Method and apparatus for providing similarity model-based fine-tuning generative ai

Abstract

Provided are a method and a service device for providing similarity model-based fine-tuning generative AI. The method is performed by a fine-tuning generative AI service server and comprises the steps of: receiving a first input message from a user; searching for a second input message by using a similarity AI model from among one or more input messages received from the user and stored in a database, the second input message having a similarity degree of a predetermined reference value or higher with respect to the first input message received from the user; and outputting, to the user, an answer to the first input message according to a result of the search.

Description

Method and device for providing AI for fine-tuning generative AI based on similarity model

[0001] The present invention relates to generative AI, and more particularly, to a method and device for providing fine-tuned generative AI based on a similarity model.

[0002] Generative AI (Generative AI) refers to artificial intelligence technology that generates content such as text, images, videos, audio, and source code. Generative (AI) models are distinguished from discriminative models, which classify or predict input data based on training criteria, in that they generate new data similar to the training data set. Generative AI models (such as applications or systems) can perform tasks such as natural language processing, text writing, machine translation, image generation, video clip generation, audio sample generation, and source code generation. Well-known generative AI models include OpenAI's ChatGPT, GPT-3, GPT-4, and DALL-E; Microsoft's Bing Chat; and Google's Bard, Midjourney, and Stable Diffusion.

[0003] Generative AI offers numerous advantages, including saving time, enhancing creativity, and being applied in a variety of fields. However, it also presents serious challenges. First, the high cost of generative AI services. Most generative AI services operate on a pay-as-you-go basis. For example, ChatGPT charges a fixed fee based on the number of tokens in input and output messages, while DALL-E charges a fixed fee based on the resolution and number of images. There is a risk that fees will increase excessively if generative AI service usage increases. Second, the use of generative AI services can lead to the leakage of sensitive information, such as personal information or trade secrets. In fact, some companies are attempting to prohibit the use of generative AI services internally or develop their own generative AI for internal use due to concerns about trade secret leaks. Third, obtaining the desired level of response from users is extremely difficult. Generative AI often provides inaccurate or unclear answers. To improve the response quality of generative AI, additional work, such as prompt engineering, is required to find the optimal combination of input values.

[0004] The problem to be solved by the present invention is to provide a method and device for providing AI for fine-tuning generation based on a similarity model.

[0005] However, the problems to be solved by the present invention are not limited to the problems described above, and other problems may exist.

[0006] As a means for solving the above-described problem, a method for providing a similarity model-based fine-tuning generation AI according to a first aspect of the present invention comprises the steps of receiving a first input message from a user, searching for a second input message having a similarity higher than a predetermined threshold with the first input message received from the user among one or more input messages received from the user stored in a database using a similarity AI model, and outputting an answer regarding the first input message to the user according to the search result, wherein the step of outputting an answer regarding the first input message to the user according to the search result comprises the steps of: if the second input message is not searched, inputting the first input message received from the user to a generation AI model of an external server; receiving a first output message as an answer regarding the first input message from the generation AI model; pairing the first input message received from the user and the first output message received from the generation AI model and storing them in a database, and outputting the first output message as an answer regarding the first input message to the user; and if the second input message is searched, outputting the first input message received from the user to the The method further comprises: extracting a second output message paired with the second input message from among one or more output messages received from the generating AI model stored in the database without inputting the second output message to the generating AI model; and outputting the second output message to the user as a response to the first input message.

[0007] In some embodiments of the present invention, the step of searching for the second input message using the similarity AI model may include, when a plurality of input messages having a similarity higher than a predetermined threshold with the first input message received from the user are searched among one or more input messages received from the user stored in the database, determining an input message having the highest similarity among the plurality of input messages having a similarity higher than a predetermined threshold with the first input message as the second input message.

[0008] In some embodiments of the present invention, the step of inputting the first input message received from the user to the generation AI model of the external server may include the steps of detecting preset security information from the first input message received from the user, preprocessing a portion corresponding to the security information in the first input message received from the user by one or more of blocking, masking, encryption, and de-identification methods, and inputting the preprocessed first input message to the generation AI model of the external server.

[0009] Additionally, the method may further include a step of recording and managing a usage log regarding the generated AI model of the external server.

[0010] In addition, the usage log may include one or more of the name of the generated AI model used, the date or time of use, the input message received from the user, the portion corresponding to the security information, the presence or absence of the preprocessing, the preprocessing type, the preprocessed input message, and the output message received from the generated AI model.

[0011] In some embodiments of the present invention, the method may further include the steps of receiving a training data set for fine-tuning including one or more paired input messages and output messages from the user, and uploading the training data set for fine-tuning received from the user to the external server to fine-tune the generated AI model.

