生成式AI与智能体知识库

GenAI & Agentic AI — From Fundamentals to Production Design

A practical engineering knowledge base for building and evaluating Generative AI and agentic systems. Follow a structured learning path, examine architecture and production trade-offs, or explore eight runnable educational case studies.

For: AI engineers, technical leads, architects, and engineering leaders assessing how AI systems should be designed, evaluated, secured, and operated.

Explore in 2 minutes

What makes this repository useful

• Architecture-first explanations: not just what a component does, but where it fits, what can fail, and which alternatives to consider.
• Production decision points: evaluations, observability, access controls, failure handling, latency, scalability, and unit economics.
• Executable learning examples: eight case-study scenarios with synthetic data and tests; examples are educational, not claims of deployed enterprise systems.
• Leadership-level trade-offs: when to use RAG versus fine-tuning, single versus multi-agent designs, shared platforms versus product-specific implementations, and when simpler automation is preferable.

Suggested first read for a CTO or AI engineering leader: AI Platform Architecture → Evaluation & Observability → Security & Guardrails → Applied Case Studies.

Scope: This is an educational and technical reference portfolio. Runnable case studies use teaching implementations and synthetic scenarios; production architecture discussions are design guidance, not evidence of live customer deployments.

Repository Structure

The numbered folders are the primary navigation. Start with 01 and move forward, or jump directly to the engineering area you need.

Explore Runnable Case Studies

The Applied AI Case Studies section turns concepts into executable examples with synthetic data, tests, architecture diagrams, safety controls, and production evolution. Start with Customer Service for retrieval and action confirmation, Agent Pipeline for orchestration, or Project Status for evidence-grounded business reporting.

The case studies are deterministic teaching implementations unless explicitly stated otherwise; the production diagrams describe possible future architectures.

What You Will Learn

Learning Journey

flowchart LR
    A[AI & ML] --> B[Transformers]
    B --> C[LLMs]
    C --> D[Prompt Engineering]
    D --> E[Embeddings]
    E --> ER[Search & Retrieval]
    ER --> F[RAG]
    F --> G[AI Agents]
    G --> H[Tool Use]
    H --> I[Memory]
    I --> J[Planning]
    J --> K[Agentic RAG]
    K --> L[Multi-Agent Systems]
    L --> MM[Multimodal & Conversational]
    MM --> AT[Agent Training & Adaptation]
    AT --> RT[Runtime / State Machines / Graphs]
    RT --> M[MCP & A2A]
    M --> N[Evaluation & Production]
    N --> O[System Design]

Start Here

Beginner

AI & ML Fundamentals → Foundation Models → Tokenization → Transformers → Large Language Models → Prompt Engineering → Embeddings → Search & Retrieval → RAG

Applied AI Engineer

LLMs → Prompting → RAG → Tool Use → Agents → Evaluation & Observability → Security & Guardrails → Reliability & Resilience → Context Engineering → Performance, Latency & Cost → AI Data Engineering → LLMOps & AI Deployment → Production

Agentic AI Engineer

AI Agents → Agent Architecture & Agent Loops → Tool Use & Orchestration → Agent Memory → Planning & Reasoning → Agentic RAG → Single vs Multi-Agent → Multi-Agent Collaboration → Multi-Agent Systems Architecture → Multimodal & Conversational Agents → Agent Training & Adaptation → Agent Runtime, State Machines & Protocols → MCP → A2A

System Design / Interview Preparation

LLM Architecture → AI Product Architecture → AI Platform Architecture → Production RAG Design → Agent Architecture → Tool Orchestration → Memory → Planning → Agentic RAG → Multi-Agent Architecture → Evaluation & Observability → Security & Guardrails → Reliability & Resilience → Context Engineering → Performance, Latency & Cost → Trade-offs

See the complete Learning Roadmap.

