Every question asked of your assistants is recorded, along with the answer, the reader's rating, and the context around it: where they were, how long the answer took, how many passages it consulted.

Spend twenty minutes in that feed and you will discover the same thing everyone discovers. What people ask is reliably different from what you assumed they would ask.

A live feed, not a report

The Conversations tab is exactly that: new conversations matching your filters appear at the top in real time. Open it during a training session or the morning after a launch and you can watch the questions arrive.

Each entry carries a rating dot (green for helpful, red for unhelpful, none for unrated), the opening lines of the question, the document, a country flag, an AUTH or ANON marker, and a relative time. Badges add the rest: RX where an answer triggered medication validation, TRUNC where it hit the length limit, SUCCINCT for short-answer mode, HIDDEN where an admin removed a question from the public list.

Select one and the detail pane gives you the whole picture: the exact time, the reader's city and country, whether they were signed in, their rating and any feedback reasons, timing and token counts, how many passages and references the answer used, and then the question and response rendered as the reader saw them.

Conversations from every channel land in the same feed. Web chat, embedded chatbots, Telegram, all together.

Two practical notes before you go hunting. Search questions… searches only the conversations already loaded, so scroll or press Load more first. And there are no date-range or rating filters here, so for anything time-based you want the CSV export.

Where to start: the red dots

If you only have a few minutes a week, this is the queue.

A thumbs-down opens a form where readers pick what went wrong, and those reasons arrive here as tags. They are more diagnostic than a star rating, because each points somewhere specific:

  • Factually inaccurate usually means a gap or outdated content.
  • Not helpful often means the question was outside the document's scope at all.
  • Off topic points at retrieval matching the wrong sections.
  • Confusing points at source material that was unclear before the AI got near it.

A recurring question that keeps collecting red dots is almost always telling you the source document needs a section adding, followed by a retrain.

What gets flagged automatically

Several classifiers run without being asked.

Every conversation is checked for profanity and junk, and anything caught is marked banned with a recorded reason, kept out of the public recent-questions list, and left out of the group-wide feed.

Two things about that are worth stating plainly. Banning is fully automatic, with no manual ban or unban, and it can occasionally catch something legitimate, such as a bare number typed as an answer to a follow-up question. And nothing is deleted. The reader's experience is untouched apart from their question not being promoted publicly.

There is a scoping consequence people trip over: the group-wide tab hides automatically banned conversations, while per-document views show them with a BANNED badge. If you want to audit what the filter caught, go in through a document rather than the dashboard.

An adverse-events classifier does something similar for questions about side effects, drug interactions or harmful reactions, and there is a toggle to see only those. Because the public recent-questions list shows only cleared questions, a brand-new one takes a moment to appear there. That is classification finishing, not a fault.

Moderating the public list

Chat pages can show readers what other people recently asked, and you decide what appears there. Open a conversation, find the Recent Q's row, and flip the button between Visible and Hidden. Immediate, and reversible.

Banned and adverse-event conversations show a fixed hidden status instead and cannot be promoted, which is the correct order of precedence: an automatic safety flag should not be overridable by a click.

There is deliberately no delete button. If something must be permanently removed, such as a reader having typed personal information into a question, that goes through Support under Account. Deletion of the record is a decision with a human in it rather than a button in a busy feed.

Readers can also share a conversation by link. There is no separate moderation queue for those, because they are the same records, so everything above already applies to them.

Export, and the two quieter logs

For anything time-based, export. Filters carry through, exports over 5,000 conversations ask for confirmation, and very large ones cap at 100,000 rows.

The rows are unusually complete: 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 and adverse-event and banned flags, whether the question re-asked a recent one, a session identifier, the documents involved, and which model answered. Enough to do real analysis somewhere else rather than a summary you have to take on faith.

Two more feeds live in the document editor. Download Activity on the Attachments tab tracks who downloaded which file, when and from where. Gallery Activity does the same for lightbox views of images and video, with a most-viewed ranking, and counts a view only after a reader has lingered about a second, so scrolling past an image is not recorded as looking at it. Both export to CSV.

The debug console, for one bad answer

When a specific answer goes wrong and you want to know why, add ?debug=true to the chat page address and reload. A Debug tab appears on the right edge.

It shows the model used, response times, token usage, reranking, your custom prompt, and, most usefully, the complete set of instructions and excerpts that were sent to the AI, with a copy button. That last panel is the single best thing to attach when reporting an answer-quality problem, because it turns "the answer was wrong" into something reproducible.

A model selector lets you re-ask against a different model in your own session only, without changing anything for readers.

One caution, since it is easy to miss. This is a technical switch rather than a role-restricted feature. Anyone who adds the parameter can open it, and it reveals the full instructions sent to the AI, including your custom prompt. Treat a custom prompt as visible to a determined reader rather than as a secret.

The dashboard glance

Most popular ranks your five most-used documents by questions asked, over 30 days or all time, with an activity bar and a shortcut into each one's configuration. The dashboard home adds a recent-activity feed and a weekly panel with questions asked, conversations all time, and satisfaction.

Treat those as a glance rather than a tool. When you want to know what people were actually trying to learn, that is Document Intelligence.