Before you can enable the AI configuration for incident categorization, you must prepare your training and inference data, review your filters, and select incidents to include in training.
Before you begin
Ensure that the AI service is available and running.
About this task
To prepare the data, complete the following high-level steps:
- Review the default training and inference filters. The training filter determines which incidents can be selected for training. The inference filter determines which incidents display AI suggestions.
- In the Tickets application, select which incidents are used for training. Select incidents that are included in the training filter.
Procedure
-
Review the training and inference filters.
Filters determine which incidents can be used for training and which incidents display AI suggestions. The filters are defined as query definitions for the incident object structure.
You can edit the query definitions to change which incidents are included in training or inference.
-
Select incidents for training.
From the incidents that are included in the training filter, you must select at least 20 incidents for training. These incidents must represent at least two different incident type and category combinations, with at least 10 incidents for each combination.
Consider the following best practices when selecting training data:
- Use unique incidents. The training process filters out duplicate incidents.
- Use diverse incidents that include a range of descriptions and address various problems.
- Use incidents that have accurate incident type and category assignments.
- Ensure that each incident type and category has an accurate description in the system.
- Ensure that each incident has an accurate summary and details.
- Although the minimum is 20 incidents, larger, more diverse, and accurate training data improves model accuracy. The ideal training filter contains 20-50 incidents per incident type and category combination.
What to do next
Configure the AI configuration. For more information, see Configuring AI-powered incident categorization.