The knowledge base, explained: the documents your AI agent is allowed to answer from, and why it matters more than the AI
People imagine an AI agent "learns" their business. It does not. It reads. Every answer it gives comes from a pile of documents you provided, the knowledge base, and if the answer is not in the pile, a well-built agent says so. Understanding that one fact explains most of what makes an agent good or bad.
Quick answer
A knowledge base is the set of documents an AI agent is allowed to answer from: your website pages, price list, policies, FAQ and taught answers. When a customer asks something, the agent searches the knowledge base for the most relevant passages and writes a reply from them. It is not training; nothing is memorised, and updating a document updates the answers immediately. A good knowledge base is small, specific and current.
How the agent uses it
- A customer message arrives.
- The agent searches the knowledge base for the few passages most related to the question.
- It writes a reply using those passages and its instructions about tone and boundaries.
- If nothing relevant is found, it says it is not sure and hands over, and logs the question.
This is why one clear price line beats twelve pages that mention prices in passing, and why a fact that lives in your head cannot be answered. The retrieval step is explained further in how retrieval works.
Knowledge base versus training
Training changes the model itself and takes data, time and money; it is how the underlying model learned language. A knowledge base leaves the model untouched and supplies your facts at answer time. For a business that means: your data is not baked into anything, updating a price is uploading a new file, and the same model serves every business with a different knowledge base. It also means your documents are not used to train shared models, which ReplyKit's data practices spell out.
What makes a good one
- Specific. Facts, prices, hours, policies, in plain lines.
- Small. Only what customers ask about. Marketing copy and internal documents add noise.
- Current. Replace documents when they change; delete the old version.
- In customer language. Answers phrased the way customers ask, which is what the search matches.
The checklist is in what to put in the knowledge base, and the price list format in the price list guide.
Sources in ReplyKit
A website crawl (up to 100 pages, useful pages first), PDFs with real text, pasted text, and taught answers from the unanswered-questions list. Each WhatsApp agent and each website assistant has its own knowledge base; nothing is shared between them unless you upload it to both.
Frequently asked questions
What is a knowledge base in AI customer service?
The documents an AI agent is allowed to answer from: website, price list, policies, FAQ and taught answers. The agent searches it for each question and answers from what it finds.
Does the agent learn from my documents?
It reads them at answer time. Nothing is memorised or trained; updating a document updates the answers immediately.
What happens if the answer is not in the knowledge base?
A well-built agent says it is not sure, hands over, and logs the question for you to teach.
How big should a knowledge base be?
As small as covers what customers ask. Specific and current beats large.
Is my knowledge base used to train AI?
Not by ReplyKit. It is used only to answer your customers.
Give it something to read
Website, price list, notes, taught answers. 7-day free trial.
Sources and further reading
- Harvard Business Review, The Short Life of Online Sales Leads (Oldroyd, McElheran, Elkington, 2011) · firms that responded within an hour were about seven times more likely to qualify the lead
- WhatsApp Business Messaging Policy · opt-in, opt-out and prohibited-use rules every business must follow
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.