What to put in your AI agent's knowledge base, and what to leave out

An AI agent is only as good as what it has been given to read. Give it your whole website including the blog and it drowns in noise; give it nothing and it answers every question with "let me check". The right knowledge base is small, specific and current. Here is what goes in, in what form, and what stays out.

By Ritchie, ReplyKit · Published 5 September 2026 · 8 min read

Quick answer

Start with your website's services, prices, FAQ, about and contact pages, add your price list and any policy documents as PDFs, then write plain text for everything a new employee would be told on day one but that is not written anywhere. Add taught answers as customers ask new questions. Leave out marketing fluff, internal documents, personal data and anything you would not want quoted back to a customer.

How the agent uses knowledgeThe four sources and what each is forThe day-one checklistWhat should never go inKeeping it currentFAQ

How the agent uses knowledge

Retrieval-based agents like ReplyKit do not memorise your documents. On every message they search the knowledge base for the few passages most relevant to what the customer just asked, and answer from those. Two consequences follow. First, clarity beats volume: one clean price list outperforms twelve pages that mention prices in passing. Second, the agent only knows what is written. If your parking arrangement lives in the receptionist's head, the agent cannot answer it. ReplyKit also tells you: every question it could not answer from knowledge appears in an unanswered-questions list, and you can teach the answer in a click.

The four sources and what each is for

Website crawl

Best for services, prices, about, contact and FAQ pages. ReplyKit reads up to 100 pages, follows links two levels deep, reads useful pages first and blog posts last, and skips login, cart and legal pages. After a crawl you see the page list and can remove any page that adds noise.

PDFs

Best for price lists, menus, timetables, brochures and policy documents. Upload the current version; when it changes, upload the new one and delete the old. Scanned image-only PDFs have no text to read, so export them again with text.

Plain text

Best for the unwritten rules: parking, deposits, what to bring, who does what, holiday hours, "we do not do X". Write it the way you would brief a new hire. Short headed paragraphs work better than one long block.

Taught answers

Best for the long tail. When the agent could not answer something, the question appears in the unanswered list. Type the answer once and it is knowledge from then on. Review the list weekly for the first month, then monthly.

The day-one checklist

  1. Services, what each includes, and who it is for.
  2. Prices, packages, deposits, and what is not included.
  3. Hours, public holidays, and any seasonal changes.
  4. Location, landmarks, parking, public transport.
  5. How to book, cancel and reschedule, and the fees for late changes.
  6. Preparation and what to bring.
  7. The ten questions you are most tired of answering, with the answers you give.
  8. The things you do not do, so the agent can say so instead of improvising.
  9. Team names and roles, if customers ask for people by name.
  10. Payment methods and delivery or collection options.

If your website already covers half of these, the crawl handles it. Write the other half as plain text; it usually takes twenty minutes and it is the highest-value twenty minutes of the setup.

What should never go in

Keeping it current

A knowledge base decays. Prices change, a stylist leaves, holiday hours arrive. Three habits keep it honest: re-crawl the website after you change it, replace PDFs rather than adding to them, and check the unanswered-questions list on a schedule. In ReplyKit, each WhatsApp agent and each website assistant has its own knowledge base, so a branch or a product line can be updated without touching the others. The test chat lets you ask the new question yourself before a customer does.

Frequently asked questions

How many pages should I crawl?

As many as are useful, up to 100. ReplyKit reads services, prices, FAQ, about and contact pages first and blog posts last, and shows you the page list so you can remove noise.

Does the agent read scanned PDFs?

Only PDFs with real text. A scanned image has no text to read; export the document again with text, or paste the content as plain text.

What happens when a customer asks something not in the knowledge base?

The agent says it is not sure and offers a handover, and the question is added to the unanswered list so you can teach the answer once.

Can I keep separate knowledge for different agents?

Yes. Each WhatsApp agent and each website assistant in ReplyKit has its own knowledge base and instructions; nothing is shared between them unless you upload it to both.

Is my knowledge base used to train AI models?

No. Your content is used to answer your customers and is not used to train shared models. See the privacy policy for the providers involved.

Paste your website and see what it learned

The page list shows exactly what was read. The test chat shows what it answers. 7-day free trial.

Start free
Written by Ritchie, ReplyKitPart of the small team in Malaysia that builds and runs ReplyKit. Writes about WhatsApp automation, AI customer service and small business lead response.

How we know this. ReplyKit's knowledge base is built from four sources (website crawl, PDFs, pasted text and taught answers), and we designed the crawler to read up to 100 pages with services, prices and FAQ pages first. The advice here is what we tell customers whose agents keep saying 'I am not sure'. Product figures in this article (limits, prices, per-reply costs) are taken from ReplyKit as it runs today and are re-checked when we update the page. Where we cite outside research, the source is linked below. We sell ReplyKit, so read our product claims with that in mind; we say where a different tool or no tool is the better choice. About ReplyKit.

Sources and further reading

  1. Personal Data Protection Department, Malaysia (PDPA 2010)
  2. ReplyKit Acceptable Use Policy · what we refuse to automate
  3. ReplyKit Privacy Policy · how customer content is processed and which providers see it

ReplyKit is an independent product and is not affiliated with or endorsed by Meta or WhatsApp. This article is general guidance for running a small business, not legal, financial or regulatory advice.