NotebookLM: Its More than AI Podcasts- Your Personalized AI Research Assistant

Think smarter, not harder. We're looking at NotebookLM from Google, but it feels like it was built by a startup. You have probably seen the audio overview features, but NotebookLM can do a lot more than this. Here's one.

All right, so we're diving deep into AI today. Or more specifically how to make those powerful but data hungry large language models. You know, LLMs everyone's talking about a.

Bit more manageable and cost effective, right?

Exactly. You sent in some really cool resources about audio.

Overviews are a feature within Notebook LM that feels like two human beings are talking in a podcast. More on that later in the video. But it can potentially be your personalized AI research assistant and has some really cool features. You can upload a number of different file types including PDFs, links to websites, YouTube videos, audio files, or you can connect your Google Drive. Since it's using the multimodal Gemini 1.5, it can also look at images and charts. Now everything is in context of Gemini, so unlike traditional Rack, it can give you insights by looking at all the documents together thanks to Gemini's 2 million token context window. And I think it is pretty impressive. I'll show you some really cool examples later in the video. And not only it can generate answers, but can show you the sources. Now in the beginning I said it feels like it's built by a startup because if you look at the website it's unlike Google. I think the only mention that we see of Google is the Gemini 1.5 and the website link is NotebookLM Google and probably this small Google logo at the end. Okay, so to get started you need to have a Google account. Just go to NotebookLM Google then click on Try Notebook. You will be presented with this interface. There are a few example notebooks already available. To get started just click on New Notebook. Now here you can provide up to 50 different information sources and these can be your Google Docs, Google Slides, links to websites and YouTube videos. Or you can drag and drop different file types. Here is how it's going to look like when you add file types. So I'll show you a few of my own notebooks that I'm using and some pretty great features that it comes with. So this notebook is called Prompt Caching. Here I included a web link to a blog post Context Caching from Gemini API. It generates a quick summary and then it also figures out what are the key topics that are being discussed. Here's another blog post, but this one is from Anthropic which talks about their prompt caching and then I added two of my own YouTube videos which talks about the same topic. And again you can see the there's a quick summary that it generated. And if you go here, there is an overall summary for all the documents that you have uploaded. And the summary is pretty good because here it says the text focuses on context caching, a new feature offered by both Google and Anthropic. And that's exactly what these videos and blog posts are talking about. It also gives you some suggestions of questions that you can ask. So it's a pretty amazing source for interacting with your documents or for learning. Now what I really like about this is that if you provide a lot of different sources, you can still create a very good summary. So Here I provided 21 of my different videos that I have created over the last year on rag. And the summary is pretty amazing. So it says the sources are all video Transcription from the YouTube channel Prompt Engineering. So I didn't tell it that I am providing these links from Prompt Engineering, but it figures out that all these links belong to this channel. And then it says the video covers how to build RAC systems using tool likes LangChain, Llama Index and Colbert. In here there's a list of different videos that talks about vision based RAC system. So this it's able to generate a pretty decent summary. But I think the main power is in some of the other things that it can do. Now to get started, let's pick one of these suggested questions. So what is the key difference between prompt caching with cloud and context caching with Gemini? So it generated the response. Now depending on the query that you have, it will take some time. These are some of the questions that I have previously asked. So I think this is the one that we are talking about. It generates the response and also highlights where the information is coming from. So this is pretty neat. Like if you click on or hover over these references, it actually tells you where exactly the information is coming from. Another feature that everybody is talking about is these audio overviews. So essentially it can generate an AI podcast between two AIs. And it's pretty amazing because the conversation sounds very human. Now it also gives you the ability to customize that so you can tell it to focus on one specific source or one specific topic that is covered in your provided sources. Or you can also tell it to adopt a certain level of expertise when it's doing discussions. Right? So you can customize it using this prompt and then just click on generate and it will generate that AI podcast. But I Want to see how good this is not only at retrieval, but creating its own concepts. Now I asked the RAG notebook, can you list all the RAG techniques, including multimodal techniques? Now here it goes through my videos and these are the different topics that I have covered over the year. So it talks about dense embeddings, I have a lot of different videos which talks about dense embedding space RAG techniques. Then I have discussed keyword based search. So it includes those hybrid search, Colbert, contextual retrieval, late chunking, contextualized late chunking, that's the Colbert approach and also lists the vision based techniques. So this is pretty good list of techniques that I have discussed over the year. Now, since it keeps all the transcription resources that you provide in context. So I wanted to see if I were to ask it to pick one technique for rag, which one it will be. So I said if you were to pick one technique, which one it will be, I need you to pick only one. And I'm actually really impressed by the answer that it generated. So it picked Colbert for enhanced rag. Now I have discussed Colbert in a number of my videos and if I were to pick only one technique, I would actually go with Colbert. So