How does Generative Actions route between plugins and knowledge?

(0) Share

Report

Posted on by TB-22061018-0

Hi everyone,

I'm currently working with Generative Actions in Copilot Studio and looking at the tracing logs for a test conversation. I noticed something that I need help understanding.

As you can see in the screenshot (I've redacted the names for privacy and highlighted the specific actions in red boxes):

Could you explain what the feature in the top red box (the plug icon) is and how to disable it?
I'm also wondering why the AI decided to trigger the top tool instead of the bottom tool (the book icon) when I asked my first question.

Even though my two questions were almost identical (the second one just has a full stop at the end), it selected the plugin first, failed to find a result, and then switched to the knowledge base (Generative Answers) on the second attempt.

Any insights on how this routing works behind the scenes or how to configure it properly to prevent this would be greatly appreciated. Thank you!

Categories:

General topics

Calling actions from Copilot Studio

Copilot Studio skills development

I have the same question (0)

All responses (2)

Answers (0)

Sort by

Prasad-MSFT Microsoft Employee on at

Like
a
(0)

Report
Copy link

Link copied!

The plug icon represents a tool/action (Generative Action, Power Automate flow, connector, etc.), while the book icon represents a knowledge source lookup (Generative Answers/Knowledge).

The AI orchestration model automatically chooses the capability it believes is most relevant based on the user's query, action descriptions, context, and confidence scores. Because the routing is probabilistic, even minor changes (such as punctuation) can result in different tool-selection decisions.

If the action is being triggered unexpectedly, review its description and scope. Broad descriptions often cause actions to be selected before knowledge sources. Refining the action description or disabling/removing the action can help improve routing behavior.

Was this reply helpful? Yes No
Suggested answer

chiaraalina 2,425 Super User 2026 Season 1 on at

Like
a
(0)

Report
Copy link

Link copied!
Hi @TB-22061018-0

When you turn on generative orchestration in Copilot Studio, the agent uses an LLM-based planning system to understand what the user is asking. It then decides which resources to use, such as tools, knowledge sources, topics, or child agents, and can carry out several steps to answer the request.

The plug icon means tools or plugins. These are actions that can do something, such as calling an API, starting a workflow, or running logic.

The book icon means knowledge sources. These are places where the agent looks for information, such as documents or help articles, so it can give a grounded answer.

So, your point is right: two very similar prompts can be handled in different ways. One prompt may trigger a plugin first, while another may search knowledge first or only use knowledge after retrying. This can happen because of the current conversation context, the flexible nature of LLM decisions, and how closely the prompt matches available tools or knowledge.

Between user input and agent output, the Copilot orchestrator works as an LLM-driven planner:

1. User input
The user sends a question or request in natural language.

2. Preliminary checks
The system applies Responsible AI and security checks.

3. Reasoning phase
The orchestrator looks at the conversation context, available data, and possible resources. It may also use Microsoft Graph data when available. The LLM then decides whether it can answer directly or needs to use tools, actions, knowledge sources, or other capabilities.

4. Function matching
The system looks for the best matching functions or actions. It compares the user’s request with function names, function descriptions, action names, and action descriptions using both keyword matching and meaning-based matching.

5. Tool use
If a tool or action is needed, the orchestrator builds the required API request and runs the selected tool.

6. Result analysis
The LLM reviews the results from tools or knowledge sources. It may continue reasoning and use more resources until it has enough information.

7. Final response
The agent combines the information, applies Responsible AI guidelines, and sends the final answer to the user.

How conversation context affects routing

The orchestrator uses the previous conversation history to help decide what to do next. This means the same or very similar question can be routed differently depending on whether it is asked in a new conversation or in an ongoing one.

For example:

In a fresh conversation, such as the first test in the Copilot Studio test panel, the orchestrator has little context. It may match the user’s question to a plugin description and try to run that plugin.

In an ongoing conversation, the orchestrator also considers what already happened earlier in the chat. If a plugin was tried and failed, the orchestrator may change its plan the next time.

Hope it helps!

Was this reply helpful? Yes No