How the AI answers

Every time a visitor sends a message, AskVoro assembles an assistant for that one turn: your published help center, plus whichever tools that visitor is currently allowed to use.

What it can draw on#

Your help center. Published articles and FAQs. This is the foundation — an AI with an empty help center can only apologise politely.

For documentation questions, the assistant calls its help-center search tool. AskVoro searches both exact wording and meaning, then combines those rankings, so a visitor does not have to use the same phrase as your article. Only published content in the public knowledge base selected by that widget is in scope.

Publishing an article or FAQ, or pushing docs with the CLI, refreshes its search chunks automatically. The content save finishes first; if the embedding service is briefly unavailable, exact-word search keeps working and the next edit or push retries the semantic index.

Your tools, if you've defined any, for looking up real data in your systems. See Custom tools.

Verification tools, if you've configured a site URL and signing secret, so the visitor can prove an email address is theirs. See Visitor verification.

The visitor's own connected accounts over MCP, if you've set that up. See MCP connections.

The toolset changes per turn#

Tools aren't a fixed menu. A tool you've marked as needing verification does not exist in the assistant's toolset until that visitor verifies — it isn't offered and refused, it simply isn't there.

One practical consequence: when a visitor verifies mid-conversation, the newly available tools appear on their next message, not the one that triggered verification.

Human handoff#

If you've left handoff enabled, the AI can offer to connect a visitor with your team when it can't help.

When someone takes over from the inbox, the AI is muted for that conversation. If a reply was already being generated at that moment, it's discarded rather than posted — so a visitor never sees the assistant and a human talking over each other.

Hand the conversation back and the AI resumes.

You can turn handoff off entirely in Settings if you'd rather it never offer.

Getting better answers#

Nearly every "the AI got it wrong" turns out to be a help center problem, and the fix is usually to write something down rather than to change a setting.

It invented an answer. Something was missing or ambiguous. Models fill gaps. Write the article, including the edge case it guessed at.

It gave an outdated answer. An old article still says the old thing. The AI has no way to know which of two contradictory articles is current — delete the stale one.

It won't answer a question you can answer. Check the article is published. Drafts are invisible to the AI.

It's vague. Your article was vague. Reading its answer is often the fastest review your documentation will ever get.

It won't look something up. Either no tool covers it, or the tool needs verification and the visitor hasn't verified.