> ## Documentation Index > Fetch the complete documentation index at: https://cntillion.com/docs/llms.txt > Use this file to discover all available pages before exploring further. # Conversations > Review transcripts, branch history and analysis across your agents **Conversations** in the workspace sidebar lists every conversation your agents have had, in one place: website chat, voice calls, and WhatsApp, email, SMS, Messenger, Instagram and Slack threads. The agent's **Conversations** page uses the same view, filtered to that agent. Both pages open directly on the conversation list, with search and filters above it. Open a conversation to review its transcript, client data and analysis. **Open logs** in the conversation panel provides the reply, takeover and pause controls for live conversations. ## The list The most recent conversations come first, by their last activity. Each row shows: | Column | What it shows | | - | - | | Agent | The agent that handled the conversation. This column appears in workspace history; an agent's own history leaves it out. | | Branch | The branch and version assigned when the conversation began, or "Draft preview" for Playground. Older conversations without a recorded assignment show "Not recorded". | | Title | A few words naming what the conversation was about, such as "Move appointment to Friday" or "Reached voicemail", written from the transcript in the conversation's language. The title appears within a few minutes of the conversation ending or going quiet for five minutes, and is written again when new messages arrive. Writing it uses a small amount of AI credits. Until then the row shows the customer's first message, or "Untitled conversation" when the customer hasn't written or said anything. The channel, contact and conversation status appear beneath it. An anonymous visitor reads "Website visitor", and Playground conversations show no contact. | | Date | The date and time of the last activity, such as "Oct 21, 6:02 PM". Dates in a previous year include the year. Hover for the full local date and time. | | Duration | How long a voice call lasted. Other conversations show "—". | | Messages | How many messages the conversation holds. | | Evaluation | Failure, Pending, Success, Scored or Unknown, from the agent's [evaluation criteria](https://cntillion.com/docs/agents/releases-and-workflows.md#evaluate-real-conversations). Blank when there are no results. | The list loads 50 conversations at a time. **Load more** adds the next 50. ## Filter and search * The date range is always above the list. It starts at **Last 30 days**; choose **All time** to include older conversations. * **Filters** opens every other filter, grouped by what it reads. Each filter you turn on becomes a chip beside the date range. Click a chip to change it, or its **×** to remove it. Filters of different kinds narrow the list together, and choosing several values in one filter matches any of them. * In workspace history, branches, versions, workflow steps and evaluation criteria belong to one agent. Changing the **Agent** or **Tags** filter removes any of these that no longer apply. * When nothing is filtered and the last 30 days hold no conversations, the page reads "No conversations in the last 30 days". **Show all time** lists the older ones. * **Keyword** searches a contact's name, phone number or email address, both speakers' message text, and an exact conversation ID. Every unquoted word must match somewhere in the conversation. Use `"double quotes"` for an exact phrase. Characters such as `%` and `_` match themselves. * Switch to **Smart search** to describe the meaning you are looking for, such as “customers unhappy about a damaged delivery”, then press Enter. Matches show the closest transcript excerpt and are ordered by relevance. Enable Smart search in the agent's [Evaluations](https://cntillion.com/docs/agents/releases-and-workflows.md#sentiment-topics-and-search) and analyze completed conversations before searching. An empty index or an unavailable embedding service shows a message; you can switch to Keyword. Agent, tag, channel, branch and date filters apply in both modes. The other filters need Keyword search. * The page's address keeps the filters, query and search mode, so you can bookmark or share it. * From a node in Workflow analytics, **View conversations that entered this node** opens history with a **Node entered** filter and the analytics date range. Clear its chip to return to all nodes. * When nothing matches, the page reads "No conversations match these filters" and **Clear filters** returns to the full list. ### Available filters | Group | Filter | What it matches | | - | - | - | | Agent | Agent, Tags | Conversations with the chosen agents, or agents with the chosen tags. These appear in workspace history; an agent's own history is already narrowed to it. | | Agent | Branch | The branch assigned when the conversation began. Includes archived branches. | | Agent | Version | A published version of the chosen branch. Choose a branch first. Draft previews are left out. | | Agent | Node entered | Conversations that reached a step of the agent's workflow. In workspace history, choose one agent first. | | Conversation | Duration | Voice calls of that length, from a preset or your own limits in seconds. Other conversations have no duration, so they are left out. | | Conversation | Channel | Where the conversation took place. | | Conversation | Modality | **Voice** for calls, **Text** for everything else. | | Conversation | Environment | **Live** for real conversations, **Draft preview** for Playground and Preview. | | Conversation | Status | Ongoing, Paused, Waiting, Taken over or Ended, as the row shows it. | | Evaluation | Evaluation | Failure, Pending, Success, Scored or Unknown, as the row shows it. | | Evaluation | Evaluation criteria | The latest result of one criterion. A criterion being evaluated again matches once the new result is in. | | Evaluation | Sentiment | Positive, Neutral or Negative, as the conversation's review shows it. | | Evaluation | Frustration | The percentage from [sentiment