Creating a Dataset of Past Conversations

A dataset decides which past conversations a historical test evaluates. You describe the sample you want, by date, channel, length and what was recorded about the conversations, and the platform tells you how many it found before you save.

Running a test against one is covered in Creating a Test. What the module measures overall is on AI Testing.

📝 Note: Only conversations held against a published version of the project can be sampled. A project that has never been published has nothing to draw on, however much traffic its Conversations page shows, because that traffic came from the draft designer.

The Datasets page

Datasets sits under AI Testing in the main navigation. The list shows each dataset with its Name, Description, Number of Targeted Conversations and Created Date.

+ New Dataset opens the form. Clicking a dataset's name reopens that same form as Edit Dataset, so there is no read-only view: opening one is editing it.

Defining the sample

Dataset Name and Description are both required. The description is for the people who will pick this dataset from a list later, so say what kind of conversation is in it.

The Sampling section is the sample itself.

Fixed Dates or Relative

Fixed Dates asks for a Start Date and an End Date, and that window never moves.

Relative replaces the two dates with a Period: Today, Yesterday, Last Week, Last Month or Last 3 Months. The window is worked out again each time the test runs, so a dataset set to Last Week always means the week before the run rather than the week before you saved it.

Pick Relative for a dataset you intend to run repeatedly and Fixed Dates for a one-off measurement of a period you want to keep coming back to.

Narrowing it further

Two optional sections sit under the sampling fields, and both work the same way: pick what to filter on, pick an operator, then pick the values.

AI Insight Filters filter on what a model extracted from the conversation after it ended, so a dataset can be narrowed to the conversations where Churn Risk came out High. Custom Attribute Filters filter on what a flow recorded itself, such as whether a payment link was created.

This is what turns a dataset from "some conversations from last month" into a question worth asking. Evaluating a random sample tells you how the project does on average. Evaluating the conversations that carried a high churn risk tells you how it does when it matters.

Both sections take more than one row, and each row can hold several values.

Confirming what was found

Save does not save straight away. It first reports what the sample actually turned up.

The dialog compares what you asked for against what exists: how many conversations you requested and how many were found, and the same for live agent handovers, each marked Met when the sample can be filled. A sentence underneath states what the dataset will hold, and Yes saves it.

This is the step that catches an empty dataset before it becomes a test that evaluates nothing. If the numbers come back short, widen the date range, drop a filter, or lower the minimum duration, and try again.

Limits