Handbook Topic Index¶

This handbook prepares you for AI engineering interviews through the layers of a system, not a catalog of products. Its nine topics cover model foundations, multimodal capabilities, protocols, application architectures, frameworks, production engineering, safety and governance, and field delivery. A useful answer explains both how a mechanism works and what limits the design.

The hierarchy is topic, module, chapter. A topic index shows how its modules relate; a module index lists its chapters. Relative links connect related explanations across topics. English is the primary manuscript, and every reader-facing page has a Simplified Chinese companion.

For a continuous reading path, start with the book contents. The book has nine parts and preserves chapter numbering within each part. The title page, preface, reading guide, and acknowledgments are separate from the technical chapters. Choose a language on the website or download that language's EPUB; neither format has its own independently maintained prose.

The central bibliography contains the numbered sources linked by superscripts throughout the chapters, plus each chapter’s reading suggestions and source qualifications.

The overall structure¶

flowchart TB
    subgraph L1["Layer 1: Models and multimodal capabilities"]
        LLM["LLM<br/>Transformers / training / inference / serving"]
        MM["Multimodal AI<br/>Vision / speech / image and video generation"]
    end

    subgraph L2["Layer 2: Protocols and interfaces"]
        TOOLS["Tools<br/>Function calling / MCP / skills / A2A"]
    end

    subgraph L3["Layer 3: Application architectures"]
        AGENT["Agents<br/>Harnesses / planning / memory / multi-agent systems"]
        RAG["RAG<br/>Indexing / retrieval / reranking / generation"]
    end

    subgraph L4["Layer 4: Framework implementations"]
        FW["Frameworks and orchestration<br/>LangChain / LlamaIndex / DSPy / Semantic Kernel"]
    end

    subgraph L5["Layer 5: Production and governance"]
        ENG["AI Engineering / LLMOps<br/>Evaluation / observability / releases / reliability"]
        SAFE["AI safety and governance<br/>Threats / isolation / red teaming / audit"]
    end

    subgraph L6["Layer 6: Field delivery"]
        FDE["FDE<br/>Discovery / acceptance / integration / delivery / reuse"]
    end

    LLM --> MM
    LLM --> TOOLS
    TOOLS --> AGENT
    LLM --> RAG
    MM --> AGENT
    MM --> RAG
    AGENT --> FW
    RAG --> FW
    FW --> ENG
    ENG --> SAFE
    ENG --> FDE
    SAFE --> FDE
    FDE -.field feedback.-> FW
    AGENT -.risk inputs.-> SAFE
    RAG -.risk inputs.-> SAFE
    RAG -.knowledge augmentation.-> AGENT

The layers organize the material; they do not require a project to adopt every component. Evaluation, safety, and customer acceptance run through design and delivery. A simple task may need only a model API call, not an agent or an orchestration framework.

Topics¶

Layer	Topic	Coverage	Chapters	Start here
Foundations	LLM	Transformers, training and alignment, decoding, quantization, MoE, serving, evaluation, and model selection	23	LLM
Model capabilities	Multimodal AI	Fusion, VLMs, Document AI, speech, image and video generation, evaluation, and serving	10	Multimodal AI
Protocols and interfaces	Tools	Function calling, tool learning, MCP, skills, A2A, transports, security, and model gateways	15	Tools
Application architectures	Agents	Architecture, harnesses, memory, planning, multi-agent systems, evaluation, safety, code editing, and post-training	25	Agents
Application architectures	RAG	Document processing, indexing, retrieval and reranking, multimodal data, generation, evaluation, updates, security, and Text-to-SQL	22	RAG
Framework implementations	Frameworks and orchestration	LangChain, LangGraph, LlamaIndex, DSPy, Semantic Kernel, lightweight agent frameworks, and migration	23	Frameworks
Production engineering	AI Engineering	LLMOps, gateways and fallback, evaluation, observability, CI/CD, SLOs, cost, and feedback data	13	AI Engineering
Safety and governance	AI safety and governance	Threat modeling, prompt attacks, supply chains, privacy, execution isolation, red teaming, governance, and audit	10	Safety and governance
Field delivery	FDE	Discovery, product collaboration, acceptance, integration, delivery, scope changes, proofs of concept, project memory, and handover	2	FDE

Where shared concepts are explained¶

A concept can appear at several layers without needing several copies of its full explanation. The following ownership keeps those explanations consistent:

