Grounding agents and AI Prompt nodes on knowledge

Grounding connects a knowledge base to an Agent node or an AI Prompt node in CloudRadial AutomationAI, so the model reasons with your own documents instead of general knowledge alone. This article explains how to attach knowledge in the Designer, how AutomationAI recalls the relevant material at run time, and how a grounded agent searches knowledge on its own. It is for workspace users who build workflows (Admin or Owner Role).

• What grounding does
• Grounding a node on knowledge
• How recall works at run time
• The knowledge search tool for agents

What Grounding Does

When a node is grounded on one or more Knowledge folders or documents, AutomationAI supplies the most relevant passages from that knowledge to the model as context for that node. Grounding is available on both the AI Prompt node and the Agent node. It lets the model answer using your runbooks, policies, and prior notes rather than guessing.

Grounding a Node on Knowledge

Open the workflow in the Designer and select the AI Prompt or Agent node. In the node's settings, use the grounding section to attach knowledge:

1. Add a Knowledge folder or a specific document from the picker
2. Each attachment appears as a chip; remove a chip to detach it
3. Use the recall preview to see the kind of context the node will pull

Attaching a folder grounds the node on the documents within it. You can attach more than one folder or document to the same node.

How Recall Works at Run Time

When the node runs, AutomationAI performs a semantic search across the attached knowledge and injects the closest-matching passages into the node's context. Matching is by meaning, not exact keywords, so relevant material surfaces even when the wording differs. Only the attached folders and documents are searched.

The Knowledge Search Tool for Agents

A grounded Agent node also gets a built-in knowledge search tool. Beyond the context injected up front, the agent can decide during its run to search the attached knowledge for more detail — the model invokes the tool itself when it needs additional context. The search respects the folder scope you attached to the node.