether responses are topically on-point but often requires a reference context or knowledge base. SimilarityEvaluator (option B) requires a reference answer for comparison. QAEvaluator (option C) is a composite evaluator for question-answering tasks that requires a ground-truth context document. CoherenceEvaluator is the only option that works purely from the conversation history itself with no external knowledge base, perfectly matching the stated constraint and the multi-turn chatbot use case.

Microsoft Learn Reference Topic: Evaluate conversational AI applications in Microsoft Foundry – CoherenceEvaluator for multi-turn chatbots