Práctica NVIDIA-Certified Associate: Generative AI LLMs (NCA-GENL)

A RAG system retrieves five chunks per query and stuffs them into the prompt. Users report that the model often ignores the retrieved context and falls back to its training-data knowledge. Which mitigation is most likely to help?

• AIncrease the number of retrieved chunks to 50
• BUse a system-prompt instruction that explicitly tells the model to ground answers in the provided context and to refuse if the context does not support the answer, plus a reranker to surface the highest-relevance chunks first
• CSwitch to a smaller base model
• DRemove all instructions from the system prompt

Atajos de teclado: Desactivados