NOTE:

You only need to configure an AI Reference dataset if you are configuring a "RAG" Suggestion Assistant, which make suggestions based on other data in your system.

Label

Notes

Name

Specify a name for this AI Reference Data Set

Description

Optional field to add in some description.

Component

Select the "main" component in the dataset.

If your dataset contains more than one component, separated by a 1-Many relationship, you will most likely need to select the "Many" component.

For example, in an AI Reference Dataset containing both Incidents and their related Actions, we have selected “Action” for the component

Component Filter

Select the Component Filter (link to refer to Component Filters) by which all the records from the selected component will be assessed from.

In our example, we selected a filter "Actions from Incidents, in the last 12 months" so that our AI Reference dataset only includes Actions that came from Incidents.

Prompt

Where your data contains structured data from multiple components, we need to translate this into prose, so that the AI can read it and understand the relationship between the data.

You can set this up with placeholder fields and the system will substitute the values from the record and send it to the engine.

Comply click "[Select Field] to see a list of available placeholders and select those that you want to include in your dataset prompt.

Example Prompt:

When the following Incident ([Incident as source of Action:Incident Number]) occurred: '''[Incident as source of Action:Summary]. [Incident as source of Action:Detailed Description]''', we took this Action ([Action Number]): '''[Action Summary]'''

In the above example, the system will create a dataset where every row looks like the following, describing the relationship between an Incident and each of the Actions that were associated with it:

When the following Incident (INC-0000172) occurred: '''Leaking fitting on relief line for acid tank. Acid truck driver was unloading tanker when the pressure relief fitting began to leak, spraying acid.''', we took this Action (ACT-0000054): '''Acid truck pressure relief fitting leaking. Needs to be replaced.'''

Last Refresh Date

The date the AI Reference Dataset was last updated.

Preview

This will give you a count of how many rows will be returned by your filter, and a snapshot of the first 10 rows that will be generated by your prompt.

NOTE:

If you include auto-numbers (like Incident Number or Action Number) in your AI Reference Dataset, and they are included in the Suggestions, the system will render them as clickable hyperlinks in the suggestion dialog, so that users can click through to read more detail.

NOTE:

AI Reference Datasets are currently limited to 5 million "tokens" - which is a measure used by OpenAI's large language model, which approximates to about 4 text characters. In practice this limit equates to approximately 90,000 rows that include a summary and a detailed description.

If your AI Reference Dataset filter returns more than 5 million tokens, the remaining rows will be excluded form the AI Reference Dataset.

Field

Explanation

Name

The name for the Event.

We recommend something like "Refresh AI Reference Data Set - DATASETNAME", where DATASETNAME is the name of your AI Reference Dataset

Event Type

Should be set to "Refresh AI Reference Data Set". On Save, this field becomes read-only.

Run Time (HH:MM) - if set, event will run daily at this time

Define what time of day you want the AI Reference Dataset Refresh to occur.

You must enter the time in 24 hour format.

Field

Explanation

AI Reference Data Set to Refresh

Select your AI Reference Dataset

Start Date

Set this to the first date you want the event to run

Frequency in Days

Enter how many days you want between runs.
Enter 7 for weekly, enter 1 for daily, etc

Error Email Address

Email Address that will be notified of any errors that have occurred

Re-build entire AI Reference Data

Ticking this box will re-build the entire AI Reference Data on the next run, rather than just updating the data set with the record changes since the last run. This will take longer to run.