Trust & accuracy

An assistant is only useful if the people talking to it can trust what it says. ChatHQ is built so that an assistant answers from your own knowledge base, says so when it does not know, and never pretends to do something it cannot do.

This section explains what the assistant does to earn that trust and what you can do to reinforce it.

What the assistant does

• Evidence verifier — Before replying, the assistant checks that its answer is supported by your sources. If nothing supports it, it says it does not know instead of guessing.
• Source citations — Every answer is linked to the content it came from, so your team can check it.
• Staying on topic — The assistant declines requests that have nothing to do with your business.
• Sensitive questions — On guarantees, policies and money, it answers only what your documents say and points people to the official source or a person.
• No fake actions — When it cannot do something, it shares the right link or steps instead of claiming it did it.
• Resisting manipulation — Messages crafted to make the assistant ignore its rules do not work.

How this fits with your settings

These behaviours are built into every assistant. Three settings shape how they show up in your conversations:

• Knowledge base — Defines what the assistant knows. Everything it says must trace back here. See Knowledge base.
• Instructions — Define the assistant's role, boundaries and tone, including what to do when the answer is not in your content. See Instructions & behavior.
• Fallback message — The reply visitors see when the assistant cannot answer. Set it under the assistant's Fallback & Privacy settings.

Related

• Reliability — How the team keeps the assistant working release after release.
• Glossary — Plain-language definitions of fallback, hallucination, prompt injection and other terms used here.