Three agents share a knowledge base: policy lookup, exception handling, and customer explanation. Retrieval often returns irrelevant partial tables, so the exception agent applies the wrong rule. How should you improve the multi-agent RAG architecture?

B is correct. Multi-agent RAG is not one-size retrieval. Bad table chunks and weak embeddings tank precision, and specialists then act on the wrong rule. Fix chunk boundaries for structured content, improve embeddings for domain vocabulary, and use metadata filters so the exception agent does not pull marketing fluff. A single-PDF-as-one-chunk blows embedding quality and retrieval. C stuffing the corpus every call is cost and latency suicide. D “ignore retrieval” abandons grounding. AI-500 wants you to connect multi-agent consumption to chunking, embedding quality, and retrieval precision—not bigger prompts alone.