Skip to content

Train an agent with files, a website or past chats

Knowledge sources are what an agent reads before it replies, beyond its instructions. They sit under Knowledge sources in the agent editor, and each one is plain text the agent reads, whether it began as a file, a web page or your old conversations.

To add one:

  1. Open the agent and press Add knowledge source.
  2. Give it a Name.
  3. Choose a Type: Upload file, Add website or WhatsApp chats.
  4. Fill in what that type asks for.
  5. Press Done, or Start reading for chats.

Open the info mark beside Upload file: it takes PDF or text, up to 5 files at a time. Text files include .txt, .md, .csv, .json, .yaml and .log. Drop the files onto the box, or press upload inside it.

Add several at once and each file becomes its own source, named with your name and the file’s name. Only the words are kept. A long file keeps its first 40,000 characters and says it was cut, and a scanned PDF with no text in it is refused.

Type the page into Website URL. Under Advanced settings:

  • Load URLs from sitemap is on to begin with, and adds the pages the site lists.
  • Additional URLs takes up to 4 more starting pages.
  • Max crawling depth runs from 1 to 3 and starts at 3. Open the info mark beside it and it says 1 reads the page alone, 3 follows links two hops out, up to 8 pages.
  • Include URL globs and Exclude URL globs fence which paths are read. Exclude wins over include, and Learn about URL globs explains the patterns.

Only pages on the same site are followed. Each page becomes its own source, holding up to 4,000 characters of it. A page that needs JavaScript to show its words can’t be read, and the message after Done counts any pages that were skipped.

Set resync schedule offers Never, Daily or Weekly, and Never is the default. On a schedule the pages are read again while the agent is published, and a page you switched off stays off. If a re-read fails, the Activity tab of a business the agent answers says so.

WhatsApp chats reads the past conversations of the business in the picker and writes up what they teach. Customer names and phone numbers are left out, and so are group chats.

Once you press Start reading, the card counts the conversations read so far against the total. It starts with the newest, up to 5,000 of them, then writes them up as one sheet of up to 12,000 characters. Stop ends the reading early and writes up what it has.

A read carries on while you save the agent or leave the editor. If the page reloads, the card offers Carry on, Use what was read, or Write it up when only the write-up is left. A big inbox can hit the hourly reading limit, and Carry on picks it up later.

An agent keeps one chats source per business, so reading the same inbox again replaces the last write-up.

An agent reads 24,000 characters of knowledge before every reply. The line at the foot of the card counts what you’ve used against that.

Sources are read from the top of the card down. The one that crosses the limit is cut where it crosses, and anything below it isn’t read. A new source is added at the bottom, so switch off what the agent doesn’t need rather than stacking more on top.

Press a source’s name to read its words. Change what’s wrong and press Save changes. A source holds up to 100,000 characters. A web page on a Daily or Weekly schedule opens read-only, because its next read would replace your edit.

The switch on each row turns a source off without deleting it. The bin removes it, and the message that follows offers Undo.

Can I paste text in instead?

Not directly. Save the words as a .txt file and upload that.

Does a chats sheet include customers’ names?

No. Names and phone numbers are left out. Press the source’s name to read it before you publish.

My website changed. Does the agent know?

Only if its pages were added with Daily or Weekly, and only while the agent is published. Otherwise add the pages again.

Why won’t Start reading begin?

A read of that inbox is already on the card, and the drawer says so in amber. Finish that one or remove it first.