Copilot Architecture

How does reasoning work?

1. Plans multiple steps to accomplish a goal,
2. Executes the next action, and then
3. Observes the results of its actions & adapts the plan.

This reasoning accuracy depends on your plugin configurations. Read the next section for configuration best practices.

How to design for Copilot

To ensure your Copilot reliably interprets your plugin, follow these best practices:

1. Write high quality names & descriptions:
   • Setup names & descriptions for all of your
     • Plugins (plugin naming best practices)
     • Plugin slots (slot naming best practices) &
     • Plugin actions (action description best practices).
   • Don’t set plugin names or descriptions to ‘TEST’ or ‘DEV 123'. Use plugin names & descriptions as close to real-life ones as possible.
   • Make all names & descriptions as descriptive as possible. e.g. name your slot as referred_candidate_email instead of slot1.
   • Make slot & plugin names unique from each other. e.g. clarify similar slot names as referred_candidate_email & referrer_email instead of email1 & email2.

2. Write faithful triggering examples
   • Don’t add semantically identical examples as both positive & negative for same plugin.
   • Write slot level feedback for Queries that spans a few diverse examples.

3. Minimize plugin complexity as much as possible:
   • Don’t pass Event slots that are not required by your APIs.
   • Don’t setup extra questions that your Plugin doesn’t actually need.

4. Finally, test like a real employee:
   • Don’t test plugins or slots with utterances such as ‘TEST’. Use examples as close to real-life ones as possible.
   • Don’t keep testing the same plugin again & again with the same utterance without clearing history. Stakeholders can use copilot clearhistory, refer to this guidance.