# AI Streaming Responses - Implementation Guide

## Overview

This document describes the implementation of streaming AI responses using Server-Sent Events (SSE) for real-time token-by-token display.

## Architecture

### Backend Components

**1. StreamingService** (`src/modules/ai/streaming/streaming.service.ts`)
- Handles Azure OpenAI streaming API
- Processes SSE stream from Azure
- Emits tokens via callback function
- Chunk types: `token`, `metadata`, `done`, `error`

**2. AI Controller** (`src/modules/ai/ai.controller.ts`)
- Endpoint: `POST /api/v1/ai/chat/stream`
- Headers: `Content-Type: text/event-stream`
- Streams response chunks to client
- Error handling with SSE events

**3. AI Service Integration** (TODO)
- Add `chatStream()` method to AIService
- Reuse existing safety checks and context building
- Call StreamingService for actual streaming
- Save conversation after completion

### Frontend Components (TODO)

**1. Streaming Hook** (`hooks/useStreamingChat.ts`)
```typescript
const { streamMessage, isStreaming } = useStreamingChat();

streamMessage(
  { message: "Hello", conversationId: "123" },
  (chunk) => {
    // Handle incoming chunks
    if (chunk.type === 'token') {
      appendToMessage(chunk.content);
    }
  }
);
```

**2. AIChatInterface Updates**
- Add streaming state management
- Display tokens as they arrive
- Show typing indicator during streaming
- Handle stream errors gracefully

## Implementation Steps

### Step 1: Complete Backend Integration (30 min)

1. Add StreamingService to AI module:
```typescript
// ai.module.ts
import { StreamingService } from './streaming/streaming.service';

@Module({
  providers: [AIService, StreamingService, ...]
})
```

2. Implement `chatStream()` in AIService:
```typescript
async chatStream(
  userId: string,
  chatDto: ChatMessageDto,
  callback: StreamCallback
): Promise<void> {
  // 1. Run all safety checks (rate limit, input sanitization, etc.)
  // 2. Build context messages
  // 3. Call streamingService.streamCompletion()
  // 4. Collect full response and save conversation
}
```

### Step 2: Frontend Streaming Client (1 hour)

1. Create EventSource hook:
```typescript
// hooks/useStreamingChat.ts
export function useStreamingChat() {
  const streamMessage = async (
    chatDto: ChatMessageDto,
    onChunk: (chunk: StreamChunk) => void
  ) => {
    const response = await fetch('/api/v1/ai/chat/stream', {
      method: 'POST',
      headers: { 'Content-Type': 'application/json' },
      body: JSON.stringify(chatDto),
    });

    const reader = response.body.getReader();
    const decoder = new TextDecoder();

    while (true) {
      const { done, value } = await reader.read();
      if (done) break;

      const chunk = decoder.decode(value);
      const lines = chunk.split('\n');

      for (const line of lines) {
        if (line.startsWith('data: ')) {
          const data = JSON.parse(line.substring(6));
          onChunk(data);
        }
      }
    }
  };

  return { streamMessage };
}
```

2. Update AIChatInterface:
```typescript
const [streamingMessage, setStreamingMessage] = useState('');
const { streamMessage, isStreaming } = useStreamingChat();

const handleSubmit = async () => {
  setStreamingMessage('');
  setIsLoading(true);

  await streamMessage(
    { message: input, conversationId },
    (chunk) => {
      if (chunk.type === 'token') {
        setStreamingMessage(prev => prev + chunk.content);
      } else if (chunk.type === 'done') {
        // Add to messages array
        setMessages(prev => [...prev, {
          role: 'assistant',
          content: streamingMessage
        }]);
        setStreamingMessage('');
        setIsLoading(false);
      }
    }
  );
};
```

### Step 3: UI Enhancements (30 min)

1. Add streaming indicator:
```tsx
{isStreaming && (
  <Box sx={{ display: 'flex', alignItems: 'center', gap: 1 }}>
    <CircularProgress size={16} />
    <Typography variant="caption">AI is thinking...</Typography>
  </Box>
)}
```

2. Show partial message:
```tsx
{streamingMessage && (
  <Paper sx={{ p: 2, bgcolor: 'grey.100' }}>
    <ReactMarkdown>{streamingMessage}</ReactMarkdown>
    <Box component="span" sx={{ animation: 'blink 1s infinite' }}>|</Box>
  </Paper>
)}
```

3. Add CSS animation:
```css
@keyframes blink {
  0%, 100% { opacity: 1; }
  50% { opacity: 0; }
}
```

## Testing

### Backend Test
```bash
curl -X POST http://localhost:3020/api/v1/ai/chat/stream \
  -H "Content-Type: application/json" \
  -d '{"message": "Tell me about baby sleep patterns"}' \
  --no-buffer
```

Expected output:
```
data: {"type":"token","content":"Baby"}

data: {"type":"token","content":" sleep"}

data: {"type":"token","content":" patterns"}

...

data: {"type":"done"}
```

### Frontend Test
1. Open browser DevTools Network tab
2. Send a message in AI chat
3. Verify SSE connection established
4. Confirm tokens appear in real-time

## Performance Considerations

### Token Buffering
To reduce UI updates, buffer tokens:
```typescript
let buffer = '';
let bufferTimeout: NodeJS.Timeout;

onChunk((chunk) => {
  if (chunk.type === 'token') {
    buffer += chunk.content;

    clearTimeout(bufferTimeout);
    bufferTimeout = setTimeout(() => {
      setStreamingMessage(prev => prev + buffer);
      buffer = '';
    }, 50); // Update every 50ms
  }
});
```

### Memory Management
- Clear streaming state on unmount
- Cancel ongoing streams when switching conversations
- Limit message history to prevent memory leaks

### Error Recovery
- Retry connection on failure (max 3 attempts)
- Fall back to non-streaming on error
- Show user-friendly error messages

## Security

### Rate Limiting
- Streaming requests count against rate limits
- Close stream if rate limit exceeded mid-response

### Input Validation
- Same validation as non-streaming endpoint
- Safety checks before starting stream

### Connection Management
- Set timeout for inactive connections (60s)
- Clean up resources on client disconnect

## Future Enhancements

1. **Token Usage Tracking**: Track streaming token usage separately
2. **Pause/Resume**: Allow users to pause streaming
3. **Multi-Model Streaming**: Switch models mid-conversation
4. **Streaming Analytics**: Track average tokens/second
5. **WebSocket Alternative**: Consider WebSocket for bidirectional streaming

## Troubleshooting

### Stream Cuts Off Early
- Check timeout settings (increase to 120s for long responses)
- Verify nginx/proxy timeout configuration
- Check network connectivity

### Tokens Arrive Out of Order
- Ensure single-threaded processing
- Use buffer to accumulate before rendering
- Verify SSE event ordering

### High Latency
- Check Azure OpenAI endpoint latency
- Optimize context size to reduce time-to-first-token
- Consider caching common responses

## References

- [Azure OpenAI Streaming Docs](https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/streaming)
- [Server-Sent Events Spec](https://html.spec.whatwg.org/multipage/server-sent-events.html)
- [LangChain Streaming](https://js.langchain.com/docs/expression_language/streaming)