Skip to content

Conversation inbox

The Inbox shows what people asked your agents and what the agents answered, on every channel, grouped by conversation. Use it to find bad answers and improve your agents.

It’s off by default. While it’s off, backcrew keeps no transcripts and the Inbox is empty.

  1. Open Settings → Conversations. Only owners can change it.
  2. Choose On, and how long to keep transcripts: 30 days, 90 days or 1 year.
  3. Select Save changes.

Transcripts hold what people type, which can include personal data. Before you turn this on, make sure your privacy notice tells the people using your agents that conversations are kept, why, and for how long.

For every request an agent answers, on every channel (the dashboard, the API, the website widget, WhatsApp, Telegram, Slack, schedules and webhooks):

  • the question and the answer (up to 20,000 and 50,000 characters),
  • the credits it cost,
  • how it answered: the tools it called (with secret-looking values masked), the passages it cited, and anything its policy blocked,
  • who asked, named as the audit log names them: a member’s email, an API key’s name, Website visitor, Telegram user of @yourbot or Slack user in your Slack workspace.

Messages from the same session are one conversation: a chat in the dashboard, a widget visitor’s chat, a Telegram chat, a Slack thread, or an API session id. An API request without a session id is a conversation of its own.

On the widget, WhatsApp, Telegram and Slack, each conversation also keeps the address to reply to (a chat, a thread or a number), encrypted, so your team can reply or take over.

Not kept: requests that failed or were refused (no credits, not deployed), and your agents’ test cases.

Transcripts are stored on backcrew’s platform and encrypted at rest. See Where your data lives.

  • After the retention period. Each message is deleted once it’s older than the period you chose. backcrew checks every hour.
  • When you shorten the period. Transcripts older than the new period are deleted when you save.
  • When you turn it off. Every stored transcript is deleted at once, and nothing new is kept.
  • When you delete a conversation in the Inbox, for example when someone asks for their data to be erased.
  • When you delete an agent. Its conversations go with it.

The agent’s own conversation memory is separate: it has its own setting per agent and isn’t affected by any of the above.

Owners and builders can read conversations. Viewers, billing and auditors can’t. Anyone in the workspace can see whether transcripts are kept, in Settings → Conversations.

The audit log records when someone opens a conversation (once per person per conversation every half hour, since an open conversation refreshes itself), takes it over, flags an answer, deletes a conversation or changes the setting. It names the agent and the conversation, never what was said.

  • Find conversations by period, agent and channel, or search the questions and answers. Flagged only shows the conversations with a bad answer.
  • Open one to read it, with How it answered under each answer.
  • Flag as bad answer, with an optional note on what was wrong. Select Remove flag to undo it.
  • Delete conversation removes it for good.

An agent with human handoff on can hand a conversation to your team. It’s marked Needs a person, and you reply from here. You can also Take over any conversation you’re watching. Tick Needs a person to see only those with a person.

  • Assign a conversation to an owner or builder (or to yourself) from the box at its top. They’re notified, and Assigned to me in the filters shows theirs. Choose Unassigned to clear it.
  • New messages in a conversation waiting for a person notify its assignee, or the agent’s builder and the owners if nobody’s assigned it, at most once every ten minutes per conversation. Choose how in Settings → Notifications.
  • The list refreshes itself every 20 seconds, unless you’ve loaded older conversations with Load more.

Under an answer, Add as test case (owners and builders) turns the question into one of the agent’s test cases, with the flag’s note as a starting point for what a good answer does. Test cases are on the agent’s page under Test cases (see Test cases). Run them by hand, or have them run after every new version, so a bad answer you fixed stays fixed.