Fabrikam Legal wants generative answers in Copilot Studio grounded only on an internal SharePoint knowledge library of approved policies. Users must not receive model answers invented from general training data when no policy chunk is found. Which configuration best meets this requirement?

Generative answers in Copilot Studio retrieve relevant chunks from configured knowledge sources and ground the LLM response on those chunks. Enterprises often require knowledge-only behavior so the model does not "hallucinate" policy from general training data.

The configuration that adds the SharePoint site as a knowledge source and configures generative answers to use knowledge sources only, with a clear fallback message when content is not found, is correct because it keeps responses grounded and controllable.

Enabling Bing search as a knowledge source so the agent can fill gaps with public legal websites when SharePoint returns nothing is wrong because Bing/public web expands the answer surface beyond approved policies and can introduce non-authoritative legal content.

Removing all knowledge sources and relying on the foundation model with a system prompt that says "only answer legal questions" is wrong because a system prompt alone does not enforce retrieval grounding; without knowledge sources, the model can still fabricate plausible policy language.

Setting the agent to classic orchestration only and deleting the generative answers node entirely is wrong because removing generative answers and knowledge defeats the goal of policy question answering; classic topics alone would require hand-authored question-and-answer content for every policy question.

Exam tip: Grounding = knowledge sources + generative answers settings that prefer (or require) retrieved content, with a safe fallback when nothing matches.