Data for an AI chatbot knowledge base should describe information your company actually intends to share with its audience. Do not upload a whole drive because the importer accepts documents. Define the audience, the owner of each item and the confidentiality boundary before selecting a file or website address. You can make those decisions in an ordinary spreadsheet, before switching on an assistant.
Choose chatbot knowledge data for its audience
The audience determines which information belongs in the knowledge base. A visitor assistant for a store needs delivery rules and product descriptions, but not staff instructions for negotiating discounts. Write a single purpose statement, such as: “We answer public questions before purchase.” Keep material outside that purpose on a separate review list instead of importing it in anticipation of hypothetical questions.
For each source, record a realistic customer question and the kind of answer it should support. If you cannot provide an example, the document probably arrived on the list by accident. The broader guide to AI in customer support helps define the assistant role, but cannot approve individual data. Do not mix public and private material simply to make the collection bigger.
Separate process knowledge from personal case details
Process knowledge can support a shared answer, while case details should remain in the appropriate service system. “Request a return using this form” differs from a message containing a customer bank account and personal reason for returning an item. Remove identifying data before considering a document for import. Hiding a name alone does not establish that the remaining document is anonymous. At a guesthouse, door codes and guest details stay out of shared knowledge in the same way, as the guide on guest questions answered from house rules explains.
When an answer depends on a signed-in account, a current order or an individual agreement, mark it for a human check. Do not describe a source collection as access to a transactional system. Meet AI knowledge does not place orders or make bookings. Check company privacy requirements and supplier agreements with the responsible person before including personal information in any material intended for the assistant. The same care applies when someone on the team wants to paste a conversation into a personal AI tool; the guide on shadow AI in customer support sets out the rules. If you are considering read access to case-specific data, the guide to AI agent permissions sets out how to scope it.
For a safe transformation, turn “Alex Taylor, order 123, wants an exchange after delivery” into the topic “How do customers request an exchange after delivery?”. Identity and order number do not help another visitor understand the process. If instructions need an example, create a fictional case and label it as an illustration rather than a record about a real person.
Give each source an include, edit or exclude decision
Every item needs an explicit decision and a brief reason. The proposed table supports a conversation between support, the information owner and the knowledge administrator. Technical accessibility does not grant permission to use material. A public page may be outdated, while an internal document may need a separate version whose content is suitable for customer answers.
Scroll the table sideways to see all columns.
| Item | Decision | Check |
|---|---|---|
| Public offer | Include the current version with service scope and conditions. Remove expired promotions. | Check that customers see the same offer when opening the source. |
| Internal instructions | Prepare a separate version approved for answers to external customers. | Remove staff privileges, negotiation exceptions and confidential operating costs. |
| Customer conversation | Use the anonymous question intent rather than importing an individual history. | The answer should explain a process, not reveal an order or account details. |
| Archived PDF | Exclude the document or replace it with an approved current version. | Record its owner, effective date and the source replacing the outdated file. |
Add an approver column so the person responsible for the information settles scope questions. Support can identify a needed question without independently changing pricing policy. You can collect question topics using the method for a website knowledge base and FAQ. At this stage you are deciding whether the source belongs in the collection, rather than drafting the final customer answer.
Choose a format that preserves the meaning
A useful format makes the text and its conditions readable together. A text page or clear document often preserves names and headings more reliably than a scanned brochure. Open the file and check whether you can select the important sentence and its qualification. If a table depends entirely on visual positioning, inspect its meaning again after the import.
PDF and DOCX can supply Meet AI knowledge, alongside FAQs, websites and sitemaps. That does not mean every document layout will be interpreted perfectly. When importing a website, exclude archives, search results and duplicate address variants where they add nothing useful. Prefer a narrow approved collection over an entire domain whose contents nobody has reviewed, and keep the scope recorded with the source.
In a spreadsheet, inspect whether the column description remains beside its value: “14 days” without a “Delivery period” heading may not identify what the number means. In a presentation, check slides with conditions placed in small footnotes. If essential information exists only in an illustration, prepare an approved text description instead of assuming the image will convey the intended meaning.
Resolve conflicts and outdated versions before import
Conflicting sources need a company decision before they reach the assistant. If a price page gives one delivery period and a brochure gives another, the model should not choose your policy. Name the authoritative source, retire the old version and record the effective date. For seasonal offers, also record when the material should be withdrawn rather than waiting for a customer complaint.
Once the list is approved, use the Meet AI knowledge panel to prepare sources and inspect answers through a private test with citations. The account supports site configuration and work on its knowledge; an active application and eligible plan are required. Import is separate from publication. Refreshing sources is a manual action, and a changed version still needs a review before it becomes the accepted reference.
Record the decision for a pair of versions, for example: “February brochure excluded; October delivery page approved by the store owner is authoritative”. This explains why a source disappeared and prevents someone else from adding it again. A file moved into a working archive should not remain an active source for a public assistant merely because it is still available internally.
Accept a source using a question it should answer
Source acceptance checks a specific fact, rather than the fluency of the response. Ask about an exception, service scope or condition stated in the material, then compare the citation with the document. If the information is present but unused, investigate why the chatbot cannot find an answer. Uploading another copy blindly makes the collection harder to manage without identifying the actual problem.
Microsoft documentation on knowledge sources distinguishes sources from the knowledge base that uses them. This selection procedure is our editorial proposal, not a description of a Meetlead integration with Azure.
Keep the acceptance result next to the source inventory: question, cited passage and approver decision. Mark uncertain material as on hold instead of assuming publication will fix it. The guide to updating chatbot knowledge covers ownership and version maintenance. During the first import, already name the person who will receive notice when an offer or process changes.
Questions before your first import
Can I include the entire terms document? You can consider a current document as a source, but inspect readability, scope and the conditions attached to answers. An individual dispute may still require a human decision. Organize the material logically and check citations about exceptions, rather than assuming that an import guarantees a correct interpretation of every clause.
Are internal documents always excluded? No, but a safe version needs a defined audience and an approved scope.
Remove information that customers must not receive. A knowledge base used by a public assistant is not a place for confidential instructions merely because a support operator also works with AI. Keep internal handling details in the appropriate restricted process. Text inside sources can also try to change the assistant's behaviour, as explained in the guide on chatbot prompt injection.
Does more data make answers better? Source quantity is not a quality measure by itself.
Additional files can introduce conflicts, retired versions and information unrelated to real questions. Add a source when it supports a needed intent and has an owner. After adding it, check an answer based on that material, including any important limits or exceptions.
