Knowledge and the help centre
Two different surfaces: what your AI agents read, and what your customers read.
Two things in the product hold knowledge, and they are not the same thing.
| Surface | Who reads it | Where |
|---|---|---|
| Knowledge bases | Your AI agents | Knowledge in the sidebar |
| The help centre | Your customers, on your own site | Help centre in the sidebar |
Start by deciding which one you want. Feeding an agent is not publishing a page, and publishing a page does not feed an agent.
Knowledge bases
A knowledge base is a named collection of documents an AI agent can search. Create one from Knowledge, then add sources to it.
Add data source offers three groups:
- URLs: a single web page or a list of them, a whole sitemap to crawl, or a YouTube video, which is transcribed.
- Files: documents (pdf, txt, docx, md), tables (csv, xlsx), or text pasted straight in.
- Connect: Zendesk help centre articles, Notion pages and databases, Confluence spaces and pages, and Shopify products.
Each document moves through a pipeline and its status is on the row: queued, parsing, chunking, embedding, quality, storing, and finally ready, or failed. The list shows the chunk count once a document is ready.
An agent subscribes to one or more knowledge bases, and the knowledge page shows which agents are subscribed to each.
The help centre
The help centre is a public site of articles on your own domain. There are two distinct surfaces and it is worth keeping them apart.
The staff side, at Help centre in the sidebar, is where your team works. It needs a signed-in session.
The public site is what a visitor sees. It resolves which help centre to serve from the host in the request, so each workspace's content is served on that workspace's own domain.
Importing
Content is imported from an existing help centre. The import page reports what arrived: how many articles were found at the source, how many landed, anything that was made more private here than it was at the source, and anything being kept that still needs a decision. Where it can, it shows the source system's own words so you can check it against what you know.
The article list
Every imported article, browsable and filterable, 50 rows a page.
One locale at a time. The locale is the list's identity rather than a filter over a mixed list, because that is how the underlying data is shaped. Mixing locales would make every count on the page ambiguous. The locale control is the first thing in the filter bar.
The list also reports how many categories hold nothing in the current locale. That is the visible symptom of an untranslated category, which is otherwise invisible until a customer finds an empty page.
Four different empty states exist and they mean four different things: no help centre at all, nothing imported yet, nothing in this locale, and nothing matching these filters.
Editing an article
Exactly six fields are editable: title, slug, visibility, editorial state, category and body. Anything else is refused rather than quietly ignored, so an edit you thought you made cannot go missing.
Send only what you are changing. An omitted field is left alone.
Two fields decide whether a customer sees an article, and both have to be right.
- Editorial state is
draft,publishedorarchived. This is what publishes an article. - Visibility is
public,signed_inorinternal. This is what decides who may read it.
An article can be published and still invisible because it is internal, and it can be public and still invisible because it is a draft.
Editing marks an article as locally edited
Once you have edited an article, a later update from the original source is held as a conflict rather than overwriting your text. You then choose: keep your version, or adopt the incoming one and discard yours.
The category tree
The staff view shows the whole tree, including containers with nothing visible in them. The public site hides those. Staff need to see an empty container in order to fix it.
Branding
The branding page sets the look of the public site.
What a visitor can and cannot reach
The public surface is read only. There is no vote, no comment and no subscription.
Everything missing answers the same way. An unknown host, a disabled help centre, a draft article, an internal article, a signed-in-only article, an article whose section is internal, and an article that does not exist all produce identical responses. Nothing there confirms that a hidden thing exists.
A staff session presented to the public surface changes nothing about what it serves.
Searching knowledge from a client
search_knowledge_base over MCP reads the knowledge base and only reads it. See
the MCP tools reference.