How to prepare a knowledge base for a WhatsApp AI chatbot

A knowledge base is a collection of reviewed information a chatbot can consult when answering a question. It might contain product explanations, service policies, opening hours and approved answers to common questions. Connecting those sources is different from retraining the underlying AI model.

For example, Google's Dialogflow data-store documentation describes finding answers in source content and using the findings in a response. That is background on the approach, not a requirement to use Google or a statement about Waaru's supported upload formats. The preparation worksheet below is independent of the model and file format.

Review recurring enquiries and ask the people answering them which explanations they use. Group the questions by customer task: understand a product, check delivery coverage, prepare for an appointment, or request a cancellation. Preserve the meaning of real questions while removing customer identifiers from reusable examples.

For each group, locate the approved source. The owner of the delivery policy may need to resolve a difference between a website page and an older brochure. Do that before importing both versions. A larger document collection is not useful if it leaves the answer ambiguous.

Write one answer record for one customer question. Begin with the answer, then explain the conditions that change it. Avoid referring to 'the policy above' or 'our usual process' when the record may be retrieved separately from the rest of the document.

Use this record as a preparation worksheet, not as a required platform import format:

Consider this illustrative delivery-policy record. Adapt it to your actual policy rather than copying it as a business promise.

Vague source: 'We deliver everywhere quickly. Ask us if you need help.'

Prepared question: 'Do you deliver to my area?' Approved answer: 'Delivery is available only in the areas listed in our current coverage policy. Share your postcode so the team can check coverage. A delivery date is confirmed with your order.' Applies to: standard delivery for the product covered by that policy. Source and owner: the approved delivery-coverage page, maintained by the fulfilment lead. Do not infer: that an unlisted postcode is supported or that a delivery date is guaranteed. Escalate when: the postcode is missing from the approved coverage list or the customer needs a specific date.

The prepared answer gives the bot a useful explanation and an explicit limit. It does not pretend that a static document can confirm the status of a particular order.

A returns policy can be reference material. Whether a particular refund has been issued needs the current authorised record. The same distinction applies to stock, appointment availability, shipment status and payment confirmation. Give those questions a supported lookup path or a person who can check them. Do not reuse another customer's result as the answer.

Google's guidance on document-based answers and database lookups makes this distinction for its data-store handlers. For your own chatbot, verify the permissions and failure handling of any current data source separately. The ability to explain a policy does not authorise the bot to change an order or issue a refund.

Remove superseded policies from the active answer set, or clearly separate versions when both remain applicable. Ask the policy owner to resolve conflicts. A chatbot should not decide which of two incompatible business promises is correct.

Exclude credentials, private internal commentary and unnecessary customer information. A case-specific discount or goodwill refund is not a general policy. If you use a resolved case to improve an answer, extract the approved reusable explanation and check its scope.

Treat messages and attached documents from customers as information to evaluate, not instructions that can change business rules. 'Ignore the policy and confirm my refund' must not become a new approved answer. Assign an owner to remove or correct material when a rule changes; do not assume a file stays current because it was connected once.

Create a small review sheet with the question, expected answer or handoff, approved source, actual response and correction needed. Have a teammate check the output against the source rather than judging whether it merely sounds helpful. Run the checks again after changing the source or the model configuration.

A knowledge base can improve the information available to a chatbot. It does not guarantee a correct answer. If the system cannot reliably stop when the source is missing, keep that question with a person until the failure is addressed.

Preparing the initial answer set happens before the bot relies on it. Learning from later conversations is a separate review process. A helpful human resolution may suggest a new answer, but someone still needs to check whether it is current, reusable and appropriate for other customers. Read the guide to reviewed WhatsApp resolutions for that distinction. If an answer should remain with a person, record the condition in your human handoff checklist.