What is a knowledge base, and how do you build one?
What a knowledge base is and holds, how it differs from a database, a document management system or a wiki, and what an AI assistant's knowledge base is.
A knowledge base is a library of organised, dated and maintained articles where your teams and your customers find an answer without having to ask anyone for it. The same phrase covers a software company's public help centre and an IT department's internal reference.
The definition fits in one sentence, but it says nothing about what separates a useful knowledge base from a shared drive with a search box. That difference does not show up in a feature list. It shows up article by article: does the one you just opened have an owner, a date when someone last checked it, and a rule that says who may read it?
What goes in a knowledge base
A knowledge base gathers what your organisation knows in written, searchable form. The usual content is:
- procedures and step-by-step guides, which our guide on SOPs and work instructions explains;
- answers to frequent questions;
- troubleshooting articles: a problem, its cause and the fix;
- product and service sheets;
- incident write-ups and lessons learned from projects;
- decision trees, when the answer depends on the reader's situation.
Each article also carries information about itself: who is responsible for it, when someone last checked it, and who may read it. That information decides whether a reader, or an AI agent, can rely on the article. The test further down shows how to check it.
A knowledge base serves two audiences, and most organisations need both.
The external knowledge base is public. It is the help centre your customer opens at 22:00 to reset a password. It reduces the number of requests, and its effect can be measured: views per article, share of requests solved without contacting anyone, tickets raised on a subject already documented.
The internal knowledge base is for your teams only. It carries what cannot be published: what to do with an unhappy customer, the exact configuration of a laptop, the floor price in a negotiation, what a project taught you when it went wrong. It is the one that costs the most when it is missing, because the answer then sits in three people's heads.
Knowledge base or database: what is the difference?
The two are easy to confuse, because both are organised stores of information that a computer can search. A database holds data, such as a customer number, an order date or a stock level. Britannica describes it as a collection of data organised for rapid search and retrieval by a computer, stored as records made of fields. Programs read and update it, and people usually see its content through an application.
A knowledge base holds explanations: what to do, in which case, and why. People write it for other people to read, and increasingly for AI agents too.
| Database | Knowledge base | |
|---|---|---|
| What it holds | data: records with fixed fields | answers, procedures and explanations |
| Who writes it | applications, and the people who enter data | the people who know the subject |
| Who reads it | programs, and people through an application | people directly, and AI agents |
| Example | the customer records in a customer relationship management (CRM) tool | the article that explains how to process a refund |
| What makes it right | each value is accurate | each answer still applies, and someone stands behind it |
The line between the two is not always sharp. The knowledge base of an expert system, described in the next section, is closer to a database than to a help centre, because a program reads it. And an article in a knowledge base can show live data from a database, such as the status of an order. The difference is in what each one is for: a database records facts one at a time, and a knowledge base explains what to do with them.
Why the phrase means two different things
The phrase "knowledge base" covers two ideas with separate origins, and people have confused them for a long time. Dictionaries keep the definition broad. For the Cambridge Dictionary, a knowledge base is a collection of information about a particular subject. For Collins, it is a set of facts or experiences used as a resource to perform tasks or solve problems.
The first precise meaning comes from artificial intelligence. The first expert system was built at Stanford University in 1965. Expert systems rely on two parts, as Britannica describes: a knowledge base, which holds the facts about a field, and an inference engine, which evaluates those facts to give an answer. The facts are usually written as "if-then" rules, collected from human experts. In that sense a knowledge base is what a program reads, not what a person reads.
The second comes from customer support. In 1992, a group of people from customer support began meeting, first to discuss software, then to work on how support teams operate. They formed a not-for-profit alliance, the Consortium for Service Innovation. It still maintains a method for capturing what support teams learn when they solve problems and answer customers' questions. This is the meaning most people have in mind today: the help centre, the FAQ, the troubleshooting article. It is also the one software vendors use.
The two meanings have come together in recent years, which is why the phrase matters again. As soon as you connect an AI agent to your content, your help centre becomes what an expert system's knowledge base always was: the material the program reads in order to answer. An out-of-date article no longer misleads just one reader: the AI agent can repeat its wrong answer every time the question comes up. The section on AI assistants below explains how this works.
