Every question asked in your chatbots is recorded, with the answer, the reader's rating, and context: where the reader was, how long the answer took, how many passages it consulted. The dashboard gives you a live view of that activity, and it is where you learn what people actually ask — reliably different from what you assumed.
Where the conversations live
Four entry points open the same viewer at different scopes.
| Entry point | Scope |
|---|---|
| Conversations tab, dashboard sidebar, System group | Everything you may see |
| Documents tab, row action menu (⋯), Conversations | One document |
| Document configuration, Conversations tab, Learning group | One document |
| Collection editor, Conversations tab, Learning group | One collection, chats spanning its documents included |
Owner admins see documents in their own owner groups; super admins see the whole platform, with an extra filter for one owner group (users, roles and invitations).
One scoping detail matters more than it looks: the group-wide tab hides conversations that were automatically banned, while per-document views include them with a BANNED badge. To audit what the filter caught, go in through a document.
Reading the feed
The tab is a live feed titled Live Conversations: new conversations matching your filters appear at the top in real time, and the toolbar shows the total, for example 1,204 conversations.
Four controls narrow it. An owner filter (All owners) for super admins with more than one owner group. A document filter (All documents) in the group-wide view. An Adverse Events toggle for conversations flagged as asking about side effects, drug interactions, or harmful reactions. And Search questions…, which has a catch worth knowing before you go hunting: it searches only the conversations already loaded, so scroll or use Load more first. There are no date-range or rating filters either; for those, export.
Each entry carries a rating dot (green helpful, red unhelpful, none unrated), the first two lines of the question, the document title, a country flag, an AUTH or ANON marker, and a relative time such as "3h ago". Badges add detail: RX for medication recommendations that triggered validation warnings, TRUNC for answers cut off at the maximum length, SUCCINCT for short-answer mode, HIDDEN for a question an admin hid from the recent-questions list, and BANNED in per-document views.
Selecting an entry opens a detail pane showing the exact time; the document, with a link control offering Copy Link and Open in New Tab; the reader's city, region and country; Authenticated or Anonymous; the Rating and any Feedback reasons; timing, token counts, and passages and references used; the Response Mode (Succinct or Chatty); medication and truncation flags; and a Recent Q's row with a visibility control. Below sit the full question and response, rendered as the reader saw them.
Conversations from every channel land in the same feed — web chat, embedded chatbots, Telegram — and the viewer does not label which.
Ratings, feedback, and automatic flags
Readers rate answers with a thumbs up or down, shown here as Helpful and Unhelpful. A thumbs-down lets them pick reasons, which appear as tags in the Feedback row: Factually inaccurate, Too short, Not helpful, Too slow, Off topic, Confusing, and Missing references. Each points somewhere: inaccuracy at gaps or outdated content, Not helpful at questions outside the document's scope, Off topic at retrieval matching the wrong sections, Confusing at unclear source material.
Several classifiers also run automatically. Every conversation is checked for profanity and junk content, and flagged ones are marked banned with a recorded reason, kept out of the public recent-questions list and excluded from the group-wide feed. Banning is fully automatic — there is no manual ban or unban — and can occasionally catch a short legitimate reply, such as a bare number answering a follow-up. Nothing is deleted, and the reader's experience is unaffected apart from the question not being promoted publicly. The adverse-events classifier likewise keeps flagged questions out of the public list; since that list shows only cleared questions, a brand-new one does not appear until classification finishes — a short delay, not an error. Medication detection marks qualifying answers RX.
Moderation and exports
Chat pages can show readers a list of recently asked questions, and you control what appears there: open the conversation, find the Recent Q's row, and select the button, which reads Visible or Hidden. The change is immediate and reversible. Banned or adverse-event conversations show a fixed Hidden (banned) or Hidden (adverse event) status instead and cannot be made public. Super admins get two shortcuts in the chat interface itself: a remove control on each recent-questions entry, and pinning.
There is no delete button in the viewer. If a conversation must be permanently removed — a reader typed personal information into it, say — contact the team through Support under Account.
Readers can also share a conversation by link, producing a read-only copy. There is no separate moderation queue for those; they are the same records, so the tools above apply. One shared from inside a collection opens that collection with the thread loaded.
For anything time-based, export. Your filters carry through; select Export CSV in the toolbar. Exports over 5,000 conversations ask for confirmation, and very large ones are capped at 100,000 rows. Each row includes the question and full response, date and time, sign-in status, location, rating and feedback reasons, response time, processing volume, passages and references used, the medication, adverse-event and banned flags, whether the question re-asked a recent one, a session identifier, the documents involved, and the model that answered.
Two other feeds live in the document editor, also part of the paid analytics features. Download Activity, on the Attachments tab, tracks attachment downloads with Downloads, Unique Users, and Unique IPs tiles and a history table. Gallery Activity, the third sub-tab of Images, does the same for lightbox views of images and video, adding a Most viewed ranking; a view counts only after a reader lingers about a second. Both have Export CSV.
Analytics panels and the debug console
On the dashboard's Documents section, Most popular ranks the five most-used documents by questions asked, switchable between 30d and All time. Rows show rank and title, question count, an activity bar, last activity, an open-chat button, and a gear opening its configuration. The panel follows the Workspace strip, so it always ranks one organization's documents, super admins included. The dashboard home adds a Recent activity feed and a This week panel with Questions asked, Conversations · all time, and Satisfaction — a glance, not a tool; for analysis use Document Intelligence.
One thing is explicitly absent. Docutrain records deeper diagnostic data about how questions are interpreted behind the scenes — when one is automatically rephrased to find a better answer, for instance — but keeps it internally rather than surfacing it in the dashboard.
For a single bad answer there is the Debug Console. Add ?debug=true to the chat page address and reload, and a hidden Debug tab appears on the right edge. It is a technical switch, not a role-restricted feature — anyone who adds the parameter can open it, and it reveals the full instructions sent to the AI, so treat it accordingly. Its sections are Model Information, Response Times, Token Usage, Statistics, Reranking, Custom Prompt, and Prompt — the complete instructions and excerpts sent to the AI, with a Copy button, and the most useful thing to attach when reporting an answer-quality issue. A Select Model control re-asks against a different model in your own session only.
A few minutes in the Conversations tab each week is enough. Start with the red dots: thumbs-down conversations with reasons are the highest-value queue. And when a recurring question keeps getting poor answers, the source document usually needs a section added, then a retrain.