Transcription
So, you've bookmarked 50 articles, 20 tweets, 15 YouTube videos, and have notes scattered across three different apps. But when you actually need that information, you can't find it. And you definitely can't see how it all connects. That's the problem Notebook LM solves. It's the best learning tool I've found for actually understanding and retaining information, not just collecting it.
The biggest problem with AI tools like ChachiPT is they hallucinate. They confidently make stuff up all the time. Notebook LM solves that, too. It has an extremely low hallucination rate because it only works with your sources. It also has one of the most viral features of any AI product where you can create an engaging podcast from your notebooks. You've probably seen that and that's cool. We will cover it. But there is much more to the platform than that.
But let's start with the basics. When you first open Notebook LM, you can create a new notebook. And this is all on the free plan. By the way, there is a paid plan, but it's pretty unnecessary for most people. Although, it does come with Google's AI premium plan if you already have that.
Once you create a notebook, you can upload all types of sources: PDFs, text files, audio files, links to websites, YouTube videos, or pull directly from Google Drive. Each source can be up to 500,000 words, and you can upload 50 sources per notebook. You might be thinking, that is way more than ChateBT or any other AI tool you're used to can handle. And you're right, it is not even close. I'll explain how that's possible in a second. I find it super interesting.
But in here, I uploaded two PDFs, some websites, and two YouTube videos from Saj at Skillap about Notebook LM. He's the one who actually made the full Notebook LM course we have on our course platform on Futuredia. So, of course, his YouTube videos are full of great information, too. But I want some more sources, and I don't want to keep searching for them. So, we also have this discover sources feature. Just type in your topic, and it will search for relevant sources you can add. Just look through and deselect any that you don't want to add. Then click import and those will be added right to your sources.
Notebook LM generated a summary and a title. You can come up here to adjust the title. Then you have this chat interface where you can ask questions. But what makes it different is it uses only your sources to answer and it cites exactly where it found the information with clickable citations that take you directly to that section. So that greatly lowers the hallucination rate. And then over here on the right, you can view your information in different formats: an audio overview that creates a podcast, a video overview, a mind map, all sorts of reports, flashcards, a quiz. All these new ways to visualize and greater understand your information. Then when you click out, all your notebooks are saved right here. You can revisit them whenever you need.
I'll dive into each of those features with tips on how to get the most out of them. But first, I want to explain how all this actually works and why it can be so much better than just using Chat GBT or Claude. I mentioned you can upload 50 sources that are each 500,000 words. That's 25 million words total. In Chatbt, you can upload about 96,000 words. So, Notebook LM handles 260 times more content. So, how is that even possible? It's because of the difference between the knowledge base and the context window. A context window is the actual working memory used to generate an answer. Notebook LM uses Google Gemini 2.5 which has one of the largest context windows available: 1 million tokens, which is about 750,000 words. Chat GPT's is only 128,000 tokens. Claude is 200,000. Grock is 256,000.
Now, tokens are what inputs get broken down into. The general rule of thumb is that one token equals about 75 words. So, 1 million tokens is roughly 750,000 words. I want to put these numbers into perspective using Harry Potter. With ChachiBT, the 96,000 words is about Harry Potter and the Sorcerer's Stone with a little room left over. Now, with Gemini's 750,000-word context window, that's the first five Harry Potter books. Then for Notebook LM's knowledge base, 25 million words, that's 23 complete Harry Potter series. All seven books, 23 times.
So, how do we jump from that 750,000-word working memory to 25 million words of storage? That's where RAG comes in: Retrieval Augmented Generation. Notebook LM doesn't cram all your documents into working memory at once. Instead, think of it in two parts. The 1 million token context window is the working memory. The 25 million-word knowledge base is the long-term memory stored as a vector database. It's pretty interesting how those work, too, but that's a tangent for another video.
So, here's what happens when you ask a question. Step one: you ask something like, "What are the main risks in my financial reports?" Step two: Notebook LM searches its entire document index for the chunks of text most relevant to your question. Step three: it loads just those specific sections into its working memory along with your question and chat history. And step four: Gemini generates your answer using only that focused, relevant information.
