"We told it not to make things up" isn’t a safety measure. It’s a request.
Every AI support tool tells its model to stick to the provided information. Models mostly comply, and the times they don’t are exactly the times that matter: an unusual question, a gap in your content, a customer pushing for a definite answer you haven’t published.
Helpie doesn’t rely on compliance. It checks.
How the check works
When Helpie answers a question, it first searches your indexed content and hands the model a numbered set of passages. The model is required to mark which passage each claim came from.
After the model replies, and before anything reaches the customer, Helpie parses those markers and validates them:
- No markers at all? The answer is discarded. It may well have been correct, but there’s no way to know, so it doesn’t go out.
- A marker pointing at a source that was not provided? Discarded. That’s the signature of a model filling a gap from general knowledge and attributing it to your site.
Only an answer that’s genuinely traceable to something Helpie found gets shown. The markers themselves are stripped from the display, so the customer reads clean prose with source links underneath.
This runs in your code, on your server, after the model has spoken. It isn’t a prompt the model can be talked out of.
The salvage pass
There’s one refinement worth knowing, because it stops the strictness becoming annoying.
Sometimes a model writes a perfectly good answer entirely from the provided sources and simply forgets to mark it up. Throwing that away and telling the customer to raise a ticket would be a bad outcome caused by a formatting slip.
So on that specific failure, Helpie asks once more for the same answer with the markers added, then re-runs the check. If it passes, the customer gets their answer. If it fails again, the handover happens.
Correct answers aren’t lost to bookkeeping. Unsupported answers still don’t get through.
Secrets never leave
A second filter runs over every reply looking for things that should never appear in a customer-facing message: API key shapes, access tokens, private key headers, and text that looks like the assistant echoing its own instructions back.
This is belt-and-braces rather than a response to a known problem. The assistant has no reason to produce any of it. The filter is there because "no reason to" and "can’t" are different guarantees.
Three things it always knows
Strict grounding creates one predictable annoyance: questions with obvious answers that aren’t written on any page. "What currency are your prices in?" isn’t usually a sentence anyone publishes.
Helpie supplies three always-available sources so those questions work without you writing anything:
- What your site does, which Helpie can draft for you by reading your richest indexed pages and the plugins you run, then leaving it for you to edit.
- A free-text block you write yourself, for anything true about your business that the indexer can’t see.
- Your store settings, read automatically: currency, timezone, date and time formats.
That last one means the assistant knows what today’s date is in your timezone and how you write dates, so it doesn’t answer an Australian customer with an American date format.
Why this is the feature that matters
Everything else about an AI support tool is a convenience. This is the part that decides whether you can leave it switched on.
A support assistant that occasionally invents a returns policy is worse than no assistant, because you now have a documented statement from your own website that you have to either honour or retract. The enforcement is what makes the rest of it safe to use.