analysis](https://cntillion.com/docs/agents/releases-and-workflows.md#sentiment-topics-and-search). Conversations that were not analyzed are left out. | | Evaluation | Rating | The customer's satisfaction survey, from 1 to 5. A thumbs up counts as 5 and a thumbs down as 1. Only answered surveys match. | | Evaluation | Comments | Answered satisfaction surveys with a written comment. | | Tools | Tools | Conversations that ran a workspace tool, or where a run of it failed. Built-in tools are not listed. | Branch and Version filters use the version pinned to the conversation, so publishing a new version or changing traffic does not change past assignments. A conversation without a recorded assignment appears only when no branch is chosen and reads **Not recorded**. Older conversations keep their original history; starting version control does not assign them a version retroactively. Playground and Preview conversations record the branch being previewed and show **Draft preview** because they use unpublished settings. ![Agent conversation history with search, the date range and the Filters button above two demo website conversations](https://mintcdn.com/centillion/8mVwbMm9sd1il7gC/images/workspaces/conversations-02-draft-preview.webp?fit=max\&auto=format\&n=8mVwbMm9sd1il7gC\&q=85\&s=f1fc91e8654dc649f681ac45634a2f2c) An agent's conversation history uses the shared search and filters, with a generated title and the branch, channel and evaluation details in each row. ## Open a conversation Click a row to open the conversation panel. The transcript is on the left, with four tabs on the right. On smaller screens, use **Transcript** and **Details** to switch panels: ![Demo refund conversation beside its summary and metadata, showing positive sentiment and a 100 percent conversation success score](https://mintcdn.com/centillion/umb2_NXHi6EGeOFB/images/workspaces/conversations-01-review.webp?fit=max\&auto=format\&n=umb2_NXHi6EGeOFB\&q=85\&s=ef10dc2aa63dccd83b8d106ccc187a9f) Conversation review with a demo chat transcript, summary, sentiment, frustration, success score and topics. * **Overview** opens with the summary and metadata. Metadata shows the conversation ID, date, agent, branch, version, status, sentiment, environment, duration and message count. When analysis is available, it also shows customer frustration, conversation success and topics. Sentiment updates when analysis finishes; pending or failed work has an explicit status. A rerun keeps the previous result visible while it runs. Voice calls also show their recorded user ID. Missing values are labelled as such. Branch and version links beside the agent open their release history. Completed voice calls generate a summary automatically. If no summary has been saved, **Generate summary** creates and saves one using the agent's model and AI credits. This requires permission to manage agents. * **Client data** shows what the conversation ran with. **Client overrides** lists the call's language, whether it was text only, and for a preview whether its tools were mocked. **Dynamic variables** lists each variable the conversation set, from initiation data, preview variables, tool results and the workflow, at its last value. Booleans read **true** or **false**, an empty value reads **Empty string**, and a missing one **null**. The copy button beside a value copies it. Sensitive workflow variables and `secret__` initiation variables show their mask, never the value. See [Personalization](https://cntillion.com/docs/agents/personalization.md) for the initialization response and override permissions. * **Analysis** opens with **Metadata**: how many times the conversation was analyzed, the AI credits all those runs used, and the credits of the latest run, including extraction and sentiment jobs. **Conversation success** averages the known binary results and normalized numeric scores, with the number of scored criteria shown alongside it. **Results** lists each criterion's latest verdict or numeric score and rationale under its **Conversation** or **Agent** scope. **Group by** switches between **Scope** and **Result**, and **Result** filters to the results you pick. **Re-run analysis** uses the branch's currently enabled criteria, data collection, sentiment, topics and search indexing. **Rerun**, beside a criterion when you hover it, evaluates that criterion alone. Both use AI credits and require permission to manage agents. A job whose first try fails is tried twice more in the background, and its result appears when it finishes. Collected data includes each typed value and its evidence; a missing value is distinct from false or zero. Sentiment shows its rationale, frustration estimate and user-turn trajectory. Topic badges summarize the customer's intent. * **Workflow** shows the workflow the conversation ran on: a preview's draft, or the version it was assigned. Steps it entered and transitions it took carry how many times, the rest are dimmed, and the summary at the top reads how many steps it visited and the one it ended on. Zoom with the controls at the bottom left. Select a step or a transition to read its settings, read-only, as the workflow builder shows them. A conversation from before version control shows the agent's current workflow and says so. ![Demo transcript with a Returns policy source link beside analysis showing 100 percent success across two scored criteria and their rationales](https://mintcdn.com/centillion/sWjfDtZ1gwwoAPXw/images/workspaces/conversations-05-success.webp?fit=max\&auto=format\&n=sWjfDtZ1gwwoAPXw\&q=85\&s=c030ff9a6d88e13e1e1dfd6ff9f8809f) Conversation success combines binary and numeric results. Knowledge sources appear beneath the replies that used them. ![Demo conversation analysis showing positive sentiment, a rising trajectory, a collected refund issue type and topic labels](https://mintcdn.com/centillion/sWjfDtZ1gwwoAPXw/images/workspaces/conversations-04-analysis.webp?fit=max\&auto=format\&n=sWjfDtZ1gwwoAPXw\&q=85\&s=0cde75e69cce8c57e4d76f21874ea7ad) Scroll the Analysis tab to review sentiment, its user-turn trajectory, collected data and topics. Use **Customize conversation