Notebooklm: a neural network for analysis of documents, self -education and investigation

Analysis of documents using neural networks is a task with a star, although, it would seem, artificial intelligence trained on a huge array should easily cope with it. However, popular chat bots in such situations hallucinate significantly more often than average and require thin bonds to obtain the correct result. Google Notebooklm effectively solves this problem.

Notebooklm It works on the basis of the Gemini Pro model. Unlike traditional chat bots that are based on general knowledge from the Internet, this model builds its answers exclusively to the documents loaded with the user. This significantly increases the accuracy of the analysis and makes the tool an excellent assistant when working with large amounts of data. At the time of publication of this text, he works with Google Docs, files in PDF, Markdown and TXT formats, audio recordings, web page links and video on YouTube. As in other Google products, NotebookLM provides for a joint work mode.

Description and basic functions 

On the main page, the new user is invited to create his first notebook and add documents to it. Each of the notebooks may contain up to 50 sources (documents, links or files) in the free version and up to 300 in a paid version.

This is how "fresh" Notebooklm looks like

And it looks like Notebooklm editor of the "Particular" Verified "" Pavel Bannikov

When working with a notebook on the left side of the screen, there is a list of loaded sources, and in the center-a chatbot analyzing sources, the user is primarily interacting with it. In the chat, you can search and ask semi -layer questions about the loaded documents. Each answer can be saved in the form of a note inside the notebook, notes can be turned into new sources. 

In the right side of the screen is the so -called “studio”, in which the user offers five automated methods for studying documents in a notebook: 

  • "Brief review";
  • "Guarder";
  • "Chronology";
  • “Frequently asked questions”;
  • "Audio transit."

"Brief review" And "Particular" - find for students and teachers. The first will help to allocate the main thing from one or all files in a notebook at once. The second will help to develop a whole training course on their basis: lecture notes, practical tasks, essay topics and questions for tests. We tested this function on Hans Rosling's book "Factological"And in 7 sec. received a well -structured lecture notes on this book with the final quiz for students.

This function is also useful for investigations. For example, when experimenting on data on several sites that are suspected of disseminating misinformation in Kazakhstan and other countries, Notebooklm generated a completely true training manual on which you can conduct an investigation, find connections between sites and theoretically even reach the owners of this network.

The function works in a similar way "Chronology": it isolated dates and related semantic elements found in a notebook. Note that the model can mistakenly compare the date and element, which is simply mentioned in the text nearby. In such cases, it is programmed to make a reservation, but it is better to double -check it yourself.

Function "Frequently asked questions" Allows you to generate a block of questions that can probably occur before or when reading the contents of the notebook. The model, respectively, will find answers to them and show them to the user. 

"Audio dispersal" - The most impressive function. It allows you to generate podcast based on a notebook, and this is done in 50 languages, including Russian (the language can be selected in the settings). The conversation is conducted by two AI characters-male and female voices. In English, they sound very naturally, in Russian the incorrect pronunciation of words skips, but the quality of the analysis allows you to put up with this. 

An example of a generated podcast

Next to the Plutstation button, there is a “set up” button. By clicking on it, you can adjust the details of the podcast and the main topics that the “leading” will discuss - you just need to set the parameters important for you in Prompta.

Suppose you analyze the financial documentation of several companies in two years (and these can be dozens of reports for the study of which you have to spend days) and you need to find potential evidence of corruption. You can set just such a clarification and get an audio transmission focused on this topic. Next, you can again switch to text mode and analyze the files that have passed the preliminary selection. 

Just a couple of days before the release of this material, the function of the audio make (so far only in beta testing mode and in English) has become interactive. Now the user can take part in a generated conversation. It is necessary to intervene in the conversation of the “leading” podcast and ask a clarifying question - the conversation will go in the direction you need (you have to wait a bit). 

In the free version, in each notebook you can store only one audio transmission, but this is not a problem - you can download the generated file, delete retelling from the notebook and set the promot to generate podcasts with new tasks. The downloaded file, if desired, can be downloaded as one of the sources. 

Restrictions

  • Each source in a notebook may contain no more than 500,000 words and/or weigh no more than 200 MB.
  • The same restriction applies to the video with subtitles: the file with them should not exceed 500,000 words. However, this is not a problem: even guests of Yuri Dudia are not able to utter so many words in three hours of interviews.
  • In one notebook you can store up to 1000 notes. 
  • Remote notes are not restored.
  • At the moment, the tables (for example, files in XSLX and CSV formats) are not supported, to analyze them will have to be previously transferred to PDF.
  • Synchronization with Google Drive in May 2025 is also incomplete - only documents and presentations. 
  • When analyzing, an audio tool does not break the replicas into speakers and does not give time codes.

Some of these shortcomings will probably be eliminated in the future. So, it is logical to assume that the tool from Google should once start working with Google Sheets, as well as tolerate time codes from the video on YouTube. But in the current configuration, Notebooklm is able to significantly reduce the time to study a large array of data. In the free version, as part of one notebook, up to 25 million words can be analyzed-these are three full 90-volume works Leo Tolstoy.

Share with your friends