技能 人工智能 LangChain生产环境就绪度清单

LangChain生产环境就绪度清单

v20260423
langchain-prod-checklist
这是一份全面的LangChain应用程序生产环境就绪度清单。它系统性地覆盖了从配置管理、错误处理、可观测性(如LangSmith)、性能优化、安全加固(防止提示注入、处理PII)到严格测试和部署策略等多个维度,帮助开发者确保AI应用具备高可靠性和可维护性,适用于应用上线前或现有系统的审计。
获取技能
131 次下载
概览

LangChain Production Checklist

Overview

Comprehensive go-live checklist for deploying LangChain applications to production. Covers configuration, resilience, observability, performance, security, testing, deployment, and cost management.

1. Configuration & Secrets

  • All API keys in secrets manager (not .env in production)
  • Environment-specific configs (dev/staging/prod) validated with Zod
  • Startup validation fails fast on missing config
  • .env files in .gitignore
// Startup validation
import { z } from "zod";

const ProdConfig = z.object({
  OPENAI_API_KEY: z.string().startsWith("sk-"),
  LANGSMITH_API_KEY: z.string().startsWith("lsv2_"),
  NODE_ENV: z.literal("production"),
});

try {
  ProdConfig.parse(process.env);
} catch (e) {
  console.error("Invalid production config:", e);
  process.exit(1);
}

2. Error Handling & Resilience

  • maxRetries configured on all models (3-5)
  • timeout set on all models (30-60s)
  • Fallback models configured with .withFallbacks()
  • Error responses return safe messages (no stack traces to users)
const model = new ChatOpenAI({
  model: "gpt-4o-mini",
  maxRetries: 5,
  timeout: 30000,
}).withFallbacks({
  fallbacks: [new ChatAnthropic({ model: "claude-sonnet-4-20250514" })],
});

3. Observability

  • LangSmith tracing enabled (LANGSMITH_TRACING=true)
  • LANGCHAIN_CALLBACKS_BACKGROUND=true (non-serverless only)
  • Structured logging on all LLM/tool calls
  • Prometheus metrics exported (requests, latency, tokens, errors)
  • Alerting rules configured (error rate >5%, P95 latency >5s)

4. Performance

  • Caching enabled for repeated queries (Redis or SQLite)
  • maxConcurrency set on batch operations
  • Streaming enabled for user-facing responses
  • Connection pooling configured
  • Prompt length optimized (no unnecessary verbosity)

5. Security

  • User input isolated in human messages (never in system prompts)
  • Input length limits enforced
  • Prompt injection patterns logged/flagged
  • Tools restricted to allowlisted operations
  • LLM output validated before display (no PII/key leakage)
  • Audit logging on all LLM and tool calls
  • Rate limiting per user/IP

6. Testing

  • Unit tests for all chains (using FakeListChatModel, no API calls)
  • Integration tests with real LLMs (gated behind CI secrets)
  • RAG pipeline validation (retrieval relevance + no hallucination)
  • Tool unit tests (valid input, invalid input, error cases)
  • Load testing completed (concurrent users, batch operations)

7. Deployment

  • Health check endpoint returns LLM connectivity status
  • Graceful shutdown handles in-flight requests
  • Rolling deployment (zero downtime)
  • Rollback procedure documented and tested
  • Container resource limits set (memory, CPU)
// Health check endpoint
app.get("/health", async (_req, res) => {
  const checks: Record<string, string> = { server: "ok" };

  try {
    await model.invoke("ping");
    checks.llm = "ok";
  } catch (e: any) {
    checks.llm = `error: ${e.message.slice(0, 100)}`;
  }

  const healthy = Object.values(checks).every((v) => v === "ok");
  res.status(healthy ? 200 : 503).json({ status: healthy ? "healthy" : "degraded", checks });
});

// Graceful shutdown
process.on("SIGTERM", async () => {
  console.log("Shutting down gracefully...");
  server.close(() => process.exit(0));
  setTimeout(() => process.exit(1), 10000); // force after 10s
});

8. Cost Management

  • Token usage tracking callback attached
  • Daily/monthly budget limits enforced
  • Model tiering: cheap model for simple tasks, powerful for complex
  • Cost alerts configured (Slack/email on threshold)
  • Cost per user/tenant tracked

Pre-Launch Validation Script

async function validateProduction() {
  const results: Record<string, string> = {};

  // 1. Config
  try {
    ProdConfig.parse(process.env);
    results["Config"] = "PASS";
  } catch { results["Config"] = "FAIL: missing env vars"; }

  // 2. LLM connectivity
  try {
    await model.invoke("ping");
    results["LLM"] = "PASS";
  } catch (e: any) { results["LLM"] = `FAIL: ${e.message.slice(0, 50)}`; }

  // 3. Fallback
  try {
    const fallbackModel = model.withFallbacks({ fallbacks: [fallback] });
    await fallbackModel.invoke("ping");
    results["Fallback"] = "PASS";
  } catch { results["Fallback"] = "FAIL"; }

  // 4. LangSmith
  results["LangSmith"] = process.env.LANGSMITH_TRACING === "true" ? "PASS" : "WARN: disabled";

  // 5. Health endpoint
  try {
    const res = await fetch("http://localhost:8000/health");
    results["Health"] = res.ok ? "PASS" : "FAIL";
  } catch { results["Health"] = "FAIL: not reachable"; }

  console.table(results);
  const allPass = Object.values(results).every((v) => v === "PASS");
  console.log(allPass ? "READY FOR PRODUCTION" : "ISSUES FOUND - FIX BEFORE LAUNCH");
  return allPass;
}

Error Handling

Issue Cause Fix
API key missing at startup Secrets not mounted Check deployment config
No fallback on outage .withFallbacks() not configured Add fallback model
LangSmith trace gaps Background callbacks in serverless Set LANGCHAIN_CALLBACKS_BACKGROUND=false
Cache miss storm Redis down Implement graceful degradation

Resources

Next Steps

After launch, use langchain-observability for monitoring and langchain-incident-runbook for incident response.

信息
Category 人工智能
Name langchain-prod-checklist
版本 v20260423
大小 4.5KB
更新时间 2026-04-28
语言