An AI knowledge base is the set of your own documents and pages that an AI agent reads in order to answer customers. For each question it finds the passages that match, then writes the reply from them, so the agent is only as accurate as what you put in and how well you keep it up to date.
That makes the knowledge base the cheapest place to improve an agent. A better document fixes every future answer on that topic, while a rewritten prompt fixes only the cases you remember. This guide covers what to include, how to shape it, how to keep it right, and how to find what is missing.
What an agent does with the knowledge base
When a customer asks a question, the agent searches the knowledge base for the passages closest in meaning to it, reads them, and composes an answer. Two consequences follow, and both shape how you should write.
The agent works with passages pulled out of your documents. A sentence that only makes sense on page three, such as "the same rule applies to these cases", loses its meaning when lifted out.
The agent cannot answer from what is missing. It may still write something plausible, which is why an instruction to say "I will check with the team" when the knowledge is silent belongs in the setup from the first day.
The mechanism has a name, retrieval-augmented generation, and this article on RAG goes into the parts. Here we stay with the practical side.
What belongs in it
Think of the first month of a new hire and write down what they would need. A starting inventory:
| Content | Examples | Why it matters |
|---|---|---|
| Offer | Products, services, packages, what is included | Most questions are "what do you do" and "what does it include" |
| Prices | Price lists, deposits, fees, payment methods | The costliest thing to get wrong |
| Policies | Cancellation, refunds, delivery, warranty, privacy | Customers quote these back to you |
| Process | How to book, how orders are fulfilled, what to bring | Lets the agent walk a customer through steps |
| Places and times | Addresses, branches, opening hours, holiday closures | High volume, and wrong answers are visible at once |
| Edge cases | "Can I change the date after paying?" "Do you ship to this country?" | The answers that only live in staff heads |
| Tone examples | Five replies your best person wrote | Sets the voice more cheaply than a long description |
Leave out material the agent should not say. Internal margins, staff gossip, drafts you have not decided on, and personal data about individual customers do not belong in a knowledge base that many conversations can draw on. Facts about one customer, such as an order or a booking, come from your systems at the moment they are needed, which is a different mechanism from the knowledge base and is covered below.
Knowledge versus live data
A knowledge base holds things that are true for everyone and change slowly. A system of record, such as a calendar, a CRM or an order database, holds things that are true for one customer and change by the hour. Mixing them is a common cause of wrong answers.
Put "we open at 9:00 and close at 18:00 on weekdays" in the knowledge base.
Do not put "Thursday 10:00 is free" there. Let the agent check the calendar when someone asks.
Do not paste a stock count into a document. Connect the store or sheet so the agent reads the number when asked.
If a fact has a time stamp that matters, it belongs in a connected system. Volty uses one Volty request each time it reads from or writes to another system, a meter described on the pricing page.
How to write for retrieval
One topic per section, with a heading that matches how customers ask. "Can I change my booking?" retrieves better than "Booking amendments, procedure 4.2".
Make each section stand alone. Repeat the product name instead of writing "it", and state the condition next to the rule.
Put the answer first, then the exceptions. "Refunds are available within 14 days of delivery. Exceptions: custom orders."
Use numbers and dates in the form customers use. If customers write "€" or "A$", so should the document.
Write questions and answers for the top 30 questions. A plain FAQ format is easy to retrieve and to review.
Keep tables for lists of prices or hours, one row per item.
Sources: where the content comes from
Most businesses already have the content in four places.
The website, including product and policy pages. This is the easiest source to keep current, because the page staff update for customers is also the page the agent reads.
Files: PDFs, brochures, price lists and slide decks.
Shared documents and spreadsheets in Google Docs and Sheets, which are handy for material that changes often.
The answers your team has written in past conversations, which you turn into short entries rather than uploading raw chat logs.
Volty reads your website, files, and Google Docs and Sheets. For an online store, connecting Shopify gives Volty the store's own product pages, indexed from its sitemap up to 5,000 pages and re-read every night, along with the policies and FAQs written in Shopify, and the answer carries the link to the product page. That connection reads the public storefront, so order lookups need a separate connection to your order system. See the Shopify integration for the details.
Keeping it accurate
Knowledge bases rot quietly. A price changes on the website, the old PDF stays in the folder, and the agent has two answers. Four habits prevent most of it.
Name an owner. One person is responsible for each document, even if many can edit it.
Change the source, then the copy. If a price lives in three places, list the three places next to the price.
Put an end date on promotions and seasonal rules, and add a recurring task to remove them.
Retire old files instead of leaving them. A superseded document left in the folder is a source of wrong answers.
A monthly review in thirty minutes
Open the conversations where the agent said it did not know or a person took over. Each is a missing or unclear entry.
Read ten replies at random against the documents. If the reply and the source disagree, fix the source.
Check that the facts that changed this month have changed everywhere.
Ask the team what customers asked this month that was new, and write the answer down.
The review should end with a short list of edits. The gaps are the training data you could not have predicted in advance.
Testing the knowledge base
Keep a fixed set of 20 questions that cover your top topics, two or three phrased badly, and a few the knowledge does not cover. Run it after every significant change and compare. Add any question that ever failed in production. Over a few months the list becomes your regression test for the agent's knowledge, and it takes five minutes to run. How to train an AI agent shows how this fits with instructions, tools and corrections.
Privacy and where the content goes
The documents you give an agent are sent, in pieces, to a language model whenever a customer asks something they relate to. So decide what is allowed to travel. For Volty, every AI model request is sent with zero data retention and data collection denied, so customer conversations are not used to train models, and from the Max plan you can choose your AI providers. Customer data is stored on AWS in Singapore. If your content includes personal data, read the guide to data privacy for WhatsApp Business first.
A school that kept its knowledge current
Happy Fish, a swim school with 21 outlets across Singapore and Malaysia, uses Volty to answer parents from the school's website, curriculum and schedules, and staff retrain the agent regularly. The IMDA story the case study cites puts the automated volume at more than 1,800 messages a day, which used to be answered by an overseas team, and service to customers moved from an average of two hours to instant. CEO Tan Jian Yong said: "Some customers even told us they didn't know they were chatting with a chatbot as the responses were so natural." More at Happy Fish's case study.

Doing this in Volty
In Volty you point the agent at your website, files and Google Docs and Sheets, and it answers from them in plain language, including questions phrased in ways no FAQ anticipated. The same knowledge serves WhatsApp, Instagram, Messenger, email and web chat, so one correction fixes every channel. Start from Volty's WhatsApp AI page; every plan begins with a 7-day free trial.
Frequently asked questions
What is an AI knowledge base?
It is the collection of your own content, such as website pages, policy documents, price lists and FAQs, that an AI agent reads to answer customers. The agent finds the relevant passages for each question and writes the reply from them, instead of relying on what the model learned in general training.
What should I put in an AI knowledge base?
Put in anything a new staff member would need in their first month: products and services, prices, policies, how-to steps, and the answers people give from memory. Leave out internal gossip, outdated material, and personal data about individual customers.
How often should the knowledge base be updated?
Change it whenever the business changes a price, policy or hour, and review the whole set monthly. If the agent reads your live website, an update there reaches it without a separate upload.
Is an AI knowledge base the same as a help center?
A help center is written for people to browse. An AI knowledge base is read by a model that picks passages, so short, self-contained sections with clear headings work best. Your help center can be part of it.
Does the AI learn from my customers' conversations?
In Volty, no AI model is trained on your customers' conversations: every model request is sent with zero data retention and data collection denied. What Volty does keep is your staff's corrections, which become Learned rules the agent follows and you can revert.










