Streaming

Lyntaris uses two streaming patterns in production.

1) Prediction SSE (/api/v1/prediction/{chatflowId})

Set "streaming": true in prediction request.

{
  "question": "Tell me a story",
  "streaming": true,
  "overrideConfig": {
    "sessionId": "session-1"
  }
}

Server responds as SSE (Content-Type: text/event-stream). Each data: line contains JSON with { "event": "...", "data": ... }.

Typical events:

• start
• token
• metadata
• sourceDocuments
• usedTools
• error
• end

Additional runtime events may be emitted (for example agentFlowEvent, usageMetadata, tts_start, tts_data, tts_end).

2) Realtime-v1 protocol (Flowise -> Unity)

Unity parser expects streamed tokens to form this final structure:

{
  "text": "Conversational text for TTS",
  "tool_calls": [
    {
      "name": "LoadFlowiseImageCommand",
      "args": {
        "arg1Name": "arg1Value"
      }
    }
  ],
  "error": "Optional error string"
}

Rules used by Unity:

• Tokens in text are sent to TTS buffer immediately.
• tool_calls are buffered, then executed when the array is complete.

Realtime queue API (external triggering)

Base path: /api/v1/openai-realtime/{chatflowId}

• POST /queue/enqueue
• GET /queue/status
• POST /queue/toggle
• POST /queue/complete
• POST /queue/cancel
• POST /queue/clear
• POST /queue/remove

enqueue requires form object with non-empty ID (case-insensitive, normalized to ID).

{
  "form": {
    "ID": "customer-123",
    "question": "Hey, how are you?"
  },
  "ccBaseUrl": "https://optional-unity-base.example.com"
}

Success response:

{
  "accepted": true,
  "taskId": "flowise-4ed0073e-b9a8-49e9-a086-819e77990780",
  "pendingCount": 1
}