To prepare a knowledge base for an AI chatbot, start from the questions customers really ask, write one clear, self-contained answer for each, include the specific facts (prices, timeframes, conditions, exceptions), remove anything outdated or contradictory, and keep confidential material out. Then test the assistant with real questions and fix the content wherever an answer is weak. Most problems people blame on "the AI" are really gaps or ambiguities in the content it was given.
This guide walks through the process step by step, with examples and a checklist.
Why content quality decides answer quality
Most website AI assistants answer by finding the parts of your content that match a visitor's question and writing a reply from them. (Our guide to an AI chatbot trained on your data explains this in more detail.) That has three consequences:
- If the answer isn't in your content, the assistant can't give it reliably.
- If your content is vague, the answer will be vague.
- If two documents disagree, the assistant may pick the wrong one.
So the work is less about configuring AI and more about good, plain writing.
Step 1: Collect the questions people actually ask
Don't start by uploading everything you have. Start with demand.
Good sources of real questions:
- Your chat and email history
- Notes from sales and support calls
- Your existing FAQ page (but check it's still accurate)
- Search terms visitors use on your site, if you track them
- Questions your team answers "every single week"
Group them into topics: pricing, ordering or booking, delivery or onboarding, policies, product or service details, account and billing, contact and hours. Most small businesses end up with a manageable list that covers the bulk of what visitors ask.
Step 2: Audit the content you already have
For each topic, check what you have and its condition.
| Question to ask | Why it matters |
|---|---|
| Is it current? | Old prices and retired policies produce wrong answers |
| Is it complete? | Missing conditions lead to misleading answers |
| Does it contradict anything else? | Conflicting sources confuse the assistant and customers |
| Is it written for customers? | Internal jargon doesn't help visitors |
| Is it public? | Anything confidential must stay out |
Delete or update before you add. One accurate page beats three overlapping ones.
Step 3: Write clear, self-contained answers
When the assistant pulls a passage to answer a question, it may only see that passage, not the page around it. Write each answer so it makes sense on its own.
Rules for chatbot-friendly content
- One topic per section, with a heading that names it ("Refunds for annual plans").
- Lead with the answer, then add detail.
- Repeat the subject instead of "it" or "this," which can lose meaning out of context.
- Spell out specifics: amounts, timeframes, steps, conditions, exceptions.
- State what you don't do, too. "We don't offer phone support" prevents the assistant from implying you do.
- Use the words customers use. If people say "cancel," don't only write "terminate subscription."
Before and after
| Vague | Clear |
|---|---|
| "Refunds are available in some cases." | "You can request a full refund within 30 days of purchase by emailing support. After 30 days, purchases aren't refundable." |
| "We ship quickly." | "Orders placed before 2pm on a weekday ship the same day. Delivery usually takes 2 to 4 business days within the US." |
| "Contact us for pricing." | "Website projects start at $3,000. Ongoing support is $200 a month. For larger projects, we quote after a short call." |
(These are example policies. Use your own.)
A simple template for each entry
- Topic: [What it's about]
- Short answer: [One or two sentences]
- Details: [Conditions, steps, exceptions]
- What we don't do: [Related things that aren't offered]
- If the visitor needs more help: [Hand off to the team, or the right contact route]
Step 4: Prepare documents and web pages
Most tools accept several formats, such as typed text, PDFs, Word files and imported web pages. Each needs a little preparation.
PDFs and Word files
- Use text-based files. Scanned images of pages may not be readable unless the tool can recognize text in images.
- Remove drafts and old versions so only the current one is uploaded.
- Keep tables simple. Complex multi-column tables sometimes lose structure; add a sentence summarizing important tables.
- Split very long documents by topic if you can, so each part stays focused.
Web pages
- Import pages with real substance: FAQs, policies, service pages, pricing.
- Skip thin or promotional pages that say little beyond slogans.
- Re-import after you update a page, so the assistant isn't working from the old version.
Typed text
Typed entries are ideal for the short, precise facts that don't live anywhere else: "Our office is closed on public holidays," "We serve customers in the US and Canada only," "Same-day appointments aren't available."
Step 5: Keep confidential information out
Treat the knowledge base as public. Anything in it could appear in an answer.
Keep out:
- Customer names, emails or order details
- Internal processes, margins and supplier details
- Discount rules your team uses at its discretion
- Employee information
- Draft plans or unannounced changes
Step 6: Plan for the questions you can't answer
No knowledge base covers everything, and that's fine if the assistant knows when to stop. Make sure:
- The assistant hands off to a person when your content doesn't answer the question.
- Out of hours, it collects an email so the team can follow up. Our guide to after-hours live chat covers this.
- Visitors can always reach a person directly. See our chatbot to human handoff checklist.
Step 7: Test with real questions before launch
Build a test set of 20 to 30 real questions and run through them all:
- Common questions with clear answers in your content.
- The same questions phrased differently, with typos or slang.
- Questions that need a combination of facts ("Can I get a refund on an annual plan after 2 months?").
- Questions your content doesn't cover. The assistant should hand off, not invent.
- Off-topic questions. It should decline politely.
- A few questions in other languages, if you have international visitors.
For every weak answer, ask: "What's missing or unclear in the content?" Fix that, then test again. Keep the test set and rerun it whenever you make big content changes.
Step 8: Maintain it like a product page
A knowledge base drifts out of date quietly. Prevent that:
- Name an owner responsible for keeping it accurate.
- Update it the same day prices, policies or services change.
- Review AI conversations regularly and turn repeated weak answers into new or better entries.
- Do a full review every few months.
Knowledge base checklist
- Real customer questions collected and grouped by topic
- Outdated and contradictory content removed or fixed
- One clear, self-contained answer per topic
- Specific facts included: amounts, timeframes, conditions, exceptions
- "What we don't do" stated where it matters
- PDFs and Word files are current, text-based versions
- Imported web pages are substantive and up to date
- No confidential or customer data included
- Handoff to a person configured for unanswered questions
- Test set run, weak answers fixed and retested
- Owner named and update routine agreed
How to do this in RetroChat
On RetroChat's Autopilot plan, the AI assistant answers from a knowledge base you build from typed text, uploaded PDFs and Word files, and imported web pages. It's restricted to your business's own topics, replies to visitors in their language, and hands off to a person when it isn't sure. It also drafts suggested replies for agents. You place it in your chatbot flow with an "AI answer" step, so your rules still handle greeting and routing. The free 14-day trial includes 10 AI replies, enough to run a handful of your test questions.
Better content, better answers
Spend your effort on the content: real questions, clear answers, current facts, nothing confidential. Then test, fix and repeat. When you're ready, you can try it on your own content in a free 14-day RetroChat trial.
Frequently asked questions
What should be in a knowledge base for an AI chatbot?
Clear answers to the questions customers actually ask: pricing, policies, services, ordering or booking steps, hours and contact routes. Include specific facts and conditions, and state what you don't offer. Leave out anything confidential.
What file formats work for an AI chatbot knowledge base?
It depends on the tool. Common options are typed text, PDFs, Word files and imported web pages. Use current, text-based files rather than scanned images, and keep each document focused on a topic.
How often should I update my chatbot's knowledge base?
Update it the same day anything customer-facing changes, such as prices or policies. Review AI conversations regularly to find gaps, and do a full review every few months.
Why does my AI chatbot give wrong or vague answers?
Usually because the content is missing, vague, outdated or contradictory. Test with real questions, trace each weak answer back to the content, fix it, and test again. Make sure the assistant hands off when it can't find an answer.