In Copilot Studio you can create custom topics to handle requests by end user for a specific scenario.
The Classic Orchestration and Generative Orchestration has different sets of trigger to sense activity requests.
The above list contains triggers that can sense activities like messages, events etc.
There is a trigger to track conversation changes, user activity and many more. In this article, I will cover couple of triggers exclusively available for generative orchestration only.
1. An AI generated response is about to be sent
2. When a plan completes
1. AI response about to be sent (trigger)
This trigger is available in Generative AI orchestration mode and can be changed by clicking on selector on the topic trigger node.
You can view the video here – https://youtu.be/7R_KBBCfP9M
This feature is available for agents with generative orchestration enabled.
The AI response generated trigger executes when LLM has generated a response for the user, but yet not sent the message out
It can be used for custom tailoring of AI response generated by the LLM.
It provide guard rails for your agent.
The schematic architecture is as shown below:
You can intercept the request and perform some of the actions listed below:
- Format the response
- Check the response and moderate it as per business logic
- Check policy violations
- Add disclaimer text
- Text redactions
- Check compliance parameters / rules
- Check for accidental PII / confidential information exposure
- Language translation on the fly
2. A Plan Completes (Trigger)
This feature is available for agents with generative orchestration enabled.
You can view the video here – https://youtu.be/JI1mTfSpW-c
It’s a special type of trigger for Generative Orchestration that executes topic or actions after an AI plan is finished running.
The trigger has an ability to intercept the request and inject custom logic in to the topic steps.
It also allows you to setup automated process to run after agent has completed executing a plan.
The schematic architecture for this trigger is as shown below:
The list of tasks that can be done post intercepting the requests are listed below:
- Setting variables
- Logging requests
- Cleanup activities
- Generating specific response based on plan output
- Run some background processes
- Author formatted response
