Zum Hauptinhalt springen

Create and Edit Knowledge Bases

This page walks you through building a knowledge from scratch and editing one later.

Before you start​

  • You have at least one embedding model configured in the workspace. See Models → Introduction.
  • You know what content you want the knowledge to hold and where it lives (a set of files to upload, a few inline snippets, …).
  • You have Editor, Manager, or Admin rights in the workspace.

Step by step — create a knowledge​

1. Open the Knowledge tab and click Create a new knowledge​

The Knowledge tab with the Create a new knowledge button at the top.

2. Give it a name and description​

  • Name — short and specific: "HR Handbook", "Product FAQ".
  • Description — a sentence explaining what the knowledge contains. This is shown to agents that have the knowledge connected, so they understand when to use it.

3. Pick the embedding model​

The embedding model translates documents and queries into the same numeric space so they can be compared. Pick from the list of embedding-enabled models in the workspace.

Important: once you pick an embedding model and add data, you cannot easily switch. Changing the model invalidates every chunk you have indexed. Pick deliberately — and pick the same model for knowledges you want to keep comparable.

4. Set the refresh policy​

The refresh policy controls how the knowledge re-reads its sources. The dropdown offers:

  • Manual / On Change — the platform refreshes only when triggered (from the editor, from a workflow, or when a source signals a change).
  • Every 2 hours / 4 hours / 6 hours / 12 hours — fixed-interval refreshes during the day.
  • Daily / Weekly — for content that rarely changes.

📷 SCREENSHOT: The knowledge create form with name, description, model selector, and refresh-policy dropdown visible.

You can change the refresh policy later. Manual / On Change is a good starting point while you are testing — short intervals consume embedding tokens.

5. Add at least one source​

The form shows an empty source list with an Add source button. Click it and pick the connector that matches your content. The available source types in the workspace include:

  • File Upload — a drag-and-drop area in the editor where you upload documents (PDF, Word, Markdown, plain text, …). Files are stored alongside the knowledge.
  • Static content — short text snippets you maintain inline in the editor. Good for FAQ-style content that does not warrant a separate document.

Additional source types may be available depending on the installation. Each connector has its own configuration fields. Fill them in and save.

You can add multiple sources to one knowledge. The platform deduplicates and re-indexes them together.

6. Save and run the first refresh​

After saving the knowledge, click Refresh to run an initial indexing. The first refresh can take a while — minutes for small knowledges, longer for big ones. Watch the Jobs view for progress.

📷 SCREENSHOT: The knowledge editor right after the first refresh: a green status, a chunk count, and a "Last refreshed: just now" label.

Editing an existing knowledge​

Click any knowledge card in the Knowledge tab to open the editor. From there you can:

  • Rename, edit the description, change the refresh policy.
  • Add or remove sources.
  • Trigger a manual refresh.
  • Open the Jobs view to inspect past refreshes.
  • Test a query — type a question and see which chunks come back. Use this to spot obviously wrong matches.
  • Duplicate the knowledge — useful if you want to fork a configuration to a related topic.
  • Delete the knowledge.

The knowledge editor with the source configuration, the model picker, and the toolbar (Jobs, Test, Duplicate, Refresh, Remove, Save).

Testing a knowledge​

Knowledge testing is the single most useful debugging tool. Type a question, get the top matches, and look at them:

  • If the matches are clearly off-topic, the embedding model or the chunking is wrong.
  • If the matches look right but the AI's answer is wrong, the prompt or the model is the issue.

Test before you give the knowledge to an automation in production.

Recommendations​

  • ✅ Start with one source and verify the knowledge works before adding more.
  • ✅ Use the Test function with five to ten realistic queries before connecting a knowledge to production automations.
  • ✅ Keep the description accurate. Agents use it to decide when to consult the knowledge.
  • ✅ Group content by topic, not by file type. A knowledge with five Word documents and three web pages on the same topic is fine; a knowledge with mixed topics is not.
  • ⚠️ Adding a source does not automatically refresh the knowledge. Refresh manually or wait for the next scheduled refresh.
  • ⚠️ Changing the embedding model invalidates everything. Plan such a change as a "create new knowledge, repoint automations, delete the old one" sequence.
  • ❌ Do not dump unrelated content into one knowledge to "save quotas". Search quality drops fast as content diversity grows.

What to do next​