Building tags with Prompts
Updated over a week ago

What’s's new prompting experience?

Prompts in are a new way to build tags that’s unique to Applying our expertise of 5 years working with customers to generate custom taxonomies, we have adapted our technology to leverage generative AI to transform what our customers and users have in mind into tags that capture it.

How to leverage the new prompting experience?

How to create an accurate prompt?

For accurate prompts, try to be as specific and not verbose as possible. Avoid using terms like Members or Customers.

❌ Bad prompt *Tickets from members that send a request because they want to cancel their subscription

✅ Better prompt Tickets related to cancel subscription

❌ Bad prompt *Customers inquiring about how to get a refund, but don't include anything about fraudulent charges

*✅ Better prompt Refund requests, exclude fraudulent charges

How to combine multiple ideas in a prompt?

The best way is to combine them by leveraging a phrase separated by comma and “or”. You can combine as many ideas as you want.

Example Tickets related to cancel account. Alternatively, include Tickets related to deactivate account or delete account.

First try with one idea and if you see that the prompt isn’t automatically capturing related ideas, you can include them in the prompt.

In the example above, it may be the case that “cancel account” already captures deactivate account or delete account.

How to exclude ideas in a prompt?

The best way is to exclude them by leveraging a phrase separated by comma and “or”. You can combine as many ideas as you want.

Example Refund requests. Exclude fraudulent charges, receiving refunds or refund dispute.

How to make a prompt more specific?

The best way is to specify what idea you want to focus on from a more general concept

Example Refund requests. Include only safety issues.

How to create a “Catch-all” prompt?

If you use "about" or "related to", it will open up the search to closely related concepts (eg: synonyms), increasing the recall. This is suited for situations where we are aiming for covering more than needed versus being more accurate. For example, in exclusion it can have good applications.

In the case below it will try to open up as much as possible to exclude anything that can be similar to fraud:

Example Refund requests. Exclude all tickets about fraud.

⚠️ When used incorrectly, it can lead to lower accuracy. For instance, the prompt “tickets about fraud” may open the search too much when trying to capture fraud.

Did this answer your question?