There are other RAG tools out there, but Notebook LM is unique because it forces citations for every answer. You can click through to the exact spot in the source document. This is why it has such low hallucination rates. It's not making stuff up or pulling from the internet. It's working with your documents, finding the exact right passages to answer your question. And that 25 million-word capacity is for each notebook you create. And you can have 100 notebooks on the free plan. That means you could upload your company's entire knowledge base, every meeting transcript from the past year, or all your research papers for your PhD. Then Notebook LM can actually understand and connect information across all of it.
So that was a long explanation, but I thought RAG was really interesting when I was first learning about it. But let's get back into the practical stuff, what you can actually do with this information. If you're using Notebook LM for research or content creation and want to go deeper into what's possible, I highly suggest you check out this free guide provided by HubSpot. It's called the Marketer's Guide to Google Gemini and Notebook LM. This is a really practical guide that shows you how to use both Gemini and Notebook LM together to streamline your entire content workflow. It goes beyond just the basics. It covers real marketing use cases like competitive analysis, SEO content briefs, and repurposing content across different formats. Inside, you'll find step-by-step workflows with actual prompts you can use, like how to turn a single piece of content into multiple formats using Notebook LM's analysis, plus Gemini's generation capabilities. There's also a whole section on advanced Notebook LM features, some of which I'm covering in this video. But my favorite part is the real-world marketing scenarios. They walk through exactly how to use these tools for things like creating content calendars, analyzing customer feedback, and even generating social media posts. It's super actionable, and again, completely free to download using the link in the description. So, thank you to HubSpot for sponsoring this video and providing valuable resources like this one.
On this main page, I have all my notebooks. You can switch over to a list view as well, just whichever one you prefer. Then, click right here to create a new notebook. Now I want to go deeper on how RAG actually works. So I have two PDFs of research papers about RAG. Just drag those in. Then I also have two YouTube videos on the topic. All you have to do is paste in the URL and it will get the entire transcript from that video. I also found a bunch of good articles. So I just paste each of those in. You can upload multiple at once. Just add a space in between each of them. And that's a good amount already, but I'll click discover sources to get some more. I'll just type in retrieval augmented generation and submit. I just want to verify where each of these are coming from. I don't really want Wikipedia. I will skip Reddit. This isn't really relevant to what I want to ask about. And the rest of these look like they're research papers. So, I'll leave all those in.
But now, I already have this summary here in the middle. And I can ask my own questions or use their suggested questions, which are actually really good most of the time. So, I'll click this one: "What are the primary benefits and potential drawbacks of using RAG with LLMs?" Now, I have my answer fully sourced. "RAG prevents LLMs from generating factually incorrect or biased responses known as hallucination by ensuring the output is grounded in factual retrieved data and evidence from external sources." And I can hover over the citation to read here or click through to see it in context.
And you can continue chatting in here. But here's something really important: it doesn't save your chat history by default. If you leave, it disappears. But if you like an answer, you need to click "Save to note" to keep it. You can see that saved it right over here. And if you came to a really thorough answer in your chat, you can even convert that saved note to a source itself.
Another great feature here is you can easily check and uncheck sources. This lets you focus on one specific source or a few sources with a particular angle. It's really useful for comparing perspectives or drilling into specific documents. But say I want to understand vector databases specifically. I have everything unchecked except a couple sources that focus on that topic. Now I can ask, "How do vector databases work?" and summarize that in one sentence. "Specialized data stores designed to efficiently index and retrieve vector embeddings, numerical representations of semantic data." I'm getting a focused answer from just these documents. Then I can toggle them all back on once I'm ready to see the full picture again.
