Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
Trigger phrases are the foundation of effective topic recognition in Copilot Studio. They train your agent's natural language understanding (NLU) model to identify when users ask about specific topics. By using trigger phrases, your agent can route conversations to the right topic with accuracy.
What are trigger phrases in Copilot Studio?
Trigger phrases train your agent's natural language understanding (NLU) model. Trigger phrases are configured at the topic level and indicate to the agent what typical user utterances should trigger a specific topic.
Trigger phrases typically capture how a user asks about a problem or issue. For example, a trigger phrase might be "problem with weeds in lawn."
When you create a new topic, you only need to provide a few sample phrases (ideally between 5 and 10). At runtime, the AI parses what the user says and triggers the topic closest in meaning to the user utterance. Learn more in Choose effective trigger phrases.
Importance of the triggering context
Copilot Studio NLU behaves differently based on the conversation state. This behavior can sometimes lead to different behaviors for the same user utterance.
The following are the different conversation states:
Start of the conversation: The agent has no context, so a user utterance is expected to either:
- Trigger a topic directly (intent recognition).
- Trigger a "did you mean" (Multiple Topics Matched) disambiguation question among intent candidates if there are multiple matching topics.
- Go to a fallback topic if the intent isn't recognized.
After a "did you mean" (Multiple Topics Matched) is triggered: NLU optimizes to match one of the suggested topics, with higher thresholds to move out of the presented options.
Switching out from a current topic: If the NLU is trying to fill a slot in a topic, and the user gives a user query that could trigger another topic (topic switching).
Punctuation
The NLU model works the same way regardless of punctuation, including question marks.
Create new trigger phrases
If possible, start with real production data instead of making up your own trigger phrases. The best trigger phrases are similar to actual data coming from users. These phrases are the ones that users ask a deployed agent.
Don't leave out specific words. The model gives less weight to unnecessary words, such as stop words. Stop words are words that the system filters out before processing natural language data because they're insignificant.
Optimize trigger phrases
The following best practices help you optimize your trigger phrases.
| Tip | Examples |
|---|---|
| Have at least 5-10 trigger phrases per topic Iterate and add more as you learn from users. |
Find my nearest store Check store location Find a store Find me your nearest location Store near me |
| Vary sentence structure and key terms The model automatically considers variations of those phrases. |
When are you closed Daily open hours |
| Use short trigger phrases Fewer than 10 words. |
When are you open |
| Avoid single-word trigger phrases This increases weight for specific words in topic triggering. It can introduce confusion between similar topics. |
Store |
| Use complete phrases | Can I talk to a human assistant |
| Have unique verbs and nouns or combinations of those | I need customer service I want to speak with a consultant |
| Avoid using the same entity variation You don't need to use all of the examples from the entity value. The NLU automatically considers all the variations. |
I want to order a burger I would like a pizza I want chicken nuggets |
Balance the number of trigger phrases per topic
Try to balance the number of trigger phrases between topics. That way, the NLU capabilities don't overweight a topic versus another based on the configured trigger phrases.
Assess your changes
After you update trigger phrases or merge or split topics, assess the changes. For example:
- You observe an immediate change in agent behavior through the test chat. For example, a topic might trigger or stop triggering based on trigger phrase updates.
- After deploying your agent and handling traffic, you see higher or lower deflection rates (non-escalation rates). Check the Analytics tab in Copilot Studio to observe these rates.
Tip
To test topic triggering and how your NLU model performs against test data in bulk, use the Copilot Studio Kit.