AI knowledge base: what an AI agent answers from

An AI knowledge base is the content an AI agent answers from, cut into short passages and indexed so the agent can find the few that fit each question. For customer support, that’s your help pages, your pricing and policies, any files you upload and the answers you write for questions nobody had covered. It isn’t what the model learned in training, and it isn’t the same thing as knowledge base software. You don’t need a help centre to have a good one, because the agent reads whatever content you give it.

Deacon is an AI customer support agent for founders who handle support themselves. It builds its knowledge base from your website, files, pasted text and your own answers, replies to your customers from it in seconds, and says so when it has nothing to go on. It also shows you what customers ask, so you know what to add next. There’s a free plan, and it needs no card.

Three things people mean by an AI knowledge base

The phrase covers three different things, and it helps to know which one you’re after.

• Knowledge base software with AI added. A help centre tool, where you write articles for customers to browse, adds AI that drafts articles, tags them or answers a search in sentences.
• The knowledge an AI agent answers from. That’s the content behind a support agent or chatbot, plus the index that lets it find the right part, and it’s what this guide is about.
• A personal knowledge base run with AI. Your own notes, searched and summarised by a model, are useful, but they do a different job.

The first two overlap without depending on each other. A help centre can be one source in an AI agent’s knowledge base, next to your website and your files. And an AI agent can answer well from a pricing page, a few docs and a list of written answers, with no help centre at all.

What an AI knowledge base is made of

Whichever tool builds it, an AI knowledge base has three layers.

1. The content. Pages, files and written answers, turned into plain text.
2. The index. The text is cut into passages, and each passage is stored so it can be found by its meaning and, in the better tools, by its exact words.
3. The rules. When a question comes in, the agent finds the closest passages and a language model writes the reply from them, under instructions about what it may and may not say.

That method is called retrieval-augmented generation, or RAG. The model’s training isn’t one of the layers. It supplies the language, but the facts should come from your content, which is why a well-built agent says it doesn’t know when your content is silent instead of falling back on what refund policies usually say. Grounding AI covers that part in depth.

One question on its way through

Here’s what Deacon’s knowledge base does with a single question. A visitor types, “Can my accountant get our invoices without logging in?”

1. The question is searched by meaning, which finds a passage headed “Billing, Invoices” although it never mentions accountants. It’s searched by keyword too, which finds passages that use the word “invoices”.
2. The best of both, up to five passages, go to the model with the question.
3. The passage says invoices are emailed to the billing email set in Settings. So the reply, written in seconds and in the visitor’s language, is yes, make your accountant’s address the billing email.
4. If no passage had said how invoices are sent, the reply would have been that Deacon doesn’t know, with a box for the visitor’s email so you can answer yourself.

Two things decide whether that goes well. The fact has to be in the knowledge base, and it has to be written so a search can find it. How to write knowledge base articles shows how to do the second.

How it differs from a help centre search

A traditional help centre search matches the words you type against the words in its articles, then hands you a list of links. Type “money back” into one whose article is called “Refund policy” and you may get nothing. An AI knowledge base is searched by meaning as well, so the words don’t have to match, and the customer gets a sentence instead of a reading list.

That gap is real. In Gartner’s survey of 5,728 customers in December 2023, 45% of those who started in self-service said the company didn’t understand what they were trying to do, and only 14% of issues were fully resolved there. Searching by meaning goes some way towards fixing the first. Neither kind of search helps when the answer was never written down.

What the agent can read, and what it can’t

An AI knowledge base only knows what reaches it as text. Here’s what Deacon takes from each kind of source, and what it leaves behind.

What’s true for one customer doesn’t belong in a knowledge base. Their invoice, their order and their usage live in your app. Deacon can’t see them, so it always declines a question about the visitor’s own account, invoice or data, whatever your content says. Your site can pass Deacon a few details about the visitor, such as their plan, and it takes them into account when choosing which answer applies.

How big it needs to be

Smaller than you’d think. What counts is whether it covers what customers ask, not how many pages it holds. For a young product, a pricing page, a getting-started guide, your refund and cancellation policies and answers to the ten questions you get most make a good start.

Deacon counts knowledge by the text it takes from your content, not by file size. A page of docs comes to about 15,000 bytes, measured in production in September 2026, so the free plan’s 1 MB holds about 67 pages, Starter’s 15 MB about 1,000 and Growth’s 50 MB about 3,330.

