groq-reference-architecture（groq 参考架构）

Groq Reference Architecture

Overview

Production architecture for ultra-fast LLM inference with Groq LPU. Covers model routing by latency requirements, streaming pipelines, fallback strategies, and integration patterns for real-time AI applications.

Prerequisites

• Groq API key
• groq-sdk npm package
• Understanding of model capabilities (Llama, Mixtral)
• Monitoring for latency and token usage

Architecture Diagram

┌─────────────────────────────────────────────────────┐
│                Application Layer                     │
│  Chat UI │ API Backend │ Batch Processor │ Agent    │
└──────────┬──────────────┬───────────────┬───────────┘
           │              │               │
           ▼              ▼               ▼
┌─────────────────────────────────────────────────────┐
│              Model Router                            │
│  ┌───────────────┐  ┌──────────────┐  ┌───────────┐ │
│  │ Speed Tier    │  │ Quality Tier │  │ Long Ctx  │ │
│  │ llama-3.1-8b  │  │ llama-3.3-70b│  │ mixtral   │ │
│  │ (80ms TTFT)   │  │ (200ms TTFT) │  │ (32k ctx) │ │
│  └───────────────┘  └──────────────┘  └───────────┘ │
├─────────────────────────────────────────────────────┤
│              Middleware                               │
│  Prompt Cache │ Rate Limiter │ Token Counter │ Log  │
├─────────────────────────────────────────────────────┤
│              Fallback Layer                          │
│  Groq Primary → OpenAI Fallback → Local Model      │
└─────────────────────────────────────────────────────┘

Instructions

Step 1: Model Router Pattern

import Groq from 'groq-sdk';

const groq = new Groq({ apiKey: process.env.GROQ_API_KEY });

type ModelTier = 'speed' | 'quality' | 'long-context';

const MODEL_MAP: Record<ModelTier, string> = {
  speed: 'llama-3.1-8b-instant',
  quality: 'llama-3.3-70b-versatile',
  'long-context': 'mixtral-8x7b-32768',  # 32768 = configured value
};

function selectModel(options: {
  maxLatencyMs?: number;
  contextLength?: number;
  needsReasoning?: boolean;
}): string {
  if (options.contextLength && options.contextLength > 8192)  # 8192: 8 KB
    return MODEL_MAP['long-context'];
  if (options.maxLatencyMs && options.maxLatencyMs < 150)
    return MODEL_MAP.speed;
  if (options.needsReasoning) return MODEL_MAP.quality;
  return MODEL_MAP.speed;
}

Step 2: Completion Service with Middleware

interface CompletionOptions {
  messages: any[];
  tier?: ModelTier;
  stream?: boolean;
  maxTokens?: number;
  temperature?: number;
}

async function complete(options: CompletionOptions) {
  const model = MODEL_MAP[options.tier || 'speed'];
  const start = performance.now();

  const response = await groq.chat.completions.create({
    model,
    messages: options.messages,
    stream: options.stream || false,
    max_tokens: options.maxTokens || 1024,  # 1024: 1 KB
    temperature: options.temperature ?? 0.7,
  });

  const latency = performance.now() - start;
  logMetrics({ model, latency, tokens: response.usage });

  return response;
}

Step 3: Streaming Pipeline

async function* streamCompletion(messages: any[], tier: ModelTier = 'quality') {
  const model = MODEL_MAP[tier];

  const stream = await groq.chat.completions.create({
    model,
    messages,
    stream: true,
    max_tokens: 2048,  # 2048: 2 KB
  });

  for await (const chunk of stream) {
    const content = chunk.choices[0]?.delta?.content;
    if (content) yield content;
  }
}

// Usage with Express SSE
app.get('/api/chat', async (req, res) => {
  res.setHeader('Content-Type', 'text/event-stream');

  for await (const token of streamCompletion(messages, 'quality')) {
    res.write(`data: ${JSON.stringify({ token })}\n\n`);
  }

  res.write('data: [DONE]\n\n');
  res.end();
});

Step 4: Fallback Chain

async function completionWithFallback(messages: any[]) {
  try {
    return await complete({ messages, tier: 'quality' });
  } catch (error: any) {
    if (error.status === 429 || error.status >= 500) {  # 500: HTTP 429 Too Many Requests
      console.warn('Groq unavailable, falling back to OpenAI');
      return openai.chat.completions.create({
        model: 'gpt-4o-mini',
        messages,
      });
    }
    throw error;
  }
}

Error Handling

Examples

Multi-Model Pipeline

async function analyzeDocument(doc: string) {
  // Fast extraction with speed tier
  const summary = await complete({
    messages: [{ role: 'user', content: `Summarize: ${doc}` }],
    tier: 'speed',
  });

  // Deep analysis with quality tier
  const analysis = await complete({
    messages: [{ role: 'user', content: `Analyze in detail: ${summary}` }],
    tier: 'quality',
  });

  return { summary, analysis };
}

Resources

• Groq API Documentation
• Groq Model Cards
• Groq Rate Limits

Output

• Configuration files or code changes applied to the project
• Validation report confirming correct implementation
• Summary of changes made and their rationale