details**, the sliders button at the top, to choose which Overview sections appear. **Summary**, **Metadata** and **Metrics over time** start checked. **Additional details** starts unchecked; turn it on to see the contact, channel, country, last activity and any linked ticket. For voice calls it also shows language, model, direction and call result. **Reset** restores these defaults. Drag the divider to resize the transcript and details. Focus the divider and use ← and → for keyboard resizing, or double-click it to reset the widths. The arrows at the top, or the ↑ and ↓ keys, move through the list. Expand the view for more space. The close button returns to the list where you left it. **Open logs** opens this conversation's operational controls. The conversation ID is part of the address, and so is the open tab, so refreshing, bookmarking or sharing it opens the same panel on the same tab. Workspace links use `/dashboard//conversations/`; agent links use `/dashboard//agent//activity/`. The tab is `?conversationTab=client-data`, `analysis` or `workflow`; Overview has none. Moving to the next conversation keeps the tab, and closing the panel drops it. Filters stay in the query string, including `branchId`. Older links with `?conversation=` still work. ### Recording and turn metrics When a call has a saved recording, the player stays at the bottom of the panel while you scroll either side or change tabs. Play or pause, skip ten seconds, seek along the timeline, change speed or volume, or download the recording. Selecting a transcript timestamp seeks to that message without starting playback. The current spoken message is highlighted as the audio plays. The timeline displays a waveform from the saved audio. Click or drag it to seek; the arrow at the right hides or shows it. If a waveform cannot be displayed, the timeline still lets you seek. Voice replies with saved timing data show **LLM** and **TTS** measurements beside their timestamp. Open the **LLM** measurement to see time to first token and the recorded input, output and cached token counts. **TTS** shows the wait until the first audio was ready. Under **Metrics over time**, **Latency** shows the measured wait before each reply. Select a graph point or open **View turn details** to jump to that transcript message and its recording position. The table also shows recorded input and output token counts. On a small screen, selecting a turn switches to **Transcript**. ![Latency graph of a demo phone call with its measured turn selected and the turn's latency shown below the graph](https://mintcdn.com/centillion/kGku8Vj2IrG9BHOA/images/workspaces/conversations-03-turn-metrics.webp?fit=max\&auto=format\&n=kGku8Vj2IrG9BHOA\&q=85\&s=420444a4b57977ce8dd65a8138fdb81e) Selecting a latency point shows that turn's wait; when the call saved its messages and audio, it also highlights the transcript turn and seeks the recording. A call without saved turn metrics shows **Not recorded** rather than an inferred value. Older calls without captured audio show **No recording is available**. Conversation audio is saved automatically, including calls started from the agent's [Preview](https://cntillion.com/docs/agents/releases-and-workflows.md#preview-a-conversation). There is no recording switch. Set the recording retention period under **Settings → Advanced → Privacy**. Caller opt-out and retention can make a recording unavailable. ### Transcript actions Hover over a message, or focus one of its controls. On mobile, open the message's three-dot menu. You can: * **Analyze** closes the review and opens Mastermind with that message and its conversation ID. * **Flag issue** adds the original message and your notes to the agent's Helpdesk inbox. An existing open ticket receives a note; otherwise a review ticket is created. It does not take over the conversation or send a customer message. Helpdesk must be enabled and your role must allow managing it. * **Create test** opens a test seeded with the conversation up to the customer's message, with its variables, channel and workflow node. Choosing the agent's reply keeps that reply under **Original response from conversation**, so you can compare a new run with it. Choose Simulation test, Next reply test, or Tool invocation test, describe what should have happened, then choose **Create** to save it in the agent's Tests. See [Start from a conversation](https://cntillion.com/docs/agents/tests.md#start-from-a-conversation). **Search transcript** at the bottom finds messages in the displayed transcript. Use the result arrows, Enter or Shift+Enter to move between matching messages. Escape closes search. When **Auto-translate transcript** is enabled in Evaluations, review starts in the viewer's supported browser language. **Source attribution** reveals saved knowledge citations beneath agent replies. The language selector at the bottom of the transcript translates both speakers' messages. Search by language name or code; each language has a country flag. **Show original** restores the saved transcript. Translation uses AI credits and requires permission to manage agents. The original stored messages are never replaced. The transcript menu also has **Copy transcript**, **Download transcript** and **Create test**. Download saves the displayed transcript as a text file, including the translation when one is selected. A link to a conversation that no longer exists, or that belongs to another workspace, shows "This conversation isn't available". ## Export Click **Select**, then tick conversations, or tick the box in the header to select every loaded row. Choose **Export** and CSV, JSON or PDF. You can export up to 100 conversations at a time, and each exported conversation names its agent. ## Who can use it Owners, Members and Viewers can see, open and export conversations. Support Associates have access to Helpdesk only. See [Members and roles](https://cntillion.com/docs/workspaces/members.md#roles). JSON exports also include the latest evaluation verdicts and scores, collected data, sentiment and topics when analysis is available. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.