Documentation

Running the knowledge base

The editorial cycle, import from a URL, and the rules that decide what the agent may cite. Practice rather than retrieval theory.

An article's life cycle

The agent sees only the published state. That is the simplest way to keep a working note from becoming an answer to a customer.

StateWho usually does itVisible to the agent
DraftThe person writing the contentNo
ReviewedA second person on the teamNo
PublishedSomeone with publishing rightsYes
ArchivedSomeone withdrawing outdated contentNo

Import from a URL

What you can give
The address of a page with textual content. Supported types are a web page, plain text and markdown text.
One page, not a whole site
The import fetches the address you give. It does not follow links and does not crawl a site.
Guard against internal addresses
The import checks the address scheme, every network address and every redirect, and counts size while downloading. It also passes no authentication headers, so it cannot reach behind your login.
What happens next
Imported content arrives as a draft and goes through the same editorial cycle as a hand written article.

How to write so the agent finds it

01

One topic per article

An article covering returns, complaints and exchanges at once will be retrieved for all three questions and serve none of them well.

02

The customer's words, not only yours

If customers say parcel and you say shipment, both words are worth having in the text.

03

Conditions stated plainly

A deadline, an exception and a threshold written out can be quoted. A general sentence cannot.

04

Dates and validity

Content that applies from a given date should say so, otherwise versioning answers what the agent knew but not since when.

The gap register as a backlog

Questions without coverage land in the gap register. It is the best source of further articles, because it comes from real conversations rather than from guesses about what customers might ask.

Start with five articles

The five most frequent questions are enough to see how the agent behaves on your content and where it sends a case to a human by itself.