And one last note here is you can come up to these settings and define your conversation style and the response length. If I click the custom style, I have the example of "Respond at a PhD student level." If you're new to a topic, you could say something like, "At a beginner level." They also suggested, "Pretend to be a role-playing game host." Or, "Help me prepare for upcoming board meeting." Or maybe you're using it as a fitness guide and exercise tracker. You could give it your goals: "to run a sub three-hour marathon" or "get shredded." Maybe you're uploading legal documents and want to say, "Help me avoid jail time for this crime." There's a lot of ways you could customize this. I'll leave it at the default for now. And they also have longer or shorter response lengths. So this customization can be really helpful for specific use cases. And a side note, Notebook LM doesn't train on your conversations either, and that is great for privacy. So this is already a big step up from traditional chat models for many applications. Notebook LM isn't for everything, but it's far better for particular things like research, learning, and knowledge synthesis. There's some other really creative and overlooked ways to use it I'll mention a little later, too.
So far, we've covered how you can question and organize information, but the studio section on the right is where things get really interesting. This is where you can dive deeper using different formats to understand more deeply and retain knowledge more thoroughly. Let's start with the extremely viral feature, the audio overview. This turns all your knowledge into a realistic sounding podcast with two AI hosts. I'm generating one and I'll come back when it's done.
All right, let me play a bit of this. >> Welcome back to the deep dive. Today we're uh really digging into retrieval augmented generation. You probably know it as RAG. >> Yeah. >> I think there's a great example uh from the sources we looked at, right, about space. If you ask a pretty simple question like what planet has the most moons? >> Ah yes, a classic base LLM one trained maybe a couple of years ago. It just uses its internal training data. >> So remembers what it learned back then. >> Exactly. It'd probably tell you quite confidently Jupiter with 88 moons. >> Because well that was the right answer when it was trained. >> But it's wrong now. >> Totally wrong now. And it wouldn't tell you why it thinks that. No source, just an assertion. >> Okay. So now ask that same question to a RAG system. What happens differently? >> Big difference. The RAG system takes your question, but first it goes out and looks things up. It retrieves fresh information from an external source, maybe like a live NASA database or something similar. >> So it finds the current count. >> Right? It finds the latest data: Saturn currently at 146 moons, though that number keeps changing. Then it takes that retrieved context, that fact, and gives it to the LLM along with your original question. >> Ah, so it forces the LLM to use the new information. >> Precisely. It grounds the response in verifiable, up-to-date external data.
So I skipped some of the intro there. They kind of always take forever to get to the point and there's a lot of fluff, which is typical of both LLMs and podcasts. So, I guess they nailed the vibe there. Once you get past that, that's just amazing. I'm still blown away by this feature, even though it's been out for a long time now. Especially if you've already read through some of this and gotten a decent understanding, this will really reinforce that learning, and they'll use a bunch of examples and analogies throughout the podcast. So, you get it in a totally different format and perspective. And if you feel like there's too much fluff or they're going too deep or not deep enough, there is a way to help with this. Before you generate, you can click this edit button and they have some customization options: there's a deep dive, a brief, critique, debate. Then you can also fully customize it, telling them specifically what to focus on in the episode. So, I definitely recommend doing that on pretty much every time you generate one of these.
The other option here for customization is this interactive mode. This lets you join the conversation. So, I'll click interactive mode, then get it playing here, and then I'll jump in. >> Welcome back to the deep dive. Today, we're uh really digging into retrieval augmented generation. >> Oh, wait. Someone wants to join. Hey, go for it. Can you explain how vector databases work in simple terms? >> Oh, that's a great question. >> That's the core component of RAG. >> And we absolutely can break that down simply. >> Okay, let's unpack this. >> Think of a vector database less like a traditional database. >> You know, with rows and columns. >> Exactly. Instead, think of it as a really specialized address book for ideas. >> An address book for meaning. >> Precisely. First, we take all your company's documents, right? >> All the PDFs, manuals, etc. >> And we chop them into smaller sections or chunks. That's the part we called chunking. >> Then a special AI model takes each chunk >> and turns it into a long list of numbers. >> That list of numbers is the vector embedding. >> It's a numerical representation of the meaning of that text. >> The vector database's job is to store those vectors efficiently. >> So all the similar ideas are grouped together. >> Yes. If two ideas are semantically similar, their vectors are mathematically close in this high-dimensional space. >> So when you ask a question, >> your question also gets turned into a vector, your query vector. >> Okay. The database then uses special algorithms like approximate nearest neighbors >> an&m the efficient search method >> to quickly find the vectors that are closest to your question's vector. >> Meaning it finds the chunks of text whose meaning is similar to your query. >> And it does this incredibly fast, even with billions of documents. >> That's why we call it semantic search. >> It's searching by meaning, not just keywords. >> And the vector database is the engine that enables that. >> It's the engine, but it's also the operational layer.
