Which dataset is your chatbot meant to use?

A chatbot or an agent works from a dataset, and a better version appears. Better is not the same as in force. After a while nobody remembers when the switch was made, or on what grounds.

What Smeldr does well is keep the decision to switch: who ratified it, what it replaced, and what it rests on. It does not store the dataset, and it does not check which one is loaded.

To keep the switch on record, you could write a Decision such as "The support chatbot uses dataset 3.3" and put the version in its text. It may also help to record the evaluation's conclusion as a Decision of its own, "Dataset 3.3 answers more support questions correctly than 3.2", and to record that the first rests on the second. The evaluation itself stays where it is. If "Dataset 3.3 answers more support questions correctly than 3.2" is later replaced, Smeldr flags "The support chatbot uses dataset 3.3", and someone checks whether 3.3 still fits.

Where the dataset lives decides how much more you can do.

If it lives in a git repository, you could pin the dataset file to a commit in the Decision, with a hash of its content, and place the Decision in an Area that covers the file. A pull request that changes the dataset then lists the Decisions in force and can wait for the Area's approver.

If it lives somewhere else, such as a shared drive, you could put the link in the Decision's external reference. Smeldr does not follow the link, so the version in the Decision's text is what people read.

Going back to 3.2 is a new Decision, and the history keeps both. Whether an agent or a person proposes the switch, a person ratifies it. Beyond that, how you name versions and where you keep the data is up to you.

What is described here is what Smeldr does. The way to use it is a suggestion. Read the other stories in How Smeldr works in practice.