# Knowledge Base

The knowledge base is your help-center content. It pays off twice: customers self-serve before opening a ticket, and your AI agents ground their answers in it and cite it. This article is for whoever writes and manages articles. Open **Helpdesk → Knowledge Base**.

## Articles

An article is a Markdown document with a title, a category, and an **audience**:

- **Public** — visible to customers in the help center **and** available to AI agents
- **Internal** — staff and AI agents only; use it for playbooks, refund rules, escalation steps, and other content customers shouldn't see

## Writing and Publishing

1. Click **New article**. Write the body in Markdown (headings, lists, bold, links, code).
2. Choose the **audience** (Public or Internal) and a **category**.
3. Save it as a **draft** while you work.
4. When it's ready, click **Publish**.

Publishing does two things: it makes the article live (public articles appear in the customer portal's help center), and it **syncs the article into your AI knowledge store**. From that point, your chat, voice, SMS, and email agents can find it and cite it when answering.

Editing a published article re-syncs it automatically, so agents never quote a stale version. **Unpublishing** or **archiving** removes it from the portal and the AI store; the draft is preserved.

## Categories

Group articles into categories (e.g. "Getting Started", "Billing", "Troubleshooting") so the help center is easy to browse. Create them with **New category**. Deleting a category leaves its articles in place, just uncategorized.

## How AI Agents Use Articles

Once published, articles become a grounded source for every AI agent through the platform's retrieval tools. When an agent answers a customer, it can pull from your articles and cite them — so the AI's answers reflect *your* policies and product, not guesses. This is the single highest-leverage thing you can do to make AI replies accurate: keep the knowledge base current.

## Deflection

When a customer starts typing a ticket subject in the portal, matching public articles surface as suggestions ("These articles might answer your question"). Many issues get resolved before a ticket is ever created. Each article also has a **"Did this answer your question?"** prompt.

## Measuring and Improving

Each published article tracks **views**, **helpful / not-helpful** votes, and how often AI agents cite it. Use these to find:

- Articles that get lots of "not helpful" votes → rewrite them
- High-traffic topics with no article → write one
- Recurring ticket subjects that a good article could have deflected → close the gap

A small, well-maintained set of articles beats a large, stale one — both for customers and for AI accuracy.