Part 3: hands-on RAG console app (#500)

[jongalloway]

Summary

Adds the new Part 3: Add RAG by hand — extends the Part 2 console app with a complete, hand-written Retrieval-Augmented Generation loop and no vector database. Resolves #500.

Second hands-on part of the workshop rewrite.

What's included

• Part 3 - Add RAG/RagChatApp/ — a net10.0 console project
  • IEmbeddingGenerator from the same AzureOpenAIClient (two abstractions, one resource)
  • Paragraph chunking of a sample document
  • Embed + store in an in-memory List (no vector DB)
  • Hand-written cosine similarity search + top-k retrieval
  • System-prompt augmentation with retrieved context, then the streaming chat loop from Part 2
  • Sample product doc (sample-docs/contoso-trailblazer-3000.md), copied to output
• Part 3 - Add RAG/README.md — step-by-step walkthrough with a data-flow diagram

Teaching flow

Reinforces the swappable-abstraction theme: Part 2 swapped the chat provider, Part 3 shows the same idea for embeddings and vector stores. The in-memory store is deliberately naive to motivate the real vector store revealed in Part 4.

Models & provider

• Provider: Azure AI Foundry (Azure OpenAI)
• Chat: gpt-4.1-mini · Embeddings: text-embedding-3-small

Validation

• ✅ Builds clean against net10.0 (dotnet build -c Release, 0 warnings, 0 errors)
• ⏳ Runtime validation against a live Azure AI Foundry resource is deferred to the E2E test pass (End-to-end test pass vs a real Azure AI Foundry resource #504)

Notes

• bin/ and obj/ are gitignored — only source is committed.
• Adapted with attribution from Steve Sanderson's dotnet-ai-workshop (chapters 2, 3, 6).