Conversations¶

View, analyze, and manage conversations between users and your AI agents.

---

Overview¶

The Conversations view provides access to all chat interactions with your agents, enabling you to: - Monitor agent performance in real conversations - Identify common user queries and issues - Review conversation quality - Export conversations for analysis - Debug agent responses

---

Accessing Conversations¶

1. Navigate to MY AGENTS
2. Select your agent
3. Click Conversations from the agent menu

The Conversations menu (the conversation list panel on the left) is collapsible — collapse it to free up screen space for the conversation detail view.

Conversations admin page — the conversation list on the left is the collapsible Conversations menu, and the Show filters button reveals filters that are collapsed by default

---

Continuous Conversations — No Hard Expiry¶

Conversations no longer reach a hard, user-visible expiry state. The chat input remains active across long idle periods, and post-processors run iteratively as the conversation continues.

Concept	Behavior
Soft expiry on the backend	When the configured stale period is reached, the backend triggers a soft expiry — Conversation Post-Processors run automatically, generating the initial outcomes (summaries, metadata extraction, CRM updates) just as before
Iterative post-processing	If the user resumes the thread later, the new message is appended to the existing context. When the conversation hits the stale period again, post-processors re-run and update / overwrite the previously generated outcomes based on the newly accumulated context
User-visible state	The chat input remains active across long idle periods — no "conversation expired" lockout

---

Replying to Message Feedback¶

Admins can respond to feedback left on a disliked message, and the user can reply in the same thread.

1. Open My Agents, then the agent card's three-dot menu → Conversations.
2. Open the conversation and find the message with feedback.
3. Choose Reply to feedback to write your response.
4. The user sees the response in Feedback Thread when they next open the conversation in the app or embedded widget, and can reply there.

Replies are not delivered live. They cannot be edited or deleted and do not support attachments. Anonymous widget users must return to the same conversation to see the reply.

When message-feedback notifications are enabled, email delivery is best effort. An admin reply can notify the user if an email address is available; a user reply can notify the configured feedback recipients. Configure the recipient-specific email templates under Settings → Templates → Email Template.

---

Conversation List¶

Select a conversation from the left panel to open its messages. The current list shows the conversation timestamp and a menu for each entry. The detail header provides Flag, Browsing Details, Analysis and Full screen.

Pinning Favourite Conversations¶

Any conversation a user owns can be pinned (favourited) directly from its menu.

Behaviour	Detail
Pin from the conversation menu	Available on any conversation the user owns
Pin indicator	Pinned conversations are visibly marked in the list
Persists across devices	Pin state is saved to the account, so it survives sessions and follows the user between devices

---

Filtering Conversations¶

Filters appear collapsed by default behind the Show filters button (visible in the Conversations Page Layout above). Expand to reveal the controls below.

The expanded panel provides Search Conversations, Date, User ID, Filters and Analysis. Use the control that matches the data you need; available analysis fields depend on the agent’s configured post-processors.

Analysis Filter¶

A new Analysis option under Filters lets you slice conversations by any Analysis (After The Conversation) field and value. The value is free text, works on a contains-match, and is case-insensitive — e.g., filter by conversation Type: Sales to surface every conversation the analyzer tagged as a sales chat.

Analysis filter — pick an Analysis field (e.g., conversation Type), enter a free-text value (e.g., Sales), and Add. Applied filters appear as removable chips below

See Post-Processors → Analysis (After The Conversation) for how the underlying fields are configured.

---

Conversation Analysis — Per-Conversation Post-Processor Inspection¶

Click the conversation’s Analysis button on any conversation to view every post-processor that ran on it, with:

• Status — Ready, Processing, or Failed
• Extracted values — fields and JSON output from each analyzer / summary run
• Error details — if a post-processor failed
• Last-updated timestamps — most recent run for each post-processor

This makes analyzer outputs actionable at the individual conversation level, not just in aggregate reports.

Conversation Analysis modal showing three Ready results, including extracted fields and a conversation summary.

---

Conversation Observability — Debug Info¶

A unified Debug info panel surfaces everything that happened on a conversation turn — the language model that produced the reply, the pre-processors that ran (with the tags they applied), the tools called, the memory state used, the raw request trace, and the request's stop reason. Diagnose truncated or off-looking responses directly instead of guessing.

What it shows¶

Section	What's captured
Language Model	The model used for the turn (e.g., Gemini 3.1 Pro), status (Completed / Running / Failed), and TTFT (time-to-first-token)
Pre-Processors	Each pre-processor that ran on the turn, with status and timing. Tags applied by these processors flow downstream and are surfaced here too
Tools	Each tool invocation — name, status (Completed / Running / Failed), and timing. Useful for spotting slow or erroring tools
Tags Applied	The tags attached to that turn, listed alongside it
Token Counts	Input and output token counts for the turn
Memory	The user/context memory state for the turn. Shows: At Message Time (snapshot when the user's message landed), Current (live state — may have been edited since), and the Memory Map (raw memory contents). An "Edited since" badge appears if memory has changed since the turn — Copy buttons let you snapshot each block
Request Trace (View screen only)	The raw JSON request sent to the model — metaData, prompt turns, etc. Not persisted: closing the page or navigating away drops it
Stop Reason (in Debug Logs)	Per-call stop reason (e.g., STOP, MAX_TOKENS) alongside request/response headers, response body, and timing. Available from the separate Debug Logs view

Debug info for an existing example turn: model, stop reason, TTFT, input/output tokens, and expanded pre- and post-processor status/timing. Available sections depend on the turn.

