The AI chat widget sends requests to a language model through the selected AI provider. You can configure general instructions (the Global prompt) and a set of quick-question buttons (Quick questions).
Configuration
| Name | Required | Description |
|---|---|---|
| Global prompt | No | General instructions for each request. The widget combines them with the prompt of the selected quick question before sending it |
| Quick questions | Yes | A set of buttons with labels and prompt text (up to 20). Add at least one question for the widget to work correctly |
Configure the following fields for each quick question:
| Name | Required | Description |
|---|---|---|
| Question name | Yes | A short label displayed in the widget’s bottom panel and in the chat |
| Prompt | Yes | Instructions sent to the model when the user selects the button. Go templating is supported |
When writing a prompt, explicitly specify the names of the Model Context Protocol (MCP) tools that the model must call to prepare the response.
Example prompt for a quick question:
1. Call the MCP tool get_external_data for the external service "Deckhouse Code" and retrieve pipelines for the project with ID {{ .entity.properties.deckhouse_code_id }}.
2. Display a table with the 10 latest pipelines.Using the widget
To use the widget, add at least one AI provider.
The chat does not retain history:
- It displays only one response to the most recent question.
- The response is not retained when the user navigates to another page or refreshes the page.
Before sending a question, the user can customize the prompt by selecting Send with prompt changes from the question button menu.