What your AI can learn from

Kai answers only from what you connect, so this is where its quality is actually decided. Open AI agents and choose Add content to see the sources.

Content your customers read too

• Help Articles — your help center. The best place to start, because one article both answers the question inside the AI and stands on its own for someone reading it.

• News & Changelogs — what you've shipped. Worth connecting because “is this available yet” is a question help centers rarely answer.

Draft articles are invisible to the AI. An article has to be published before it can appear in an answer — the most common explanation for “we documented that, why doesn't Kai know it?”

Content only the AI sees

Snippets are text written for Kai and not published anywhere. This is where things go that are true but not worth a public article: current pricing edge cases, the wording you use for a known issue, internal policy on refunds.

They're also the fastest fix for a wrong answer. Writing a three-line snippet takes a minute; writing a good article takes longer and can wait until you know the question is common.

Material you already have

Four sources exist so you don't have to rewrite what's written.

• Websites — crawls any public site and uses what it finds. The usual first move for a product whose documentation lives outside Gleap.

• Notion — connects your workspace, for teams whose real documentation lives there.

• Files — PDF, Word, PowerPoint, Excel and Markdown, converted to text on upload. Good for manuals and spec sheets nobody will ever port into articles.

• YouTube — analyses a video or a whole playlist. If your onboarding is a video series, this makes it answerable in text.

Connecting existing material is quick, which is exactly why it's worth checking afterwards. A crawled marketing site teaches Kai your marketing claims, and it will repeat them to customers as fact.

Technical and live sources

• Code Repositories — connect a GitHub repository so your code can serve as context. This is what lets an AI answer questions about behaviour nobody wrote down.

• API Data Sources — fetches content from an API of yours, for knowledge that changes too often to keep as text.

Knowledge is not the same as tools

The same picker also offers API actions and MCP servers. Those aren't knowledge — they let the AI do something rather than know something. If you want an answer, add a source; if you want an action, add a tool.

Where to start

Publish help articles for the questions you already answer most often, and add snippets for everything that's true but not publishable. That combination covers more real questions than any amount of connected material, because it's written to answer rather than to describe.