You’ve trained, configured, and tested your AI. You’ve set it live, people have been using it, and conversations are coming in. It is now time to review and analyze these conversations so you can improve the answers whether it means adding or updating training sources or adjusting the configuration.
Conversations list view
In the Conversations tab of your Unless dashboard, you will find a list view of all conversations from the last 30 days. For each conversation, you will also see the date and time of the last message, the number of messages within the conversation, and the rating it received from the user.
Filtering options
You can then filter the conversations by rating or number of responses to make it easier to review. Reviewing these conversations lets you know if the AI is responding in the way you would expect. It also allows you to see if you need to add more training sources or update (or remove) existing ones.
Ps. You may have noticed the Negatively rated conversations block in the AI zone which will bring you to a filtered version of this Conversations list as can be seen in the screenshot below.
Additionally, going through these conversations can provide insight into what your users and customers are asking and what their needs are. This can inform the content you create in the future as well as any product decisions you make whether it be additional articles, FAQs, user guides, onboarding procedures, and more.
Tip: Longer conversations tend to provide more insight so we recommend filtering for at least 2 or more responses though you could always opt for a higher number too.
Conversation details page
Clicking Details for a conversation will bring you to the details view which consists of all messages within that conversation along with any thumbs up/down ratings attached to responses.
Viewing a source
It is possible to take a quick glance at the sources used for a response. And you can click each listed source to open them up and review the content within that source. If there are any inconsistencies between the AI answer and what you’d expect, this is a quick way to troubleshoot. Any wrong or old information source can be removed and updated versions can be added.
Flagging a response
If you identify any technical issues, you can also flag a response, and fill in the details about what happened and what you would expect to happen instead. We can then check the conversation, sources, configuration, and logs to see if something needs to be adjusted, if there’s a bug that needs fixing, etc.
When training and configuring your AI, make sure to also take a look at the quality control center. You can read more about that in this article.