Documentation

Chat Logs

View, filter, share, and translate conversation history.

Chat Logs show every conversation visitors have with your agent. Review what's being asked, identify content gaps, and share conversations with your team.

Viewing Conversations

  1. Open your agent in the dashboard
  2. In the sidebar, expand Activity and click Chat Logs
  3. Browse conversations by date
  4. Click a conversation to view the full exchange

Each conversation in the list shows:

  • A preview of the latest message
  • The last activity time

Filtering Conversations

Date Range

Filter conversations by time period:

  • Today
  • Yesterday
  • Last 7 days
  • Last 30 days
  • This month
  • Last month
  • Last 3 months
  • Last 12 months
  • Custom date range

Conversation Details

Each conversation opens with two tabs:

Chat — full message history (visitor and agent), timestamps for each message, tool/action indicators (for example when Collect Leads or another action ran), response feedback badges on rated assistant messages, and attached file metadata (file name) when visitors upload files in the widget.

Details shows the channel, conversation status, reference ID, message count, started timestamp and last activity. Contact details, outcome and collected information appear when available.

New website widget, Help Page and Center Stage conversations also capture estimated country and device/browser from the first request. Country uses the existing Cloudflare country header when available. Device and browser use the request's user-agent information. No raw IP address or raw user-agent string is stored for these details. Later messages do not replace the initial snapshot.

Older conversations and requests without usable headers show Not available. Playground, API, Instagram and WhatsApp conversations do not capture visitor context from request headers. Country can reflect a VPN location, and browsers can report incomplete device information, so these details are estimates.

Improve an answer

Every assistant reply that follows a visitor message shows an Improve button. It drafts a better answer for you:

  1. It re-reads the sources your agent already has for that question.
  2. If none of them answer it, it looks at pages on your own website that are not yet indexed and reads the most likely one.
  3. It drafts a short answer from what it found, your agent's instructions, and the current reply, then checks the draft against that evidence. A draft that is not better or not supported is sent back with the reason, up to three rounds.

The modal shows the current score, the suggested answer in an editor, a Source support score, and the page the answer came from when it came from your website. Edit the suggestion if you like, then Add as Q&A to save it as a revision. Tick Also add this page as a source to queue the page for indexing under Sources. Once a reply has a revision the button reads Revised and opens that Q&A pair for editing.

Limits worth knowing:

  • Each reply is researched once. Opening the modal again shows the stored suggestion; the button still reads Improve.
  • Research stops after 45 seconds and only looks at pages on the same site as your agent's website URL.
  • When nothing new is found, the modal says so and shows the current answer tightened, ready to edit. It never blocks you from writing the answer yourself.
  • The suggestion is grounded in your content and instructions, not in general knowledge. It is a starting point for you to check, not a fact check.

Share Conversations

Share a conversation with your team:

  1. Open the conversation
  2. Click Share
  3. Your email client opens with a pre-filled message containing the conversation transcript and a link back to the chat log in your dashboard
  4. Send to colleagues

The link in the email requires a teammate to be signed in to your workspace to view the conversation. The transcript itself is included in the email body.

Translate Conversations

Translate a conversation to English:

  1. Open the conversation
  2. Click Translate
  3. Translated messages replace the originals in the view

Useful for:

  • Supporting multilingual visitors
  • Understanding foreign language queries
  • Training content in other languages

What You Can Learn

Common Questions

Identify patterns:

  • What do visitors ask most often?
  • Are there recurring topics?
  • What questions lead to follow-ups?

Content Gaps

Find missing information:

  • Questions the agent couldn't answer
  • Topics needing more training content
  • Areas where Q&A pairs would help

Response Quality

Evaluate performance:

  • Are answers accurate?
  • Is the tone appropriate?
  • Are responses the right length?
  • Did visitors mark assistant replies as good or bad?

Response Feedback

When visitors rate an assistant reply in the widget, Chat Logs show that rating on the assistant message as a Good response or Bad response badge. API v2 feedback on assistant messages appears the same way.

Use these badges while reviewing conversations to find answers that should become Q&A pairs, source updates, or instruction changes.

Quality and Relevance Scores

Each assistant reply carries a quality number from 0 to 1. It is judged after the reply is sent and measures how well the reply answers the visitor from the agent's sources. Greetings, language switches and clarifying questions score high because they make no claim that needs a source. A reply that answers without any supporting source, or contradicts the sources, scores low. Green is 0.6 and above, amber is between 0.35 and 0.6, red is below 0.35. Hover the badge for the explanation. A high score is not a guarantee of correctness.

Each visitor message carries a relevance number on the same scale. It measures how relevant the message is to your business, and the hover text names what kind of message it is: a question, chit-chat, spam or abuse, off topic, or a request for a human.

Conversations recorded before quality judgment was introduced keep the similarity number they were recorded with. Its hover text says so. Those rows are never rewritten.

Weak assistant replies also show a reason chip beside the quality badge. Hover or focus it to learn whether the visitor rated the reply badly, no source covered the question, the reply missed the question or contradicted sources, its claims lacked support, or an older source match was weak.

Links from Knowledge gaps carry a conversation filter in the URL. The conversations parameter accepts up to 50 comma-separated conversation IDs and combines with the date and weak-answer filters.

Privacy Considerations

What's Logged

  • Conversation text (questions and answers)
  • Timestamps
  • Response feedback on assistant messages
  • Lead information (if collected)
  • Source references
  • File attachment metadata, such as file name and type

What's Not Logged

  • IP addresses
  • Device information
  • Location data
  • Personal cookies
  • Raw uploaded file contents from widget attachments

Data Retention

You can delete individual conversations from the chat logs view via the conversation's more-options menu.

Using Insights

Improve Training Content

  1. Find questions with poor answers
  2. Add relevant sources or Q&A pairs
  3. Test in Playground
  4. Monitor future conversations

Optimize AI Instructions

  1. Identify response tone issues
  2. Update AI instructions
  3. Test changes
  4. Review results in logs

Next Steps