---

The core process of the document summary feature is executed during the stage of inference, where the input document is analyzed and its key elements are extracted and synthesized to generate a summary.

In this inference process, the AI model utilizes its extensive knowledge base accumulated up to the cutoff date (Note 1) to provide the most appropriate interpretation and response based on the input information.

The content of the input document and the user’s subsequent operation history are not retained as new training data for the AI model.

The input is used solely and temporarily for the specific task of generating the summary of the document, and it is deleted after the process is completed.

---

To enable an AI model to acquire advanced information processing capabilities, a process called “training” is essential.

Training refers to the process of feeding a large volume of well-organized and high-quality datasets into the AI model, allowing it to automatically extract underlying patterns and rules within the data.

For example, by training an AI model with various PDF documents and their corresponding summaries, the model can autonomously learn relationships such as:
“PDFs with a specific structure often contain certain types of important information.”

This training process requires significant computational resources, considerable time, and expert knowledge, and therefore involves a corresponding level of cost.

On the other hand, information entered by an unspecified number of users may be biased, contain noise, or be of low value.

Using such information directly for training could degrade the model’s performance or result in unintended behavior.

Our PDF summarization feature leverages the inference capabilities of a pre-trained AI model, enabling efficient and highly accurate summary generation.

---

In our document summarization feature, the information entered by users is not directly used for the continued training of the AI model.

The inference process is executed based on the model’s existing knowledge base.
Therefore, even documents that contain highly confidential information can be used with confidence.

The underlying service, Amazon Bedrock (Note 2), clearly states in its service terms (Note 3) that input data will not be used for training purposes.

Generative AI technology holds the potential to greatly improve information accessibility and contribute to the creation of new value.

We hope that by gaining a correct understanding of these technical characteristics, you will be able to use this feature as a reliable tool.