What is a knowledge base and how a model answers from it
Why this matters
- To understand where the model gets its answers about tariffs, products and internal company rules
- To see how much of the budget goes into preparing documents rather than into the model itself
- To ask a contractor what backs up an answer and who keeps the source current
What is a knowledge base in plain words
Picture the information desk at a railway station. The clerk has a binder: timetables, instructions, service phone numbers. A passenger asks, the clerk flips through and answers in half a minute. If the binder has not been updated in a month, the answer comes out just as confidently — and turns out to be wrong.
A knowledge base is that binder, gathered in one place and put in order. Inside are policies, instructions, product descriptions, answers to frequent questions, contract templates. Order matters more than volume here: every document has a date, an owner and access rights.
The model plays the clerk. It holds no information about the company: prices, stock and policies all appeared after its training. The knowledge base slips the right pages under its hand at the moment the question is asked.
How it works
The figures below are planning benchmarks for a company of this profile, not a report on a specific rollout. Each comes with the method used to measure it, so it can be repeated on a company's own data.
The pairing of document search with a model's answer is called RAG. Search follows meaning: the question "the March invoice never arrived" finds the section on resending documents, even though none of those words appear in it. That is the work of semantic search and a vector database.
Access rights move into the base along with the documents. Otherwise a manager receives an excerpt from a document they are not meant to see. This is checked before launch: ten questions asked in the role of an employee without elevated rights.
An example from practice
A service company, 900 tickets a month, policies scattered across email and three folders. The time to find an answer was measured with a stopwatch on 30 typical questions over two weeks: 4–12 minutes, with a median of about six. For a third of the answers, operators double-checked with colleagues.
The first setup was assembled from 60 documents covering the ten most frequent topics. The timelines quoted by vendors of such systems run 4–8 weeks to a working version in a single area. The same 30 questions were run again: the count covered the share of correct answers and the share of answers where the citation led to the current edition.
The second figure proved more important than the first. An answer without a verifiable citation gets re-checked by hand anyway, and the saving disappears.
What the business gets out of it
- The answer stops depending on who is on shift. A new hire answers from the same documents as someone in their third year.
- Knowledge stays with the company. When a department head leaves, the rules do not leave with them: they are written down and tied to the process.
- Model costs become a countable figure. What goes into the model's window is not the whole archive but a few retrieved pieces — the volume can be estimated with the load calculator.
- Checking answers becomes cheap. A source citation in the answer takes an operator seconds instead of a walk over to a colleague.
When this can be skipped
Volume is worth checking first. If there are few questions and all of them are covered by a single page of rules, it is cheaper to put that page in order and place it directly into the prompt.
A separate base is also unnecessary where documents already live in one system with decent search and nobody complains about finding things. The work then comes down to tagging rights and dates. A full setup is assembled when there are several sources and they contradict one another.
What to check
First: every section needs a named owner and a review deadline. Tariffs and limits are reviewed more often, instructions less often. Without this the base goes stale within a quarter, and answer quality drops for no visible reason.
Second: superseded editions are physically taken out of circulation. As long as the current and the cancelled policy sit side by side, search honestly finds both, and the model answers from whichever turned up.
Third: ask the contractor to show a measurement on 30 questions with known answers — before the base was connected and after. Two shares are worth watching: correct answers, and answers citing a current document. That measurement can be repeated on a company's own data, and again a month after launch.
Frequently asked questions
What is a knowledge base in plain words?
It is a company's documents gathered in one place: policies, instructions, product descriptions, answers to frequent questions. Every document has a date, an owner and access rights. The model pulls facts from the base at the moment it answers, because it holds no knowledge of the company's prices and rules on its own.
How does a knowledge base differ from a folder of files
A folder stores; a knowledge base answers. In a knowledge base each document is cut into pieces of a workable size, every piece carries a source and a date, and outdated versions are taken out of circulation. Search follows the meaning of the question rather than an exact word match. Without that groundwork, a model finds the superseded edition of a policy just as readily as the current one.
How many documents are needed to start
One area is enough: 30–80 documents that cover the most frequent questions is the usual range. The check is a month of support tickets — the top ten topics account for most of the volume. The base is widened only after the share of correct answers in that area has been measured and found acceptable.
Who updates the knowledge base after launch
The process owner, not the contractor. Each section gets a named responsible person and a review deadline: tariffs and limits more often, instructions less often. Corrections made by operators return to the base through a separate procedure. Without an assigned owner the documents go stale within a quarter, and answer quality drops for no visible reason.
How can it be measured that the knowledge base works
Take 30 typical questions with known correct answers and run them before and after. Count the share of correct answers, the share of answers citing a current document, and the time to answer. The same set is run once a month: the gap shows which sections are due for an update.
Let’s discuss your project?
Tell us about your process — we’ll suggest where AI pays off fastest.
Related articles
What is RAG: how to make a model answer from your documents
RAG is a way of working with AI where the system first finds the right passage in a company's own documents and then builds an answer from it.
What Semantic Search Is and What It Does for a Business
Semantic search finds documents by the meaning of a query: the system returns the right document even when it contains none of the words from the question.