Skip to content

Conversations

The Conversations view lets you browse all conversations that have taken place (or are currently happening) across your projects.

Conversation List

The main list shows conversations with their status and key details. Use search and filters to narrow results by project, status, date range, or user.

Conversation States

Each conversation has a status that reflects where it is in its lifecycle:

StatusMeaning
InitializedCreated but not yet started
Awaiting user inputWaiting for the user to speak or type
Receiving user voiceThe user is actively streaming voice audio
Processing user inputThe system is analyzing the message and running actions
Generating responseThe AI is composing a reply
FinishedThe conversation ended normally
AbortedThe conversation was terminated abruptly
FailedThe conversation ended due to an error

Conversation Details

Click on any conversation to open its detail view. The detail view has three tabs:

Events Timeline

A chronological log of everything that happened during the conversation. Each event is displayed as a card color-coded by type:

EventColorWhat It Records
MessageBlueUser and AI messages — the actual conversation transcript
ClassificationYellowWhich classifier ran, which actions matched, and extracted parameters
TransformationVioletWhich context transformer ran and which fields it set
ActionPurpleWhich action executed and its effects
CommandIndigoClient commands received (e.g., button presses from the application)
Tool CallPinkAI tool invocations with parameters, result, and success/error status
Conversation StartGreenWhen the conversation began and which stage it started in
Conversation ResumeCyanWhen a paused conversation was resumed
Jump to StageTealNavigation from one stage to another
ModerationAmberModeration check result — whether the message passed or was blocked
Conversation EndGrayGraceful end — how and why the conversation concluded
Conversation AbortedOrangeAbrupt termination
Conversation FailedRedFatal error

Non-message events are indented slightly to visually separate them from the main conversation flow.

Event Card Buttons

Each event card may show action buttons in its top-right corner depending on what data is available:

ButtonWhen VisibleWhat It Opens
View System Prompt (document icon)When a system prompt snapshot was capturedThe full prompt the AI received when generating the response
View Filler Prompt (wand icon)When a filler response was generatedThe filler prompt instructions sent to the filler LLM
View Raw Response (scroll icon)When a raw LLM response was capturedThe unprocessed JSON response from the language model
View Stage Variables (braces icon)When a variable snapshot is availableThe state of all stage variables at that point in the conversation
Create Issue (bug icon)Always (on non-archived conversations)Opens the Issue creation form pre-filled with the conversation ID, event index, and stage — so you can file a bug report directly from the event that caused the problem

The Create Issue button is especially useful for quality review — you can browse events, spot a problem, and create a linked issue without leaving the detail view.

Performance

The Performance tab shows per-turn latency data for the conversation as a waterfall chart. Each turn is visualized as a horizontal bar broken into color-coded phases:

PhaseColorWhat It Measures
ASRBlueSpeech-to-text transcription time
ModerationYellowContent moderation check duration
ProcessingPurpleOverall input processing (classification, transformers)
KnowledgeTealKnowledge base retrieval time
ActionsOrangeTime spent executing action effects
FillerGrayFiller response generation time
LLMGreenMain language model response generation time
TTSPinkText-to-speech synthesis time

Below each bar, individual phase durations are listed alongside TTFT (time to first token) and, for voice turns, First audio (time until the first audio chunk was delivered).

Metadata

Read-only system fields: conversation ID, project ID, user ID, stage ID, status, and timestamps.

Lifecycle

A typical conversation follows this path:

  1. Start — A client application connects and begins a conversation, specifying the user and starting stage.
  2. Input loop — The user sends messages (voice or text), the system processes them and the AI responds.
  3. Stage navigation — Actions move the conversation between stages based on what happens.
  4. End — The conversation concludes via:
    • An End Conversation action effect (graceful ending — the AI may say goodbye first).
    • An Abort Conversation action effect (immediate termination with no AI response).
    • The user disconnecting.
    • A system error.

Conversations can also be resumed — if a user disconnects and reconnects, the conversation picks up where it left off. Conversations in Awaiting user input state have a Resume button in the detail view that opens them directly in the Playground.

Tips

  • Use events to debug — When a conversation doesn't go as expected, the event log shows exactly what the classifier matched, which actions fired, and what effects ran.
  • Check classification events — If the wrong action fires, look at the classification result to see why.
  • Use Create Issue from event cards — When you spot a problematic event, create the issue directly from that card. It pre-fills all the context so you don't have to copy IDs manually.
  • Monitor failure patterns — If you see conversations frequently ending in "Failed" status, investigate the event logs for common error causes.

Released under the Apache-2.0 License.