Core Knowledge Base

1. Foundations

• AI & ML Fundamentals
• AI Foundation Models
• Tokenization
• Transformers
• Large Language Models
• Pre-training vs Fine-tuning

2. LLM & GenAI Engineering

• Prompt Engineering
• Embeddings & Vector Databases
• Search & Retrieval Engineering — Deep Dive
• Retrieval-Augmented Generation
• LLM Inference
• Fine-Tuning
• LLM Hallucination
• AI Frameworks

3. Agentic AI

• AI Agents
• Agent Architecture & Agent Loops — Deep Dive
• Tool Use & Agent Orchestration — Deep Dive
• Agent Memory Architecture — Deep Dive
• Planning & Reasoning Patterns — Deep Dive
• Agentic RAG — Deep Dive
• Single-Agent vs Multi-Agent
• Multi-Agent Collaboration
• Multi-Agent Systems Architecture — Deep Dive
• Multimodal & Conversational Agents — Deep Dive
• Agent Training & Adaptation — Deep Dive
• Agent Runtime, State Machines & Communication Protocols — Deep Dive — naive-agent failure modes, FSMs, termination, graph orchestration/LangGraph concepts, Google ADK runtime/sessions/events, MCP + A2A integration
• Agent Skills & Capability Engineering — Deep Dive — skills vs tools/workflows/agents, progressive disclosure, discovery, composition, contracts, MCP/A2A relationships, security and lifecycle
• Voice & Realtime Agent Engineering — Deep Dive — ASR/TTS, speech-to-speech, VAD/endpointing, interruption, WER, acoustic robustness, voice RAG, realtime latency, evaluation and cost

4. Agent Runtime & Interoperability

• Model Context Protocol — MCP
• Agent-to-Agent — A2A

5. Production AI Engineering

• Evaluation & Observability — Deep Dive
• Security & Guardrails — Deep Dive
• Reliability & Resilience — Deep Dive
• Context Engineering — Deep Dive
• Performance, Latency & Cost — Deep Dive
• LLMOps & AI Deployment — Deep Dive — versioned behavioral releases, evaluation-gated CI/CD, progressive delivery, rollback, hosted/self-hosted inference, RAG/index and agent migrations, GPU serving, operational readiness, supply-chain security, and delivery economics.
• AI Data Engineering — Deep Dive — source-to-AI data architecture covering batch/stream/CDC, multimodal/document processing, chunking, embeddings and index generations, contracts, ACLs, datasets, agent state/memory, lineage, freshness, deletion, recovery, scaling, and data economics.
• Production Engineering Hub

6. Architecture & System Design

• AI Product Architecture
• AI Platform Architecture — Deep Dive — control/data/evidence/developer planes, model gateway and serving, shared RAG/tool/state capabilities, evaluation/observability, multi-tenancy, security, reliability, capacity, cost, developer experience, and platform governance.
• Production RAG System Design
• System Design Hub

The system-design material covers requirements, architecture, data/control flow, retrieval, orchestration, memory, security, evaluation, observability, scalability, latency, cost, failure handling, and explicit trade-offs.

7. Design Patterns

The Design Patterns catalog organizes reusable RAG, agent, memory, orchestration, and reliability patterns.

How Topics Are Developed

Major engineering topics progressively follow a common structure:

1. What is it?
2. Why does it exist?
3. Mental model
4. How it works
5. Architecture
6. Step-by-step flow
7. Practical example
8. Implementation
9. Design decisions and trade-offs
10. Failure modes
11. Evaluation
12. Security
13. Performance, scalability, latency, and cost
14. Production considerations
15. Interview questions
16. Key takeaways

Architecture-First Learning

The goal is not only to know definitions. A reader should be able to answer:

• How does this component work?
• Where does it fit in an architecture?
• When should I use it?
• What breaks in production?
• How do I evaluate it?
• How does it scale?
• What alternatives exist?
• What trade-offs would I discuss in a system-design interview?

Contributing

Contributions that improve technical accuracy, diagrams, examples, system-design discussions, implementation quality, or production guidance are welcome.

Goal: bridge the gap between learning GenAI concepts and engineering production-grade intelligent systems.