Skip to main content
Search

Check your knowledge base with your AI

How your own AI checks the whole knowledge base the way customers ask, finds what the bot cannot answer, contradictions and duplicates, and fixes them after you approve.

Your own AI, connected to the knowledge base, can check all of it in one go: it asks the questions your customers ask, sees what the bot finds, looks for contradictions, duplicates and structure problems, and proposes fixes. Nothing is written until you say yes.

What you need

  • Your AI connected to the knowledge base: Claude, ChatGPT, Cursor or another app that supports MCP. See Connect your AI (MCP).
  • A panel account with the Manage AI features permission to allow the connection.

Start the check

  • In Claude: the knowledge server offers a ready request called Check the knowledge base (audit_knowledge_base). Pick it from the list of prompts or commands of the connected knowledge server and send it.
  • In ChatGPT, Cursor and other apps: open Admin > AI > Connect your AI, copy the request from the block 4. Check the knowledge base and paste it into the chat with your AI. The panel shows it in your panel's language.

The AI then works on its own and reports back. A full check of a knowledge base with a few dozen articles takes a few minutes.

What the AI checks

  • Questions customers ask. It reads your articles and writes 30 to 50 questions the way customers word them: how to contact you, prices, delivery, returns, opening hours, common problems and whatever your articles cover. It runs them through the same search the bot uses and, for doubtful ones, asks your assistant directly, the same check as the Test tab.
  • Contradictions. It gets a list of every email address, phone number and amount of money in the knowledge base with the text around it, and points out the same thing with different values in different articles: the bot would repeat either one.
  • Duplicates. Paragraphs repeated in several articles take places among the pieces the bot reads.
  • Structure. Long articles without headings, very long sections, very short articles and articles with the same title.
  • The basics. Whether there is an article about your company and how to reach you, and whether your email addresses appear only on legal pages such as the privacy policy.
  • Imports. Articles that failed to index, and imported pages not refreshed for 30 days.

You get a table: question, what the bot found, what is wrong, followed by the contradictions and the other findings.

What the AI changes and what it does not

  • Every change is shown first as a dry run, and the AI writes only after you confirm. Every write can be undone, see Connect your AI (MCP).
  • It may split an article into sections with headings worded as customer questions, merge duplicates into one place, fix outdated and contradicting facts, and write an article about your company if there is none.
  • It does not add question-and-answer copies of a fact that is already in the knowledge base just to make the bot find it. If the fact is there and the bot misses it, that is a search problem on our side: the AI gives you the question and the article, and you send them to support@epavdesk.com.
  • It does not edit articles that come from imported web pages or from a Confluence wiki: the next import or sync would overwrite them. It tells you what to change on the website or in Confluence instead.

What the import already does for you

  • Questions of collapsible FAQ blocks are kept. On many websites a question in a FAQ is a button you click to open the answer. Importing such a page keeps every question as a section heading, so the bot finds the answer by the question. Pages imported before 7 October 2026 have only the answers: import them again.
  • One contacts article per website. Email addresses and phone numbers from imported pages, the footer included, are collected into an article named like Contacts of Example Shop, one per website and language, with your company name as the site shows it in page titles (otherwise the domain) and the labels the site gives the addresses, such as Legal enquiries. General addresses come first and legal or data protection addresses last. If only one ordinary address has no label, it is marked as the general address for any question. The article updates itself when the pages are imported again; do not edit it by hand.

How often to run it

After a large import, after you change your website or prices, and about once a month. A knowledge base that changes often benefits from a check after every change.

Was this article helpful?