How to Build an AI Knowledge Base for Your Store

Build an AI knowledge base by collecting approved store documents, splitting them into clear sections and connecting them to your agent. Keep changing product information in the catalogue feed. Assign a document owner to review policy changes and fill gaps from customer questions.

You’ll need access to the current policies, product manuals and the agent’s document settings. A customer service lead must approve the facts before customers receive them. Use the sections below in this order.

1. Choose the documents and remove old versions.
2. Rewrite unclear sections and staff replies into reusable policy answers.
3. Load the documents, connect the catalogue and set the agent’s tone.
4. Test questions against the approved answers, then assign ongoing reviews.

What Makes an AI Knowledge Base Different

An AI agent knowledge base is the set of store documents an agent searches before it replies. A help centre gives people pages to browse. The agent needs passages that make sense when retrieved separately from the rest of a page.

An AI-powered knowledge base can find passages by meaning, beyond exact keyword matches. A customer asking “Can I send these shoes back?” can retrieve a section about returns. The answer still depends on the policy’s conditions. Retrieval doesn’t turn a vague policy into a reliable rule.

Keep each document’s title and source with its text. If two documents disagree about a return window, resolve the conflict before loading them. Let the agent hand an unresolved case to staff instead of choosing between incompatible rules.

Which Documents to Include, Payments Too

Include the store documents customers need to make a purchase and understand what happens afterwards. Start with this inventory.

• Shipping regions, delivery estimates, dispatch times and collection rules.
• Returns eligibility, exclusions, return costs and refund processing.
• Sizing charts with units and instructions for taking measurements.
• Care instructions for the materials or products they cover.
• Warranty terms and the route for reporting a faulty product.
• Store locations, opening hours, contact details and service availability.
• Accepted payment methods and buy-now-pay-later terms.

Leave out drafts, staff-only notes and outdated promotions. Keep an approved promotion only while its terms apply. Remove customer names and order details from reusable documents.

For Afterpay questions, document whether your store accepts it and where it appears at checkout. State any store-specific purchase restrictions and how cancellations or refunds are handled. Record the regions and currencies your store supports. Add any minimum or maximum purchase rules it actually applies.

Separate your store’s terms from decisions the payment provider makes. The agent mustn’t promise that Afterpay will approve a customer’s purchase or change their repayment schedule. An answer about payment eligibility needs the relevant approved provider terms as well as your store’s checkout rules.

Loading Product Manuals

Load product manuals, assembly guides and specification sheets so the agent can answer setup and usage questions before and after purchase. Link each document to the exact models or product identifiers it covers.

Use text-based PDFs or web pages where your chosen system supports them. A scanned image needs text extraction and a reading check before it becomes a usable document. Confirm that headings, diagrams’ instructions and table values survive the import.

For a hypothetical desk range, attach the assembly guide to the desk model, not every item in the furniture category. State which fittings belong to that model. Keep a safety instruction with the step it governs, and avoid merging instructions from different editions.

After loading a manual, ask a question whose answer appears in it. The agent must give the instruction for the requested model. If the answer uses another model’s instructions, check the document’s product mapping and imported text before adding more content.

Writing for AI Retrieval

Write short sections with clear headings and one fact per sentence so the agent can retrieve the rule a customer needs. Keep a condition next to its exception. Replace “as above” with the actual subject when a passage might be read alone.

Consider this hypothetical policy paragraph before restructuring:

Dispatch is usually quick but custom pieces take longer. Express delivery is available for some items depending on where you live.

A usable layout separates the unanswered questions:

Standard dispatch

Ready-made items leave the warehouse within two business days.

Custom dispatch

Allow ten business days for custom pieces to leave the warehouse.

Express eligibility

Express delivery is available for ready-made items sent to metropolitan addresses.

Those are invented example terms, not promises for your store. Replace them with the approved facts. The worked example shows how to organise a page.

For each section, ask what the customer needs to decide. “Dispatch” must distinguish warehouse processing from delivery after dispatch. Keep those durations separate so the agent doesn’t present a dispatch estimate as an arrival promise.

Turning Canned Responses Into Agent Answers

Rewrite canned responses into policy facts before adding them to a chatbot knowledge base. Staff macros often contain greetings, merge fields and decisions about one customer’s order. Strip those elements and retain the reusable rule.

Here’s a hypothetical refund email template staff might use after approving a change-of-mind return:

Thanks for contacting us. We’ve approved your return under our 30-day change-of-mind policy. We’ll refund your original payment method within five business days of inspecting the returned item.

