A AtomicX Chat
Help Sign in

Updated October 30, 2026

View conversations

Overview

The Conversations page is where you review the chats visitors have had with your assistant. The list of conversations is on the left. The conversation you select opens on the right, with its AI analysis, the full transcript, the tools the AI called, visitor feedback, and the prompts the AI followed.

Open it from Chat → Conversations. Owners, Admins, and Users can all use this page. On a narrow screen the list and the details are stacked, with the details below the list.

Each row in the list shows:

Column What it shows
Conversation The visitor's first message, or (untitled) if they haven't written yet. Underneath: the widget or workflow, the status (Active, Completed, Abandoned, Handed Off, or Closed), and Human once a team member has taken over
User The visitor's name, if they gave one (hidden on very narrow screens)
Started The time for today's conversations, otherwise the date, in your account's time zone
Analysis Whether the conversation has an AI analysis (see below)

The Analysis column shows one of:

  • A green ✓: the analysis is ready (hover it for "Analysis completed").
  • An amber Queued, Running, or Failed badge: the analysis is waiting, being written, or couldn't be written.
  • A round wand button (Generate analysis): no analysis yet. Click it to request one without opening the conversation.

Steps

Find a conversation

  1. Go to Chat → Conversations.
  2. Use the filters above the list:
    • Search: matches the visitor's name, their email, or any message text.
    • Status: pick one status or All Statuses.
    • Workflow: show only conversations from one workflow.
    • Date Range: choose Today, Yesterday, Last 7 days, Last 30 days, or All time.
  3. Click Apply filters. Click Clear to reset the filters.
  4. The list shows 25 conversations at a time. Use Previous and Next below it. The text between them shows the page you're on and the total number of matching conversations, for example Page 2 of 4 (87).

The list shows Today by default. If a conversation you expect is missing, widen the Date Range.

Export conversations

Choose your date range and any search, status, or workflow filters, then click Export CSV or Export JSON. The download includes every matching conversation across all pages. For example, Last 30 days exports conversations started during the last 30 days. Export uses the selected filters even if you haven't clicked Apply filters yet.

The files contain only message text in order, labelled User, Assistant, or Admin. CSV has one transcript per row. JSON contains an array of conversations, each an array of messages with speaker and content. IDs, timestamps, visitor profile details, AI analysis, internal system messages, and message metadata are excluded. Information visitors or staff wrote in the messages remains in the text.

Review a conversation

Click a row. It's highlighted, and the details open on the right. When the assistant was interrupted while answering (for example by a server restart), an Interrupted replies box appears above the details (see Interrupted replies below). The details have these sections:

  1. Header: the visitor, your account and the widget or workflow, when it started (in your time zone, plus the visitor's own time when it differs) and how long it lasted, its status, and Bot or Human.

  2. Analysis: the AI analysis (see below).

  3. Messages: every message in order, with the sender's name and the time. Replies show the same product cards (photo, price and link) and page links that the visitor saw below them; click a card to open the page. On replies written by an AI Chat node, click Show AI Details to see:

    • Whether the Regular or Premium model wrote it, the temperature, and the tokens used.
    • The Agent that wrote the reply and its version.
    • Skills used and Tools used (MCP tools). A red chip with a cross means that call failed.
    • The System Prompt, the AI Reasoning, and the Knowledge Base Context it used, with its Retrieval similarity. This percentage shows how closely the retrieved text matched the visitor's question. It is not a measure of whether the answer is correct: a reply can have high similarity and still be wrong, so read the context and the reply together.
    • The knowledge base search outcome, when the reply used a knowledge base: Evidence found, Degraded ranking (the search used its slower backup method instead of the search index, so the ordering may be less precise), No relevant evidence (nothing in the knowledge base matched closely enough), Only stale evidence (nothing current matched, but website pages the knowledge base no longer trusts would have: pages whose latest scan failed or that haven't been scanned for over 30 days; those pages are never used in answers, so the visitor gets the same reply as for no relevant evidence. Rescan the website to bring them back), or Retrieval unavailable (the search couldn't run).

    Show AI Details also appears on the fixed reply a visitor gets when the knowledge base search was unavailable or the answer was withheld for lack of evidence (including when only stale evidence matched), so you can see which one happened. See Create a knowledge base.