How an AI knowledge base grows

Every question the agent couldn’t answer marks a missing piece, which makes an AI knowledge base easier to grow than a help centre. You don’t have to guess what to write next. The customers tell you.

In Deacon, the questions your content didn’t cover wait in Conversations, each with an Answer this button. Write the answer once and press Save & check. Deacon saves it as a Q&A, asks itself the question again and shows you the reply it now gives. When it answers, it looks at similar open questions and tells you how many more your answer covers. Every answer you write becomes part of what Deacon knows, so it needs you less over time. When your answer and an older page both reach the model and disagree, Deacon is told to prefer yours.

Topics show the shape of the whole thing. They group what people ask, count the people behind each group and label most of them, as a how-to, a feature ask, a bug report or a pricing question. A topic with 20 people behind it and no good answer is the next thing to write.

How to keep it true

An AI knowledge base repeats whatever it holds, so an out-of-date page is as dangerous as a made-up answer. Customers hear the old price in a confident voice.

• When you ship, list the pages the release touches and update them the same day.
• Refresh the agent’s copy. In Deacon, nothing is re-fetched on a schedule, so fetch the changed page again and press Retrain agent. The old version keeps answering until you do, so nobody reads an update you’re halfway through.
• Where two pages say the same thing, keep one, because two copies drift apart.
• Read a week of conversations after a big release. The questions about the new feature show what your pages left out.

How Deacon builds and answers from your knowledge base

Deacon isn’t knowledge base software. It doesn’t publish articles for customers to browse, and you don’t need a help centre to use it. It’s an AI customer support agent that builds its knowledge base from what you already have, answers your customers from it and shows you where it falls short.

You add four kinds of source, your website, uploaded files, pasted text and the answers you write. A crawl lists up to a hundred pages at a time and every page starts unticked, so you choose exactly what Deacon learns from before pressing Retrain agent. Despite the name, no model is trained. Retrain agent cuts the pages you chose into passages and indexes them. A single page added by its address, a file or pasted text starts answering as soon as it’s been read.

Deacon then answers your customers in seconds, at any hour and in the language they write in, from the passages it finds for each question. An answer taken from a page on your site usually links to it. When your content doesn’t cover a question, Deacon says it doesn’t know instead of guessing and asks for the visitor’s email. If they leave it or ask for a person, the owners on your team get an email, and your reply from the dashboard appears in their chat if it’s still open. There are no flows or decision trees to build, and every feature is on every plan.

You also learn what customers need. Analytics › Widget shows how many people wrote in, the share Deacon answered and the languages they wrote in, and every conversation is saved and searchable by what visitors asked. It’s customer feedback you never had to ask for.

Questions about AI knowledge bases

Do I need knowledge base software to have an AI knowledge base?

No. The agent needs your content as text, not a help centre to hold it. Deacon reads pages on your website, files and pasted text, and adds the answers you write. If you already have a help centre, add its public pages like any others.

Is an AI knowledge base the same as training an AI on my data?

No. Training, or fine-tuning, changes the model itself. A knowledge base leaves the model as it is and hands it the passages that fit each question, so correcting a fact means editing a page, and the agent can point to where an answer came from. Deacon’s Retrain agent button updates the index and trains no model.

Can an AI knowledge base be free?

Deacon’s is. The free plan holds 1 MB of knowledge, about 67 pages of docs, crawls up to 300 pages a month and gives you 50 answers a month and one seat. It needs no card, has no end date and isn’t deleted if you go quiet. Every feature is on it. See the plans

What happens when the knowledge base doesn’t have the answer?

Deacon says it doesn’t know, in the visitor’s language, and offers to take their email so you can reply. It won’t fill the gap with a price, a date or a policy detail your content doesn’t state. Why chatbots make things up explains why that matters.

Does every answer say which page it came from?

No. An answer from a page on your site usually shows a link to it. An answer from a file, pasted text or a written answer shows the visitor no link, because there’s no public page behind it. In your dashboard, each answer shows the source Deacon named, or “Your content” when it named none.

Can my team use it as an internal knowledge base?

No. Some AI knowledge base tools answer staff inside Slack or an intranet. Deacon answers your customers on your website, and your team works in its dashboard, reading what customers asked and replying to them.