Nate B Jones demonstrates Google's NotebookLM as a free personal AI note and retrieval system
A solo walkthrough of NotebookLM, explaining why it outperforms other tools for organizing documents and retrieving information accurately.
Summary
Nate B Jones of AI News & Strategy Daily makes the case that Google's NotebookLM is currently the best free tool available for anyone who needs a personal retrieval-augmented generation (RAG) system — that is, a way to store, search, and converse with large collections of documents without writing any code. He demonstrates the tool using a 62-page article he wrote about Microsoft Copilot, showing how NotebookLM can summarize it, surface non-obvious use cases, and cite its sources precisely. He argues that NotebookLM's defining strength is its exceptionally low hallucination rate — it retrieves accurately rather than generating speculatively — but that this same design choice means it is not a strong writer or deep thinker. His recommended workflow is to use NotebookLM for precise retrieval and then pipe that output into a more cognitively capable LLM for analysis and writing.
Key Takeaways
FULL TRANSCRIPT
The core problem: getting more than fits in a chat window
Nate B Jones: What do you do when you want to give your LLM — your AI — more than will fit in a chat? And how do you keep it consistent information that you can trust? If you have a lot of information and you have to squeeze it into an AI, I get this question a lot. People ask: "How do I organize all my notes? How do I learn a new topic? Nate, you have so many articles — what do I do with them?"
The answer is a personal retrieval-augmented generation system, but for people who are non-coders, that's not super accessible. I have been looking and looking and looking for the past few months for the best answer for this. And I have to tell you — the best answer is free, and it's Google.
Google has launched NotebookLM, and nothing comes close to beating it for the ability to learn complicated subjects, the ability to have lots of documents in one place, and the ability to have reliable search across those documents. I have to be honest with you: the LLM search with the lowest hallucination rates right now is NotebookLM. There is nothing else that comes close. It is extremely precise about what it recalls.
Live demo: a 62-page Microsoft Copilot article
But let me show you — I always think examples are more helpful here. This is an actual document from one of my articles a few months back where I talked about Microsoft Copilot. The document is 62 pages long. It is a long read. What if you don't have all the time in the world and you want to just understand what's relevant for you?
NotebookLM is honestly the best tool out there for figuring that out. And you'll note that in this case, even though I could add lots of other sources, I've chosen just to have a conversation with this one document so that I can get exactly what I want. I could add more if I wanted — I'm not going to. It gives me a summary of what's in the document. This is very accurate — I wrote it, so I should know. It gives me specific use cases for non-coders because I asked for that, and so it goes through and gives me a really great answer based strictly on the document. It cites everything it's getting, so you can actually see where it's citing from, which is phenomenal.
Then I can ask: give me some non-obvious use cases. And it's going to give me specific use cases also found in the document. This is a way to get gold out of longer documents — not just for me, but for any situation where you have a longer academic paper, a complicated subject to learn about, transcripts and notes for a particular project or maybe with a client, and you just need to get them organized somewhere and talk about them and get very accurate retrieval.
And by the way, if you're not a reader, look at how you can work with this in non-reading ways. Audio overviews work really well. There's a video overview. There's a mind map option, reports, flashcards, quizzes. This is designed to help you learn. We are sleeping on NotebookLM right now, and that's why I want to talk about it.
Why NotebookLM works: accuracy plus a project mindset
So what is it that makes me confident I can recommend this to people who just need a personal RAG system, a personal system for notes? Fundamentally, it is the combination of accuracy and a project mindset. Everything in NotebookLM is a project, and so you can apply any project orientation you want. You could have a project per client, a project per subject — you can do both. All you have to do is put the links you care about, put the articles you care about into that project. It is so easy to add more documents. It will accept just about any file type uploaded, and it will also make it very easy for you to link any file you want on the internet. So if you are behind a paywall, you can just download an academic journal of some sort, upload it to NotebookLM, and it works great. That is why I think it's the most flexible tool for that use case.
