chathq.ai logo
chathq.ai

Evidence verifier

AI models are good at producing confident text, including text that is wrong. ChatHQ assistants are built to answer only what your content supports. Before a reply is sent, the evidence verifier checks it against the sources it was built from. If the check finds no support, the assistant does not guess.

How it works

  1. Retrieve — The assistant finds the passages in your knowledge base that are relevant to the question.
  2. Draft — It writes an answer using those passages, not general knowledge.
  3. Verify — Each claim in the draft is checked against the retrieved passages. Claims with no support are removed. If nothing supported remains, the assistant falls back instead of answering.
  4. Reply — The answer is returned together with the sources it used and a confidence score.

This verification runs on every message, in the widget and through the API.

When the assistant says it does not know

If the question cannot be answered from your content, the assistant shows your fallback message instead of inventing a reply. The default fallback apologises and suggests contacting your support team. You can change the wording under the assistant's Fallback & Privacy settings.

A fallback is the right outcome when the answer truly is not in your sources. It is a problem when the answer was there and the assistant missed it. In that case, the fix is usually in the content:

  • Make sure the relevant document is uploaded and has finished processing.
  • Use clear, well-structured text. Headings and short paragraphs are easier to match than long walls of text.
  • Split very large files into smaller, focused documents.

See Knowledge base for processing times and content tips.

Note:

API responses include a confidence score between 0 and 1 and a fallback_triggered flag, so your own systems can tell a grounded answer from a fallback. See the REST API.

Reinforce it in your instructions

Grounding is always on, but your instructions can make the behaviour explicit for visitors. A line such as "Only answer from the provided context. If the answer is not there, say so and suggest contacting support." tells the assistant how to phrase the fallback in your voice. See Instructions & behavior.

Test it

Open the assistant from the Assistants page and use the test chat. Ask questions you know are covered by your content and a few that are not. A grounded assistant answers the first group with sources and falls back on the second.

Related

Was this article helpful?

Build AI assistants you can actually rely on

Turn your company's knowledge into assistants your team and customers can rely on.