So, as you can see, you can completely guide the conversation by jumping in. And you can keep jumping in like that and steer it in different directions as you go. It's pretty amazing how it can just take a turn like that and still respond in that really natural-sounding way. This honestly blows my mind.
But next up, after the audio overview is a video overview. Again, you can customize this if you want. I will just generate it as is. I will come back to that one since it takes a while. We'll move on to mind map. This can be really helpful for visualizing how certain concepts connect. So it splits this up into some different options. Let's look at the core components: the retrieval, the generation, and the knowledge base. Then you can dive deeper anywhere you want. Then we can close that one back up and take a look at the benefits: improved accuracy and reliability, reduces hallucination, uses up-to-date information, generate specific, diverse, factual language. There are just a lot of ways to explore here. This is especially useful for understanding the relationships between different components or for just getting a bird's-eye view of complex topics. This will be more or less helpful depending on what topic you're looking into.
The next one here is reports. Under that, you have a briefing doc, a study guide, a blog post, or you can create your own. "Craft a report your way by specifying structure, style, tone, and more." It even generates some example custom report ideas based on your sources. And these suggestions are usually pretty solid. I'll do a technical report if we want to go super deep. An explanatory article to just learn the fundamentals. Concept overview. I'll do that one just so you can get an idea of what this does. A study guide is one that's really helpful a lot of the time. So, I've got one of their defaults and a custom one generating. The study guide was done first, so we'll take a look at that. We have a quiz with short answer questions, some essay questions, then the answer key below that. Then down further, we have a glossary of key terms. And this all looks really thorough. This is an amazing learning resource. And here we've got the report. This is a very in-depth report based on those 17 sources we uploaded. So you can see there's just tons of different ways you could use these reports.
Then next up, we've got flashcards. I mean, this is pretty self-explanatory, but we'll generate some so you can see what they look like. "What is retrieval augmented generation?" And you can click down here to see the answer. It's got a nice short definition there. Then you can click "Explain" down here if you need to understand it further. Pops you right over into the chat. And back to these flashcards, we've got looks like 67 total flashcards here. So, this is super, super helpful if you're studying for something or if you just want to retain the information about something you're learning. And down here, we also have a quiz. Another one that's self-explanatory. I'll go ahead and generate that so you can see it. And just like up above, you can also customize the flashcards and quiz if you want to. We've got a 12-question quiz here. It's multiple choice. If you're stuck, you can click the hint down here. Looks like a pretty well-crafted question here, and it gives you immediate feedback. Let's try getting one wrong. There we go. So it says "Not quite" and then shows you the right answer. And again, it pops up with "Explain" when you get an answer wrong down here.
Now, back to the main studio page here. The video overview is done. Let's take a look at this. >> All right, let's talk about one of the biggest things happening in AI right now: retrieval augmented generation, or as everyone calls it, RAG. You can kind of think of it as the secret sauce that's making AI not just clever, but actually accurate and a whole lot more useful. So, you know, large language models, LLMs, are just incredible. They can write code. They can whip up a poem, explain quantum physics. It's amazing. But >> All right, so you can see it has the nice voice going over explaining all of these slides. I'm not going to make you watch this whole thing. Let's just skip forward and view some of the slides. Looks like it's using the same type of information from the podcast here. So it does have some really good-looking slides. It's basically like an automated version of PowerPoint that automatically generates the presentation, speaking along with it, too. >> First up, indexing. You take all your documents, break them up into smaller bite-sized chunks, and organize them into a special searchable library. >> So, this isn't going to win any design awards, but for automatically generated educational content, this is pretty solid. But this is the feature in this section that I use the least. It's cool it can do it, but it's a really slow-paced and boring way to take in the information for me.