Now, could I have recommended a custom-coded implementation using Obsidian and a local language model? Yes — I know people who do that, and they are engineers. It is super helpful if you are an engineer and you're willing to build your own retrieval-augmented generation system using a local AI. But most of the people I talk to who desperately want something they can use at home to organize stuff and talk with AI about material that is much bigger than a chat window — they are not going to code. And so I wanted to find something that's useful.
Honest assessment: strengths and weaknesses
I want to be honest with you about where it works and where it doesn't. Where are the weaknesses? Because this gets back to the memory problem. There is no perfect solution. This is the best solution we have at this time, but there is no perfect solution out there right now.
What is good about NotebookLM is, as I've been mentioning, the accuracy — which I think is a super important feature — the ability to upload lots of sources (you can do dozens and dozens of sources in one project), the ability to add lots and lots of projects, and the multimedia outputs: the audio, the visual options I showed you. Those are all great.
The drawbacks are not killer issues, but they're kind of unfortunate. I find personally that the most effective use for NotebookLM is actually to organize projects by theme and then to take selective searches — like I showed you with the Copilot document — and output those selective searches as slices of accurate context into a thinking LLM to finish up work. Because NotebookLM is not a great writer. NotebookLM is not super heavy on thinking. It tends to be focused on retrieval.
In fact — and two people don't know this — those two things are kind of opposites. If your LLM thinks more, it is more likely to draw from its parametric weights — its own internal LLM training data — to answer your questions, instead of retrieving accurate information from your documents and putting it in front of you. Google has correctly chosen that for this application, you want to retrieve accurately. And so you get a very accurate slice of data back, but not a lot of thinking with the data.
I look at NotebookLM as a chance to extremely accurately summarize and extract slices of information I care about, and then I use copy and paste to pull that out and put it into an LLM when I need to think about it more. That is not a workflow I see very often, but I think it's super important. You need to be able to understand that AI doesn't all do the same stuff. We talked about Perplexity and how Perplexity is AI-native. Well, in the same way, this is a retrieval-native system. It is focused on retrieving from a RAG, but it's very tightly constrained — even more tightly constrained than Perplexity. So it's super accurate, and that means it's not super cognitive. It doesn't think a ton. And that's the drawback you get.
Scale considerations: who it's best for
I think if you're trying to build an evergreen note system for all of your notes and your notes go back for years, it will be difficult to pull them into a system like this. That is a situation where investing in a custom-built solution is going to be more meaningful for you, because if you have tens of thousands of notes, it just isn't at that scale.
Where NotebookLM really shines is at a smaller scale where you have dozens to maybe a hundred or two sources and you want to look at these related sources in a very coherent way and retrieve things. In my experience, that is actually most people.
For example, if you're trying to understand what your client has done in the last six months, you can upload recent client emails and documents, or link them from Google really easily. You don't necessarily need 20 years of client history, because you have that in your head if you're a longtime business owner — and you can bring that to bear. You get 80% of the value with much less effort than custom coding.
Similarly, if you're trying to build a knowledge system about AI, you can have the flexibility to add a particular project for a given AI topic, learn about that, have all of that knowledge in a project, and then move on to the next thing when AI evolves. You stay very focused on the cutting edge but have all the old project files there.
The one drawback that matters most
The one drawback that I think is most painful about NotebookLM — and I want to be honest about this — is that right now it doesn't save your chats. When you're chatting it feels really fluid, but you had better copy and paste out what you care about, because otherwise you're going to have to recreate that chat the next time you go in. I don't know why Google did this. I think it's silly, but there we are. It doesn't save your chats.
Despite all of that, it is easily the best RAG system out there, all things considered. It is the least technical. It is drag and drop. No code required. And it's free — which is not trivial when everybody else is asking you to pay for things. So I think it's super relevant.
I hope you've enjoyed this quick tour through NotebookLM. I'm going to put together a guide of use cases and also some suggested prompting tips, because when you're doing retrieval prompting it's very different from doing cognitive prompting. I'm going to get into that in the Substack as well.