8.1 Configure the Generative Answers Node

8.1 Configure the Generative Answers Node

Quick Answer: Add Advanced → Generative answers in a topic, open Properties, choose knowledge (agent-wide or Search only selected sources), set content moderation, control general knowledge and web options, and decide whether the node sends a message or stores the answer in a variable for custom formatting and cards.

AB-620 leaf skill Configure generative answers node sits inside Configure topics. Exam scenarios often ask whether an agent should answer from SharePoint only, escalate when moderation blocks content, or behave differently under classic conversational boosting versus generative orchestration.

Why the generative answers node exists

Authored topics match intents and run deterministic nodes. Users still ask free-form questions that no trigger phrase covers. A Create generative answers node lets that topic (or a system topic such as Conversational boosting) search knowledge sources, ground a model response in retrieved chunks, and return a summarized answer—often with citations back to the source document or page.

You can use generative answers as:

• Fallback when no topic matches (Conversational boosting / unknown intent path)
• Intent-specific retrieval inside a deliberate topic (policy FAQ, product manuals, onboarding guide)
• Focused override that ignores noisy agent-level sources for a sensitive conversation path

Adding the node

1. Open Topics → the topic you need.
2. Add node → Advanced → Generative answers.
3. A Create generative answers node appears.
4. Open … → Properties to configure knowledge, classic data, advanced options, and moderation.
5. For Input, makers commonly bind System.Activity.Text (latest user message) so the node answers the current utterance.

Save the topic and republish the agent after meaningful knowledge or moderation changes, then retest with a new conversation so stale session state does not hide configuration mistakes.

Knowledge sources you can use

Microsoft documents internal and external sources. Capabilities and limits differ by orchestration mode.

Supported file/content patterns commonly include modern SharePoint pages and docx / pptx / pdf. Always verify current limits for unstructured folders and large files in Microsoft Learn before exam day—numbers change, but the idea of per-mode caps does not.

Authentication and least privilege

When the source requires identity (SharePoint, Dataverse, many connectors), the agent calls on behalf of the chatting user. Configure agent authentication (manual Microsoft Entra ID with delegated Graph permissions such as Files.Read.All and Sites.Read.All when SharePoint content is required). The exam expects: users only see knowledge they are allowed to open—not a shared maker service account that leaks restricted libraries.

Search only selected sources

The Knowledge sources section includes Search only selected sources. This is a top AB-620 trap.

Important behaviors:

• Selected sources replace agent knowledge for that node; they do not “add on top.”
• Explicitly selected sources ignore trigger conditions that might otherwise limit when an agent-level source is eligible.
• There is no automatic fallback to agent-level knowledge if selected sources return nothing. (Agent-level generative answers fallback is a separate system-topic behavior, not a silent second search inside this node.)
• Turning a source off at agent level removes it from nodes that inherit the agent set; nodes with explicit selections keep searching those selected items as configured.

Answer using only custom data from a variable

For pipeline-fed context (API payload already in a variable):

1. Turn Search only selected sources on and leave knowledge sources unselected (no enterprise search).
2. Under Classic data, set Custom data to the variable holding content.
3. Keep Web search and Allow the AI to use its own general knowledge off so the answer depends only on your variable.

That pattern is pure grounding on maker-supplied text, not a free web answer.

Classic conversational boosting vs generative orchestration

Think of two layers: how the agent plans (classic vs generative orchestration) and how knowledge is searched inside boosting or a generative answers node.

Classic orchestration + conversational boosting

• Topics match by trigger phrases; tools run when topics call them.
• Default Conversational boosting system topic hosts a generative answers path as fallback when intents do not match.
• Knowledge searches obey classic caps (examples from Microsoft Learn: roughly four public website URLs, four SharePoint URLs per generative answers node, two Dataverse knowledge sources with table limits, limited connector counts—memorize the pattern of lower fixed caps).
• Custom data and Bing Custom Search remain relevant via classic data / node configuration.

Generative orchestration

• Planner chooses topics, tools, and knowledge using names, descriptions, and context.
• Higher practical limits (for example many more website/SharePoint URLs in generative mode; unlimited-style connector/Dataverse patterns in generative mode—again, know “generative scales higher than classic”).
• When more than about 25 distinct knowledge sources exist, an internal model filters which sources to search; strong descriptions matter.
• Uploaded files are treated specially relative to the “25 sources” filter (files are not counted the same way as discrete knowledge entries).
• Generative orchestration does not treat custom data / Bing Custom Search the same as first-class agent knowledge—embed those via a generative answers node and classic data when needed.

