Option A is correct because instrumenting the 'generate_answer' function with the 'GENERATION' span type is essential for correctly capturing and evaluating the LLM's output for metrics like 'answer_relevance' and 'groundednesS. Registering them as distinct ' TruApp' versions or runs allows for side-by-side comparisons. Option C is correct because instrumenting the retrieval component with 'RETRIEVAL' span type enables the calculation of 'context_relevance' , which directly assesses the quality of the search results and is crucial for RAG evaluation. Option D is correct as creating separate runs with specific configurations (like ' and explicitly computing desired metrics such as 'answer_relevance' and 'groundedness' is the standard way to set up systematic evaluations and comparisons in AI Observability. Option B is incorrect; while cross-region inference might be necessary for model availability, it doesn't directly enable the comparison 'feature' within AI Observability. Option E is incorrect because 'prompt_tokens' and 'completion_tokens' track cost, not directly the quality aspects like 'answer_relevance' and 'groundednesS , which are key for RAG performance evaluation.