[Organizational AI Utilization #144] The Wall of NotebookLM Operations. What is the Correct Way to Divide 'Notebooks'?

Hello! This is Terada.

Currently, I serve as the Representative Director of the AI Digital Community (ADC), a community for AI practitioners such as AI promoters at digital-related companies, as well as the Representative Director of FURIKAKE Partners Inc., which supports the 'xAI transformation' of client businesses and AI Portalize Inc., which provides products that support organizational AI utilization, supporting organizational AI utilization from various perspectives! Also, until recently, I served as the Representative Director of D-Marketing Academy Inc., a generative AI and digital marketing talent development training service for corporations, and worked on promoting organizational AI utilization at the AI Promotion Office of CARTA HOLDINGS. In any case, I have been deeply immersed in supporting practical AI utilization for the past two or three years.

In companies that are advancing internal AI adoption, cases of utilizing Google's "NotebookLM" have increased significantly. It is a simple yet powerful tool for creating AI bots and as a learning environment for internal documents.

However, once you actually start operating it, there is a problem you will almost certainly face.
That is the issue of "In what units should notebooks be divided?"

Therefore, this time, I would like to write about the "philosophy of notebook creation" and the "criteria for division" when operating NotebookLM.

The "How much information should I include?" problem encountered in operations

Before entering the main topic, I will organize the premises briefly.
NotebookLM is a tool where AI learns from documents uploaded by the user and provides answers based on that content.

For example,

• Loading sales materials to create sales pitches

• Inputting legal PDFs to check if internal materials are compliant

• Loading YouTube video links to create summaries

are common ways to use it.
It is very convenient, but as the "sources" that serve as information origins increase, the difficulty of management rises.

For example, imagine a case where you try to create an FAQ bot for sales.
Even if you call it "for sales," the information needed in the field is diverse.

• Frequently asked questions from customers (for external use)

• Knowledge that new employees should know (for internal training)

• Internal application flows and administrative procedures (business workflows)

Should all of these be packed into one notebook, or should they be divided?
What should be considered here is the "effort of updating" in NotebookLM.
With the current specifications, if you want to update the content of an uploaded file, you need to either delete it and add it again, or perform a re-sync action even with Drive integration.

How to balance management effort and answer accuracy.
Therefore, I would like to recommend the idea of dividing by "three axes."

※ Please also see this article regarding the units in which NotebookLM should be divided.

Axis 1: Who is using it?

The first axis is simply to categorize by 'user'.

• Is it for sales representatives?

• Is it for new employees?

• Is it for administrative departments?

If the users are different, the tone of the requested answers and the scope of necessary information will naturally change as well.
By separating notebooks according to user attributes, you should be able to minimize discrepancies in the answers.

Axis 2: What is it used for?

The second is to categorize by 'purpose of use'.

• Is it for 'external' use (customer responses, presentation materials, etc.)?

• Is it for 'internal' use (internal regulations, application flows, etc.)?

There is a risk if you mix these up.
For example, you might want to generate a draft response for a customer, but the AI might generate an answer that includes internal terminology or sensitive internal information.

When you need to separate information, I think it is safer to clearly divide the notebooks.

However, there are some points to keep in mind.
If you divide them too finely, such as by 'specific laws' or 'expense reimbursement vs. attendance management,' it becomes difficult to find where everything is.

If the amount of information in the sources is not that large, there is no problem with grouping them together to some extent.

Axis 3: Balance between capacity and accuracy

The third axis is the technical issue of 'capacity and accuracy'.

You can add many sources to NotebookLM, but if you exceed a certain amount, information loss or hallucinations (plausible lies) may occur.
If you feel that the accuracy is low because there is too much information, you need to make adjustments.

• Reduce the number of sources

• Consolidate multiple documents into a single summary file before uploading to reduce the load

In this way, it is also very important to decide on segmentation from the perspective of 'is this a size that the AI can process correctly?'

*Please also see this article regarding the fact that NotebookLM source information is not automatically updated and the countermeasures for it.

First, create a broad framework based on "Who and For What Purpose"

How should notebooks be divided when utilizing NotebookLM?

When in doubt, I recommend organizing them using the following steps.

1. Who: Define who the users are

2. For what purpose: Separate whether it is for external or internal use

3. Accuracy/Capacity: Check if there is a decrease in accuracy due to overloading information

First, try creating a broad framework based on "Who and For What Purpose."

Then, after actually using it, if there are issues with accuracy, perform splitting or information compression from the perspective of "Capacity."

I believe that by proceeding with this procedure, you can build an AI environment that is easy to operate and has high accuracy.

Thank you for reading! Here are past AI-related articles!