  4. Tool executions: every skill and tool call in the conversation, as name · Skill or Tool · succeeded or failed, with the time of the reply. None. means the AI didn't call any.

  5. Feedback: the thumbs up and thumbs down visitors gave, the reply they rated, and when. None. means no one rated a reply.

  6. Prompts in effect: each distinct system prompt the AI used in this conversation, with the agent name and version. None captured. means the replies came from before prompts were recorded, or from nodes that don't use the AI.

Private prompts, reasoning, knowledge-base context, tool diagnostics, and operator notes are available only through authorized staff views. The visitor widget receives the reply, display sources, forms, options, and workflow presentation; it does not receive the stored AI diagnostic metadata.

Interrupted replies

If the server stops while the assistant is answering a visitor message, the reply may never reach the visitor. The Interrupted replies box lists each such message with the time it arrived and what happened to it:

  • Interrupted. It is answered automatically if the visitor's chat resends it and nothing was sent to the AI yet. The chat widget resends a message whose connection dropped. If nothing reached the AI before the interruption, the message is answered again automatically, once, and it disappears from this box.
  • Not retried: the request had already been sent to the AI. or Not retried: AI usage had already been recorded. The AI may already have written (and charged for) an answer that was lost, so it is never generated again automatically.
  • Not retried: a reply, booking, integration or workflow call had already been recorded. Part of the answer already took effect, so repeating it could book, send or charge twice.
  • Not retried: a newer message or a team member took over first. The conversation moved on before the reply finished.
  • Not retried: automatic retries were used up. The message was interrupted again after being answered again automatically.

For any message marked Not retried, read the messages around that time. If the visitor still needs an answer, click Take over and reply from Live monitor. Check any external system (calendar, CRM, store) before repeating an action.

Get an AI analysis

You usually don't need to ask. Once a conversation has ended, or has been idle for 30 minutes, it is analyzed automatically (up to 100 conversations per account per day), and its findings feed Improve your chat. Use the steps below for a conversation that is still going, or to run a fresh analysis.

  1. Select the conversation and find the Analysis section. No analysis available. means none has been requested yet.
  2. Click Generate analysis (or the wand button in the list's Analysis column).
  3. The analysis is queued and written in the background, so you can keep working. The badge in the list and the line under Analysis change from Queued to Running, and then to the finished analysis. The page updates by itself.
  4. When it's done you'll see:
    • A line such as Completed · revision 2 · analyzed by AtomicX: the status and which analysis this is for the conversation.
    • A short summary.
    • Scores out of 5 for Relevance, Helpfulness, Tone, Resolution, and Error handling: green for 4 or 5, amber for 2 or 3, red below 2.
    • What went well (✔), What went wrong (✖), and Suggestions, each naming what to change (for example the knowledge base or the workflow) and how.
  5. Click Regenerate analysis at any time for a fresh one. Each new analysis gets the next revision number; the page always shows the latest. Regenerating also refreshes the issue for that conversation in Improvements.

If writing the analysis fails, it's retried automatically, up to 3 attempts. After that the status shows Failed with the reason in red. Click Regenerate analysis to try again.

Take over or archive

Select the conversation, then use the buttons at the top right of its details:

  • Take over appears on active conversations that the AI is still handling. It hands the conversation to you and stops the AI from replying. You then reply from Live monitor.
  • Archive closes the conversation and sets its status to Closed.

Conversation language

The selected conversation's details show Chat language. The chat starts in the website's language and switches to the language the visitor writes in; the current language is saved with the conversation. The transcript retains each message in its original language, including human replies; switching language does not rewrite earlier messages. See Chat languages and translations.

Tips

  • The list and the open conversation update by themselves when visitors write or a conversation changes. Turn on Auto-refresh (30s) as a fallback, or click Refresh to update now.
  • Request analyses straight from the list: click the wand on several rows, then come back when they show a green ✓.
  • Every Generate analysis and Regenerate analysis is recorded in your account's audit log with who requested it.
  • Use the Suggestions to improve your workflows, agents, and knowledge bases.
  • Filter by Handed Off to review the conversations your team handled.
  • If a reply looks wrong, check Tool executions and Prompts in effect, or Show AI Details on the reply, to see whether a call failed or the AI followed different instructions than you expected.
An unhandled error has occurred. Reload ×