One idea builds on the next.
This route places prerequisites before the lessons that use them. Topic series help you browse; article IDs are stable labels, not reading order. Reaching a page does not mean you have completed its prerequisite exercises.
For inference infrastructure, read 00-1 and 00-2, then jump to I-00. Application readers can finish at 10 and take the inference route when needed. Each lesson names the work to bring forward, the artifact it adds, and where that artifact is used next.
Recommended reading route
00-1 How Language Models Generate Text
Read first: none
00-2 Prompts, Tokens, and Context
Read first: 00-1
00-3 Why Fluent Answers Can Be Wrong
00-4 From Model Output to Controlled Action
01 Designing an AI System
02 Reliable Pipelines and Validation
Read first: 01
04 Memory, State, and Knowledge
05 Choosing Between Workflows and Agents
03 Grounding Answers with RAG
11 Evaluating AI Systems
12 Identity and Authorization for Agents
06 Agent Loops and Tool Reliability
07 Hybrid Retrieval and Evidence Packets
08 Reading Documents Beyond OCR
Read first: 03
09 Grounding Multimodal Evidence
00-5 Lab: A Reliable Support Workflow
10 Lab: Build a Governed Travel Copilot
I-00 Inside an LLM Inference Request
I-01 Continuous Batching
Read first: I-00
I-02 Managing the KV Cache
Read first: I-00
I-06 Distributing a Model Across GPUs
I-04 Prefix Caching and Request Routing
I-03 Separating Prefill and Decode
I-05 Deploying Mixture-of-Experts Models
Read first: I-06
I-07 Benchmarking Inference Capacity
Optional exercise: source checks and retries
This teaching script uses fictional travel data to demonstrate source checks, keyword retrieval, and lookup after a timeout. A handwritten dictionary connects English and Traditional Chinese terms. The simulated supplier creates records only in a local SQLite database.
Identity and approval are simulated in code. It runs no language model, booking service, or payment. Use it to inspect program flow and duplicate-operation handling; it does not establish the quality or safety of a production AI system.
Download the Python exercise (English and Traditional Chinese instructions)
python3 lab.py all --out out
python3 -m unittest discover -s . -vRequires Python 3.9+. Unpack the archive and run inside travel-copilot. The script prints JSON; confirmed refers only to the local simulated record. Run it again using the same out directory to check that the record count stays at one.