Knowledge base overview

Ochre's knowledge base is two products in one. It is a public help center your customers read, and it is the source of truth your AI agent uses to draft replies. The same article that ranks on Google answers a chat at 2am.

What it does

The knowledge base stores every article your team writes, syncs from Notion, pulls from GitBook, crawls from a public URL, or uploads as Markdown. Articles live in categories. Each article has SEO metadata, visibility rules, and analytics. When you publish, the article becomes part of three surfaces at once.

• Public help center. Your customers browse and search at your help domain.
• AI agent context. The agent retrieves relevant articles when drafting replies. See The brain: KB graph.
• In-app search. Agents in the inbox pull article snippets while replying.

You do not pick which surface. Publishing fills all three.

Where articles come from

Ochre supports four ingest paths. You can use any combination.

Authored in Ochre

Write directly in the editor. Markdown, slash commands, and inline images. This is the fastest path. See Authoring articles.

Synced from Notion

Paste a Notion integration token and pick the pages you want. Ochre walks the page tree on a schedule and updates articles in place. See Connecting Notion.

Synced from GitBook

Drop in a GitBook personal access token and select spaces. Same idea, different source. See Connecting GitBook.

URL crawl or Markdown upload

Point the BFS crawler at a public docs URL, or upload .md files directly. Useful when your content lives on a custom CMS or a static site we do not have a first-class connector for.

The Notion and GitBook walkers cap at 500 pages per run and 4 levels of depth. They retry once on rate limits or transient server errors. Most teams never hit the cap.

Confluence is not supported. Mirror Confluence pages into Notion or GitBook if you need them in the brain.

Structure

Articles live in categories. Categories support one level of nesting, so you can have a parent like "Billing" and sub-categories like "Invoices" and "Refunds". That is it. No infinite trees, no orphan pages.

The flat structure is deliberate. Customers do not read past two clicks. See Categories and sub-categories.

Lifecycle

Every article has a status:

• Draft. Visible only to your team. Not indexed, not searched, not used by AI.
• Published. Live on the help center, indexed if you allow it, fed to the AI agent.
• Archived. Removed from public surfaces but kept for history. Not used by AI.

You can flip an article back and forth. Archiving is reversible.

URLs

Published articles render at two equivalent URLs out of the box:

• Apex: https://ochrehq.com/help/<workspace-slug>/<article-slug>
• Subdomain: https://<workspace-slug>.ochrehq.com/help/<article-slug>

Set up a custom domain (help.yourcompany.com) and your articles serve from there too, with SSL handled for you.

Cross-article links inside Ochre use the apex form (/help/ochre/<slug>) because it is portable.

Visibility

By default, the help center is public. You can also gate it. Gated help centers require a magic link sent to the visitor's email. The session token is HMAC-signed and lives 30 days. Useful for partner docs or paid product tiers. See Article visibility and gating.

SEO

Each article has its own SEO fields: title, description, OG image, robots directive, canonical URL, slug, and keywords. Defaults are sensible. Override them when you have a reason. Full reference: Article SEO.

Branding

Match the help center to your product. Colors, fonts, hero, footer, and a sanitized custom CSS slot. See Help center theme.

Discoverability

Ochre generates a sitemap when you flip the workspace-level toggle. See Help center sitemap. Submit it to Google Search Console and Bing Webmaster Tools.

Measurement

Every article tracks views, helpful and unhelpful votes, and the search queries that surfaced it. See Article analytics and Article voting.

The vote signal flows back into AI confidence. Articles voted unhelpful repeatedly get downweighted in retrieval. Do not ignore the votes.

A reasonable starting point

If you are building from zero, do this in order.

1. Create five categories that match your product surface area.
2. Write the ten articles your support team retypes most often.
3. Connect Notion or GitBook only if you already maintain docs there. Do not dual-author.
4. Set the help center theme.
5. Publish with the sitemap on. Submit it to Search Console.
6. Watch the votes. Rewrite the bottom three each month.

Most of the value is in the first ten articles. Get them right, and the AI agent gets dramatically more accurate. See AI overview for how that loop closes.