How to Build a Knowledge Base That Makes Your AI Chatbot Accurate

Every AI chatbot tutorial skips to the fun part — connect the tool, paste the embed code, watch it answer. Then real customers arrive and the bot invents a return window, quotes a discontinued product, or promises free shipping you never offered.

The problem is almost never the tool. It’s the knowledge base. A chatbot can only be as accurate as the documents you feed it, and most businesses feed it a sitemap scrape and hope.

We built a complete knowledge base for a fictional international electronics retailer — Novavolt Electronics — specifically to test what a support bot does well and where it breaks. Here’s the document set that worked, why it’s structured the way it is, and the four mistakes that cause most wrong answers.

The five document types a support bot needs

1. A product catalogue with real specs

Not “we sell laptops.” A structured list where every product has a name, SKU, price, key specs, warranty length and stock status. Ours looks like this:

SKU: NV-APH100
Name: AuroraPhone X100
Price: $649
Display: 6.5-inch OLED, 120Hz
Battery: 4,800 mAh, 45W fast charging
Warranty: 24 months
In stock: yes

The stock field matters more than it looks. Without it, the bot happily recommends the out-of-stock StreamCam 4K Pro. With it, the bot can say “expected back 5 December” instead of selling something you can’t ship.

2. A shipping policy with actual numbers

“Fast international shipping” is useless. The bot needs the table:

Region	Standard	Express
US & Canada	3–6 business days	1–2 business days
East Africa	7–12 business days	4–6 business days

Plus the awkward truths: free shipping threshold ($75 in our fictional store), signature required over $500, no PO boxes, and — critically — who pays customs. For Kenya our policy says the carrier bills duties on delivery, estimated 25–40% of item value. If that sentence is missing, the bot will improvise an answer that costs you a refund later.

3. A returns policy with the exceptions spelled out

The return window (30 days) is the easy part. What trips bots up:

• Non-returnable categories. Our policy lists opened earbuds (hygiene seal) as non-returnable. Customers ask about this constantly, and the answer is surprising enough that a bot without the document will guess “sure, 30 days.”
• Who pays return shipping. Change-of-mind vs defective are different answers in our policy — the customer pays one, we pay the other.
• The exchange rule. We don’t do direct exchanges; it’s return-and-reorder. A bot that doesn’t know this invents an exchange process that doesn’t exist.

4. Troubleshooting guides for your top products

Write these as ordered steps, not paragraphs. Our EchoBuds guide starts:

1. Put both buds in the case, close the lid, wait 10 seconds.
2. Clean the charging contacts with a dry cotton swab.
3. Forget the device in Bluetooth settings.
4. Hold the case button 10 seconds to factory reset.

Ordered steps give the bot a procedure to quote instead of a blob to summarise. End each guide with the escalation path — “contact support, this may be a hardware fault covered by warranty” — so the bot knows when to stop.

5. An FAQ that captures how customers actually ask

This is where regional and policy quirks live. Ours includes:

• “Is the AuroraPhone X100 dual-SIM?” (yes — nano-SIM + eSIM, not two physical SIMs)
• “Do you take M-Pesa?” (yes, for Kenya deliveries)
• “Will the TV work on 240V?” (yes, 100–240V multi-voltage)

These are the questions where a confident wrong answer does real damage. If customers in your market ask about a local payment method or voltage, put it in the FAQ.

The four mistakes that produce wrong answers

1. Feeding the bot your whole website. Your About page, your blog, your careers page — none of it answers “where is my order.” Scrape only pages a customer would quote to a support agent.

2. Contradicting yourself. If the returns page says 30 days and an old blog post says 60, the bot picks one at random. Before you upload anything, search your own docs for the policy and make sure there’s exactly one answer.

3. Missing the negative space. The bot needs to know what you don’t do: no PO boxes, no exchanges, no phone support on Sundays. If the documents don’t say it, the model fills the gap with whatever sounds reasonable.

4. Pricing buried in prose. “Starting at just $79 for a limited time!” belongs in an ad, not a knowledge source. Put prices in the structured catalogue where the bot reads them as facts, not marketing.

How to check it works

Don’t just ask the bot “what’s your return policy” — it will always sound right. Ask trap questions: a product that doesn’t exist, a discount code you never issued, a service you don’t offer. The correct answer to all three is “I don’t have that information” — not a confident invention.

We use a fixed set of 18 questions, including three traps, and re-run it every time the knowledge base changes. The exact set — and how to score the answers — is in 18 questions to test your AI chatbot.

The bottom line

An hour spent writing down your real policies — including the awkward ones — does more for chatbot accuracy than any prompt engineering or model upgrade. Tools like Chatbase will happily ingest whatever you give them; the accuracy ceiling is set by you.