Conversations: your agent's inbox
Everything visitors ask your agent lands in Conversations. Here you read back what was asked, what the answer was and how well that answer scored. Two tabs keep things clean: Integrations shows the real conversations from your website and other channels, Playground the conversations your own team had while testing.
The overview: seeing what needs attention
Every conversation sits in the list with its start question, the number of messages, the three answer scores and the visitor's feedback. The Review column does the groundwork for you: conversations where the answer fell short automatically get a label such as "Missing content" or "Weak answer". That way you start with the conversations where there is something to gain, instead of at the top of the list.
Looking for one specific conversation? The search bar looks at the content as well as the title: type a word that appeared in the question or the answer and you will find the conversation. If nothing matches, the list says so, instead of suggesting you have no conversations yet.
The filters cut the list down further: by period, by theme or topic, by feedback (positive, negative or none at all), by interrupted answers or the content policy, by what still needs review, or by conversations your team flagged.
One conversation up close
Open a conversation and you see the full story per question: the generated answer, which sources were cited and which were only consulted, and the evaluation scores for faithfulness, relevance and source relevance. The side column holds the practical facts: which agent answered, how fast, and from which page the question came. Want to know exactly which steps the agent took? Open the trace.
If an answer did not finish, the conversation says why. A visitor hitting stop halfway through shows as "stopped by the user"; if something broke technically, the conversation says that instead. That saves you digging: the first is normal behavior, the second is worth a look.
Reviewing together
The inbox is built to go through with your team. Notice something? Flag the answer so a colleague can find it back. Below every conversation you can leave internal comments; visitors never see those. If an automatic diagnosis is off the mark, you dismiss it with one click. And if a conversation sits under the wrong topic, an Editor changes the category so the analysis stays clean.
How to run such a review round, from giving feedback to reading the scores, is covered in Testing and evaluating.
Visitor feedback
Visitors can rate an answer in the widget with a thumbs up or down, and add a comment. That feedback lands with the conversation, so you not only see that someone was unhappy, but also about what. The Negative feedback filter puts those conversations at the top of your work list.
Who can do what
Everyone on your team with the Annotator role or higher can read, review and flag the inbox. Exporting or deleting conversations and changing categories starts at the Editor role. That way a large group can read along without anything happening to the data. More about the roles in Managing users and roles.
How long conversations are kept is up to you, through the retention period of your environment; see Hosting and data storage. Want to analyze the inbox outside vragen.ai? Export it as a spreadsheet: Exporting and analyzing the inbox.