Your AI shouldn’t guess.
It should cite.
HalluciBlock only answers from the facts you teach it. Every reply carries its sources — and when there’s no fact to stand on, it says so instead of inventing one.
“Can I return this 40 days after I bought it and get a full refund?”
“Yes — absolutely. We offer a 60-day, no-questions-asked refund on all orders, so you’re well within the window.”
“Our policy is a full refund within 30 days with proof of purchase — 40 days is outside that window. I can connect you with the team about an exception.”
A bot you can vouch for — in three moves.
Teach it
Paste a doc, type a fact, or point it at a URL. HalluciBlock breaks it into atomic, checkable facts and stores them. No fine-tuning, no GPUs, no wait — it knows it the moment you hit save.
It verifies before it speaks
Every answer is assembled from stored facts and checked against them. In Extractive mode the reply is your fact, word-for-word — so “zero hallucinations” is a property of the system, not a hope.
Ship it & learn
One line of HTML puts it on your site. Then watch the questions roll in: anything it couldn’t answer becomes a gap you close with a single fact.
Not a smarter guesser. A bot that won’t.
Every question your bot couldn’t answer lands in an inbox. Close it with one fact — your knowledge grows from real demand, not guesswork.
Checked deterministically — answer equals source, byte-for-byte. No LLM judge, no probability.
⛔ refuses instead of bluffing
Free UK delivery over £25.
Re-teach something and it flags the change — never silently overwrites.
<script src="…/widget.js" data-key="sk_…"></script>Don’t take our word for it. Interrogate one.
This is a real HalluciBlock bot taught the facts of a sample coffee store — and nothing else. Ask what it knows, then try to trip it up. It will refuse before it invents.
👋 This is a real HalluciBlock bot trained only on a demo store’s facts. Ask it something — then try a question it was never taught and watch it refuse instead of inventing an answer.
“Zero hallucinations” is measured, not marketed.
We publish a live grounding & faithfulness report from adversarial runs — jailbreaks, off-topic traps, the lot. If a release ever regressed, the numbers below would move on their own.
read the full report →Start free. Pay when it earns its keep.
Citations, the widget, and the grounding guarantee are in every plan.
- 200 chats / day
- 50 trainings / day
- Embeddable widget
- Answer citations
- Shared infrastructure
- 5,000 chats / day
- 1,000 trainings / day
- Fast managed cloud inference
- Analytics dashboard
- Knowledge gaps & conflict review
- Remove “powered by”
- Email support
- Unlimited usage
- Dedicated / self-hosted GPU
- Your own model, in your region
- Fully isolated — data never leaves your infra
- Extractive (provably 0) mode
- SSO, SLA, audit logs
Billed securely via Stripe · cancel anytime from the portal.
Answered from facts.
How is this different from ChatGPT or a generic chatbot?+
Generic models generate plausible text and will confidently make things up. HalluciBlock only answers from facts you explicitly taught it, shows the sources, and refuses when it has none. You own the knowledge.
Do I have to train or fine-tune a model?+
No. You “teach” facts (paste text, type them, or point it at a URL) and they’re stored and indexed instantly. No GPU training step, no waiting.
What happens when it doesn’t know something?+
It says so — and logs the question as a knowledge gap. You add a fact and it can answer next time. You improve it from the questions people actually ask.
How does it get on my website?+
Paste one <script> tag before </body> and a branded chat bubble appears. Or call the streaming chat API directly with a key.
Is my data private?+
Each tenant’s knowledge is isolated. Enterprise runs on a dedicated cluster in a region you choose, with your own model — nothing shared.
Can I cancel anytime?+
Yes — from the Stripe customer portal in your billing settings. You keep access until the period ends.