Train the AI agent on your knowledge

An AI agent is only as good as what it has been given. The knowledge sources are that input — and the design assumption throughout is that grounded answers beat confident ones.

There are four kinds of source you can add, and one that exists in the schema but cannot be created yet.

Who this is for

Admins setting up the AI agent, and whoever owns support content.

Before you start

• Permission to manage AI agents and their knowledge.
• Something to feed it — a help center, documentation files, or a website worth crawling.
• A realistic view of what your content actually covers.

A sync never leaves the agent worse off
Each sync writes a new generation of the knowledge and then swaps to it in one step. That means a sync in progress is invisible to anyone asking a question, and a sync that fails leaves the previous good version answering rather than emptying the agent. You are not risking your existing knowledge by re-syncing.

Set up knowledge

1. Connect your help center
   If you already have a help center, this is the highest-value source — it is written for customers and already reviewed. Publishing or editing an article re-syncs the connected source automatically.
2. Upload files
   Documentation, policies, product notes. Anything your team already relies on is a candidate.
3. Add question-and-answer pairs
   For the things that are not written down anywhere. This is the fastest way to fix a specific wrong answer — write the correct one down explicitly.
4. Crawl your website
   A website source pulls in public pages. Useful for product and pricing information that lives on your site rather than in your docs.
5. Watch the sync status
   A source is pending, processing, ready or failed. Only ready means it is being used, so check rather than assume after adding something.
6. Test with real questions
   Ask the questions your customers actually ask, including the awkward ones. Gaps show up immediately and each one is a piece of content to write.

What you get

• An agent answering from your material rather than from generic knowledge.
• Help center edits flowing through to the agent automatically.
• A direct way to correct a wrong answer, by writing the right one as a Q&A pair.
• Safe re-syncing, because a failure leaves the last good version answering.

Frequently asked questions

Which source should I add first?

Your help center, if you have one. It is already written for customers, already reviewed, and it stays in sync automatically as you edit articles.

How quickly does an article edit reach the agent?

Editing or publishing an article triggers a re-sync of the connected source, so it flows through without you doing anything.

The agent gave a wrong answer. How do I fix it?

Add a question-and-answer pair with the correct answer. That is the most direct correction available, and it works for things that are not written down anywhere else.

Can it learn from my CRM records?

Not today. A CRM source kind exists in the schema but has no implementation behind it, and creating one is refused rather than silently accepted. The sources you can add are files, Q&A pairs, a help center and a website crawl.