Companion Code Repository

From Problems to Patterns: Enterprise RAG Engineering with .NET By Najaf A. Shaikh

Building Grounded, Secure, and Production-Ready Retrieval-Augmented Generation Systems

---

The runnable code for every chapter of the book. One solution, Contoso-Knowledge.slnx, holds a folder per chapter and the shared libraries the chapters build on.

Get the code

The book's paths start at a folder named sample-code, so clone into one:

git clone https://github.com/codeshayk/enterprise-rag-dotnet-samples sample-code
cd sample-code

Requirements

• .NET 10 SDK (the book targets net10.0, C# 14). global.json accepts any 10.0 feature band from 10.0.100 up and selects Microsoft.Testing.Platform as the test runner.
• Python 3 for the two scripts in eng/, which use the standard library only.
• Docker, optional, for the chapters that start PostgreSQL with pgvector, Qdrant or Redis.
• Ollama, optional, for new model calls: the chapters' experiments ran llama3.2 on the CPU.

No GPU or API key is needed. Every sample defaults to a scripted model that runs offline, and the model calls the book made are recorded under runs/, so a rerun with nothing changed replays them. Each chapter that needs a real model shows how to switch to Ollama, Azure OpenAI or OpenAI.

Build and test

dotnet build Contoso-Knowledge.slnx
python eng/get-models.py                               # 32 model files, 2.8 GB, checked against models/models.json
dotnet run --project Shared/Contoso.Corpus.Generator   # the synthetic corpus, the same 43 files on every run
dotnet test --solution Contoso-Knowledge.slnx

Tests never call a chat model. A test that needs a downloaded model, the generated corpus or Docker says so and skips when it's missing. The Build and test workflow runs these four steps on a Windows runner for every push and pull request. A newly published package advisory stays a warning, so it can't break your build; the Package audit workflow fails on it instead, on every change and once a week.

Run one chapter's sample from this folder, for example:

dotnet run --project Chapter01-Why-RAG/Ch01.WhyRag

Layout

Folder	Contents
ChapterNN-*/	Each chapter's checkpoint: its programs and its tests
Shared/Contoso.Corpus	Contoso Industrial's facts, people, permissions and documents, and the golden sets
Shared/Contoso.Corpus.Generator	Writes the synthetic corpus to corpus/ with the truth each file was rendered from
Shared/Contoso.AI	Model provider selection and the scripted model
Shared/Contoso.Knowledge.*	The libraries the chapters build, from parsing and chunking to retrieval, generation, evaluation, security and operations
models/models.json	Every model file (ONNX models, tokenizers and Tesseract's language data) with its source, license and SHA-256
runs/	Recorded model calls, answers and the run outputs the chapters replay and quote
eng/	get-models.py downloads the models; generate-solution.py regenerates Contoso-Knowledge.slnx after you add a project

Each chapter's folder is a self-contained checkpoint that references Shared/ libraries only. The one exception is Chapter 20's experiment tool, which reuses Chapter 17's agent lab.

Configuration

Model settings live in the AI section of each program's appsettings.json, and can be overridden with environment variables (AI__Provider=Ollama) or arguments (--AI:Provider=Ollama).

Provider	Settings
Scripted (default)	none; a deterministic stand-in that answers from the evidence it's given
Ollama	OllamaEndpoint (default http://localhost:11434), OllamaModel (default llama3.2; run ollama pull llama3.2 first)
AzureOpenAI	Endpoint, Model (the deployment name); signs in with DefaultAzureCredential
OpenAI	Model, ApiKey (keep it in User Secrets or an environment variable, never in a file)

License

The code is under the MIT License (see LICENSE). The models and packages it downloads keep their own licenses, which models/models.json and Appendix B of the book list.