it says while no single technique is universally best, the sources suggest Colbert offers a compelling combination of accuracy, efficiency, explainability and making it a versatile choice for various applications. We are talking about 21 different videos which talks about different techniques. And SYL was smart enough to select codebud. Now it does give some reasons which is basically improved retrieval accuracy. And this comes from the multi vector representation which captures fine grained semantic relationships within the text. And this is something that cold but is able to do because it's looking at token level embeddings rather than chunk level embeddings. Now it at the end refers to one of my more recent videos. It says while coldbud exhibits considerable strength, it's important to acknowledge that late chunking, especially when combined with specialized embedding models like those from Jenna AI, is presented as a technique capable of achieving state of the art results. However, this approach requires substantial computational resources which might not be feasible for all users. So I think it's pretty amazing because it's able to do deductions since it's looking at everything in context rather than doing standard rag. And that's why I think it can be pretty good tool for learning so you can ask questions, especially more insightful questions. And since it's looking at everything, it probably will be able to generate better answers compared to a standard RAG system. But There are a few other things that you can do with the help of notebook so you can create FAQs, study guides, table of content. There's an option for timeline. So this is an interesting one. Let's see what it does. Now timeline is going to be different for different type of information that you provide. So it created a timeline of events. It says this timeline tracks the evolution of RAG techniques and tools as presented in the provided sources. This is not a chronological timeline of events, but rather a progression of ideas and their implementation. So early stages traditional rac this is probably some of my early video that simply talks about chunking and then doing embedding. Challenges with the traditional RAC system, issues arises with purely formulated user queries leading to inaccurate retrieval hallucination. Then information embedded in images and tables is largely inaccessible. So traditional RAG also struggles with capturing context across multiple chunks. These are valid issues that I have discussed in my videos. Now there were videos about agentic rag which introduces an agent to analyze and refine user query, improving retrieval accuracy and reducing hallucination. This is actually pretty neat because it kind of started the traditional RAG system, then talks about agentic rag Colbert which is pretty great for preserving global information. Now it also talks about my local GPT project and then vision based RAG plus contextual retrieval. So these are some of my later videos. So I really like the way that it put together them in a kind of a chronological order. But I think the most impressive thing was that in earlier videos I talked a lot about the bloke because he was this person who is quantizing all of these open source models and putting them together. So I have probably mentioned him in a number of different videos, although it got the last name wrong. It says they a prominent figure in the open source LLM community, known for quantizing and optimizing large language models for efficient local use. And like, if I were to put together a list of people who are really important, I think the bloke is going to be one of them. So it's really good to see that the Notebook LM was able to identify him as a prominent figure. These are some of the frameworks and projects that I have mentioned in my videos. Local GPT is my own project, Cold Poly. I am a big advocate of vision based retrieval and it mentions Llama Index, LangChain. These are different projects or frameworks that I have talked about. So I think it does a pretty good job at creating this timeline plus the different technologies that are being used in RAC systems. Now if you are studying for something you can really use this as a study guide. So if you ask it to create a study guide, it will do that for you. So based on the content of these 21 videos it came up with some questions. The first one is one of what are the main two components of traditional RAC system and what are the respective functions. Then it also gives you answers at the end. Now you can customize this, but I think it can be a great tool for learning and studying, especially for students. A couple of other neat features. One is you can share your notebook by clicking on this share button. The other neat feature is that you can share your audio overview with everyone. Just click on this button, then make it publicly accessible and then either you can click on the preview. So here you can just click on this and listen to it or you can copy the share link and share it with people.

All right, so we're diving deep into AI today or more specifically how to make those powerful but data hungry large language models LLMs everyone's talking about.

So if you're interested I'm going to put a link to this prompt caching audio overview as well as the one for my rack YouTube videos. This basically talks about everything that is in my YouTube videos that I've created over the last one year related to Rack. I would love to have two features if the notebook and the team can put them together. The first one is going to be obviously an app that is going to be extremely helpful and the second, which probably is going to be helpful but may not be as critical is the ability to do audio chat. I can see myself using NotebookLM more and more and it's probably one of the best use case for long context LLM such as a 2 million context window. Anyways, they also have a business version of NotebookLM which I think is going to be extremely helpful for small and medium sized businesses because you can put all your documents or contacts into different notebooks and your teams can use them. At least based on my results, the retrieval accuracy is pretty great. Anyways, let me know what you think. I'm going to be posting the links to these audio overviews in the description of the video. I hope you found this video useful. Thanks for watching and as always see you in the next one.