Skip to main content
The AI advisor uses Server-Sent Events (SSE) for real-time streaming. Every /ai/stream request opens a persistent HTTP connection; the backend pushes events as they happen rather than waiting for a complete response.

​
Event format

Each event is a JSON object on a single line, prefixed with data: and followed by a double newline: 
data: {"event": "phase", "phase": "thinking"}\n\n
data: {"event": "tool", "name": "query_transactions", "status": "calling"}\n\n
data: {"event": "token", "text": "You spent "}\n\n
data: {"event": "token", "text": "$420 on food "}\n\n
data: {"event": "token", "text": "last month."}\n\n
data: {"event": "done", "text": "You spent $420 on food last month.", "widgets": [], "suggestions": [], "agent_used": "cashflow", "tools_used": ["query_transactions"]}\n\n
The Flutter AiService parses each line, strips the data: prefix, and decodes the JSON.

​
Event types

​
phase

Indicates which stage the pipeline is in. Flutter updates the loading indicator text. 
{ "event": "phase", "phase": "thinking" }
{ "event": "phase", "phase": "tool_call" }
{ "event": "phase", "phase": "generating" }
PhaseUI label
thinking”Thinking…” (animated dots)
tool_call”Analyzing…” (replaced by tool event label)
generating”Generating…”

​
tool

Fires when a tool call starts. Flutter shows a dismissible badge in the chat bubble. 
{ "event": "tool", "name": "query_transactions", "status": "calling" }
{ "event": "tool", "name": "get_budgets", "status": "calling" }
The Flutter _toolLabels map in chat_provider.dart converts tool names to readable labels: 
final _toolLabels = {
    'query_transactions': 'Analyzing transactions',
    'get_portfolio':      'Loading portfolio',
    'get_budgets':        'Checking budgets',
    'fetch_stock_price':  'Fetching live price',
    'run_anomaly_detection': 'Scanning for anomalies',
    'scan_subscriptions': 'Finding subscriptions',
    // ...
};

​
token

Incremental text from Claude’s streaming output. Flutter appends each token to the current message bubble in real time. 
{ "event": "token", "text": "You spent " }
{ "event": "token", "text": "$420 " }
{ "event": "token", "text": "on food last month, " }

​
done

Final event. Contains the complete assembled text plus any widgets and follow-up suggestions. 
{
  "event": "done",
  "text": "You spent $420 on food last month, up 12% from $278 in February.",
  "widgets": [
    {
      "type": "chart",
      "chart_type": "bar",
      "title": "Food spend — Feb vs Mar",
      "data": { "labels": ["February", "March"], "values": [278, 420] }
    }
  ],
  "suggestions": [
    "What were my top food merchants?",
    "Am I over my food budget?"
  ],
  "agent_used": "cashflow",
  "tools_used": ["query_transactions", "get_budgets"]
}

​
error

Replaces done if the pipeline fails. Flutter shows an error bubble with fallback suggestions. 
{
  "event": "error",
  "message": "I hit a snag processing your request. Please try again.",
  "suggestions": [
    "Try asking about my spending",
    "How's my portfolio doing?"
  ]
}

​
Widget response schema

When the answer is better expressed as a chart or card, Claude returns a structured widget object in the done event’s widgets array. Flutter’s widget_renderer.dart renders each widget natively.

​
Chart widget

{
  "type": "chart",
  "chart_type": "bar | line | pie | sparkline",
  "title": "Spending by Category — March",
  "data": {
    "labels": ["Food", "Transport", "Shopping"],
    "values": [420, 180, 310]
  }
}

​
Summary card

{
  "type": "card",
  "title": "Net Worth",
  "value": "S$15,270",
  "delta": "+$420 this month",
  "delta_positive": true
}

​
Alert card

{
  "type": "alert",
  "severity": "warning | critical | info",
  "title": "Budget exceeded",
  "description": "Shopping budget is 116% used with 10 days left."
}

​
Multi-widget

{
  "type": "multi_widget",
  "widgets": [
    { "type": "card", ... },
    { "type": "chart", ... }
  ]
}

​
HTTP headers

The /ai/stream endpoint sets three headers critical for SSE to work through nginx: 
headers = {
    "Cache-Control": "no-cache",
    "Connection": "keep-alive",
    "X-Accel-Buffering": "no",   # disables nginx response buffering
}
return StreamingResponse(generator(), media_type="text/event-stream", headers=headers)
X-Accel-Buffering: no is essential — without it, nginx buffers the entire response before forwarding, which defeats the purpose of streaming.

​
Flutter SSE parsing

// ai_service.dart
Stream<AiEvent> streamAsk(String query, List<Message> history) async* {
    final request = http.Request('POST', Uri.parse('$baseUrl/ai/stream'));
    request.body = jsonEncode({...});
    final streamedResponse = await client.send(request);

    await for (final chunk in streamedResponse.stream.transform(utf8.decoder)) {
        for (final line in chunk.split('\n')) {
            if (!line.startsWith('data: ')) continue;
            final json = jsonDecode(line.substring(6));
            yield AiEvent.fromJson(json);
        }
    }
}
The provider in chat_provider.dart listens to this stream and updates state for each event type, triggering selective widget rebuilds via Riverpod.

​
Non-streaming endpoint

POST /ai/ask runs the same pipeline but collects all SSE events internally and returns a single AgentResponse. Use this for programmatic calls where you don’t need real-time streaming: 
curl -X POST http://localhost:8000/ai/ask \
  -H "Authorization: Bearer <jwt>" \
  -H "Content-Type: application/json" \
  -d '{"query": "What is my net worth?", "user_id": "uuid"}'
Response: 
{
  "text": "Your net worth is S$15,270...",
  "widgets": [],
  "suggestions": ["Show me my assets vs liabilities"],
  "agent_used": "wealth",
  "tools_used": ["get_portfolio", "query_transactions", "get_debts"]
}