HelpBrain
A support chatbot that says “I don’t know”
Hallucination is not a defect that gets patched out. It is what a language model does when it is asked to answer something it has no source for. The only reliable fix is to stop asking it to — which means detecting the absence of a source and refusing.
Behaviour
How it behaves
Three things it does that you can check, rather than three adjectives.
Every answer is scored for confidence before it is sent. Below the threshold the bot does not attempt an answer at all — it says it does not know and offers to pass the question to a human.
Answers are generated only from content retrieved out of pages you published, and each one cites the URL it came from, so anything it says can be checked against the source in one click.
Retrieved content and visitor messages are fenced as data in every prompt, with per-request random delimiters, so text on a page cannot issue instructions to the bot.
Live demo
Try it on your own site
Paste a URL. It trains on a capped sample of that site and answers your questions for real — no account, no card.
HelpBrain AssistantLive demo
Paste any website to see it live
We'll crawl it — then ask it questions in seconds.
Or try one of these:
- We crawl a capped sample of the site
- A real bot trains on it in about a minute
- Ask it anything — answers cite the page they came from
Powered by HelpBrain
FAQ
Questions
Can you guarantee it never makes something up?
No honest vendor can guarantee a model's output. What we can do is not ask it to answer unsupported questions, check every real answer against your own pages every night, show you every failure, and credit back the month if it states something your content does not support.
Is refusing bad for the customer?
A refusal that collects their email and reaches a human beats a confident wrong answer that sends them away satisfied and misinformed.
Every price is on the pricing page, and the accuracy policy explains the credit-back.