Build the knowledge base the AI answers from.

#The Knowledge page

The page has four parts: a banner when something needs attention, the pages waiting for a decision, the list of Sources (where articles come from), and the table of Articles, which shows each article’s address and can be sorted, filtered and cleaned up in bulk (see organise the articles). A workspace can have at most ten sources. Removing a source deletes its articles, and the AI stops answering from them at once; the site or help centre itself is not touched.

#Which source should I use

#Crawl a help site

Revvly reads your site politely: https only, it follows the site’s robots.txt and crawl delay, skips pages marked noindex, and stops at a page limit (up to 2,000 pages; the source shows “Stopped at the page limit” when it did). It reads public pages only and needs no login or token. The schedule is chosen when you add the site; the page does not yet let you change it afterward. The paths to leave out can be changed at any time with Pages to leave out on the source (see keep a section out).

#Add a document

Give it a title, paste the text (Markdown or plain text, up to 200,000 characters), and choose Who can read it. Leave that empty to show it to everyone. The roles you choose are final from the start. You paste text; uploading a file such as a PDF or a Word document is not offered.

#Import an Intercom help centre

Choose Intercom, give the source a name, choose where your Intercom workspace is hosted (United States, Europe or Australia) and paste an Intercom access token that can read articles. Revvly checks the token before saving it, then reads your published articles. Drafts count as not there.

The Import menu has a second choice, Past conversations, which brings in your Intercom conversations as tickets with their history rather than articles. It opens the same conversation import described in Import your Intercom conversations.

Intercom has no idea of who an article is for, so imported articles go through the same role matching as a crawled page, and an article with no public address is held for your review in case it is internal.

#Import a Zendesk help centre

Zendesk says who may read each article. An article shown to everyone is imported for everyone. One limited to a user segment waits in the review queue for you to choose roles, and is not used until you do. If someone restricts an article in Zendesk later, the next read pulls it out of use until you decide. The same Zendesk help centre and language can be connected once per workspace.

#Push through the API

An API source is filled by your own server, so there is nothing to add on this page. It appears in the Sources list as an API source once your server pushes its first article, with no Read now button and no schedule, because Revvly only changes it when your server calls the API. The full steps, with the key you need, are in Push your help articles to Revvly.

#Who may read an article

Each article has an audience: everyone, or one or more roles, and optionally some plans. The roles are the ones you define with the Roles button on the Knowledge page. Your system tells Revvly each customer’s role value, attributes and plan when they sign in to the chat (see Sign users in to the support chat), and each role’s rules decide whether the customer has it. A customer can have several roles and sees articles open to any of them, plus everything open to everyone. A customer who matches no role, or a visitor with no role, sees only what is open to everyone.

#Article and source status

A source is Ready, Reading (the page refreshes its numbers every few seconds meanwhile), Failed or Off. The Articles column reads “12 of 14”: searchable out of all, with the number needing a decision and the number that could not be read underneath. The Used column counts how often the AI retrieved the article.

#Organise the articles

When a source brings in hundreds of pages, the Articles table is where you sort out what the AI should and should not answer from. The first column is the article’s address, its path on your site, such as /blog/some-post. A pasted document has no address and says Document. Click a column heading to sort by Address, Title, Source, Status or Updated; click it again to reverse the order. The table shows 50 articles at a time with Previous and Next, and says how many articles match in all.

#Change many articles at once

Tick rows, or tick the box in the header to take the whole page. When the page is ticked and more articles match your filters, a line offers Select all 212 matching, which takes every article the current filters match, not just the page. A bar then shows how many are selected, with three actions:

Every action sends the count you were shown. If the list changed in between (a read added pages, or someone else deleted some) and the real number is different, Revvly refuses and changes nothing, and tells you how many match now; look at the table again and repeat. One action takes at most 500 ticked rows, or 5,000 articles by filter; narrow the filters and repeat for more.

#Set who reads a section by its address

When your help site keeps each audience in its own folder, say so once instead of choosing roles article by article. Each rule has one or more conditions on the address path (starts with, ends with, contains, or the opposite of each; all conditions must hold) and the roles that may read what it matches, or Everyone. Rules run in order and the first rule that matches decides, so put the narrow ones on top: for example /docs/user/applicant/ for Applicants above /docs/ for Admins. Each rule shows how many pages it matches as you type.

#Keep a section out for good

Imported a whole site and got the blog and comparison pages too? Add their paths as pages to leave out on the source, one per line, such as /blog and /compare. A rule matches every address that starts with it (an asterisk matches anything), so /blog also matches /blogger. Reads skip those pages, and the next read that finishes completely removes the ones already stored, so you do not have to delete them one by one. The rules are listed under the source’s name. For Intercom and Zendesk the rule is matched against each article’s public address, for example /hc/en-us/articles/123.

#Read now, or on a schedule

Read now on a source starts a read at once. It is offered for crawled sites and the two help centres, not for documents or API sources. A scheduled source (Every day, Every week) is read again when its last read is older than that; Revvly checks every few minutes. Only when I ask sources are read only when you press the button.

Before a read you start (Read now, or connecting Intercom or Zendesk), Revvly asks whether to Update who can read from the role hints. It is ticked by default. Ticked, articles whose readers Revvly guessed are guessed again from your roles’ current hints, even when their text did not change. Unticked, they keep the readers they have; new articles are still guessed. Articles whose readers you chose yourself are never changed either way, and neither are Zendesk’s own user segments. Scheduled reads follow your last choice.

An article is removed only when a read that finished completely did not see it. A read that stopped at the page limit, failed or was interrupted proves nothing about the pages it did not reach, so it removes nothing.

#Keeping it healthy

When something needs attention, a red banner at the top of the Knowledge page says how many sources could not be read, how many articles could not be updated, how many of those still answer from an earlier version, and how many need a decision. The same total shows as a red badge on Knowledge in the sidebar, so you see it from any page. Nothing is sent by email or Slack yet; the badge is the alert.

#Not built yet

• A box to test what a given role and plan would retrieve for a question.
• Changing a source’s schedule after you add it.
• Reading an article again from the table for many rows at once.
• Replacing the token of an existing help centre source.
• Email or Slack alerts for failed sources.
• Uploading files (PDF, Word) or images; articles are text. Zendesk languages other than the one you choose, and Intercom collections, are not read.

Related

• Push help articles to Support ›
• How the AI first line works ›
• Set up Support for your team ›