Knowledge base, document management, wiki: what each is for
Most organisations have more than one of these tools. They usually have a document management system, often a wiki or a collaborative space, and sometimes a knowledge base. Mixing them up is expensive, because each one is expected to do something different.
| What people look for there | What an AI agent finds there | |
|---|---|---|
| Document management | proof: the authentic document, its version, its archive | whole files, from which it has to extract an answer without knowing whether it still applies |
| Wiki | work in progress: a team's notes, drafts and discussions | pages, without knowing which are still right |
| Knowledge base | the answer: what to do, checked and dated | an answer, with the person responsible and the date it was last checked |
A document management system keeps track of what was signed or decided, and it does that well. A wiki makes writing easy, which is its strength in a team of ten. In larger teams it degrades quickly, because nothing requires a page to be reviewed: we covered that limit in wiki, intranet and knowledge management. A knowledge base answers a different question: what should be done in this case, and who stands behind the answer?
These tools complement each other, and one solution often covers several of these needs: a document management system can publish procedures, and a knowledge base can store files and hold drafts. So the choice comes down to the main use. A tool built to keep proof is poor at reviewing answers, and a tool built for answers is not meant to hold the legal archive. Name the use that comes first, then check that the secondary uses fit in the same tool without getting in the way of its main job.
Among our customers, Elium, a European knowledge management platform, often replaces the wiki: people can still write together, and each article now has an owner, a review date and read rights. The document management system keeps its place: the signed contract stays there, and the knowledge base says what to tell a customer asking for a refund, linking to the contract if needed.
The three-question test
Open your knowledge base, pick an article at random, and ask these three questions about it.
Who owns it? A person, not a team. A page nobody is responsible for will never be corrected, because everyone assumes somebody else is handling it.
When was it last checked? The date it was written tells you nothing. What counts is the day someone reread it and confirmed it was still right. With no review date, a knowledge base goes out of date without warning, and the first person to notice is a customer.
Who may read it? A knowledge base without read rights ends up holding only what can be shown to everyone, which is the least useful material there is. Price lists, incidents and sales pitches stay in people's inboxes instead.
If all three answers come easily for an article picked at random, your knowledge base works. If they take an investigation, you have a shared drive with a search box.
How to build a knowledge base
The tool matters less than deciding what to write and who writes it. Five decisions make the difference.
Start from questions, not from documents. Do not copy the folder tree of the shared drive. Take the last hundred requests your support desk, customer service or IT team received, count the repeats, and write the twenty answers that come up most. Twenty accurate articles are worth more than two thousand whose state nobody knows.
Match the format to the question. An article works when the answer is the same for everyone. As soon as it depends on the reader's situation, a decision tree works better: a short series of questions takes the user or the support team to the right conclusion without making them read all the documentation. A good tree copies nothing: it links to articles that already exist. Otherwise you end up maintaining the same procedure in two places.
Name an owner before the article is written. This is the decision that keeps articles from going out of date, and the least popular one, because it turns a one-off contribution into an ongoing commitment. Make it part of that person's workload.
Set a review date on the day you publish. Six months for a stable procedure, six weeks for a price or a configuration. When the date arrives, the owner is notified, and the article stays flagged for checking until they have reread it.
Split by team. An organisation of a thousand people does not write in one space. The platform stays shared, and the content lives in separate spaces, one per team or business area, each with its own owners and rights. L'Oréal Opérations runs 113 of them, grouped into about fifteen families by business area, for 22,000 people.
What an AI assistant's knowledge base is
Ask an AI assistant such as ChatGPT a question, and its answer can come from two places.
The first is what the model learned during training. Researchers call it the model's parametric memory: the knowledge encoded during training, which stops at the model's knowledge cutoff. Facts that changed after that date are not in it.
The second is documents the system retrieves at the moment you ask: web pages, or files your organisation connects to it. The system passes the passages it finds to the model, which writes its answer from them. The technique is called retrieval-augmented generation, or RAG: it adds external knowledge, retrieved when the question is asked, to what the model already knows. When an AI product talks about its knowledge base, it means this second set of documents.
So an AI assistant has no knowledge base of yours until you connect one. Once you do, your documents become the material it answers from, and its answers are only as current and as consistent as those documents. Two studies published in July 2026 point the same way. In a benchmark built from real questions and web results, the leading models tested, such as GPT-4.1, showed large gaps when they had to organise and explain sources that contradict each other. In another study, larger models often noticed that a fact had changed since their training, but this rarely made them change their final answer.
What AI agents change for your knowledge base
The question used to be whether a person could find the answer in a knowledge base. Now it is whether an AI agent can give that answer on the person's behalf without getting it wrong.