[0012] In addition, the step of uploading the training data set for the fine-tuning to the external server and fine-tuning the generated AI model may include the steps of detecting preset security information from the training data set for the fine-tuning received from the user, the step of preprocessing a portion corresponding to the security information within the training data set for the fine-tuning received from the user by one or more of blocking, masking, encryption, and de-identification methods, and the step of uploading the pre-processed training data set to the external server and fine-tuning the generated AI model.

[0013] In addition, the method further includes a step of increasing the amount of input messages of the training data set for the fine-tuning using the generated AI model of the external server, and the step of uploading the training data set for the fine-tuning to the external server to fine-tune the generated AI model may include uploading the increased training data set for the fine-tuning to the external server to fine-tune the generated AI model.

[0014] And, the method further includes a step of storing a training data set for the augmented fine-tuning in the database, and the step of searching for the second input message may include searching for a second input message having a similarity higher than a predetermined threshold with the first input message received from the user among one or more messages received from the user stored in the database and one or more input messages of the training data set for the augmented fine-tuning using the similarity AI model, and the step of extracting the second output message may include extracting a second output message paired with the second input message among one or more output messages received from the generation AI model stored in the database and one or more output messages of the training data set for the augmented fine-tuning.

[0015] As a means for solving the above-described problem, a similarity model-based fine-tuning generation AI providing device according to a first aspect of the present invention includes a memory storing one or more instructions, and one or more processors executing the one or more instructions stored in the memory, wherein the one or more processors execute the one or more instructions to receive a first input message from a user, search for a second input message having a similarity higher than a predetermined reference value with the first input message received from the user among one or more input messages received from the user stored in a database using a similarity AI model, and output an answer regarding the first input message to the user according to the search result, wherein outputting the answer regarding the first input message to the user according to the search result comprises: if the second input message is not searched, inputting the first input message received from the user to a generation AI model of an external server, receiving the first output message as an answer regarding the first input message from the generation AI model, pairing the first input message received from the user with the first output message received from the generation AI model, and storing the paired first input message received from the user and the first output message received from the generation AI model in a database, and It includes outputting the first output message to the user as a response to the first input message, and when the second input message is searched, extracting a second output message paired with the second input message from among one or more output messages received from the generated AI model stored in the database, and outputting the second output message to the user as a response to the first input message.

[0016] Other specific details of the present invention are included in the detailed description and drawings.

[0017] According to the present invention described above, an input message received from a user and an output message received from a generation AI model are paired and stored in a database, and when a new input message is received from the user thereafter, a similar input message is searched for among existing input messages stored in the database using a similarity AI model, and an output message corresponding to the searched similar input message is provided to the user as a response without using the generation AI model, thereby reusing the response result of the generation AI model, thereby reducing the number of uses of the generation AI model and the associated costs.

[0018] In addition, according to the present invention described above, before transmitting an input message received from a user or a training data set for fine-tuning to the generating AI model in the process of using the generating AI model, security information is detected from it and preprocessed by blocking, masking, encryption, anonymization, etc., and usage logs related to the generating AI model are recorded and managed, thereby preventing situations in which sensitive information such as personal information or corporate trade secrets are leaked in the process of using the generating AI service.

[0019] In addition, according to the present invention described above, by increasing the amount of input messages of a training data set for fine-tuning using a generative AI model, fine-tuning the generative AI model using the training data set for fine-tuning, and simultaneously loading the training data set for fine-tuning into a database, the generative AI model can be made to generate answers at a level desired by the user without prompt engineering work, or the user can obtain basic answers stored in the database.

[0020] The effects of the present invention are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the description below.

[0021] Figure 1 is a block diagram schematically illustrating a generation AI providing environment according to one embodiment of the present invention.

[0022] FIG. 2 is a block diagram schematically illustrating a generation AI providing device according to one embodiment of the present invention.

[0023] Figure 3 is a flowchart schematically illustrating a method for providing AI generation according to one embodiment of the present invention.

[0024] Figures 4 and 5 are flowcharts schematically showing detailed steps of step S500 of Figure 3.

[0025] Figures 6 and 7 are conceptual diagrams schematically illustrating an output message generation process by a generation AI providing method according to one embodiment of the present invention.

[0026] Figure 8 is a block diagram schematically illustrating a generation AI providing device according to another embodiment of the present invention.

[0027] Figure 9 is a flowchart schematically illustrating a method for providing AI generation according to another embodiment of the present invention.

[0028] Figure 10 is a conceptual diagram schematically illustrating a security information filtering process by a method for providing a generation AI according to another embodiment of the present invention.

