S
Stitex
AI adoption

Building a Support Chatbot That Doesn’t Hallucinate

Short answer: for a support bot to stop reciting scripts without inventing answers, it has to reply from your own documentation through RAG and escalate honestly when it cannot. Here is what separates that kind of bot from “just ChatGPT” and from the old scripted ones.

July 21, 20268 min readStitex Technologies

Two extremes that annoy customers

Old scripted bots march people through “press 1” menus and cannot parse a real question. A bare ChatGPT is the opposite: it sounds smooth but does not know your product and confidently invents things — terms, prices and features that do not exist. Both end the same way, with an irritated customer asking for a human.

The setup that works: RAG plus honest escalation

A support bot worth deploying answers from your knowledge base: documentation, procedures, the history of resolved tickets. Technically that is RAG — the bot first retrieves the relevant passage from your material, then answers from it rather than from memory. The mechanics are covered in how to train an AI on your own data.

The rule that matters
If the answer is not in your material, the bot says so and hands the question to a person instead of composing something plausible. This is enforced hard: “I don’t know” beats a well-phrased lie every time.

What it takes off your team

Support problemWhat the bot does
Customers wait hours for a replyanswers instantly, around the clock
70% of tickets are routinecloses them from the documentation itself
Knowledge lives in people’s headsanswers the way your best specialist would
A genuinely hard caseescalates to a human with the full conversation history

Where it runs and what happens to the data

  • Channels: a website widget, Telegram, email — all sharing one knowledge base,
  • Guides the customer to the next step: a request, a payment, a booking,
  • Shows what people ask about most, which is free product research,
  • Conversations contain personal data, so processing runs on our infrastructure or on your own server.

More about the product on the AI Support page.

Frequently asked questions

How do you stop a bot from making things up?

It has to answer from your knowledge base through RAG rather than from general knowledge, and it has to say “I don’t know” when the answer is not in the material. That rule is enforced strictly: no source, no answer — the question goes to a person instead.

How fast can it launch?

If the documentation already exists, three to seven days. Most of that time goes not into the technology but into gathering and tidying the material the bot will answer from.

What happens with complex tickets?

It does not try to solve them. It escalates, handing the specialist the conversation history along with its own reading of the problem, so the human starts with context instead of from scratch.

We will build it on your documentation

Show us your knowledge base or your site and we will put together a working demo of AI support on your own material.