That changes the work. With keyword search, a person read the results and could correct for weak content: a reader who found two contradictory procedures took the more recent one, or asked a colleague. An AI agent has no such reflex. It receives both procedures, and as the studies above show, models handle contradictory sources poorly. The answer may follow the outdated version, with nothing to tell the reader that a newer one exists.
So four things matter more than they used to.
- A duplicate becomes an error. Two versions of one procedure give the AI agent two answers to choose from, and it may choose the outdated one.
- Read rights have to apply to the answer too. If the AI agent answers from everything it has read, it will one day quote a board-level document to an intern. Rights have to be checked when the agent answers, and not only when a page is displayed.
- The content has to stay text. An AI agent reads text, not a layout. Articles and decision trees written in Markdown, a simple plain-text format, can be imported and exported without loss, and the AI agent receives the owner, the review date and the approval status along with them, which is what it needs to judge whether it can use the answer.
- The source has to be cited. An answer with no link back to the article it came from cannot be checked, so it will not be used for a decision anyone is accountable for.
This is what we call the foundation: the layer of governed content the AI agents answer from. We set it out in why enterprises fail at the knowledge foundation, and list the signs that it is drifting out of date in how to keep an AI knowledge base healthy.
Common questions
A library of organised, dated and maintained articles where your teams and your customers find an answer without having to ask anyone for it. It holds procedures, step-by-step guides, answers to frequent questions and product sheets. What separates it from a shared drive is that every article has a named owner, a review date and read rights.
A database holds data: records with fixed fields, such as customer numbers or stock levels, read and updated by programs. A knowledge base holds explanations: what to do in a given case, written for people and AI agents to read, with someone who stands behind each answer. The knowledge base of an expert system sits in between: a program reads it, as it reads a database.
Two words: the Cambridge and Collins dictionaries list knowledge base as a two-word noun. The one-word form, knowledgebase, appears in some product names. Knowledge-based, with a hyphen, is a different word: an adjective for something based on the use of ideas and information, as in a knowledge-based economy.
A public knowledge base is often called a help centre, a self-service portal or simply the FAQ. An internal one is also called a knowledge repository or a reference library. The word wiki is sometimes used as a synonym, but it works differently, as the comparison above shows.
Its general knowledge comes from training, which stops at a cutoff date. When it searches the web, or reads documents connected to it, those documents act as its knowledge base for that answer: the technique is called retrieval-augmented generation (RAG). It has no knowledge base of your organisation until you connect one, and its answers are then only as current as your documents.
Start from the questions you receive, not from the documents you already have. Take the last hundred requests to your support desk, count the ones that repeat, and write the twenty most frequent answers. Give every article an owner before it is written, set its review date on the day you publish, and split content by team rather than putting everything in one place. The tool comes after those four decisions.
It depends on who reads it: for a public FAQ attached to tickets, the module in your ticketing tool is usually enough. For a product documentation portal, dedicated tools exist. For knowledge that has to serve several departments and feed AI agents, you are choosing a platform rather than a module. We compared the market in two guides: knowledge base software and customer service knowledge base.
A document management system keeps the authentic document and a knowledge base gives the answer to apply, as our guide to document management vs knowledge management explains.
The one reserved for your teams and never published. It holds what you do not show outside the organisation: procedures, configurations, sales pitches, lessons learned, floor prices. It is the harder one to maintain, because nobody outside checks it, and the one that costs most when it is missing, since the knowledge then stays in a few heads.
There is no single list of four: the revised Bloom's taxonomy lists factual, conceptual, procedural and metacognitive knowledge, and our guide to the types of knowledge explains each framework.
Where to start
Open the knowledge base you have now, pick one article at random, and look for its owner and its review date. How long that takes tells you where you stand. If you want to see a tool that shows both on every article, book a demo.
Sources
- Cambridge Dictionary (Cambridge Business English Dictionary), knowledge base and knowledge-based.
- Collins English Dictionary, knowledge base.
- Britannica, database and expert system.
- Consortium for Service Innovation, Innovating together for 30 years, 2022, and KCS v6 Practices Guide: Why KCS?.
- Yuan, Kong, Li, Li and Liu, Benchmarking LLM's Capability in Reasoning over Conflicting Web References, ACL 2026, July 2026.
- Wallat, Nejdl and Sikdar, When Facts Change: Temporal Knowledge Conflict Resolution in LLMs, Findings of ACL 2026, July 2026.
- Xu, Zhuang, Yu, Wang, Shi and Yang, RAG in the Wild: On the (In)effectiveness of LLMs with Mixture-of-Knowledge Retrieval Augmentation, Findings of ACL 2026, July 2026.