Read a conversation

What it is

The conversation detail page is the full record of one session: a header with its metadata, session statistics, a visitor card when your site identified the person, and the complete transcript in time order.

It is where you go to answer "what exactly did we tell this person, and where did that come from?" — with enough context attached that you can act on the answer rather than just wince at it.

When you would use it

• A customer quotes something your chatbot said and you need the original.
• A thread showed up with a thumbs-down or a low confidence score.
• You are working a review queue and this is the next item.

Where to find it

Open Dashboard → Chatbots → your chatbot → Conversations, then use View details in a row's actions menu.

Clicking the row body does something different — it opens the session in the Test Chatbot drawer so you can continue it. Use View details when you want to read.

Steps

1. Open Conversations and find the thread.
2. Choose View details from the row's actions menu.
3. Read the header. Session status, channel badge, language, and a copyable session ID.
4. Scan the statistics. Messages, answered rate, average confidence, average response time.
5. Check the visitor card, if there is one — it only appears when your own site passed the person's details.
6. Read the transcript. Long messages have a Show more control.
7. Open the sources on any answer you want to verify.
8. Toggle Reviewed when you are done with the thread.

The header

• Session status — Active or Completed. Active means the session was still open at last contact; it is not a to-do state.
• Channel badge — Website widget, Shared link, Test chat, or a messaging app. See The conversations list.
• Language — the language of the conversation.
• Session ID with a copy button. This is the reference to quote in an internal ticket or to support; it identifies the thread without pasting the customer's words anywhere.
• Reviewed toggle and Open full chat.

The four statistics, and how to read them

Statistic	What it counts	What it tells you
Messages	Total messages exchanged	Length. Very long threads often mean the chatbot never quite landed the answer
Answered rate	Share of messages that got an answer	The clearest single quality number for this thread
Avg confidence	Average confidence across knowledge-grounded answers. Greetings and small talk are excluded	How well-supported the substantive answers were
Avg response time	Average time to generate a reply. Instant cached replies are excluded	Whether this conversation was slow, honestly measured

Two exclusions worth knowing, because they stop the numbers lying to you. Confidence ignores "hi" and "thanks", which would otherwise drag the average around meaninglessly. Response time ignores replies served from cache, which would otherwise make a chatbot look faster than it is.

The transcript

Messages appear in time order, styled like the live chat — the chatbot's replies render markdown the same way a visitor saw them, and the direction follows the conversation's language.

On the chatbot's messages you may see:

• A status chip — Answered, Partial answer, or No answer, with the confidence percentage.
• A sources control when knowledge was used. Opening it lists each source with its match scores, the search query that was actually used, and a link to the original page where one exists. Weak matches are labelled as such. See Sources and citations.
• A feedback chip if the visitor rated the reply. See Conversation feedback.
• A cached label when the reply was served from cache rather than generated fresh.
• "No response was generated for this message" where a turn produced nothing.

Long transcripts are paginated. If you are hunting for one phrase across many threads, search from the list instead — it searches message content, and you can scope the search to visitor messages or chatbot replies.

What "Open full chat" does

It reopens this session in the Test Chatbot drawer so you can keep asking as the owner.

Be clear about what that is and is not:

• It is your session now. The visitor does not see your messages and is not notified. This is not live chat.
• It spends message credits per turn, like any other answer, at the model that conversation uses.
• It is useful for reproducing a bad answer with the exact conversation history in place — which often explains a reply that looked bizarre in isolation.

Reviewed

Reviewed means a human on your team has looked at this. It is invisible to the visitor and changes nothing about the chatbot. It is entirely separate from the visitor's thumbs up or down: the thumb is their opinion of the answer, the flag is your record of having dealt with it.

What you will see

Header, statistics, visitor card (only when identified), then the transcript in a card. If the session cannot be found — a stale link, or a thread removed with its chatbot — you get a clear not-found state with a way back to the list.

Limits and plan notes

You need permission to view conversations. Toggling review needs permission to manage them.

There is no per-thread export or download. If you need the text elsewhere, select and copy it from the transcript. See Export and privacy.

Treat every transcript as real customer data. If your deployment has a retention window, this thread will be deleted automatically when it expires.

Common problems

The transcript is blank.

Either you opened a stale link for a session that no longer exists, or you are on a transcript page beyond the last one with messages. Go back to the list and reopen it.

There is no sources control on an answer.

That turn used no knowledge — it was a greeting, a fallback, or a question nothing matched. Sources only exist where retrieval actually found something.

A source says it is no longer available.

The knowledge it referenced was updated or removed after the conversation happened. The transcript is a historical record; the knowledge base moved on.

There is no visitor card.

Your site did not pass an identity, which is the normal case for an anonymous website visitor. To get names, see Name visitors on the widget.

Can the visitor tell I marked it reviewed, or that I read it?

No. Both are invisible to them.

Can I reply to the visitor from here?

No. Agentency stores conversations but does not provide a console for typing back as the chatbot. If a person needs to get involved, build that path — see Handoff to a human.