Part 4 of 5 — RAG Explained

Products are spellings of the same pipe.

Part 2 drew index → retrieve → generate. Shipping that pipe inside a company takes engineers, ACLs (Access Control Lists), and eval. Shipping it as a product wraps the same pattern in uploads, citations, and a UI. The names change. The loop does not.

This part maps a small set of verified products onto the pattern — not a ranking of “best RAG tools.”

Google NotebookLM

NotebookLM is Google’s document-grounded notebook: you add sources (PDFs, docs, pasted text), and the product retrieves from those sources when you ask questions or generate summaries. Official positioning emphasizes sources you provide and citations back to those sources — the consumer spelling of retrieve-then-generate over an approved mini-corpus, not open-web guessing.

For a reader coming from Part 1, NotebookLM is the “your data, not the whole internet” idea in product form. It is not an agent with arbitrary tools unless Google’s product surface adds actions beyond generation from sources. Treat it as grounded Q&A and synthesis, not a replacement for your internal ACL-aware index.

Enterprise search + RAG (Azure AI Search pattern)

Large shops often already have enterprise search. Microsoft documents retrieval-augmented generation with Azure AI Search as: index your content in the search service, retrieve relevant passages at query time, pass them to a model for generation. Same pipe as Lewis et al.; different ops owner (search team + app team).

That pattern fits when your corpus is already in SharePoint, blob storage, or a custom indexer — and you need permissions, skillsets, and hybrid keyword + vector retrieval in one place. It is infrastructure RAG, not a notebook UI.

What these are not

  • Not a bake-off. No star counts, no “NotebookLM beats X.”
  • Not agents. Neither example is “dynamic multi-step workflow with arbitrary tools” by default. Part 5 is the decision frame.
  • Not your compliance program. Uploading HR PDFs to a consumer notebook without policy review is still a policy mistake.

Map products to the diagram

NotebookLM:     your uploads ──▶ index in product ──▶ Q&A with citations
Azure RAG:      corp content ──▶ AI Search index ──▶ retrieve ──▶ Azure OpenAI (typical)
Your pipeline:  [Part 2 diagram] — you own every box

Pick the implementation that matches who runs the index and where ACLs (Access Control Lists) must live. Skip the product if it cannot meet your data boundary.

Takeaway: NotebookLM and enterprise search RAG are named instances of retrieve-then-generate — useful to recognize the pattern, not to declare a winner.

Sources

Part 3: Enterprise ops
Part 5: RAG vs agents vs automation