AI knowledge base: what to include and how to build one

A chatbot is only as good as the information behind it. If the shipping page says five days and the bot says two, the customer believes the bot, and the business has made a promise it never wrote. An AI knowledge base is where you decide what the agent knows. This guide explains what goes into one, how to write entries a chatbot can use, how to test it before customers do, and how to keep it current. It is part of the guide to AI for business.

What is an AI knowledge base

An AI knowledge base is the set of facts a chatbot or agent is allowed to answer from. When a customer asks a question, the system searches the knowledge base for the passages that match, hands them to a language model, and the model writes a reply from them. The approach is often called retrieval-augmented generation, and the chatbot glossary entry explains it in more detail.

Two consequences follow:

• If the fact is not there, the agent cannot know it. A good agent says so instead of guessing.
• If the fact is wrong or old, the agent repeats it confidently.

What to put in it

For an online store, five kinds of content cover most conversations:

• Business profile: who you are, what you sell, contact details and opening hours.
• Policies: shipping costs and times, returns, exchanges, warranty and payment methods.
• Product information: name, price as shown on the site, specs, materials, sizes and a link to the page.
• FAQs: the questions customers actually ask, each with a short answer.
• Rules: what the agent must not promise or discuss, and the tone to use.

Leave out anything you do not want a customer to hear read back: internal margins, supplier names, draft policies.

How to write entries a chatbot can use

Write for retrieval, not for a brochure:

• One question, one answer. An entry about returns should not also cover shipping.
• Put the fact first. "Returns are accepted within 14 days of delivery" beats a paragraph of context.
• Use the customer's words. If customers write "swap", add "swap" next to "exchange".
• Keep numbers in one place. A shipping price repeated in ten entries will be out of date in nine.
• Prefer a FAQ to a long article. A matching FAQ is used almost word for word, while a long article may be trimmed.
• Say what is not covered. "We do not ship to PO boxes" prevents a wrong yes.

A starting set for a new store

If you are starting from nothing, write these first. They cover most of what customers ask:

• Shipping: cost, delivery time by region, and what happens when a parcel is late.
• Returns and exchanges: how many days, who pays for shipping, and how to start.
• Payment: accepted methods and whether installments are available.
• Sizing and materials: one entry per product family, not per product.
• Contact: opening hours, phone and WhatsApp, and how long a reply takes.
• Promotions: how coupons work and whether they combine with sale prices.

Ten well-written entries beat a hundred pasted paragraphs.

Build it from your site

Most stores already have the content, scattered across pages. In StoreChart, the AI knowledge base starts with a scan of your site, which covers up to about 250 pages and extracts products, articles and FAQs. Then you:

1. Review what the scan found. Edit or switch off anything that is wrong or out of date.
2. Complete the business profile, with a phone or WhatsApp number first, because that is where the chat points when it has no answer.
3. Add rules, up to 20, such as "Do not promise delivery before a holiday".
4. Check the preview. Ask the questions customers ask and read the answers.

Changes reach the agent's answers within about a minute, with nothing to publish.

Test before customers do

Write down twenty real questions from your last month of chats, including awkward ones: a typo, two questions in one message, a question your policy does not answer. Ask them all and mark each answer as correct, vague or wrong. Fix the knowledge, not the symptom: a vague answer usually means a missing FAQ.

Example: a question about sale items

A fashion store reads its conversations and sees that visitors ask "Can I return something I bought on sale?" The chat answers vaguely from the general returns page. The owner adds an FAQ with a two-line answer, files it under Returns with the tag Sale, and adds the rule "Do not promise a cash refund, only store credit". She checks the wording in the preview, and a minute later the chat on the site answers with the new text.

The figures in this article's examples are illustrative only and do not reflect actual customer data.

Keep it current

A knowledge base decays. Prices change, products leave the range and a seasonal policy ends.

• Re-scan after changes: the scan is a snapshot of the day it ran, and the chat does not check your store live.
• Switch off instead of deleting seasonal entries, so you can bring them back.
• Read conversations weekly, and turn each missing answer into an entry.
• One knowledge base per agent: StoreChart's website chat and its WhatsApp AI agent keep separate knowledge, so update important policies in both places.

Organise it so your team can maintain it

A knowledge base that only one person understands decays as soon as that person is busy. Give each entry a category and a few tags, such as shipping, returns or product care, so anyone can find and update it. Decide who owns each area: the person who handles returns also owns the returns entries. When a policy changes, change the knowledge base the same day, because the agent will repeat the old version until you do.

Common mistakes

• Pasting the whole site: more text is not better knowledge. Outdated pages drag wrong answers in.
• No rules: the agent has no limits on what it may promise.
• Never testing: the first customer finds the gap.
• Forgetting live data: a knowledge base holds no stock or order status, so those questions need your team.