技能 人工智能 Groq应用参考架构设计

Groq应用参考架构设计

v20260423
groq-reference-architecture
本指南提供了一套基于Groq LPU推理API的生产级参考架构蓝图。它系统地展示了如何根据延迟、质量或成本需求实现模型路由、流式数据管道,并构建了多提供商故障转移机制(如Groq到OpenAI备份)。适用于设计和优化复杂的、高弹性的AI应用系统。
获取技能
278 次下载
概览

Groq Reference Architecture

Overview

Production architecture for applications built on Groq's LPU inference API. Covers model routing by latency requirements, streaming pipelines, multi-provider fallback, and the middleware layer that ties it together.

Architecture Diagram

┌──────────────────────────────────────────────────────────────┐
│                     Application Layer                         │
│  Chat UI  │  API Backend  │  Batch Processor  │  Agent       │
└─────┬─────┴──────┬────────┴────────┬──────────┴──────┬───────┘
      │            │                 │                 │
      ▼            ▼                 ▼                 ▼
┌──────────────────────────────────────────────────────────────┐
│                    Groq Service Layer                         │
│  ┌─────────────┐  ┌────────────┐  ┌─────────────────────┐   │
│  │ Model Router │  │ Middleware │  │ Fallback Chain      │   │
│  │             │  │            │  │                     │   │
│  │ speed →     │  │ Cache      │  │ Groq (primary)      │   │
│  │   8b-instant│  │ Rate Guard │  │   ↓ 429/5xx         │   │
│  │ quality →   │  │ Metrics    │  │ Groq (fallback model)│  │
│  │   70b-versa.│  │ Logging    │  │   ↓ still failing    │   │
│  │ vision →    │  │ Retry      │  │ OpenAI (backup)     │   │
│  │   llama-4   │  │            │  │   ↓ also failing     │   │
│  │ audio →     │  │            │  │ Graceful degrade    │   │
│  │   whisper   │  │            │  │                     │   │
│  └─────────────┘  └────────────┘  └─────────────────────┘   │
└──────────────────────────────────────────────────────────────┘

Project Structure

src/
├── groq/
│   ├── client.ts            # Singleton Groq client
│   ├── models.ts            # Model constants and capabilities
│   ├── router.ts            # Model selection logic
│   ├── middleware.ts         # Cache, rate limit, metrics
│   ├── fallback.ts          # Multi-provider fallback chain
│   └── types.ts             # Shared types
├── services/
│   ├── chat.ts              # Chat completion service
│   ├── transcription.ts     # Audio transcription (Whisper)
│   ├── extraction.ts        # Structured data extraction
│   └── batch.ts             # Batch processing service
└── api/
    ├── chat.ts              # HTTP endpoint
    ├── transcribe.ts        # Audio endpoint
    └── health.ts            # Health check

Instructions

Step 1: Model Registry

// src/groq/models.ts
export interface ModelSpec {
  id: string;
  tier: "speed" | "quality" | "vision" | "audio";
  contextWindow: number;
  maxOutput: number;
  speedTokPerSec: number;
  inputCostPer1M: number;
  outputCostPer1M: number;
  capabilities: ("text" | "tools" | "json" | "vision" | "audio")[];
}

export const MODELS: Record<string, ModelSpec> = {
  "llama-3.1-8b-instant": {
    id: "llama-3.1-8b-instant",
    tier: "speed",
    contextWindow: 131_072,
    maxOutput: 8_192,
    speedTokPerSec: 560,
    inputCostPer1M: 0.05,
    outputCostPer1M: 0.08,
    capabilities: ["text", "tools", "json"],
  },
  "llama-3.3-70b-versatile": {
    id: "llama-3.3-70b-versatile",
    tier: "quality",
    contextWindow: 131_072,
    maxOutput: 32_768,
    speedTokPerSec: 280,
    inputCostPer1M: 0.59,
    outputCostPer1M: 0.79,
    capabilities: ["text", "tools", "json"],
  },
  "meta-llama/llama-4-scout-17b-16e-instruct": {
    id: "meta-llama/llama-4-scout-17b-16e-instruct",
    tier: "vision",
    contextWindow: 131_072,
    maxOutput: 8_192,
    speedTokPerSec: 460,
    inputCostPer1M: 0.11,
    outputCostPer1M: 0.34,
    capabilities: ["text", "tools", "json", "vision"],
  },
  "whisper-large-v3-turbo": {
    id: "whisper-large-v3-turbo",
    tier: "audio",
    contextWindow: 0,
    maxOutput: 0,
    speedTokPerSec: 0,
    inputCostPer1M: 0,
    outputCostPer1M: 0,
    capabilities: ["audio"],
  },
};