[0029] FIG. 11 is a block diagram schematically illustrating a generation AI providing device according to another embodiment of the present invention.

[0030] Figure 12 is a flowchart schematically illustrating a method for providing AI generation according to another embodiment of the present invention.

[0031] Figure 13 is a flowchart schematically illustrating a method for providing AI generation according to another embodiment of the present invention.

[0032] Figure 14 is a flowchart schematically illustrating a method for providing AI generation according to another embodiment of the present invention.

[0033] Figure 15 is a conceptual diagram schematically illustrating a process of fine-tuning a generated AI model by a method for providing generated AI according to another embodiment of the present invention.

[0034] Figure 16 is a block diagram schematically illustrating a generation AI providing device according to another embodiment of the present invention.

[0035] The advantages and features of the present invention, and the methods for achieving them, will become clearer with reference to the embodiments described in detail below together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below and may be implemented in various different forms. These embodiments are provided solely to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the scope of the present invention, and the present invention is defined solely by the scope of the claims.

[0036] The terminology used herein is for the purpose of describing embodiments only and is not intended to limit the present invention. In this specification, the singular also includes the plural unless specifically stated otherwise. As used herein, the terms "comprises" and / or "comprising" do not exclude the presence or addition of one or more other components in addition to the mentioned components. Like reference numerals refer to like components throughout the specification, and "and / or" includes each and any combination of one or more of the mentioned components. Although "first", "second", etc. are used to describe various components, these components are not limited by these terms. These terms are only used to distinguish one component from another. Therefore, it should be understood that a first component mentioned below may also be a second component within the technical spirit of the present invention.

[0037] Unless otherwise defined, all terms (including technical and scientific terms) used herein may be used in their common sense to those skilled in the art to which the present invention pertains. Furthermore, terms defined in commonly used dictionaries are not to be interpreted ideally or excessively unless explicitly and specifically defined otherwise.

[0038] In describing the present invention, if it is judged that the detailed description of related known technology is obvious to a person skilled in the art and may unnecessarily obscure the gist of the present invention, it will be omitted.

[0039] "Generative AI model" refers to an artificial intelligence model that generates any type of data or content, such as text, images, videos, audio, or source code. When categorizing generative AI models by content type, the generative AI model can be any of the following: a conversational model, an image generation model, a video generation model, an audio generation model, or a source code generation model. The generative AI model may have a natural language-based conversational user interface, but is not limited thereto, and may also have a user interface based on any data type. The generative AI model may be a unimodal model that receives a single type of data as input, or a multimodal model that receives two or more types of data as input. In addition to chatbot-type models such as ChatGPT, generative AI models include language models that process natural language, such as GPT-3, GPT-3.5, and GPT-4.

[0040] "Similarity AI model" refers to an artificial intelligence model that determines or infers the similarity of any type of data or content, such as text, images, videos, audio, and source code. For example, if a generative AI service is provided as a natural language-based conversational interface, the similarity AI model may vector embed text messages and determine the similarity between text messages using cosine similarity, but is not limited thereto. The similarity AI model may also determine the similarity between text messages using BERT, Sentence BERT, or any other natural language processing model as a base model. The similarity AI model may be implemented in various forms according to any similarity determination method well known in the art to which the present invention pertains.

[0041] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings.

[0042] Figure 1 is a block diagram schematically illustrating a generation AI providing environment according to one embodiment of the present invention.

[0043] Referring to FIG. 1, a generation AI providing device (100) is connected to an external server (200) via a network. The generation AI providing device (100) may be provided in the form of a server that provides a generation AI service online as illustrated in FIG. 1, but is not limited thereto.

[0044] The network may include a wired network, a wireless network, or a combination thereof, capable of transmitting and receiving any form of signal, data, or information.

[0045] The external server (200) is a server operated by a company that develops and supplies a generative AI service, and includes a generative AI model (210) therein. The generative AI providing device (100) can input a message to the generative AI model (210) of the external server (200) and receive a message regarding the response from the generative AI model (210). The message can include any type of data or content, such as text, images, videos, audio, and source code. The generative AI providing device (100) can communicate with the generative AI model (210) through a natural language-based conversational user interface or a user interface based on any data type. Alternatively, the generative AI providing device (100) can communicate with the generative AI model (210) through an API (Application Programming Interface) provided by the external server (200). For example, OpenAI's ChatGPT, GPT-3, GPT-3.5, GPT-4, Naver's HyperCLOVA, Kakao's KoGPT, etc. can be used as generative AI models (210), but are not limited thereto.

