Step 7 — NotebookLM vs your own RAG

You've built a RAG pipeline by hand (embeddings → pgvector → retrieve → prompt → generate). Google NotebookLM delivers a similar result with a few clicks. This step is about deciding which fits your situation.

One-table comparison

Item	NotebookLM	Self-built RAG (LM Studio + pgvector)
Setup	5 min (just need an account)	1–2 days
Sources	50 PDFs/Docs/web/YouTube/audio (300 on Plus)	unlimited (disk-bound)
Citations	auto (page / timestamp)	implement yourself (chunk-id metadata)
Podcast	auto (Audio Overview)	build via TTS API
Mind map	auto	✗
Exposure	Google cloud	can run on internal network
Model swap	✗ (Gemini only)	LM Studio · Gemini · OpenAI
API	✗ (as of 2026-05)	yours
Training policy	Free tier may use inputs	stays internal
Cost	$0 (Free) ~ $20/mo (Plus)	server + LLM API
Korean	strong on text, EN podcast better	depends on the model you pick

When to use NotebookLM

• Public-friendly material — books, papers, open docs, course PDFs.
• Need results fast — proof-of-concept, learning.
• Want both citations and podcast.
• Personal study or small team.
• Topic-per-notebook (within the 50-source cap).

When to use self-built RAG

• Material can't leave the building — internal manuals, customer data, contracts.
• API automation — embed RAG into a product.
• Free model choice (Korean-tuned, internal hosting).
• 50-source cap or chat quota too small.
• You need fine-grained search · reranking.

Experiment — same material, both pipelines

1. Index 5 PDFs in your self-built chatbot from Step 4.
2. Upload the same 5 to a NotebookLM notebook.
3. Run the same 5 questions (summary / comparison / contradiction / citation chase / one-card synthesis).
4. Compare answer quality, citations, and latency.

Typical contrasts:

• Citations — NotebookLM tends to be more page-precise.
• Depth — self-built can dial top-k / rerank; NotebookLM exploits the long context.
• Speed — NotebookLM is instant; self-built has indexing time.
• Korean — self-built lets you swap models.

Hybrid pattern

Common production combo:

1. Mock with NotebookLM — agree on sources / validate scope.
2. Self-built RAG for production — embed in your stack.
3. Cross-check: feeding the same sources to both is a built-in consistency test.

Deeper

• NotebookLM note
• Gemini API
• LLM landscape

Next courses

• architecture-patterns — fitting RAG into a larger system.
• ai-agent-tooling — pairing RAG with Cursor / Claude Code.