KR-20260064661-A - System and method for automated AI model routing based on security sensitivity levels of input prompts

Abstract

The present invention relates to an AI model routing system (100) and a method for processing prompts by analyzing a user's input prompt in an enterprise environment to automatically determine a security sensitivity level, and dynamically selecting one of an internal AI model (150), an external API-linked AI model (160), and a RAG mode (170) according to the level. The prompt analysis unit (120) calculates a security sensitivity score of the input prompt (20) through a keyword extractor (121), a classification model (122), and a grade determiner (123), and the routing determination unit (130) selects an optimal processing path through sequential judgments of a first judgment step (131), a second judgment step (132), and a third judgment step (133). The RAG mode (170) utilizes a high-performance external AI while blocking the external transmission of internal original data through an isolation processing unit (172) and an anonymization processor (173). According to the present invention, there is an effect of simultaneously optimizing the security of enterprise data and the AI processing performance.

Inventors

• 안범주

Assignees

• 안범주

Dates

Publication Date

20260507

Application Date

20260402

Claims (1)

1. As a computing system including a processor and memory, An input receiving unit that receives an input prompt from a user; A prompt analysis unit that analyzes the above input prompt to determine a security sensitivity level; A routing decision unit that selects one of an enterprise-exclusive AI model, an external API-linked AI model, and a search augmented generation (RAG) mode based on the above security sensitivity grade; and A response generation unit comprising processing the input prompt and generating a response through the selected model or mode. AI Model Routing System.

Description

System and method for automated AI model routing based on security sensitivity levels of input prompts System and method for automated AI model routing based on security sensitivity levels of input prompts The present invention relates to an artificial intelligence (AI) model operation system, and more specifically, to an AI model routing system and method that analyzes user input prompts in an enterprise environment to automatically determine a security sensitivity level, and, according to the level, automatically selects an optimal processing path among an AI model dedicated to the enterprise, an AI model linked with an external API, and a Retrieval-Augmented Generation (RAG) mode to process the prompts. With the recent rapid advancement of generative AI technologies, including Large Language Models (LLMs), the adoption of AI in enterprise environments is becoming increasingly active. However, when companies utilize AI in their operations, processing all prompts in a single manner is rarely the optimal approach. Corporate business data possesses a wide range of security sensitivities, ranging from information requiring extreme confidentiality—such as semiconductor circuit design drawings, unfiled inventions, core process technologies, customer personal information, and internal contract documents—to publicly available information. According to conventional technology, companies have operated AI by choosing one of two main approaches. The first is a method of processing prompts through APIs from external AI service providers. While this approach has the advantage of easily utilizing the latest high-performance AI models, it presents a fundamental security vulnerability in that a company's confidential information is transmitted to and processed on external servers. If input data from an external AI service is used to train its models, there is a risk that the company's core technological information could be exposed to competitors. Secondly, there is an approach where a company builds its own in-house AI model to process all prompts exclusively using this internal model. While this method is excellent in terms of data security, it suffers from performance inferiority compared to the latest high-performance AI models and entails massive costs for building and maintaining the proprietary model. Furthermore, it results in inefficiency by consuming expensive internal resources even for processing general public information. Meanwhile, Search Augmented Generative (RAG) technology is garnering attention as an intermediate method that utilizes information from internal corporate databases as context for AI models without directly transmitting the original data externally. However, conventional technologies have not existed that automatically select and route these three processing methods—internal AI models, external API AI models, and RAG modes—within a single integrated system based on the security sensitivity of the prompts. Furthermore, conventional AI routing technologies have primarily focused on intent-based routing aimed at minimizing processing costs or optimizing response speeds, and no technical concept has been proposed to classify the security sensitivity of corporate data into multiple levels and dynamically select heterogeneous AI processing paths accordingly. Therefore, there is an urgent need for technology that can simultaneously optimize data security and AI processing performance by automatically determining the security sensitivity of input prompts in an enterprise environment and dynamically selecting the optimal processing path among internal AI models, RAG mode, and external API AI models based on the determination result. FIG. 1 is a block diagram illustrating the overall configuration of an AI model routing system (100) according to one embodiment of the present invention. FIG. 2 is a flowchart illustrating the processing flow of an AI model routing method according to an embodiment of the present invention. FIG. 3 is a block diagram illustrating the internal configuration and isolation processing structure of a RAG mode (170) according to one embodiment of the present invention. FIG. 4 is a conceptual diagram illustrating a threshold value of a security sensitivity score and a three-stage routing decision structure according to an embodiment of the present invention. Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the attached drawings. However, the technical concept of the present invention is not limited to some of the described embodiments but can be implemented in various different forms, and within the scope of the technical concept of the present invention, one or more of the components among the embodiments may be selectively combined or substituted. Furthermore, unless explicitly defined otherwise, terms used in the embodiments of the present invention should be interpreted in the sense generally known to those skilled i