[0046] Although FIG. 1 illustrates one external server (200) and one generation AI model (210), the invention is not limited thereto, and the generation AI providing device (100) may be connected to multiple external servers (200), to an external server (200) including multiple generation AI models (210), or to a combination thereof. In other words, the generation AI providing device (100) may utilize various generation AI services of any company.

[0047] FIG. 2 is a block diagram schematically illustrating a generation AI providing device according to one embodiment of the present invention.

[0048] Referring to FIG. 2, a generation AI providing device (100) according to one embodiment of the present invention includes a user interface unit (110), a message processing unit (120), a similarity AI model (130), a database unit (140), and a generation AI interface unit (150).

[0049] The user interface unit (110) is a component for interfacing with a user or a user group (hereinafter referred to as “user”). The user interface unit (110) receives an input message from a user and outputs an output message to the user as a response to the input message.

[0050] The message processing unit (120) is connected to the user interface unit (110), the similarity AI model (130), the database unit (140), and the generation AI interface unit (150). The message processing unit (120) receives an input message received from a user from the user interface unit (100), and uses the similarity AI model (130) to search for an input message having a similarity level higher than a predetermined standard with the input message received from the user among the input messages stored in the database (140). The message processing unit (120) determines a method for generating a response to be output to the user based on the search result.

[0051] If the search fails, the message processing unit (120) inputs the input message received from the user to the generation AI model (210) through the generation AI interface unit (150). Then, the message processing unit (120) receives an output message from the generation AI model (210) through the generation AI interface unit (150) and outputs the received output message to the user as a response to the input message through the user interface unit (110). The message processing unit (120) can pair the input message received from the user with the received output message and store them in the database (140).

[0052] If the search is successful, the message processing unit (120) extracts an output message paired with an input message having a similarity level higher than the predetermined standard from among the output messages stored in the database (140) without using the generation AI model (210), and outputs the extracted output message as a response to the input message to the user through the user interface unit (110).

[0053] The similarity AI model (130) is connected to the message processing unit (120) and the database unit (140). The similarity AI model (130) receives an input message received from a user from the message processing unit (120) and searches the database (140) according to a request of the message processing unit (120). The similarity AI model (130) searches for an input message having a similarity level higher than a predetermined standard with the input message received from the message processing unit (120) among the input messages stored in the database (140). The similarity AI model (130) outputs the search result to the message processing unit (120). For example, the search result may include one or more of a logical value (TRUE or FALSE) depending on whether the search was successful, a similarity calculation result, and the contents of an input message having a similarity level higher than a predetermined standard, but is not limited thereto. The similarity standard can be set and modified by the user.

[0054] The database unit (140) stores one or more input messages received from a user and one or more output messages received from the generation AI model (210). The database unit (140) stores one or more data sets in which input messages and output messages are paired. The database unit (140) may include a RAM (Random Access Memory), a ROM (Read Only Memory), an EPROM (Erasable Programmable ROM), an EEPROM (Electrically Erasable Programmable ROM), a flash memory, a hard disk, a removable disk, a CD-ROM, or any other form of computer-readable recording medium well known in the art to which the present invention pertains. The database unit (140) may be implemented as a relational database model or a graph database model, but is not limited thereto. Hereinafter, the database unit may be abbreviated as “database.”

[0055] The generation AI interface unit (150) is a component for interfacing with the generation AI model (210) of the external server (200). The generation AI interface unit (150) is placed between the message processing unit (120) and the generation AI model (210) and can communicate with the generation AI model (210) of the external server (200). The generation AI interface unit (150) performs the role of transmitting data, messages, information, etc. between the message processing unit (120) and the generation AI model (210).

[0056] FIG. 3 is a flowchart schematically illustrating a method for providing a generation AI according to one embodiment of the present invention, FIGS. 4 to 5 are flowcharts schematically illustrating detailed steps of step S500 of FIG. 3, and FIGS. 6 to 7 are conceptual diagrams schematically illustrating an output message generation process by a method for providing a generation AI according to one embodiment of the present invention.

[0057] The generation AI providing method illustrated in FIG. 3 can be performed by the generation AI providing device (100).

[0058] Referring to FIG. 3, in step S300, the generation AI providing device (100) receives a first input message from the user. Here, the “first input message” is one of the input messages received from the user and is used to distinguish it from the “second input message” described below.

[0059] Next, in step S400, the generation AI providing device (100) searches for a second input message having a similarity level higher than a predetermined standard with the first input message among one or more messages received from a user stored in a database (140) using a similarity AI model (130).