Turn that reply into a short refunds document:

Customers can request an eligible change-of-mind return within 30 days of delivery. Approved refunds go to the original payment method. We process the refund within five business days of inspecting the returned item.

State the window, method and timeframe once. Keep the return conditions in the same document, and distinguish store processing from time a payment provider takes to display the refund. Use your store’s approved durations before loading the text.

A change-of-mind window mustn’t limit a customer’s rights for faulty goods. The ACCC explains that consumer guarantees can’t be taken away by store policies. Keep faulty-product procedures separate and route disputed cases to staff.

Keep order-specific canned responses with people. An apology for one late order, a goodwill refund approval or a reply about a disputed parcel depends on that case. The agent answers the general policy and hands the conversation to a person when a decision is needed.

FAQ Content With FAQ Schema

One set of question-and-answer content can support your agent and a public FAQ page. FAQ schema, also called FAQ structured data, labels the questions and answers for systems that read structured markup. The agent still needs access to the answer text through its configured document source.

Google previously restricted FAQ rich results to authoritative government and health sites. It has since ended the feature entirely, as recorded in Google Search Central’s documentation updates. For a store, FAQ schema describes the content for machines that use it. It doesn’t create a Google FAQ rich result or guarantee AI citations.

To add FAQ schema, first publish the complete question and answer as visible page content. This hypothetical example uses the same dispatch rule in the visible answer and markup:

How long does standard dispatch take?
Ready-made items leave the warehouse within two business days.

Add a JSON-LD block, a script containing structured data, to that page’s HTML:

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [
    {
      "@type": "Question",
      "name": "How long does standard dispatch take?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Ready-made items leave the warehouse within two business days."
      }
    }
  ]
}
</script>

A free FAQ schema generator can draft this markup. A person must check every answer against the visible text and validate the syntax before publishing. Update both representations when the policy changes. The guide to creating a store FAQ page covers the public page itself.

Giving the Agent Your Tone of Voice Guide

Keep tone rules in a separate document so a policy update doesn’t change how the agent sounds. The guide must describe the brand’s normal words, formality and response style.

For a hypothetical Australian homewares store, the guide might ask for warm, brief replies and everyday language. Use “basket”, “delivery” and “we”. Avoid “valued customer” and promotional praise. Use contractions, and explain the next action without blaming the customer.

Include three sample replies that show the style in different situations:

• “The linen cover is machine washable on a cold, gentle cycle.”
• “Your postcode is outside our express delivery area. Standard delivery is available.”
• “I’ll pass this damaged-item claim to our team so they can help with the next step.”

These sample facts are hypothetical too. Tone rules must never let the agent soften an exclusion, promise an exception or invent a delivery estimate. Test the same factual answer with the brand’s preferred wording and confirm the rule stays intact.

Connecting the Product Catalogue

Connect the product feed for prices, stock and product attributes instead of repeating those facts in policy documents. A copied price can survive in an old document after the catalogue changes.

Documents explain how the store handles delivery or returns. The catalogue identifies the product and supplies its current commercial details. Use product identifiers to connect a manual or category guide to the correct items, and remove product-specific stock statements from reusable answers.

Test a product question alongside a policy question. The product answer must match the feed, while the policy answer must match the approved document. The guide to product feed optimisation explains how to prepare the catalogue data the agent draws on.

Keeping Answers Current

Review documents whenever a policy or promotion changes, with a monthly review to catch missed updates. Give each document one accountable owner and keep the approved source version outside the agent as well.

When a policy changes, update the source and its public page before replacing the agent’s copy. Remove superseded versions from retrieval. Then repeat the questions affected by the change before exposing the revised answers to customers.

Use a short acceptance check:

1. Ask a returns question using different wording from the document.
2. Confirm the answer states the approved window and eligibility conditions.
3. Ask about an excluded case and confirm the agent doesn’t promise an exception.
4. Try a question the documents don’t answer. Expect a handover to staff without an invented rule.

If an old answer persists, check whether the previous document remains searchable and whether the replacement has finished processing. If a condition disappears, inspect the section boundaries. Keep the exception with the rule and repeat the failed question. Use the AI answer testing guide to expand the acceptance check.

Filling Gaps From Unanswered Questions

Review unanswered-question logs each month and turn repeated gaps into approved document updates. Group questions by the missing fact, such as rural delivery eligibility or a product’s cleaning instructions.

For each group, decide whether the answer belongs in a policy, a manual or the catalogue. Ask the responsible person for the fact, add it to the existing source and remove any conflicting text. Keep private order problems in the staff workflow.

Replay the original question and one differently worded version after the update. Both answers must match the approved fact and retain its conditions. Leave the gap assigned to its owner until that check passes.