Knowledge base basics

Your knowledge base is what Aigento's AI reads before it writes anything. A
thin knowledge base produces vague drafts; a good one produces drafts that
quote your actual policy. It is the highest-leverage thing you can maintain.

Anatomy of a document

Two of these get confused with each other, so it is worth being precise:

Status is lifecycle. Only active documents are indexed and searchable.
A draft is a work in progress and is invisible to search and to AI drafting
until you activate it. archived takes a document out of circulation without
deleting it.

Visibility is audience. A customer_facing document may be quoted in a
reply a customer will read. An internal document is never quoted. Suggested
drafts never see it: the database query itself excludes it. Smart Reply, the AI
that answers tickets on its own, reads internal documents that match the
ticket as guidance for deciding what to do, and is told never to quote them or
say they exist. Your logged-in agents can search them inside the app. Put your
escalation rules and "how we handle this" notes in internal documents, and
switch on Always give to the AI for a short guidelines document Smart Reply
should read on every ticket (up to five, internal and active only). Keep
anything a customer must never see, such as margin figures or legal notes, out
of documents the AI reads: for Smart Reply that boundary is an instruction to
the model, not a database filter.

A document shows in every chat widget unless you pick widgets under Shown
in widgets on the document. Then only those widgets' Help tabs show it, and the AI in your other website widgets leaves it out. Email, the
public help center and what your agents see are unaffected.

How indexing works

Documents are not searched by keyword — they are searched by meaning. To make
that work, Aigento breaks each document into passages and computes an embedding
for each one.

• A background job checks about once a minute for documents that need
  indexing, and processes them in small batches.
• A document is re-indexed when its content changes, when the workspace's
  embedding model changes, or when Aigento's indexing format is upgraded.
  You never have to trigger this by hand.
• Long documents are split into semantically coherent passages — a policy
  stays with its exceptions, a procedure stays with its steps, a question stays
  with its answer — each with a short heading. Short documents (under roughly
  800 characters) are indexed whole, because there is nothing useful to split.
• Splitting never rewrites your words. The passages are copied verbatim from
  what you wrote.

Because search happens over passages rather than whole documents, a long
document is not penalised: the relevant section can match on its own.

Indexing only runs for workspaces that have AI enabled with retrieval turned
on.

Feedback, Knowledge Gaps and Suggestions

Aigento tries to notice when the knowledge base is failing you.

• Feedback. Agents can mark a suggested document as helpful or unhelpful
  from the ticket sidebar.
• Knowledge Gaps. Every night, Aigento looks back over the last 7 days of
  tickets that Smart Reply handed to your team because the knowledge base had
  no answer (or because it had no tool for the action), tickets where the best
  knowledge base match was weak, and resolved tickets with no strong match. It
  groups the ones that ask about the same thing and lists each topic with a
  priority in the Knowledge Gaps panel. A gap needs at least three similar
  tickets, so one odd question does not create noise.
• Suggestions. Roughly hourly, Aigento looks at tickets resolved in the
  last 24 hours that had at least two outbound messages, checks whether the
  knowledge base already covers the topic, and — if it does not — drafts a
  candidate new article from how the ticket was actually resolved. You review
  it before anything is published.

The article "Knowledge Gaps and Suggestions" explains where each comes from and
what to do with them.

Importing content

You can point Aigento at a URL and have it crawl the pages and propose
knowledge base articles from what it finds. Nothing is written to your
knowledge base until you accept a proposal — accepting is what creates (or
updates) the document and queues it for indexing.

Content crawled from your own public site is treated as customer_facing,
which is correct: it is already published on the open internet. Anything you
would not put on your website should be written by hand as an internal
document instead.