RAG, đánh giá chất lượng và ranh giới agent

Nội dung luyện tập

• distinguish retrieval from model training
• interpret embedding similarity limits
• verify citations support the claim
• keep query and document embeddings compatible
• preserve meaning when chunking tables
• combine lexical and semantic retrieval
• calculate retrieval precision and recall
• budget context with output reserve
• enforce retrieval authorization outside model
• trace answer freshness to index provenance
• contain instructions in retrieved content
• validate llm output beyond json shape
• locate retrieval loss before reranking
• keep final evaluation independent
• separate groundedness from world correctness
• detect regressions hidden by aggregate evaluation
• calibrate automated evaluators
• version all components affecting rag quality
• bind human approval to concrete tool action
• bound agent resource consumption

Mỗi câu đúng 1 điểm. Mục tiêu 70%. Đáp án và giải thích mở sau khi nộp bài.