Trust & safety

How does Kav stop the bot from making up answers?

Numbers come from your systems, answers only from approved content, and a confidence gate hands off instead of guessing. Here is each safeguard against invented answers.

By The Kav team Published 2 min read

A generic chatbot invents answers because nothing stops it: it has a plausible guess and no rule that says "don't". Kav is built so that the places an invented answer could come from are closed off, one by one.

Numbers never come from the model

Amounts, dates, reference numbers and statuses reach the customer through a display template, filled straight from your own system's response. The model chooses the template and writes the sentences around it — it never writes the figure. A separate check blocks any sentence that isn't grounded in what your system actually returned.

Answers come only from content your team approved

The assistant answers knowledge questions from your own articles and pages, not from whatever the model happens to know. Content read from your site is only ever a draft: a named person submits it and a different person approves it before the assistant may use it. Every article carries an owner, a version and a last-verified date, and flips back to "needs review" when its source changes.

Search that is checked twice before it answers

Finding the right passage takes several stages: a glossary of your own terms, keyword search and meaning-based search side by side, merged rankings, and a reranking model that reads the question against every passage. Then two gates:

  • Confidence gate. If confidence is still too low, Kav doesn't answer. It says it couldn't find a reliable source and offers a person instead of a guess. Your team sets that threshold in the Studio, and lowering it requires an explicit confirmation.
  • Grounding check. The drafted answer is checked against the passages that were actually retrieved. A sentence the source doesn't support doesn't ship.

Hand-written answers are delivered word for word

For a question you want answered exactly one way, your team writes an answer override. Once approved, it's delivered verbatim — never paraphrased. When the model helps draft one, any number, date or phone number that isn't in your own notes gets the draft rejected.

The right tool answers, not the nearest article

When a question belongs to a specific capability — a balance, a status — your organization can configure that article search is left out of that turn, so an unrelated article can't answer instead. If that capability fails, the conversation takes the normal decline-or-handoff path. It never falls back to a guess from unrelated content.

When in doubt, a person

If the assistant can't answer reliably, the conversation moves to a person in the right team's queue, with everything said so far. Declining is a designed outcome, not a failure to hide.