What Are AI Knowledge Sources?
Knowledge sources are the data that the AI agent draws information from when answering user questions. The system uses the RAG (Retrieval-Augmented Generation) mechanism — instead of relying solely on the AI model’s built-in knowledge, it searches for the most relevant fragments from your database and passes them as context for the response.
This allows the AI agent to:
- Answer based on current data from your company
- Know the specifics of your products, processes, and procedures
- Cite specific articles from the knowledge base or ticket content
What Sources Can Be Connected?
Each vector database can have multiple sources of different types connected:
| Source Type | Description | Example Use Case |
|---|---|---|
| Knowledge Base | Articles from the knowledge base (KB) | FAQ, product documentation, guides |
| Helpdesk (Desk) | Helpdesk tickets with comments | Previous ticket solutions |
| Email Mailbox | Email messages | Client correspondence history |
| Team | Team posts | Internal documentation, notes |
| Project | Project posts | Project documentation |
How to Add a Source
- Go to Noe → Vector Databases and select a database
- Click the NoeAI Sources button
- Click + Add Source
- Select the source type (e.g., Knowledge Base, Helpdesk)
- From the list, select the specific source to connect
After connecting, the source will appear in the “Connected Sources” list.
Managing a Source
For each connected source you can:
Indexing
- Index all — adds all elements from the source to the vector database (in the background). You’ll receive a notification when complete
- Delete all — removes indexed entries from this source
- Disconnect source — disconnects the source and deletes associated entries
Settings
- Auto-indexing — new and edited elements are indexed automatically after saving
- Index from date — only elements created from the given date will be included
Auto-indexing
When you enable this option in the source settings:
- New entry in knowledge base / new ticket / new post → automatically goes to the vector database
- Entry edit (content or title change) → old embedding is removed and a new one is created
- Works for knowledge bases, helpdesk, and posts. Auto-indexing is not available for email mailboxes (too high volume)
Excluding and Editing Content
In the source management view, you see a list of all elements with information:
- Green icon next to the number → element indexed (click to see the vector entry)
- #ID number → link to the original source (e.g., KB entry)
- “Excluded” badge → element skipped during indexing
- “Custom content” badge → original content overwritten
For each element you can:
- Exclude — the element won’t be indexed (e.g., outdated information)
- Edit content — replace the original content with your own version (e.g., simplify or supplement)
- Restore default — undo exclusion or editing
Indexing Errors
If indexing fails for some elements (e.g., empty content, embedding API error), you’ll see a red box on the source management page with a list of problems — element number and error description.
Indexing errors are also visible in account activity (Noe module) — with a link to the vector database and the problematic element.
Most common causes:
- Empty content — the entry has no content to index
- Token limit exceeded — the text is too long for the embedding model (solution: enable chunking in vector database settings)
- API error — connection problem with the embedding provider
Example: Helpdesk Desk with AI
When you enable AI in the Desk helpdesk settings:
- The system automatically creates a vector database assigned to that Desk
- Desk tickets are added as a source by default — the AI agent learns from ticket history
- You can connect additional sources (e.g., a knowledge base with FAQ or another Desk)
This way, the AI agent responding in the helpdesk widget:
- Knows answers from previous tickets
- Has access to knowledge base articles
- Can use knowledge from multiple sources simultaneously
Indexed Entries
In the “Indexed entries” tab, you see all elements currently in the vector database. You can filter them by source type, edit content, or delete them.
Each entry with custom (edited) content is marked with a Custom content badge.
The counter “12/12 indexed” shows unique sources — if one entry was split into chunks, it still counts as one indexed element.