Concept	Main explanation	Referenced by	Difference in focus
Chain of thought	LLM	Agent planning, RAG generation	LLM covers mechanisms; Agents explains their role in planning
Hallucination	LLM	RAG generation, Agent safety	LLM covers causes; RAG covers generation constrained by retrieved evidence
KV cache / prompt caching	LLM	Agent context, AI Engineering	LLM covers mechanisms; application and engineering chapters cover usage
VLMs / speech / generative models	Multimodal AI	Agent computer use, multimodal RAG	Multimodal AI covers model capabilities; applications show how they fit into tasks
Function calling / MCP	Tools	Agent harnesses, frameworks	Tools covers interfaces and protocols; Agents covers execution; frameworks provide implementations
Runtime / harness	Agents	Frameworks, AI Engineering	Agents covers the general runtime; frameworks cover implementations; engineering covers operations
Memory	Agents	LangChain modules	Agents explains layers and trade-offs; framework chapters show specific implementations
Vector retrieval	RAG	Frameworks	RAG covers indexing and retrieval; frameworks cover component abstractions
Observability and releases	AI Engineering	Application topics	Engineering connects production feedback to change management; applications define domain signals
Cross-layer safety governance	AI safety and governance	LLM, Tools, Agents, RAG	Local controls belong to each layer; governance connects threat models, red teaming, and audit
Delivery in customer environments	FDE	All technical topics	Technical topics explain capabilities; FDE combines them to deliver measurable outcomes
Agent post-training	Agents, Chapter 25	LLM, Tools, AI Engineering	LLM covers optimization; Tools covers call data; Agents connects failure analysis, training examples, and acceptance
Structured-data questions	RAG, Chapter 22	Tools, FDE	Complete-data aggregation calls for a constrained query, not a total guessed from a few retrieved chunks

Connecting topics through interview questions¶

For foundational questions, explain the concept accurately. For system design, also explain where data comes from, who may take an action, and what happens after a failure. These paths help practice questions that span chapters.

Practice question	Suggested chapters	What the answer should distinguish
How would you choose an approach for an internal knowledge assistant?	RAG, fine-tuning, and long context, RAG evaluation, offline evaluation	Missing knowledge versus unwanted behavior, and evidence from a representative business query set
How can an agent continue after a tool timeout?	Harness responsibilities, checkpointing and recovery, retries and idempotency	Restoring computation is not the same as preventing duplicate business writes
Who is responsible for permissions after MCP is connected?	MCP and function calling, tool protocol security, least privilege and identity	Protocol capabilities, proposed calls, and backend authorization
Why might a high-scoring model still be unsuitable for production?	Model evaluation, output contracts, SLOs and incident response	Benchmark scores, structural validity, business correctness, and service reliability
How do you turn a vague customer request into a deliverable project?	FDE foundations, field lessons, versioning	Who approves changes, what a proof of concept establishes, and how the customer accepts and takes over the system
Why can an agent change code without fixing the problem?	Code search, editing, and verification, agent post-training	Search omissions, edit failures, weak verification, and model errors before deciding whether training is appropriate
How can an assistant count orders accurately rather than guess a total?	Text-to-SQL, tool schemas, offline evaluation	Business definitions, table relationships, query authorization, and independent result checks

For project questions, choose a concrete change: what was not working, which alternatives you tried, and why you made that decision. Reviewing the code, failure examples, and before-and-after results together reveals gaps in your understanding more effectively than memorizing an architecture diagram.

Frequently asked questions¶

Who is this handbook for?¶

It is primarily for readers preparing for roles in AI application development, agents and RAG, model engineering, and FDE. Beginners can follow the suggested paths to learn the foundations. Experienced readers can start with a practical problem and use the links to revisit evaluation, reliability, permissions, and other easily missed concerns.

Should I start with LLMs, agents, or RAG?¶

Start with LLMs to understand model capabilities and limitations. Start with Agents and Tools to build systems that call tools and carry tasks across multiple steps. Start with RAG for private or changing knowledge. Bring in production engineering and safety during design to establish evaluation, permission, and reliability requirements.

How often is the content updated?¶

There is no fixed release schedule. Significant changes to protocols, frameworks, or model capabilities lead to updates in the relevant chapters. Page dates come from the history of their current file paths and may not indicate the original publication date. Check time-sensitive claims against the versions and official references cited in the text.