Conversations

Open Admin → Conversations to review accessible conversations and their quality or safety signals. The tabs shown depend on your permissions; this review area is separate from the conversation list under Work → Chat.

Review tabs

Tab	Scope and access
Conversations	Conversation history available in the current access mode; requires Conversations.View.
Feedbacks	Ratings and comments across agents you can review; requires Reporting.Feedback and write access to the agent.
Flagged messages	Individual messages recorded by safety checks; requires Conversations.View, Organization.AdminMode, and active Admin mode.

Selecting the Admin navigation tab does not activate Admin mode. A hidden tab can indicate missing permissions. Agent-specific Feedbacks remains a separate context.

Feature Description

Overview of All Conversations

An authorized user can review the conversations available in their current access mode, sorted by date. The displayed information in the overview includes:

• Date and time of the conversation start
• User's email
• Subject / topic of the conversation (e.g., “Specified deadline for termination of employment")
• Number of messages in the conversation
• Name of the agent with whom the communication took place (e.g., “HR Agent”)

Use search and pagination to find the relevant conversation. Review individual ratings in Feedbacks and safety findings in Flagged messages.

For safety review, open the separate Flagged messages tab. It lists individual messages and their findings rather than only filtering conversation rows.

Conversation Detail

Clicking on a row opens the conversation detail at /internal/conversations/[id].

The detail shows the full message history, including user inputs, assistant responses, and any visible execution context. Use this view when you need to understand exactly what the user asked, how the agent responded, and whether the answer relied on tools, uploaded files, or prior context.

When reviewing a conversation detail, check:

• the first user request and the agent selected for the conversation,
• follow-up messages that changed the context,
• files or images attached by the user,
• response formatting and cited information,
• errors, missing answers, or interrupted generation,
• whether the user submitted feedback on a response.

Long internal conversation details can show the same transcript outline as chat once the thread contains at least five user turns. Use it to jump between user requests while reviewing the surrounding answers and tool activity. The full behavior, keyboard controls, and surface-availability matrix are documented under Navigate a Long Conversation.

Feedback Option

Individual responses can be rated positively (thumbs up) or negatively (thumbs down), and a comment can be added. This information is subsequently displayed in the Feedback section and helps fine-tune the accuracy of responses.

Feedback from a conversation is useful for:

• identifying weak answers,
• improving agent prompts,
• checking whether the right data collection was attached,
• deciding which instruction changes an authorized editor should make in Configuration.

Feedback remains available while Evolution is unavailable in the current Web App. It does not automatically update the agent's instructions.

Central feedback

1. Open Admin → Conversations → Feedbacks.
2. Search by feedback text, conversation title, agent, or user. Open Filters to combine Rating, Agent, and User, or clear the filters.
3. Review the rating, feedback date, user, agent, and comment. Use Previous, Next, and Rows per page for longer lists; changing search or filters returns to the first page.
4. Use the row's View action to open the related conversation. The agent link opens the agent detail.

The User column identifies the feedback author when recorded. Do not assume it is always the person who started the conversation. Older records can use a legacy creator identity; unavailable or temporary identities can appear as Unknown user. The Message column contains the feedback comment, not the whole conversation transcript.

The list and its filter choices respect organization scope and agent write access. Reporting.Feedback alone does not expose every agent's feedback. An empty list can mean no matching feedback or insufficient access. A loading error is a separate state; use Retry before treating it as an empty result.

For the selected agent's own list and feedback detail, see Agent Feedbacks. Reviewing or submitting feedback does not automatically train the agent or update its instructions.

Flagged messages

Enable Admin mode, then open Conversations → Flagged messages. Search and page through the list of individual flagged messages. A conversation can have more than one row because several of its messages can be flagged.

Review Message date, User, Agent, Source, Action, and Reason. Open the reason to inspect Safety findings, including the stored reason, reported action, source, check, and scan time. The row's View action opens the related conversation so you can review context before deciding on a follow-up.

A safety finding is separate from a user's thumbs-up/down feedback. Older records may have a reason but no structured source, action, or scan details; missing values are shown as unavailable. Do not infer a full check history from those fields.

This tab requires Conversations.View, Organization.AdminMode, and active Admin mode. Turning Admin mode off removes the flagged-message view. An empty result or a hidden tab does not prove that the organization has never recorded a flagged message.

Sharing

If sharing is enabled by organization security settings, a conversation can be shared through a link. Sharing is useful for support, review, or handoff, but it should be used carefully when the conversation contains sensitive data.

Before sharing:

• confirm that public conversation sharing is allowed,
• remove or avoid sharing conversations with confidential information,
• share only with the intended audience,
• use Audit Log if you need to verify sharing-related changes.

Realtime Context

Realtime and voice sessions are stored as part of the same conversation model. A realtime session is not a separate object that replaces conversation history; it is a temporary transport for interacting with an agent while preserving the conversation thread.

When reviewing a realtime conversation, check the same operational signals as standard chat: selected agent, user context, transcript/messages, tool executions, approvals, feedback, and any artifacts generated during the session.

Audit and Tool Activity

Conversation review often connects to Logs:

• Use Tool Executions to inspect tool calls created by the conversation.
• Use Audit log to check administrative changes that may have affected access, sharing, or agent behavior.

Access Rights

Conversation permissions and the current access mode govern which records you can review. Administrative review of other users’ conversations requires the applicable permissions and Admin mode; a role name alone does not grant every action.

Approval and rejection check access to the requested conversation and that the pending message belongs to it. Before approving, review the tool, its arguments, and the target account. Tool permissions and organization policy are checked again at execution time; an earlier approval request cannot override access that has since been revoked.