But really, it's the combination of all of these things that makes Notebook LM so powerful for learning. Seeing your knowledge in all these different forms is extremely valuable for retention and deeper understanding. You can ask all your questions in chat, generate a mind map to visualize relationships, study with flashcards, listen to a podcast on a drive, even calling into that podcast to ask questions, then quiz yourself when you get home. This is honestly one of the main AI tools I recommend to just about everyone because it fundamentally changes how you learn and retain information.
Now, I'll quickly go over the pro features for anyone who's curious about the next tier. If you remember, the free version already gives you 50 sources per notebook. The paid Notebook LM Plus subscription bumps this up to 300 sources. That gives you the massive 150 million-word capacity I mentioned earlier. But now, a pro tip: you can use if you hit the 50 source limit but haven't maxed out your 500,000-word limit on each file, just combine multiple documents into a single source file. A lot of times you can merge 10 or more short sources into one file and still not hit the word count limit. By doing that, I personally have never had a use case where the free tier capacity wasn't enough.
On Pro, you also get big increases to your daily limits, like five times more audio overviews, video overviews, reports, quizzes, and flashcards. I never hit those limits myself, so they're really for heavy-duty daily users. There are also a couple of advanced features for sharing and teamwork. For advanced sharing, you can give other people access to an entire notebook or set it to chat-only access so they can query your documents without editing anything, kind of like how you share Google Docs. Then for notebook analytics, you can view data on how many users accessed your notebook per day and how many queries they made. This is great if you're using it for teaching a class or managing a small project team.
So, while these pro features are available and they are automatically included if you subscribe to the larger Google AI Pro or Google AI Ultra plans, realistically most individual users won't need them. I actually have a paid Google AI plan for tools like V3 and Gemini Pro, but I wasn't even thinking about it when I signed up and I used a different email than the one I was already using for Notebook LM. But honestly, I haven't even bothered to switch it over because I haven't needed any of those paid features. That's why I say for most people, the free plan is more than enough.
And finally, I want to just mention there are a lot more use cases for Notebook LM than just learning and research. For work, that could be uploading all your meeting transcripts, all your SOPs, or competitor research. Or I have a notebook with all of my previous YouTube scripts. I compiled them all into one giant PDF so I can go back and search for things I've talked about before or, you know, what sponsor was in a specific video. That's been really helpful for me. Other YouTube-specific things, I have a database of title formats with frameworks and templates. I can use a similar type thing for hooks for short-form or intros. Some other ideas like for personal use: if you journal, you could upload all of your journals and get insights on them. Or you could upload all your recipes and get unique combinations of them or to meal plan based on your favorites. Or another cool one is to have like a homeowners folder or a tech support folder where you upload all your product manuals to your tech, your dishwasher, washing machine, blender, things like that. Then you can ask in there anytime something goes wrong and it will reference the answers from those user manuals. Or it could be a finance coach, legal document analysis, travel planner, fitness tracker. There are tons and tons of possibilities. I utilize a lot of these. I'm actually working on a whole separate video showcasing the more underrated and overlooked use cases. So make sure to subscribe for that.
And if you want to go deeper into learning AI, we've built a full course platform at Futureedia with over 500 lessons across over 20 AI courses. You'll find full learning paths on ChatGPT, prompt engineering, automation, custom GPTs, video generation, coding with AI, and we have one on Notebook LM already, and other Google products like we just released one on Nano Banana. And there's a lot more. It's all included in one subscription. You can get a 7-day free trial using the link in the description. Or if you want to learn more about some free Google AI tools, check out this video right here.