Understanding Modern LLM Systems

You can follow a conversation about RAG, embeddings, and agents well enough to nod along. Drawing the system on a whiteboard, or saying which part is broken when the answers come back wrong, is another matter.

This is a 67-page field guide that closes that gap. The full text is free to read online at llmknowledge.org. What you get here is the typeset edition: PDF and EPUB. (A Chinese edition is available separately on Gumroad.)

Why pay for something that's free?

Typography built for reading straight through and for printing. EPUB that goes to a Kindle via Send to Kindle, or to iPad, Kobo, Apple Books. Offline, annotatable, yours. Free updates for life. And it keeps the online version free for everyone else.

What's inside

One question, followed end to end through a complete LLM system, with every mechanism it passes through explained on the way. The throughline is a working analogy: a secretariat preparing a briefing for an executive. Assignment, archive room, drafting, sign-off, filing.

• Before We Begin: What Problem Is This Whole System Actually Solving?
• Chapter One · RAG Core Flow: How a Question Becomes a Cited Answer
• Chapter Two · Model Adaptation: Turning a General Model into a Usable System
• Chapter Three · Deep Dive: Four Core Technical Mechanisms
• Chapter Four · Agents: When the Model Starts Deciding and Acting on Its Own
• Chapter Five · The Compliance Perspective: Where Does Confidential Data Actually Go?
• Chapter Six · The Security Perspective: When Someone Means to Cause Harm
• Chapter Seven · Beyond RAG: The Full Landscape of LLM Systems
• Appendix · Toolkit Overview, plus a Quick Reference card

Every concept follows the same shape: why it is needed, a concrete analogy, then the precise technical definition. No math, no code, nothing skipped.

New in the 2nd edition

• A full new chapter on security. Chapter Six walks the whole pipeline again from an attacker's point of view: prompt injection, the EchoLeak zero-click exploit against Microsoft 365 Copilot (CVE-2025-32711), and a class of attack that breaches nothing at all. It just keeps your system spending money and compute.
• Reasoning models join RAG, agents, and fine-tuning in the "beyond RAG" chapter. Chapter Two now covers reinforcement fine-tuning, part of the same family of techniques that taught models like DeepSeek-R1 to reason.
• 49 footnotes citing 46 sources. Vague claims were replaced with named papers, CVE records, and vendor documentation.
• The 2026 tool landscape is refreshed: Voyage now under MongoDB, Jina under Elastic, AutoGen in maintenance mode, and more.

What this is not

Not a survey assembled from blog posts. Seven original diagrams, and 49 footnotes linking to primary sources: the ReAct and Self-RAG papers, the PagedAttention paper behind vLLM, the OWASP Top 10 for LLM Applications, NCSC guidance on prompt injection, the CVE record for EchoLeak, and the EMNLP paper showing that fine-tuning on new facts increases hallucination.

Who it's for

• People who need to speak the language without reading the papers. Sales, business development, and product managers holding their own in conversations about RAG, agents, and MaaS.
• People new to AI-adjacent roles. A complete mental map of the system in an hour or two, instead of assembling one from scattered searches.
• Engineers from outside ML. Backend and product engineers who need the architectural logic without going back to study deep learning.
• Anyone doing compliance or security review. Chapter Five traces exactly where confidential data is exposed at each stop, including the two points almost everyone misses. Chapter Six covers what happens when someone attacks the system on purpose.

Who it's not for

If you work on LLM training or inference optimization, this will read as introductory. It is built to connect concepts into a framework, not to serve as a research-grade manual.

Questions

Does the EPUB work on Kindle?
Yes. Download it and upload at amazon.com/sendtokindle. Also opens in Apple Books, Kobo, and most e-readers.
Do I get updates?
Yes, free. Every new version appears in your Leanpub library, and major updates come with an email.
Is the Chinese edition a translation?
Both editions were written and revised in parallel, not machine-translated.
Refunds?
60 days, full refund, handled by Leanpub.

About

Written by koe. The guide grew out of needing to explain these systems to non-engineers repeatedly, and finding nothing that sat between the papers and the hand-waving.