Exam trap: Enabling generative orchestration does not remove the value of a topic-scoped generative answers node. Use the node when you need a predictable, moderated, source-locked answer inside a regulated flow—even if the planner could also search knowledge on its own.

Grounding, general knowledge, and ungrounded responses

Grounding means the model’s answer is based on retrieved knowledge (and tools), not free invention. Related agent settings (Generative AI / Knowledge pages):

Responsible AI pattern: keep moderation High (default), prefer enterprise sources over open general knowledge for HR/legal/medical-adjacent content, and disclose that generative answers may be AI-generated (Conversation Start is a common place).

Content moderation on the node

Moderation levels range from Lowest to Highest (product default High):

• Lower → more answers, higher risk of undesirable content
• Higher → fewer answers, stricter filters

Precedence at runtime:

1. Topic-level setting on the generative answers node wins when set
2. Else agent-level Generative AI moderation
3. Prompt tools can further override completion moderation when configured

If answers suddenly vanish after raising moderation, diagnose content blocked by content moderation before blaming SharePoint indexing.

Citations and customized delivery

By default, the node can Send a message with the generated answer and platform-managed citations. Advanced pattern for Adaptive Cards or branded layouts:

1. Expand Advanced on the node.
2. Create a global variable to store the generated answer.
3. Clear Send a message so the node does not immediately post the raw answer.
4. Author a Message or Question node (or Adaptive Card) that renders the variable with Power Fx.

Teams citation caveat: When you customize delivery, citation links may not auto-render the same way. Microsoft documents that Teams automatically returns citation links mainly for non-customized answers; customized Teams experiences must explicitly render citations. Also note channel limits (Teams citation count and title/snippet length caps).

Limitation: Citations from knowledge generally cannot be fed as structured inputs into subsequent tools/actions—plan tools to re-query systems of record rather than parse citation objects.

End-to-end scenarios

Scenario A — HR policy isolation

Topic Leave policy should never mix marketing web pages. Turn Search only selected sources on, select only the HR SharePoint site, moderation Highest, general knowledge off. Message node summarizes remaining leave after a Dataverse tool; generative answers handles “what is parental leave eligibility?” from the policy library.

Scenario B — Conversational boosting fallback

Classic agent relies on authored sales topics. Unmatched questions hit Conversational boosting’s generative answers node over product PDFs. If boost fails, Fallback system topic escalates. Maker tunes classic caps so only four public URLs and approved SharePoint paths are searchable.

Scenario C — Custom data only

A prior agent flow returns a JSON-like string of ticket notes into Global.TicketContext. Generative answers node uses custom data variable only, no web, no general knowledge, stores answer in Global.TicketSummary, then Adaptive Card shows summary + “Open ticket” button.

Scenario D — Generative orchestration with 30+ sources

Descriptions on each knowledge source disambiguate “warranty” vs “shipping” content. Planner filters sources; sensitive legal topic still uses an explicit generative answers node with Search only selected sources for the legal library.

Troubleshooting checklist

1. Confirm authentication and user can open the SharePoint/Dataverse content in a browser.
2. Check Search only selected sources vs agent knowledge inventory mismatch.
3. Lower moderation temporarily in a dev agent to see if filters hide answers—then restore policy-compliant levels.
4. Verify indexing/file type support (docx/pptx/pdf/modern pages).
5. For customized answers in Teams, verify citation rendering code/card fields.
6. Ensure Input is the user question you think it is (System.Activity.Text vs a stale variable).
7. Republish and test a fresh session; review activity map under generative orchestration.

Success checklist

• Node added under Advanced with correct input binding
• Knowledge scope intentional (inherit vs selected-only vs custom data)
• Moderation set and understood relative to agent default
• Classic vs generative implications documented for the solution
• Grounding/citation behavior matches channel and compliance
• Sensitive topics do not accidentally search the open web

Mastering the generative answers node is how AB-620 candidates turn static topics into grounded, moderated, source-aware conversations without authoring hundreds of FAQ topics by hand.