Where to look when your Now Assist in Virtual Agent answers are off

[Claire_Conant]

Now Assist in Virtual Agent is answering, but you're left questioning why the answers are off. Maybe it's pulling from the wrong article, skipping a topic you intended, or giving an out-of-date answer. Each situation could be caused by a few things, but they're usually traced back to configuration settings, not to the model itself. That pattern holds across most environments—the model is doing what it's been set up to do, and the setup is where the fix lives.

Does this apply to you?

This applies to any instance running Now Assist in Virtual Agent on Zurich or Australia. The configuration paths referenced here (the search application configuration, Virtual Agent Designer, and the VA Bot to Bot Provider record) require that AI Search and Now Assist in Virtual Agent are both activated. If either isn't active on your instance, some sections won't apply.

Start here: match your symptom to the fix

• The right source isn't used at all → check the assistant's search source scope, and that AI Search is activated.
  See: The Virtual Agent skips the intended article or returns the wrong one
• The right source is used, but the answer is weak or partial → the article needs restructuring for retrieval.
  See: The article is in scope, but it's written in a way retrieval can't use
• A topic isn't suggested when it should be → strengthen the topic description.
  See: A topic you expected never gets suggested
• The wrong topic wins, or the user gets an ambiguous list → adjust promotion and ordering.
  See: Two topics match but the wrong one wins
• A change you made didn't take → confirm the topic is published, active, and tested in the right language.
  See: You changed a topic but the old behavior persists
• The answer is right but renders wrong through the API → check the synthesized response citation strategy.
  See: Answers arrive jumbled or lose their citations

For how Now Assist in Virtual Agent grounds and routes a request in the first place, read Now Assist in Virtual Agent doesn't fire topics. It runs plans. The options below make more sense once you've seen that model.

The Virtual Agent skips the intended article or returns the wrong one

You know the answer lives in a specific knowledge article, but Now Assist skips it or returns something weaker. This is the most common response-quality issue, and it's rarely because the model fails to understand the question.

Why it's not working	Try this instead
Now Assist retrieves only from the assigned search sources defined in the search application configuration. A source might be a knowledge base, a catalog, or another indexed data source, and the configuration may cover all your content or just part of it. If your intended article's source isn't among them, retrieval won't find it, no matter how well written it is.	Check the search application configuration for the chat experience the assistant uses and add the article's source if it's missing. Then, confirm that AI Search itself is activated for Virtual Agent, since the AI Search Fallback setup topic depends on it. If AI Search was never requested for the instance, fallback answers won't generate at all.

The article is in scope, but it's written in a way retrieval can't use

You've confirmed that the source contains your intended article, but the answer comes back vague, partial, or pulled from the wrong section. The article matches, yet retrieval can't isolate the right part of it because the article isn't structured for machine retrieval.

Why it's not working

Try this instead

Now Assist parses knowledge articles as HTML and text and chunks them by their headings, so answer quality depends heavily on structure. An article written as one long block, or one that buries the answer under weak headings, gives retrieval little to anchor to.

This fix lives in the content, not the configuration: clear, descriptive headings, one primary intent per article, and the answer stated plainly near a heading that names it. If a single article tries to answer several unrelated questions, splitting it into focused articles tends to improve what comes back. Route this to whoever owns the knowledge base, rather than treating it as an assistant setting.

A topic you expected never gets suggested

You built an LLM topic to handle a specific request—say, password resets or PTO balances—but Now Assist answers some other way and never surfaces your intended topic.

Why it's not working	Try this instead
With LLM topic discovery, it's the topic description that the model matches against instead of intents or keywords doing that work. A thin description, like "This topic is about the holiday calendar," doesn't give the model enough to match on.	Open the topic Properties tab in Virtual Agent Designer and expand the description using plain-language detail about its purpose and the requests it should catch. Spell out what users ask, the variations they use, and the tasks the topic covers.

One caution: NLU utterances don't work as LLM descriptions. They're built around keywords and phrasing, while the LLM works from meaning and logic, so a list of sample utterances tends to underperform a written explanation of what the topic does.

Two topics match but the wrong one wins

When two topics could both answer the question, the assistant has to pick which surfaces first—and that's where it goes sideways. Either the wrong topic keeps winning, or the user gets a list of options to choose from.

Why it's not working	Try this instead
This issue comes down to promoting a topic—it tells the assistant which topic to put first. If a less relevant topic is promoted ahead of the one you want, or the one you want isn't promoted at all, the experience skews accordingly.	In Assistant Designer, check the Promoted status of the competing assets and reorder them so the stronger match leads. If two topics genuinely overlap in scope, tightening their descriptions so each clearly owns its lane reduces the collisions at the source. Overlapping descriptions are what produce the ambiguous list in the first place.

You changed a topic but the old behavior persists

You updated a topic description or flow to fix a bad answer, tested it, and nothing changed. This one tends to send admins debugging in the wrong place because the configuration looks correct.

Why it's not working	Try this instead
For changes to take effect, a topic has to be published, active, and discoverable. NLU topics have one more requirement: the model also must be trained and published for the session language. LLM topics skip that step since they're matched on their description rather than a trained model. Either way, an edit saved but not published won't show up in conversations.	Confirm that the topic is published, and that you're testing in the same language and domain the topic is bound to. In Virtual Agent Designer, the Asset library test option and the Include topic discovery check box in the test window show what the assistant actually sees when it evaluates a user's question, which separates a publish problem from a matching problem quickly.

Answers arrive jumbled or lose their citations

With Now Assist in Virtual Agent, you expect the answer to come back the way it looks in the portal: composed, readable, with citations the user can follow. Instead, the answer arrives jumbled, missing its citations, or oddly formatted.

Why it's not working	Try this instead
A synthesized response is the written answer Now Assist composes from what it retrieves, along with its citations. How those citations appear is a presentation setting, controlled on a configuration record rather than in the topic itself—so it affects how the answer is displayed, not what the answer says.	Before you suspect the answer content, check the synthesized_text_citation_strategy property on the VA Bot to Bot Provider record. This property sets whether citations appear as endnotes, inline links, both, or are suppressed for the downstream system to render itself.

Where to go from here

Most response-quality issues come down to reach (is the source in scope), readiness (is the content retrievable), or routing (which topic wins), and each of those is something you can check and adjust yourself.

If you've worked through the matching symptom and the answer is still wrong, capture the specific utterance and the retrieved result and take it to the Community forum, where other admins on the same release can compare behavior. The NLU prediction feedback tables (open_nlu_predict_intent_feedback and open_nlu_predict_entity_feedback) are useful evidence to bring, since they show what the assistant actually predicted and selected for a given utterance.

More info on this topic

• Now Assist in Virtual Agent doesn't fire topics. It runs plans:  Read about the grounding and routing model these scenarios build on.
• Improving the user experience with AI Search: Confirm how AI Search fallback and search sources determine what an assistant can retrieve.
• LLM description and instruction guidelines for Virtual Agent topics: Use this when strengthening a topic description so LLM discovery can match it.
• Promote or demote LLM assets in Assistant Designer: Use this to control which assets the assistant suggests and in what order.
• Resolve NLU topic discovery issues: Go here for the publish, language, and binding checks when a topic isn't discovered.
• Enable synthesized response in Virtual Agent API: Use this to set how composed answers and citations render through the API.