How to Reduce Chatbot Hallucinations With Better Sources
Your chatbot just told a customer that returns are accepted within 90 days. Your policy says 30. Or it invented a shipping option you don't offer, or quoted a price that hasn't been real since last spring.
That's a hallucination — the bot confidently making something up. It is one of the most common reasons small business owners give up on AI chat. The good news: most hallucinations aren't a model problem. They're a source problem. Feed the bot better information and the made-up answers mostly stop.
This guide walks through how to reduce chatbot hallucinations with better sources, step by step. We'll cover what actually causes the problem, how to clean up your knowledge base, and which tools make this easier — Chatbase, Botpress, and Tidio among them. Everything below comes from vendor documentation and pricing pages rather than paid hands-on testing.
What a hallucination actually is (and isn't)
A hallucination is when the bot answers with something that sounds right but isn't in your content. It's not the same as a wrong answer caused by bad source material — that's your fault, not the AI's. The distinction matters because the fixes are different.
When a customer asks "do you ship to Canada?" and your help center says nothing about Canada, the bot has two choices: say "I don't know" or guess. Most models guess, because guessing feels more helpful. That's the whole mechanism.
So the first rule of reducing hallucinations: give the bot a clear, complete answer for every question you can predict. Silence in your sources is an invitation to invent.
Start by auditing what the bot is actually reading
Before you touch anything, find out what sources your bot is trained on. In Chatbase, that's the "Sources" tab — you can see every document, URL, and text snippet you've uploaded. In Botpress, it's the knowledge base attached to each AI agent. In Tidio, Lyro pulls from your FAQ entries and connected help content.
The classic failure here is a bot trained on a sitemap that quietly includes an old blog post or a retired pricing page. The bot quotes that number for months, and nobody notices until a customer complains.
Pull up your source list and ask three questions for each item:
- Is this still accurate today?
- Is it written for customers, or is it internal jargon?
- Does it contradict anything else in the list?
Any "no" or "maybe" is a source to fix or delete. Contradictions are the worst offender — if two documents say different things, the bot picks one at random, and it may pick the wrong one every time.
Write answers, not documents
Here's the most common mistake: business owners upload their entire 40-page employee handbook and expect the bot to figure it out. It won't do that reliably. Long documents bury the answer in the middle of paragraphs, and the bot has to summarize — which is exactly where hallucinations creep in.
Better approach: write short, direct Q&A pairs. One question, one answer, 2-4 sentences. "What is your return window?" followed by "30 days from delivery, unused items only, with receipt." That's it.
A shop that bakes custom cakes, for instance, tends to have about sixty questions that actually matter — hours, allergens, lead times, delivery zones. Written out as clean pairs, that is a small enough corpus to read in one sitting and a far better source than a website full of marketing copy. The mechanism is the same regardless of industry: when you replace a messy brochure site with specific answers, the wrong-answer rate has less room to happen.
If you're using Chatbase, this is the fastest path. Upload a plain document of Q&A pairs and the bot picks them up in minutes. Botpress gives you the same thing but with more control — you can attach specific knowledge bases to specific workflows, so your sales bot and your support bot don't mix sources.
**Keep one source of truth per topic.** If your return policy lives in three places — the website FAQ, a PDF, and a help article — you have three chances for them to disagree. Pick one canonical location and point everything else to it, or delete the duplicates.
Tidio handles this well for small stores because Lyro AI answers from your FAQ section, and there's only one FAQ section. You update it once and the bot follows. That simplicity is a feature, not a limitation.
For bigger operations, Botpress's knowledge base feature lets you tag and version content, which helps when multiple people edit. It's more work to set up, and honestly, it's overkill for a five-person shop.
Tell the bot what it doesn't know
This is the single most underused fix. Most chatbot platforms let you add a fallback instruction — something like "if the answer isn't in your sources, say you'll connect the customer with a human and don't guess."
An explicit "don't guess" instruction is the single most underused fix here. It won't eliminate invented answers, but it changes the bot's default behavior from "always answer" to "answer when confident," which is the behavior you actually want in a customer-facing bot.
Pair that with a real handoff path. Chatbase includes a helpdesk with human handover from its Basic tier up. Tidio routes to live chat, which is its core product. Botpress can trigger a workflow that hands off to your team. If the bot can't answer and there's no human to catch it, customers just leave.
Test with the questions customers actually ask
Don't test with "what are your hours?" Test with the weird ones. The question a customer asked last Tuesday that stumped your team. The one about combining a discount with a gift card. The one about whether your product works with a specific older device.
Build a list of 20-30 real questions from your inbox and support tickets. Run them through the bot every time you update sources. Note which ones it gets wrong and why — usually it's a gap, not a bug.
Keep that list somewhere you can reuse it. The pattern tends to be consistent: bots fail on edge cases where the source material is vague, and succeed where someone wrote a specific answer.
Which tool fits which situation
**Tidio** is the easiest starting point for small e-commerce stores. Its own documentation describes live chat working in under 10 minutes, and Lyro AI answers from your FAQ automatically. The catch: Lyro conversation credits run out quickly on cheaper plans, and advanced analytics sit on higher tiers. If you're a Shopify store with a modest FAQ, start here.
**Chatbase** is for teams who want to train a bot on their own docs and embed it anywhere. Upload your content, get a bot in minutes, access advanced models. Downsides worth knowing: there's no free plan anymore — the free tier was withdrawn in October 2026, so testing means the 7-day trial or paying for Hobby at $40/mo. Human handover requires the $150/mo Basic tier, which is a real jump.
**Botpress** is the most powerful option here and the hardest to learn. It's LLM-agnostic — you can bring your own model — and the workflow engine gives you fine control over which sources feed which conversation. The free tier includes 100 conversations/month with 3 seats and 3 AI agents. But there's a genuine learning curve, and conversation packs scale costs at $65 per 100 on Plus. Only pick this if you have someone technical or a real need for control.
A quick maintenance habit
Sources rot. Prices change, policies update, products get discontinued. Set a monthly reminder to review your bot's source list the same way you'd review a menu or a price sheet.
Also check the bot's unanswered questions log — most platforms surface this. Every "I don't know" is a source you're missing. Add the answer, and the next customer gets it right.
Frequently asked questions
How often should I update my chatbot's sources?
Monthly is a good baseline for most small businesses. Update immediately whenever you change pricing, policies, hours, or product availability — those are the topics customers ask about most and the ones a hallucination will make most embarrassing.
Can better sources eliminate hallucinations completely?
No. Better sources reduce them substantially, but no chatbot is perfect. The realistic goal is fewer wrong answers, more honest "I don't know" responses, and a reliable path to a human when the bot isn't sure.
Do I need a technical person to fix my chatbot's sources?
For Tidio and Chatbase, no — both are built for non-technical users, and editing FAQ content or uploading documents takes minutes. Botpress is the exception; its workflow engine and knowledge base setup really do benefit from someone comfortable with technical tools.
What's the fastest fix if my bot keeps inventing answers?
Add an explicit instruction telling the bot not to guess when the answer isn't in its sources, and route those conversations to a human. Then write a short Q&A entry for each question it got wrong. That combination fixes most cases within a day.
Ready to pick one?
Marcus Reed — Marcus Reed is the senior editor at ChatbotRadar. He tracks live chat and AI chatbot tools for small teams, verifying pricing on vendor sites and trialling tools on their free plans. How he tests · Report an error
Disclosure: this article contains affiliate links. If you buy through them we may earn a commission at no extra cost to you. Pricing and features are re-checked regularly but can change — always confirm on the vendor's site.