The short answer
Models make things up because they are built to complete text, not to retrieve facts. When the fact is not in front of them, they complete anyway, fluently, and fluency is what makes the error dangerous. A knowledge base fixes it by changing the job: find the relevant passages in your documents first, hand them to the model, and instruct it to answer only from what it was handed and to show its source. When nothing relevant is found, it says so. That pattern is usually called retrieval-augmented generation, and it is the difference between an assistant you can put in front of staff and one you cannot.
Why it happens
Think of the model as having read most of the public internet and remembered the shape of it, not the content. Ask it for the opening hours of a well-known institution and it will produce something shaped like opening hours. Ask it for your company’s leave policy and it will produce something shaped like a leave policy, drawn from thousands of other companies’ policies. It has no way of knowing that yours is different, and no signal that tells it to stop.
This is not a bug that the next version removes. It is what the technology is. Newer models make things up less often and are better at saying they are unsure, but a model answering from memory about a business it has never seen will always be guessing. The design has to give it something to read.
Why a better prompt does not fix it
“Only answer if you are certain” sounds like a solution. It is not, because the model has no reliable sense of certainty; it has a sense of plausibility, and a plausible wrong answer feels certain from the inside. Prompting reduces the worst behaviour and is worth doing. It cannot supply the fact that is missing, and it cannot make the model check its answer against a document it was never given.
The same applies to “we trained it on our data”. Fine-tuning changes the model’s style and habits; it does not reliably install facts, and the facts change next week when the policy is updated. Retrieval is the right tool because it reads the current document at the moment of the question.
What a knowledge base actually does
- Your documents (policies, procedures, product information, contracts, past advice) are split into passages and indexed by meaning, so a question about “annual leave carry-over” finds the paragraph that says “untaken entitlement”.
- When a question arrives, the most relevant passages are retrieved, and only the ones the asking person is allowed to see.
- The model is handed those passages with the question and instructed to answer from them, quote or link the source, and say when the passages do not contain the answer.
- The answer, the sources and the question are logged, so a wrong answer can be traced to the passage that caused it, and the document fixed.
The two design decisions that decide whether it works
Permissions. A knowledge base that indexes everything and shows everything is a leak with a search box. The retrieval step must respect who is asking: the staff handbook for everyone, the salary bands for the people who may see salary bands. This is designed in from the start, using the permissions your systems already have, and it is the reason a knowledge assistant is built rather than switched on.
“I don’t know”. The assistant must be allowed, and instructed, to say the documents do not answer the question. An assistant that always answers is back to guessing. The measure of a good knowledge base is not how often it answers; it is how often it is right when it does, and how honestly it declines when it cannot be.
What it is worth
A knowledge base pays in the time people stop spending asking each other, in new staff being right on day one, in customer questions answered from the actual policy, and in a category of error that stops happening: the confident wrong answer that reached a customer. It also stops the quiet leak of staff pasting company documents into public tools to get an answer, because there is now a better place to ask.
It is usually the second AI project in a business, after the automation that removes the worst administration, and it is the foundation the agents are built on, because an agent that answers customers needs the same source of truth.
How to test a knowledge assistant before you trust it
Before an assistant goes in front of staff or customers, write thirty questions. Twenty should be answerable from your documents, with the answer written down by a person who knows. Ten should not be answerable from anything you hold, including a few that sound like they should be. Run all thirty. Score the twenty for accuracy and for whether the source shown is the right one. Score the ten for whether the assistant declined, and how. An assistant that answers all thirty is guessing on ten of them.
Then test permissions with two accounts: one that may see a sensitive document and one that may not. Ask both the same question about it. If the second account gets an answer, or even a hint, the retrieval step is not respecting permissions and the assistant is not ready, however good the twenty answers were. Repeat the thirty questions each quarter, because documents change and the assistant should change with them.
Where to go from here
The Knowledge AI page on this site describes how we build one: the documents, the permissions, the retrieval, the citations, the honest “I don’t know”, and the record. The AI Opportunity Score below will tell you in five minutes whether your information is digital and findable enough to start, or whether that is the first job.
Published 12 September 2026 · Be AI