Kelu records every conversation so you can see what readers ask, how well the assistant answers, and what content is missing. Everything here is in the dashboard sidebar.
PageWhat it answers
DashboardHow is the whole workspace doing?
ConversationsWhat exactly did readers ask, and what did they get?
AnalyticsUsage, feedback and audience for one knowledge base or all of them
Coverage GapsWhat could the assistant not answer?
Knowledge MaintenanceDrafted answers for those questions
TracesWhat happened inside one answer?

Analytics

Open Analytics, pick a knowledge base (or all), and a period: last 7, 30 or 90 days. The Overview tab shows:
CardWhat it shows
Total Conversations, Unique Users, Total MessagesVolume for the period
Citation CoverageShare of answers that cited at least one source
Positive Feedback, Feedback CountThumbs up and down from every surface
Daily ActivityConversations and messages per day
Self-serve rateConversations that were on topic, cited a source, and were not argued with or rated down. Split by intent
Support effort avoidedHours saved, assuming 10 minutes per self-served question and 15 minutes per deflected ticket
AudienceCountries, browsers, devices and operating systems, and how many people who opened the widget went on to ask
Ticket DeflectionResults from your form deflectors, across the whole workspace
The last four cards appear only once they have data. Self-serve rate needs conversations to be classified, which happens a few minutes after they end. Audience data comes from browser surfaces only. Slack, Teams, MCP and server-side API questions carry none, so the card shows what share of conversations it covers. No IP address is stored. The Sources tab lists each source with its documents, chunks, citations, sync health and last sync. A source that is never cited adds sync work without improving answers.

Conversations

Conversations lists every conversation. Click one to read the full transcript with its citations. Filter the list by:
  • Intent — what the reader was trying to do: Troubleshooting, Guiding, Lookup, Off topic or Other. Kelu sets it about five minutes after a conversation goes quiet.
  • Signals — Pushed back (the reader argued with the answer) and Frustrated.
  • Tags — your own labels. Click Tags to create them, then open a conversation and click a tag to add or remove it.
Intents and signals are set by Kelu and cannot be edited. Tags are yours.

Find and fix knowledge gaps

Coverage Gaps shows questions the assistant told readers it could not answer from your sources. Pick a knowledge base and a Week, Month or Quarter, and use the arrows to move between periods.
  • Topics groups similar unanswered questions. Each topic has a Finding (what readers asked and why it failed), a Suggestion (what to write), and the nearest existing pages.
  • Question log lists each distinct question, most asked first, including answers readers rated unhelpful. Turn off Unanswered only to see questions you have already answered.
Topics appear once a period has at least 3 conversations with an unanswered question. Questions that are not documentation gaps, like greetings or tests, are marked Not a documentation gap and hidden unless you show them. To act on a topic:
  • Answer in Q&A or Answer this opens the Q&A editor with the question filled in. Saved answers are used right away.
  • Resolved or No action needed files the topic. Reopen brings it back.
  • Export CSV downloads the topics for the period. Copy copies one topic as markdown, ready for a docs ticket or an AI agent.
Re-analyse is available only when new unanswered questions have arrived since the last analysis.

Draft answers with Knowledge Maintenance

Knowledge Maintenance writes draft answers for unanswered questions, using only your own content.
1

Pick a knowledge base

A scan works on one knowledge base at a time.
2

Click Draft answers

A scan takes up to 8 of the most asked unanswered questions from the last 90 days. It first asks your knowledge base again. If the question can now be answered, it moves to Answered now and nothing is drafted.
3

Review each draft

Drafts appear under Proposed. Facts your content does not contain are marked [[NEEDS INPUT: …]]. Fill them in, then click Review & add to save the answer to Q&A. You cannot save while a marker is left.
Dismiss moves a draft to Dismissed; Restore brings it back. Adding an answer needs the Edit Sources permission on the knowledge base.

Trace every AI interaction

Traces records each answer: the question, the output, status, latency and token counts. A search trace also lists the results it returned.
  • Filter by kind: Chat, Search, MCP, A2A, Deflection or Agent.
  • Sort by Latency, tokens In and Out, or Time.
Traces are deleted with your conversations when a retention window is set, and are not recorded at all under zero data retention. The same data is available from the API:
GET https://app.kelu.dev/api/v1/knowledge-bases/:id/traces?kind=chat&sort=latency&order=desc
GET https://app.kelu.dev/api/v1/traces/:id

Usage

Usage shows the questions used against your plan’s monthly allowance. See Billing & Plans.