Most of the work is already done if your website is reasonably current. Before the import, decide what belongs in the knowledge base, remove duplicates, keep one format per document, and make sure your contact details are easy to find.
What should I do before importing content into an AI assistant?
- Ask the people who answer the phone and the inbox what they are asked most. Twenty questions usually cover most of the volume.
- For each question, find the page or document that holds the answer.
- If there is none, write it. A short page with the plain answer is enough.
- Keep the list. It is your test set after the import.
What content should go into the knowledge base?
- Your public website, as it is. The crawler reads it; nothing has to be exported.
- Documents visitors already receive: product sheets, price lists, terms, forms, manuals, FAQs.
- Pages in every language you serve. The assistant answers in the visitor's language, and it finds the right passage faster when the content exists in that language too.
What should I leave out of the knowledge base?
- Outdated versions. If three price lists exist, only the current one goes in.
- The same document in two formats. A PDF and its Word original double the index for nothing.
- Scans when a text version exists. Scans are read with OCR and work, but the text version is more precise.
- Anything with personal data about customers or staff that is not meant for the public.
Can internal documents be added to the knowledge base?
- Yes. Handbooks, process guides, onboarding material and internal FAQs work well when the assistant is used by employees on an intranet page.
- For a public website assistant, add internal documents only if their content may be shown to visitors. The assistant does not know the difference between internal and public.
- If you need both, keep two knowledge bases: one for the website, one for the team.
How should documents be named and structured?
- Keep contact details on a page of their own. When a question cannot be answered, the assistant points to a person, and it finds that person on your contact page.
- Give documents descriptive file names and folders. "Price list 2026" in a folder called "Fees" helps the search more than "final_v3.pdf".
- Write the answer, not only the policy. "Returns are accepted within 30 days" is retrievable; a paragraph that implies it is not.
- One topic per page. Long pages with many unrelated sections retrieve worse than several short ones.
Do scanned PDFs work?
- Yes. Scans are read with optical character recognition during the import.
- A text version of the same document is more precise, so prefer it when one exists.
- Check one or two scanned answers in the preview; poor scans produce poor text.
How do I check the content after the import?
- Ask your twenty questions in the preview.
- Read the answers and the sources they cite.
- Fix the content where an answer is wrong or missing, re-import, ask again.
- Go live and watch the list of questions that need attention for a week. It is the fastest way to find out what your visitors wanted and could not find.
How long does preparing content take?
- Collecting the questions: an hour or two with the people who answer them.
- Sorting the documents: about an hour for most organisations.
- The import itself: minutes, and it runs while you do something else.
The assistant does not make your content better. It makes it visible, including the gaps.