How to use this book
This is a reference book, not a course you must finish in order. Start with a customer problem, read the smallest useful chapter, and keep the code nearby.
The standalone examples require Python 3.10 or newer. On Windows, activate a Python 3 virtual environment and confirm python --version before using python. On macOS or Linux, use python3 and check python3 --version.
Choose your path
Section titled “Choose your path”| If you need to… | Read this first | Then read |
|---|---|---|
| Understand the vocabulary | LLM foundations | Glossary |
| Build a grounded assistant | RAG, end to end | Chunking |
| Move RAG from demo to production | Production retrieval | Evaluation |
| Understand vector search at scale | Cosine to HNSW | Choosing retrieval storage |
| Choose vector, graph, or both | Vector versus graph search | Choosing retrieval storage |
| Research across several sources | Agentic RAG | Evaluation |
| Build a bounded agent | Agent systems without chaos | LangGraph |
| Connect an assistant to tools | MCP tool selection | Prompt injection |
| Connect agents or render agent-driven UI | Protocol map | A2A and A2UI |
| Resolve agent disagreement | Agent conflict resolution | Evaluation |
| Build a stateful workflow | LangGraph | State vs memory |
| Separate context from memory | Message history and memory | State vs memory |
| Give a coding assistant the right repository context | Prompt versus context engineering | Interview answer |
| Follow an applied AI roadmap | Applied AI engineering roadmap | Context engineering glossary |
| Prove a system works | Evaluation | Production checklist |
| Operate and debug a live system | LLMOps | Semantic caching |
| Protect tools from injected content | Prompt injection | Agent systems |
| Build a portfolio project | Projects that prove engineering | Production checklist |
| Prepare for an interview | Interview room | RAG debugging |
Read a full lesson in three passes
Section titled “Read a full lesson in three passes”You do not need to memorize the whole page. Carry three questions through it:
- What is happening? Start with the one-sentence mental model and the real customer situation.
- How does it move? Follow the numbered flow and code from input to decision.
- How would I know it worked? Read the failure table, evaluation rule, and field questions.
Most full lessons use the elements below. When one appears, use it this way:
| When you see… | Use it to… |
|---|---|
| A numbered flow | Retell the system in order without framework vocabulary |
| A comparison table | Choose between designs or locate the stage that failed |
| A worked case | Attach the abstract idea to one believable situation |
| Code | Check the data and control flow instead of trusting prose |
| Field questions | Turn the lesson into a customer or interview conversation |
| Sources | Verify the claim and check whether it has changed |
A sensible first week
Section titled “A sensible first week”Begin with the applied AI engineering roadmap, then read LLM foundations, RAG, and evaluation. Choose production retrieval if the system answers from documents, or agent systems without chaos if it must choose tools and actions. Then run the examples in src/examples.