Skip to content

How to Read This Book

The book has nine parts: large language models, multimodal AI, tools and protocols, retrieval-augmented generation, agents, frameworks and orchestration, production engineering, safety and governance, and field delivery. On a first reading, follow the contents in order. When a section points ahead, first understand the question at hand and use the reference if you need more detail.

Chapter numbering restarts at 1 in each part. Chapter 1 in Part 1 discusses language models; Chapter 1 in Part 5 discusses agents. References such as "Chapter 14" or "Section 14.2" without another topic name use the numbering of the current part. Cross-part references name the topic and include a link. The book does not impose a single global chapter sequence, and section numbers stay with their chapters.

This is the English edition. A complete Simplified Chinese companion uses the same chapter identities and numbering. Each edition can be read on its own; you do not need the other language to follow the explanations.

In the ebook, use the table of contents to move between parts and chapters. Links within the book lead to related discussions. Each diagram presents a compact view; read the nearby explanation for conditions, branches, and limitations. A diagram can also be opened separately as one complete image, with a return link to the original location. Formulas have similar links to their complete images. The text, diagrams, formulas, and bibliography are available offline; visiting an external source requires an internet connection.

When preparing an answer, close the book and explain the central distinction in your own words, then check what you missed. Saying "RAG can reduce hallucinations" is only a start. You also need to explain why retrieval can find the wrong evidence, why generation can misuse valid evidence, and how to distinguish those failures. That is more useful than listing components.

For a workflow or code example, introduce a failure: a timeout, stale evidence, revoked permission, or a successful tool action whose response was lost. Does the design still hold? Where must it record state, stop, or ask a person to intervene? Read setup snippets and pseudocode in their stated context rather than treating them as deployable systems. Some examples intentionally retain Chinese input text or business data; translating those values would change the example's conditions.

Superscript reference numbers link to the shared bibliography at the end of the book, where you can find the original papers, specifications, and product documentation. General further reading is distinguished from evidence for a particular passage. Version, date, and applicability limits are part of an answer, not details to discard when explaining it aloud.