How to build a chatbot knowledge base that answers correctly

A chatbot knowledge base is the set of information an AI chatbot uses to answer questions about your business. Think of it as the chatbot’s reference shelf: your website pages, product catalog, pricing, shipping and refund policies, FAQs, and the short answers your team gives every day.

Without one, a chatbot can only answer from the general knowledge built into its AI model. That model does not know your prices, your return window or which plan includes which feature, so it either refuses or, worse, guesses. With a knowledge base, the chatbot looks up the right facts before every reply and answers from them.

This is what separates a modern conversational AI assistant from an old decision-tree bot. You do not script every path; you give it good material and it finds the answer.

The technical name for this is retrieval-augmented generation (RAG). It sounds complex, but the idea is simple and happens in four steps.

1. Ingest. Your content is collected: a crawler reads your website pages, a store connection pulls in products, and you add anything else by hand (policies, Q&A pairs, internal notes).
2. Chunk. Long pages are split into short passages of a few paragraphs each. Small pieces make it possible to find the exact paragraph that answers a question, not just the page it lives on.
3. Retrieve. When a visitor asks something, the system searches the chunks for the most relevant ones. Good systems combine meaning-based search (“refund” matches “money back”) with plain keyword search (exact product names, codes and SKUs).
4. Answer and check. The AI writes a reply using only the retrieved passages. A careful setup then checks the reply against those passages before sending, so a made-up price or feature never reaches the customer.

The quality of each step adds up. Thin content, missing pages or stale prices in step one cannot be fixed by a smarter model in step four.

Start with what customers actually ask. For most businesses that is:

• Website pages. Home, features or services, about, contact. These carry most of the “what do you do” answers.
• Pricing and plans. Every price, what each plan includes and its limits. This is the most-asked topic and the most damaging one to get wrong.
• Product catalog. For stores: names, prices, sale prices, availability and links. Import it from your store rather than copying it by hand, so it stays in sync.
• Policies. Shipping, returns, refunds, warranty, privacy. Write the actual numbers (“30 days”, “2–5 business days”), not vague promises.
• FAQs. Your existing FAQ page, plus the questions your team answers by email every week.
• Q&A pairs. Short, exact answers for questions where wording matters: “Do you ship to Canada?”, “Is there a free trial?”, “Can I cancel anytime?”.
• Rules and tone. What the bot must never do (promise delivery dates, give legal advice) and how it should sound.

Leave out what would confuse it: old blog posts with outdated prices, internal-only pages, duplicate pages and anything you would not want a customer to read.

1. List your top 30 questions. Pull them from your inbox, chat history, sales calls and search console. These are your test set.
2. Fix the source pages first. If the answer to a top question is not clearly written on your website, write it there. The chatbot and your human visitors both benefit.
3. Connect your website. Let the chatbot crawl your site, or submit a sitemap so it finds every important page.
4. Connect your catalog. If you sell products, connect your Shopify or WooCommerce store so product names, prices and links come straight from the source.
5. Add Q&A pairs for the tricky ones. Anything with a precise answer, legal wording or a common misunderstanding gets its own pair.
6. Write the rules. Tone of voice, things to avoid, when to hand over to a person.
7. Test with your 30 questions. Ask each one the way a customer would, including typos and vague wording. Fix the content, not the question, when an answer is wrong.
8. Go live and review weekly. Read the conversations, especially the ones the bot handed off or could not answer. Each one points at a missing or unclear piece of content.

A knowledge base goes stale quietly. Prices change, a product sells out, a policy is updated, and the chatbot keeps quoting the old version. A few habits prevent this:

• Re-crawl your website after any change to pricing, plans or policies.
• Sync your product catalog from the store instead of maintaining a copy.
• Delete or exclude pages you have retired, rather than leaving them online.
• Put dates on time-limited offers so they stop being mentioned when they end.
• Review unanswered questions monthly and turn the common ones into content.

The biggest risk with any AI chatbot is a confident wrong answer: a price that does not exist, a feature you do not have, a refund you never offered. A good knowledge base setup guards against this in layers.

• Answer from retrieved content only. The model is told to use the passages it found, not its general memory, for anything about your business.
• Check before sending. The reply is compared with the knowledge base; claims that are not supported get fixed or removed.
• Allow “I don’t know”. When nothing relevant is found, the bot should say so and offer a person, not improvise.
• Hand off with context. When a human takes over, they should see the whole conversation so the customer does not repeat themselves.

Sukar is live chat with an AI customer service agent built in, and the knowledge base is set up for you rather than as a separate project:

• Automatic crawl. When you add your website, Sukar crawls it and builds the knowledge base from your pages. Refresh a source any time you update your site.
• Store import. Connect Shopify, WooCommerce or YouCan and your products, prices and links are imported so the AI can recommend the right item.
• AI Playbook. Pick your business type and goal, set a voice, and add facts and rules such as “never promise delivery dates”.
• Q&A pairs. Add exact answers for the questions where wording matters.
• Reply checks. Replies are checked against your knowledge base before they are sent; when the AI cannot answer, it offers your team instead of guessing.

The size of the knowledge base depends on your plan:

Because the same knowledge powers sales conversations too, the AI can answer a pricing question and then help the visitor pick a plan or product. Read more about the AI sales agent and how AI sales agents work.