Step 2: Model Router

// src/groq/router.ts
import { MODELS, ModelSpec } from "./models";

interface RoutingRequest {
  maxLatencyMs?: number;
  needsVision?: boolean;
  needsTools?: boolean;
  needsJSON?: boolean;
  contextLength?: number;
  costSensitive?: boolean;
}

export function selectModel(req: RoutingRequest): ModelSpec {
  if (req.needsVision) return MODELS["meta-llama/llama-4-scout-17b-16e-instruct"];

  if (req.costSensitive || (req.maxLatencyMs && req.maxLatencyMs < 100)) {
    return MODELS["llama-3.1-8b-instant"];
  }

  if (req.needsTools || req.needsJSON) {
    return MODELS["llama-3.3-70b-versatile"];
  }

  // Default: speed tier
  return MODELS["llama-3.1-8b-instant"];
}

Step 3: Middleware Layer

// src/groq/middleware.ts
import Groq from "groq-sdk";
import { LRUCache } from "lru-cache";
import { createHash } from "crypto";

const cache = new LRUCache<string, any>({ max: 500, ttl: 10 * 60_000 });

export async function completionWithMiddleware(
  groq: Groq,
  model: string,
  messages: any[],
  options?: { maxTokens?: number; temperature?: number; stream?: boolean }
) {
  const temp = options?.temperature ?? 0.7;

  // Cache check (only for deterministic requests)
  if (temp === 0 && !options?.stream) {
    const key = createHash("sha256").update(JSON.stringify({ model, messages })).digest("hex");
    const cached = cache.get(key);
    if (cached) return cached;
  }

  // Metrics
  const start = performance.now();

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

  const latency = performance.now() - start;

  // Emit metrics
  emitMetrics({
    model,
    latencyMs: Math.round(latency),
    tokens: (response as any).usage?.total_tokens ?? 0,
    cached: false,
  });

  // Cache deterministic responses
  if (temp === 0 && !options?.stream) {
    const key = createHash("sha256").update(JSON.stringify({ model, messages })).digest("hex");
    cache.set(key, response);
  }

  return response;
}

function emitMetrics(data: any) {
  // Plug in your metrics system: Prometheus, Datadog, etc.
  console.log(`[groq-metrics] ${JSON.stringify(data)}`);
}

Step 4: Fallback Chain

// src/groq/fallback.ts
import Groq from "groq-sdk";

export async function completionWithFallback(
  groq: Groq,
  messages: any[],
  options?: { primaryModel?: string; maxTokens?: number }
) {
  const primary = options?.primaryModel || "llama-3.3-70b-versatile";
  const fallbackModel = "llama-3.1-8b-instant";

  // Attempt 1: Primary model
  try {
    return await groq.chat.completions.create({
      model: primary,
      messages,
      max_tokens: options?.maxTokens ?? 1024,
    });
  } catch (err: any) {
    if (err.status !== 429 && err.status < 500) throw err;
    console.warn(`Primary model ${primary} failed (${err.status}), trying fallback`);
  }

  // Attempt 2: Fallback model (different rate limit pool)
  try {
    return await groq.chat.completions.create({
      model: fallbackModel,
      messages,
      max_tokens: options?.maxTokens ?? 1024,
    });
  } catch (err: any) {
    console.warn(`Groq fallback also failed (${err.status})`);
  }

  // Attempt 3: Graceful degradation
  return {
    choices: [{
      message: {
        role: "assistant" as const,
        content: "Service temporarily unavailable. Please try again in a moment.",
      },
      finish_reason: "stop" as const,
    }],
    model: "fallback",
    usage: { prompt_tokens: 0, completion_tokens: 0, total_tokens: 0 },
  };
}

Step 5: Streaming Pipeline

// src/groq/streaming.ts
import Groq from "groq-sdk";

export async function* streamCompletion(
  groq: Groq,
  messages: any[],
  model = "llama-3.3-70b-versatile"
): AsyncGenerator<{ type: "token" | "done" | "error"; content?: string; error?: string }> {
  try {
    const stream = await groq.chat.completions.create({
      model,
      messages,
      stream: true,
      max_tokens: 2048,
    });

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

    yield { type: "done" };
  } catch (err: any) {
    yield { type: "error", error: err.message };
  }
}

Integration Patterns

Pattern When to Use Groq Feature
Direct completion Simple request/response chat.completions.create
Streaming SSE Real-time chat UI stream: true
Tool calling Agent with function execution tools parameter
JSON extraction Structured data from text response_format: json_object
Batch processing High-volume document processing Queue + rate limiting
Audio transcription Voice input audio.transcriptions.create
Vision analysis Image understanding Llama 4 Scout/Maverick

Error Handling

Issue Cause Solution
429 on primary model RPM/TPM exceeded Fall back to different model
High latency Wrong model tier Route to 8b-instant for latency-critical paths
Context overflow Input > 128K tokens Truncate or chunk input
Vision errors Wrong model for images Use Llama 4 Scout full model path

Resources

Next Steps

For multi-environment deployment, see groq-multi-env-setup.

信息
Category 人工智能
Name groq-reference-architecture
版本 v20260423
大小 10.97KB
更新时间 2026-04-28
语言