The knowledge base can automatically answer user questions using AI. This works through the KB widget — the same one that displays articles.
How It Works
When a user types a question in the widget and clicks “Ask,” the system searches the vector database (RAG), finds the best matching article fragments, and generates a response via an LLM. The user sees the answer directly in the widget, with the source (link to the article).
In addition to AI, the widget still allows regular article searching via the “Search” button.
Enabling AI
- Go to the knowledge base settings (Knowledge Base > list > click the base > Edit)
- Go to the NoeAI tab
- Check the “AI enabled” checkbox
- Select a vector database or leave empty — the system will create one automatically and add articles from this knowledge base as a source
- Optionally select an AI connector (chat) and set the timeout
- Save
After enabling AI, the widget placeholder changes to “Ask or search…” and two buttons appear: Ask (blue) and Search (white).
System Prompt
The default system prompt instructs AI on how to answer questions. You can change it in the NoeAI tab in the knowledge base settings, in the “System prompt” field. If left empty, the prompt from the vector database will be used.
AI Logs
Every question asked through AI in the widget is logged. Logs contain the question, answer, model used, response time, and user feedback (thumbs up/down).
AI logs can be found on the knowledge base page — the “AI Logs” button in the top bar (visible when AI is enabled). You can filter by status, feedback, and source type.
Regular widget searches (without AI) are also logged — with the number of results found. This way you know what users are searching for and whether they find answers.
Feedback
After receiving an AI response, the user sees two buttons — thumbs up and thumbs down. Clicking saves the rating to the log. This helps evaluate response quality and improve the knowledge base.
Requirements
AI requires a configured AI connector (e.g., OpenAI, Claude, Gemini) in the Connectors section. The vector database also needs an embeddings connector.