FDE 101

A practical fieldbook for forward-deployed engineers building reliable AI systems.

Read the book · Start here · Practice interview questions · Use the production checklist

FDE 101 explains the systems around an LLM: context, retrieval, tools, orchestration, evaluation, security, observability, and delivery. Chapters use short explanations, concrete comparisons, citations, and complete code listings that you can inspect in the repository.

The material is written for software engineers, forward-deployed engineers, and interview candidates who want to move from a working demo to an operable production system.

What is inside

Area	Questions the book answers
Foundations	What does an LLM receive, what belongs in context, and when should you fine-tune?
Production RAG	How should you parse, chunk, retrieve, rerank, cite, abstain, and debug?
Agents and MCP	When should a model choose a tool, and where must deterministic controls remain?
LangChain	How do messages, models, tools, retrievers, runnables, middleware, and structured output fit together?
LangGraph	How do state, reducers, edges, checkpoints, interrupts, retries, streaming, and subgraphs work?
Evaluation and LLMOps	How do you measure retrieval and generation separately, gate releases, and reproduce failures?
Security	How do authorization, least privilege, approval, and prompt-injection defenses constrain a model?
FDE practice	How do you design projects, explain trade-offs, and answer production interview questions?

Suggested reading paths

Learn production RAG

1. RAG, end to end
2. Vector search foundations
3. Production retrieval
4. Evaluation
5. Production RAG checklist

Learn agents and LangGraph

1. Agent systems without chaos
2. LangGraph
3. Message history and memory
4. Prompt-injection boundaries
5. Agent-versus-workflow interview answer

Prepare for interviews

Start in the interview room. Practice defining the concept, giving a small example, naming the trade-off, and explaining how you would verify the result.

Editorial principles

• Start with the smallest honest implementation.
• Separate retrieval failures from generation failures.
• Keep permissions, calculations, and consequential actions in deterministic code.
• Add complexity only after an evaluation identifies a specific failure.
• Put citations next to the claims they support.
• Prefer provider-neutral examples unless a vendor API teaches an important contract.

Run locally

The repository pins pnpm@10.10.0. Use pnpm rather than npm.

pnpm install
pnpm dev

Open the local URL printed by Astro. Create and inspect a production build with:

pnpm build
pnpm preview

Standalone listings under src/examples/ require Python 3.10 or newer. They are educational examples; the documentation site itself does not require Python.

Repository map

src/
├── content/docs/   Book chapters, glossary notes, field guides, and interviews
├── examples/       Complete code listings imported into the chapters
├── components/     Small editorial and navigation components
└── styles/         Paper-and-ink typography, responsive layout, and print rules

Code examples are imported into MDX with ?raw, so the code shown in a chapter is the same file that can be checked or executed from src/examples/.

Add or update a chapter

1. Add a Markdown or MDX file under src/content/docs/.
2. Include a unique title, description, contentType, topics, and lastVerified value in its frontmatter.
3. Put complete examples under src/examples/ and import them into MDX with ?raw.
4. Cite official documentation or primary sources beside the claim they support.
5. Run pnpm build before review.

Deployment

Every push to main runs the GitHub Pages workflow. The Astro configuration derives the Pages base path from GITHUB_REPOSITORY, builds static HTML, generates the sitemap, and creates Pagefind search data.

For a custom deployment, set SITE_URL to the public origin. Set BASE_PATH only when the site is served from a subpath.

$env:SITE_URL='https://docs.example.com'
pnpm build

Package installation uses pnpm with the official npm registry configured in .npmrc.