Debug Logs — per-call wire detail¶

The companion Debug Logs view captures the raw provider-level calls behind each turn, with per-call stop reason alongside full request/response wire details. Open My Agents > agent menu > Request Debug Logs; each entry expands into one or more Calls showing:

• Provider (e.g., google) and request ID (e.g., fetch_f3357751-021)
• Stop status (e.g., STOP) and timing (e.g., 2.9s, in 5374 / out 344 tokens, optional cacheR 4046 for cached reads)
• Tabs: Request · Response Headers · Response Body · Stop Reason
• Request timestamp + a Copy JSON action

The log list provides Filter by Date, Refresh, Show only logs having errors and a text search. Log details depend on the selected record.

Current Debug Logs list with date, error-only and search controls. This example agent has no retained records.

Access — existing Debug permission¶

Debug info is gated by the existing Debug permission in the agent role editor — the Debug capability under Agent > Beyond the editor in the current role matrix, which controls debug-mode access in the chat (e.g., Show Context Sent to API). No new permissions were introduced for this surface.

• Admins have Debug by default
• Any other role granted the Debug capability sees the ⓘ icon and can open Debug info on both the View and Conversations screens

When to use¶

• Truncated response? Check the Stop Reason in Debug Logs — MAX_TOKENS means the model hit its output limit; raise the language model's Max Output Tokens.
• Wrong-feeling tool behavior? Inspect the Tools section — confirm the right tool fired and check its timing for slowness.
• Memory not picked up? Compare At Message Time vs Current memory — the "Edited since" badge tells you whether the answer was grounded in an older state.
• LLM not behaving as expected? Open the Request Trace on the View screen to see the exact payload sent to the model (system prompt + turn history + generated_tags + metadata).

---

Viewing Conversation Details¶

Click on a conversation to view:

Message Thread¶

• Complete conversation history
• User messages
• Agent responses
• System messages (escalations, errors)
• Timestamps for each message

Metadata¶

• Session ID
• User information (if available)
• Agent version used
• Models invoked
• Tools called
• Response times

Analytics¶

• Average response time
• Tool usage count
• Knowledge base queries
• Tokens consumed
• Cost per conversation

---

Conversation Actions¶

Export Conversation¶

Export individual conversations in multiple formats: - JSON: Full conversation data with metadata - PDF: Formatted conversation transcript - CSV: Tabular message data - Text: Plain text transcript

Share Conversation¶

Share conversation links with team members: 1. Click Share button 2. Copy shareable link 3. Set access permissions (view only, edit) 4. Set expiration (optional)

Flag for Review¶

Flag conversations that need attention: - Quality issues - Agent errors - Inappropriate responses - Training opportunities

Add Notes¶

Add internal notes to conversations: - Document context - Note follow-up actions - Tag team members - Link related conversations

---

Bulk Actions¶

Select multiple conversations to: - Export in bulk - Tag for categorization - Delete (compliance) - Archive

---

Search Conversations¶

Search across all conversations:

Full-text Search - Search message content - Find specific keywords or phrases - Use quotes for exact matches

Advanced Filters - User ID - Agent version - Date range - Message count - Response quality

---

Conversation Analytics¶

Aggregate Metrics¶

• Total conversations
• Average messages per conversation
• Average conversation duration
• Completion rate
• Escalation rate
• User satisfaction score

Trends¶

• Conversations over time
• Peak usage hours
• Common topics
• Escalation patterns

---

Privacy & Compliance¶

Data Retention¶

Review the workspace's Data Purge Settings before relying on older conversations or report data. A value shown in an example workspace is not a platform-wide retention default. Confirm your organization's retention and deletion requirements with its administrator.

User Privacy¶

Use synthetic conversations when preparing examples or support evidence. Remove personal details, credentials and private document content before sharing screenshots or logs. See DLP Policies for policy configuration; do not assume a configured policy covers every channel or exported record.

Access Control¶

Review the member's effective workspace role and scope before granting conversation access. A visible control in an administrator screenshot does not establish that every role can use it.

---

Troubleshooting¶

Conversation Not Showing¶

Issue: Recent conversation not visible in list

Solution: - Refresh the page - Check date range filter - Verify user has proper permissions - Allow 1-2 minutes for conversation to sync

---

Cannot Export Conversations¶

Issue: Export button is grayed out or fails

Solution: - Verify you have export permissions - Check if conversations contain PII (may require admin approval) - Try smaller date range (large exports may timeout) - Contact support if issue persists

---

Best Practices¶

Regular Reviews¶

• Review conversations weekly
• Identify patterns and common issues
• Use insights to improve agent instructions
• Flag excellent interactions for training

Quality Assurance¶

• Sample conversations randomly for QA
• Check agent adherence to instructions
• Verify tool usage is appropriate
• Ensure knowledge base retrieval is accurate

User Feedback¶

• Monitor satisfaction ratings
• Follow up on low-rated conversations
• Incorporate feedback into agent improvements
• Thank users for positive feedback

---

Related Topics¶

• Agent Reports - Aggregate analytics and metrics
• Agent Builder - Improve agent based on conversation insights
• Version Control - Track changes to agent configuration