Knowledge Base

The knowledge base is the library of support content the AI assistant reads from when it drafts responses, suggests resolutions, or triages cases. A well-maintained knowledge base is what makes the Copilot useful — it's the difference between vague AI fluff and accurate, on-brand customer answers.

What's in the knowledge base

The knowledge base is one object — Knowledge Article — and every article is an ordinary record you create from Service → Knowledge. There is no separate index and no file store behind it: what you see in the list view is what the Copilot can read.

Each article is filed under one Category, and the seven categories are the whole vocabulary:

Alongside the category, each article carries Tags (Auth, SSO, Mobile, Email, Reports, Performance, Data Import, Webhooks), a Status (Draft → In Review → Published → Archived), an Audience (Public or Internal), and a Language. Articles get an Article Number — KA-0001 — automatically, and that number is what the Copilot cites.

Finding articles

The Knowledge list ships with four tabs:

• All — everything, grouped by category, newest publication first.
• Published — published articles only, most-viewed first.
• My Drafts — your own drafts and in-review articles.
• Review Queue — published articles ranked least-recently-reviewed first, so the top of the list is what is most likely stale.

The object's declared search targets are the Title, Summary and Article Number, so those are the fields global search and the lookup picker read — the words a colleague would search for belong there rather than buried in the body.

How the AI assistant uses it

One skill reads articles: Customer 360°. Ask it for a customer's history and, alongside the account, contacts, open deals and cases, it queries published articles matched on the category or tags of the cases it just read, and cites each one by number (KA-0007).

Two things it is worth being precise about, because both used to be described the other way round:

• Case Triage does not search for articles. It declares two tools, describe_object and get_record, and no query tool at all. It returns a priority with its reason and a pointer at Escalate or Close — no article list.
• Email Drafting does not quote articles. A draft is grounded in the records the skill reads — the contact, plus the related account or opportunity — and nothing in its instructions reaches for an article. When a reply should quote one, take the article Customer 360° matched and paste it in.

Matching is by category and tags, not by meaning. An article filed under the wrong category will not surface however well its body answers the question — which makes categorisation, not prose, the highest-leverage thing a knowledge manager does.

Curating the knowledge base

The knowledge base is only as good as what you put in it. A typical content-curation workflow:

1. Capture what's working

When an agent resolves a tricky case with a great answer, copy that answer into a new article. The Originating Case field links it back to the case it came from, so the next reader can see the real conversation behind it.

2. Triage incoming questions

Every week, look at the most common case subjects. The top 10 questions deserve dedicated articles — if 30 customers asked the same thing, the 31st should be answered from an article.

3. Retire stale content

If a product changes, the corresponding articles need updating. Review Queue ranks published articles by how long it has been since anyone touched them; work it from the top. Publishing stamps the publication date once and refreshes the review date on every later edit, so the ranking reflects real attention.

4. Watch what gets cited and read

Every published article carries Helpful and Not Helpful buttons. Pressing one records your Article Feedback — one row per reader per article — and the article's Helpful Votes / Not Helpful Votes are recounted from those rows immediately. Change your mind and press the other button: your own verdict moves, it does not add a second vote, so the two numbers mean how many people, not how many clicks. The feedback rows themselves are listed on the article, so an author can see who said an answer did not work and read the comment they left.

An article voted down repeatedly is doing damage and should be rewritten rather than left up.

Tips for service managers

• ✅ Make KB updates part of the weekly team meeting — "what 3 new articles do we need this week?"
• ✅ Reward agents who contribute KB content; it scales their expertise.
• ✅ Audit the articles Customer 360° cites — pick 10 at random per week and check they actually fit the cases they were matched to.

Tips for admins

• Publish it or it does not exist. The Copilot queries published articles only; a draft is invisible to it.
• Set the audience deliberately. Public articles are the ones written for customers; internal ones are agent-only. Note that the customer portal does not serve articles today — the guest profile cannot read the object at all, by design — so Public is a statement of intent, not a publication channel.
• Categories are metadata. Adding an eighth category is a change to the object, not a setting — see Customization › Extending Objects.
• HotCRM ships no separate sales, product or competitive knowledge base to wire up; there is this one object and its categories. AI Copilot › Knowledge Bases has the detail on what the Copilot can and cannot do with it.