The quality of an AI assistant depends as much on what it is given to read as on the model that writes. A knowledge base brings these documents together; the system then splits them into passages and indexes them so it can find the right passage when a question arrives, using RAG.
What to include and what to leave out
A useful base starts with documents that already answer customers’ questions: the website FAQ, service pages, terms and conditions, practical guides and internal procedures that can be shared with customers. A plumber would include their service area and call-out charges; an accountancy firm, the documents required for each deadline; an estate agency, its property listings.
Leave out anything a visitor must not read: margins, internal notes, customer data and unapproved legal documents. In principle, anything included can be cited to a visitor.
The three qualities of a good base
- Its documents are selected. Fewer, accurate documents are better than an entire folder uploaded without sorting: two conflicting versions of the same rule produce conflicting answers
- It is kept up to date. A price, opening time or procedure that changes gets a review date; when it is due, someone checks it
- It grows from real questions. An unanswered question reveals what is missing far better than a list imagined in advance
How the assistant uses it
For each question, the system finds the passages closest in meaning (not just those with matching words), gives them to the model and asks it to answer from them. Each passage used can appear alongside the answer through source citation. This makes the assistant’s answers verifiable and also limits invented answers.
Knowledge base, FAQ, help centre: what is the difference?
A FAQ is a list of questions and answers published on a website; a help centre is a collection of articles for customers. Both can be included in the knowledge base, which is broader: everything the assistant is allowed to cite, published or unpublished, arranged so it can find the right passage.
In KernChat, knowledge is arranged in spaces that can be reused across assistants, prepared with the organization during setup (see the process).