iTranslated by AI

The content below is an AI-generated translation. This is an experimental feature, and may contain errors. View original article
📚

A Technique for "Divide-and-Conquer" Reading of O'Reilly Books Prone to Being Left Unread, Using NotebookLM

に公開
2026/11/17

TL;DR

  • When you load a thick technical book (such as an O'Reilly book) into NotebookLM as-is, there are issues like the summary accuracy dropping and slide generation being spread across the whole book, blurring the focus.
  • By splitting the PDF by chapter using a Python script and feeding it into NotebookLM, I became able to generate high-quality explanations and slides specialized to each chapter.
  • Even for difficult technical books, starting from a "visual big picture" lets you keep reading without getting discouraged, which helps you get through your backlog of unread books.

Problems I ran into when loading books into NotebookLM

NotebookLM is super convenient, isn't it?
It strongly supports your learning, whether by loading PDFs or websites and having grounded conversations with expert knowledge, or by using the Studio feature to create slide shows and quizzes.

However, if you load an entire book as-is, a slide show or quiz targeting the whole book gets created, and it tends to become a "blurry-focused" output that forcibly summarizes a long context.
You can control it to some extent with queries, but even then, information from other chapters gets mixed in, and processing time becomes longer, which is a drawback.

Motivation

My motivation was to deeply study long technical books (such as O'Reilly books) chapter by chapter or section by section using NotebookLM's Studio features (audio overviews and slide creation).
So I decided to take the approach of splitting the PDF by chapter and loading them in.

Solution: Creating a PDF splitting script

Splitting manually is a lot of work, so I created a Python script to automate it.

GitHub repository

The tool I created is published here.
https://github.com/katsuki-a/pdf-splitter

Implementation highlights

Using Python's pypdf library, you can retrieve the PDF's table of contents (outline) information.
Using this, I created a script that automatically splits the PDF file at major divisions such as "Chapter 1", "Chapter 2", etc.

Below is an excerpt of the core logic code.

from pypdf import PdfReader, PdfWriter

def split_pdf_by_outline(pdf_path: str, output_dir: str):
    reader = PdfReader(pdf_path)
    # Get the outline (table of contents)
    # * In reality it has a hierarchical structure, so recursive retrieval is needed, but this is simplified here
    outline_items = reader.outline 
    
    sections = []
    for item in outline_items:
        # Get the title and page number from the outline item
        # The retrieval method may differ depending on the pypdf version
        try:
            page = reader.get_destination_page_number(item)
            title = item.title
            sections.append({"title": title, "start": page})
        except:
             continue
             
    # Split and save for each section
    for i in range(len(sections)):
        start = sections[i]["start"]
        # For the last chapter, go to the end of the file
        end = sections[i+1]["start"] if i < len(sections) - 1 else len(reader.pages)
        title = sections[i]["title"]
        
        writer = PdfWriter()
        # Write out the specified range of pages
        for p in range(start, end):
            writer.add_page(reader.pages[p])
            
        with open(f"{output_dir}/{i:02d}_{title}.pdf", "wb") as f:
            writer.write(f)

In the actual code, I handle the hierarchical structure of the table of contents (subsections) and sanitize file names (such as replacing / with _).

One thing that concerns me here is the question, "Is there any legal problem with splitting a PDF on my own or uploading it to NotebookLM?"

As long as it is "within the scope of private use," an exception under copyright law is likely to apply, but attention specific to cloud usage is required.
*Please make the final judgment at your own responsibility.

1. Interpretation of private copying and cloud usage

Under Article 30 of the Copyright Act, processing (splitting) an e-book you purchased for your own learning is recognized as "reproduction for private use."
Regarding uploading to a cloud service as well, as long as "the user themselves is the subject" and the state is "non-public (accessible only by yourself)," it tends to be interpreted as within the scope of private copying (or as not constituting copyright infringement).

2. Service terms and data privacy

In addition to the law, it is necessary to comply with each service's specifications and terms (ToS).

  • O'Reilly Japan: Ebooks are DRM-free, but they are not copyright-free.
  • NotebookLM: As of the time of writing, uploaded personal data is not used to train the underlying models, but please check your organization's policies and the latest terms.

Boundaries that must be strictly observed (prohibited acts)

The following acts that go beyond "private use" carry a high risk of copyright infringement and are strictly forbidden.

  • Using sharing features: Making it viewable by third parties via NotebookLM's "share link," etc.
  • Redistributing files: Transferring or sending the split files to others.
  • Unauthorized use in an organization: Using a personal book as shared knowledge for a team, etc., for business purposes.

Let's use it only within the scope of "extending your own bookshelf into a digital notebook."

Examples of using the created Notebooks

By loading the PDFs split by chapter as described above, you can study with a focus on each chapter.

NotebookLM screen with split PDFs loaded

I felt that books with highly independent chapters, such as "Software Engineering at Google," are particularly compatible.
The slides NotebookLM generated also concisely summarize each chapter's key points in about 15 pages, and just looking at them helps you grasp the overall flow.

Example slide 1 generated by NotebookLM

Example slide 2 generated by NotebookLM

Example slide 3 generated by NotebookLM

As a recommended reading method, you should first look at the slides created with NotebookLM, grasp the overview as a visual image, and then read the main text.
This allows you to keep reading without getting lost, even in places where bullet points make it hard to see ahead or where it's hard to grasp the overall storyline.

"Divide and conquer" of reading through PDF splitting

Also, although I split the PDF for the purpose of loading it into NotebookLM, it turned out that there were benefits for the reading experience itself as well, as follows.

  • It creates a clear boundary, like "Today I'll read only this chapter (this file)."
  • It becomes easier to estimate "how many pages this chapter has" and "how many minutes it will take to read."

Reading that feels endless when you pick up a physical book becomes less psychologically daunting when split digitally, which helps you keep reading steadily.

Summary and thoughts

I think NotebookLM is the best tool for picking up just a part of a book through dialogue with AI without reading the whole thing, having it summarize, or grasping an overview. Beyond that, there are also features like Deep Research, and situations where AI is used to quickly grasp information in the world and apply it to business are increasing.
On the other hand, even in such a world, to train your own thinking ability, the reading experience of inputting a long context into your own head in order from the beginning is also important.

What I think is good about books lies in their "length."
I feel that experiencing a context that cannot be fully conveyed by short blog posts or AI summaries—accumulated context refined into a storyline through the author's deep thinking and trial and error—deepens your own thinking and becomes material to reinforce your verbalization and narratives.

I feel that this method helps support such "heavy reading" and improves the retention rate of knowledge.

Discussion