3.3.4. The Generative Answers Node
💡 First Principle: The generative answers node is RAG with a steering wheel: retrieval and generation happen per-node, under configuration you control — which sources, what instructions, how strict the moderation — so different moments of one conversation can have entirely different grounding rules.
The node's configuration surface is the skill. Data sources: scope this node to chosen knowledge — specific documents, SharePoint locations, indexes, connectors — independent of what the agent globally knows (and overriding it: node-level sources take priority; agent-level knowledge is the fallback). Custom instructions: shape tone, format, and constraints for the generated answer ("answer in one paragraph, cite the policy section, refuse if the question is about pending litigation"). Content moderation level: trade recall for precision — higher moderation answers less but errs less, the setting to raise when wrong answers cost more than no answer. Plus citations by default and configurable behavior when retrieval comes back empty — the hook fallback logic hangs on.
That empty-retrieval hook matters architecturally: "no answer found" can route onward — to escalation, to another topic, to a human — turning the node into a conditional answering step rather than a dead end. Classic pattern: try scoped answers first; if nothing, fall through to search a broader source or transfer to an agent.
⚠️ Common Misconception: The generative answers node only works with public websites. Websites were the demo, not the boundary; the node grounds on SharePoint, documents, Dataverse, Azure AI Search, and the rest of the knowledge family — scoped per node, cited, and moderated.
Reflection Question: Legal wants strict, citation-heavy answers from the compliance library; the same agent's FAQ topic wants friendly answers from the help center. One agent, two generative answers nodes — enumerate the three settings that differ between them.
