Google NotebookLM — source-grounded Gemini notebook (RAG-shaped tool)
NotebookLM is Google's source-first AI notebook. It only answers from the PDFs, Google Docs, web pages, YouTube videos, and audio files you upload — so hallucinations are far rarer than a generic chatbot. Where 04-gemin…
Table of contents
NotebookLM is Google's source-first AI notebook. It only answers from the PDFs, Google Docs, web pages, YouTube videos, and audio files you upload — so hallucinations are far rarer than a generic chatbot. Where 04-gemini-api is the raw API and 08-google-ai-studio is the generic playground / Build, NotebookLM is a notebook with RAG built into the UI.
1. Identity
- Site:
notebooklm.google.com(or NotebookLM Plus inside Google Workspace) - Launch: 2023-07 (Project Tailwind) → 2024 GA → 2025 mobile apps
- Models: Gemini 1.5 / 2.0 / 2.5 (auto-upgraded over time)
- Capacity: up to 50 sources / notebook (each 500MB / 500k words on Free; 2M words on Plus)
- One-line: "a Gemini notebook that only reads your sources."
2. Supported source types
| Type | Notes |
|---|---|
| OCR runs automatically | |
| Google Docs | Drive integration |
| Plain text (.txt, .md) | UTF-8 preferred |
| Web URLs | HTML parser |
| YouTube URLs | Caption-based (rejects videos with no captions) |
| Audio (.mp3, .wav, …) | Transcribed (Speech-to-Text) before indexing |
| Pasted text | 10k-word chunks |
3. Core features
3.1 Citation-first answers
Every sentence in an answer carries a clickable citation pin (page / timestamp). You see exactly where it came from — the headline difference vs. ChatGPT/Gemini chat.
3.2 Audio Overview (auto podcast)
Generates a 5–15 minute podcast where two AI hosts discuss the source. Lectures, papers, contracts, manuals — turned into a commute-friendly listen. (English first; other languages improving.)
3.3 Mind Map
Auto-builds a hierarchical mind map of the source set — handy for studying or summarizing.
3.4 Notes (Studio)
Save chat answers as notes inside the same notebook. Notes can be re-fed as input (recursive synthesis).
3.5 Sharing (Plus)
Share whole notebooks. Permissions split into Viewer (ask-only) and Chat (ask + add notes).
4. Free vs Plus
| Item | Free | Plus (Google One AI Premium / Workspace add-on) |
|---|---|---|
| Notebooks | 100 | unlimited |
| Sources / notebook | 50 | 300 |
| Daily chat | 50 | 500 |
| Daily Audio Overviews | 3 | 20 |
| Sharing | ✗ | ✓ |
| Cost | $0 | $20/mo+ |
5. Use cases
| Scenario | Example |
|---|---|
| Learning | 50 lecture PDFs + videos → study notes + podcast |
| Research | 30 papers → comparison table + open-question map |
| Legal / policy | 100-page contract → clause-level Q&A + risk summary |
| Manuals | Internal guides → self-serve Q&A for new hires |
| Meetings | A year of minutes → decision trail per topic |
| Interviews | 10 hours of audio → insight extraction |
| Books | One non-fiction → per-chapter summary + applications |
6. Limits
- Not used for training — uploads aren't used to train models (per policy). On Free, chat inputs may be used (anonymised) for quality improvement → use Plus for sensitive material.
- Refuses out-of-source questions — common-knowledge questions get "not in your sources." Intentional.
- 50-source cap (300 on Plus) — large corpora need partitioning.
- No public API as of 2026-05 — to automate similar RAG, build with Gemini API + your own RAG (02-rag-pgvector).
- Korean text works well; Korean Audio Overviews still trail English.
7. Similar tools
| Tool | Strength | Weakness vs NotebookLM |
|---|---|---|
| ChatGPT (with files) | Knowledge + sources | weak citation surface |
| Claude Projects | 1M-token context | no podcast |
| Perplexity Spaces | Web + your sources | no mind map |
| Notion AI | Notes integration | thin RAG |
| Self-built RAG (pgvector + Gemini) | Full control · self-hostable | build/run cost |
8. Tips
- PDF quality — OCR matters. Text-extractable PDFs first; scanned PDFs need separate OCR.
- Split notebooks — one notebook per topic. Mixed topics → noisier answers.
- YouTube captions — many Korean videos have no auto-caption; verify before adding.
- Save Audio Overview transcript — Save to note makes the transcript searchable.
- Try Plus free — 1-month Google One AI Premium trial covers Plus features.
9. Self-hosting alternatives
NotebookLM itself isn't self-hostable. To build a similar workflow:
- Vector DB: pgvector (02-rag-pgvector) or Qdrant
- Embeddings: Gemini text-embedding-004 or OpenAI text-embedding-3-small (05-embeddings-deep)
- LLM: Gemini API (04-gemini-api) or LM Studio (01-local-llm-lmstudio)
- Citations: attach chunk id + page numbers as source metadata → UI links to the original
- Podcast: ElevenLabs / Google TTS + two-persona script generation
Self-built wins for control and private data; NotebookLM wins for speed-to-use on personal study.
10. Further reading
Terms in this content
More in ai
All in this category →- Google AI Studio — Gemini-powered AI Web IDE + app builder
- LLM Landscape — Closed · Open · Korean-Specialized · Evaluation · Pricing
- AI Agents — Definition · Patterns · Frameworks · Autonomy
- Embeddings Deep — Models · Dimensions · Benchmarks · Cache
- Gemini — Google's Multimodal LLM Lineup
- Prompt Design — Message Roles · CoT · ReAct · Sampling · Injection
Related posts
Google AI Studio — Gemini-powered AI Web IDE + app builder
Google AI Studio (aistudio.google.com) is the in-browser home for the Gemini family. It started as a prompt playground, but since 2024 it has grown into a notebook + app builder + auto-deploy environment — squarely in t…
Gemini — Google's Multimodal LLM Lineup
Gemini is the model series Google DeepMind released in late 2023. Multimodal input — handling images · audio · video · code alongside text — and the very long context that came in with 1.5 are the often-cited features.
RAG and Search — Chunking · Embedding · Reranking · Hybrid
RAG (Retrieval-Augmented Generation) is an approach to supplement an LLM's limited training cutoff and memory limits with external knowledge. By separating retrieval and generation, you can handle new material without r…
pgvector and RAG
The trend of combining large language model responses with external knowledge has settled under the name RAG (Retrieval-Augmented Generation). At its core sits vector similarity search, and PostgreSQL's pgvector extensi…