Client (London) ──[10ms]──> Edge Function (London) 
                              │
                              └──[80ms]──> DB (US-East)
                              │
                              └──[80ms]──> DB (US-East) (Query 2)
                              │
                              └──[80ms]──> DB (US-East) (Query 3)
                              
Total Latency: 250ms+ (Slower than running the function directly in US-East!)

import { createOpenAI } from '@ai-sdk/openai';
import { streamText } from 'ai';

export async function generateStreamingResponse(prompt: string) {
  try {
    // 1. Primary: Self-hosted local vLLM / Ollama instance for sub-50ms TTFT
    const localHost = createOpenAI({
      baseURL: process.env.LOCAL_VLLM_ENDPOINT,
      apiKey: 'none',
    });

    return await streamText({
      model: localHost('qwen-2.5-coder'),
      prompt,
    });
  } catch (err) {
    console.warn('Local model cluster unavailable. Gracefully falling back to frontier cloud API...');
    
    // 2. Fallback: Cloud API gateway
    const cloudGateway = createOpenAI({
      apiKey: process.env.CLOUD_AI_API_KEY,
    });

    return await streamText({
      model: cloudGateway('gpt-4o-mini'),
      prompt,
    });
  }
}

1/4 🛠️ Pragmatic web architecture from @leerob:
Don't fall into the "Edge Database Trap." Moving your serverless function to the Edge sounds fast, but if your database is in us-east-1, multi-query waterfalls across the Atlantic will kill your TTFB.

2/4 The Golden Rule:
Compute belongs next to data. Use Edge workers for auth verification, geo-routing, and static headers. Use regional serverless compute for database queries.

3/4 Open Models vs. Cloud APIs:
The frontier isn't just closed APIs anymore. Teams are running quantized Qwen and Llama models on their own bare-metal clusters to guarantee sub-100ms Time-To-First-Token (TTFT) and fixed cost structures.

4/4 Follow @leerob for hands-on, production-grade Next.js architectures, React tips, and developer experience insights.