Back to Details

ai-engineer

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

AI Engineer

AI工程师

Purpose

用途

Provides expertise in end-to-end AI system development, from LLM integration to production deployment. Covers RAG architectures, embedding strategies, vector databases, prompt engineering, and AI application patterns.
提供端到端AI系统开发的专业知识，从LLM集成到生产部署。涵盖RAG架构、嵌入策略、向量数据库、提示工程和AI应用模式。

When to Use

适用场景

  • Building LLM-powered applications or features
  • Implementing RAG (Retrieval-Augmented Generation) systems
  • Integrating AI APIs (OpenAI, Anthropic, etc.)
  • Designing embedding and vector search pipelines
  • Building chatbots or conversational AI
  • Implementing AI agents with tool use
  • Optimizing AI system latency and cost
  • 构建由LLM驱动的应用或功能
  • 实现RAG（检索增强生成）系统
  • 集成AI API（OpenAI、Anthropic等）
  • 设计嵌入与向量搜索流水线
  • 构建聊天机器人或对话式AI
  • 实现具备工具调用能力的AI Agent
  • 优化AI系统的延迟与成本

Quick Start

快速入门

Invoke this skill when:
  • Building LLM-powered applications or features
  • Implementing RAG systems with vector databases
  • Integrating AI APIs into applications
  • Designing embedding and retrieval pipelines
  • Building conversational AI or agents
Do NOT invoke when:
  • Training custom ML models from scratch (use ml-engineer)
  • Deploying ML models to production infrastructure (use mlops-engineer)
  • Managing multi-agent coordination (use agent-organizer)
  • Optimizing LLM serving infrastructure (use llm-architect)
调用此技能的场景：
  • 构建由LLM驱动的应用或功能
  • 结合向量数据库实现RAG系统
  • 将AI API集成到应用中
  • 设计嵌入与检索流水线
  • 构建对话式AI或Agent
请勿调用的场景：
  • 从零开始训练自定义ML模型（请使用ml-engineer）
  • 将ML模型部署到生产基础设施（请使用mlops-engineer）
  • 管理多Agent协作（请使用agent-organizer）
  • 优化LLM服务基础设施（请使用llm-architect）

Decision Framework

决策框架

AI Feature Type:
├── Simple Q&A → Direct LLM API call
├── Knowledge-based answers → RAG pipeline
├── Multi-step reasoning → Chain-of-thought or agents
├── External actions needed → Tool-use agents
├── Real-time data → Streaming + function calling
└── Complex workflows → Multi-agent orchestration
AI Feature Type:
├── Simple Q&A → Direct LLM API call
├── Knowledge-based answers → RAG pipeline
├── Multi-step reasoning → Chain-of-thought or agents
├── External actions needed → Tool-use agents
├── Real-time data → Streaming + function calling
└── Complex workflows → Multi-agent orchestration

Core Workflows

核心工作流

1. RAG Pipeline Implementation

1. RAG流水线实现

  1. Chunk documents with appropriate strategy
  2. Generate embeddings using suitable model
  3. Store in vector database with metadata
  4. Implement semantic search with reranking
  5. Construct prompts with retrieved context
  6. Add evaluation and monitoring
  1. 采用合适的策略分割文档
  2. 使用适配的模型生成嵌入向量
  3. 存储到带元数据的向量数据库中
  4. 实现带重排序的语义搜索
  5. 结合检索到的上下文构建提示词
  6. 添加评估与监控机制

2. LLM Integration

2. LLM集成

  1. Select appropriate model for use case
  2. Design prompt templates with versioning
  3. Implement structured output parsing
  4. Add retry logic and fallbacks
  5. Monitor token usage and costs
  6. Cache responses where appropriate
  1. 根据使用场景选择适配的模型
  2. 设计带版本控制的提示词模板
  3. 实现结构化输出解析
  4. 添加重试逻辑与降级方案
  5. 监控令牌使用量与成本
  6. 在合适的场景缓存响应

3. AI Agent Development

3. AI Agent开发

  1. Define agent capabilities and tools
  2. Implement tool interfaces with validation
  3. Design agent loop with termination conditions
  4. Add guardrails and safety checks
  5. Implement logging and tracing
  6. Test edge cases and failure modes
  1. 定义Agent的能力与可用工具
  2. 实现带验证的工具接口
  3. 设计带有终止条件的Agent循环
  4. 添加防护机制与安全检查
  5. 实现日志与追踪功能
  6. 测试边缘案例与故障模式

Best Practices

最佳实践

  • Version prompts alongside application code
  • Use structured outputs (JSON mode) for reliability
  • Implement semantic caching for common queries
  • Add human-in-the-loop for critical decisions
  • Monitor hallucination rates and retrieval quality
  • Design for graceful degradation when AI fails
  • 随应用代码一同对提示词进行版本控制
  • 使用结构化输出（JSON模式）提升可靠性
  • 为常见查询实现语义缓存
  • 针对关键决策加入人工介入环节
  • 监控幻觉率与检索质量
  • 设计AI故障时的优雅降级方案

Anti-Patterns

反模式

Anti-PatternProblemCorrect Approach
Prompt in codeHard to iterate and testUse prompt templates with versioning
No evaluationUnknown quality in productionImplement eval pipelines
Synchronous LLM callsSlow user experienceUse streaming responses
Unbounded contextToken limits and costImplement context windowing
No fallbacksSystem fails on API errorsAdd retry logic and alternatives
反模式问题正确做法
提示词硬编码在代码中难以迭代与测试使用带版本控制的提示词模板
未做评估生产环境中质量未知实现评估流水线
同步调用LLM接口用户体验缓慢使用流式响应
无限制的上下文受令牌限制且成本高昂实现上下文窗口管理
未设置降级方案API出错时系统崩溃添加重试逻辑与替代方案