[0060] The first input message and the second input message may be completely identical messages, or may be similar messages, although not identical. For example, if the similarity between the first input message and the second input message is 100%, the first input message and the second input message may be completely identical messages. For another example, if the similarity between the first input message and the second input message is 80%, the first input message and the second input message may be similar messages, although not identical. The criteria for classifying whether messages are similar may be preset in various ways depending on the embodiment.

[0061] Next, in step S500, the generation AI providing device (100) outputs a response regarding the first input message to the user based on the search results. At this time, depending on whether the second input message is searched, the generation AI providing device (100) generates an output message regarding the first input message in different ways and provides the generated output message to the user. This will be described in more detail with reference to FIGS. 4 to 7.

[0062] Figure 4 illustrates a method for generating and outputting an output message when the second input message is not searched.

[0063] Referring to FIGS. 4 and 6, in step S510, the generation AI providing device (100) inputs the first input message received from the user to the generation AI model (210) of the external server (200).

[0064] Next, in step S520, the generation AI providing device (100) receives a first output message as a response to the first input message from the generation AI model (210).

[0065] Next, in step S530, the generation AI providing device (100) pairs the first input message received from the user with the first output message received from the generation AI model (210) and stores them in the database (140). In step S540, the generation AI providing device (100) outputs the first output message to the user as a response to the first input message.

[0066] Referring to FIG. 6, if a second input message having a similarity level higher than a predetermined standard with the first input message is not searched, the generation AI providing device (100) transmits the first input message received from the user to the generation AI model (210) as is, and transmits and outputs the first output message received from the generation AI model (210) to the user as is.

[0067] In this case, the cost of using the generation AI model (210) is inevitably incurred. If the subsequent input messages received from the user are new input messages, i.e., if the similarity with the previous input messages received from the user is less than a predetermined standard, the generation AI providing device (100) has no choice but to use the generation AI model (210). However, the generation AI providing device (100) stores the output messages received from the generation AI model (210) in the database (140) in preparation for reuse.

[0068] Figure 5 illustrates a method for generating and outputting an output message when the second input message is searched.

[0069] Referring to FIGS. 5 and 7, in step S550, the generation AI providing device (100) does not input the first input message received from the user to the generation AI model (210), but extracts a second output message paired with the second input message from among one or more output messages received from the generation AI model (210) stored in the database (140).

[0070] Next, in step S560, the generation AI providing device (100) outputs the second output message to the user as a response to the first input message.

[0071] Referring to FIG. 7, if a second input message having a similarity level higher than a predetermined threshold with the first input message is searched, the generation AI providing device (100) does not use the generation AI model. The generation AI providing device (100) extracts a second output message paired with the second input message from the database (140) and then outputs the second output message to the user as a response to the first input message. This is because if the contexts of the first input message and the second input message are sufficiently similar, there is no need to use the generation AI model (210) redundantly, and it is possible to reuse the second output message as a response.

[0072] In this case, unlike the previous case, no cost is incurred because the generation AI model (210) is not used. If subsequent input messages received from the user are similar to input messages received from the user in the past, i.e., if the similarity with the previous input message received from the user is above a predetermined standard, the number of times the generation AI model (210) is used can be reduced and related costs can be saved by reusing the output message stored in the database (140) within the generation AI providing device (100).

[0073] Meanwhile, if a plurality of input messages having a similarity level higher than a predetermined standard with the first input message are searched among input messages stored in the database (140), the generation AI providing device (100) may determine the input message having the highest similarity level among the plurality of input messages as the second input message. In addition, the generation AI providing device (100) may extract a second output message paired with the determined second input message from the database (140) in order to provide a response to the user.

[0074] The device and method for providing AI-generated content according to one embodiment of the present invention may be particularly effective for a user group comprising two or more users requiring questions and answers in the same context. Multiple users can share and use the AI-generated content provider (100), thereby saving costs related to the AI-generated content provider (210) and effectively sharing the content and knowledge provided by the AI-generated content provider (100).

[0075] Figure 8 is a block diagram schematically illustrating a generation AI providing device according to another embodiment of the present invention.

[0076] Referring to FIG. 8, a generation AI providing device (100`) according to another embodiment of the present invention includes a user interface unit (110), a message processing unit (120), a similarity AI model (130), a database unit (140), a generation AI interface unit (150), and a security information detection unit (160). Compared to the generation AI providing device (100) described with reference to FIG. 2, the generation AI providing device (100`) further includes a security information detection unit (160).

[0077] The security information detection unit (160) is arranged between the message processing unit (120) and the generation AI interface unit (150), and is connected to the message processing unit (120), the similarity AI model (130), the database unit (140), and the generation AI interface unit (150). The security information detection unit (160) detects preset security information from the input message received from the user by the message processing unit (120) during the process of inputting the input message to the generation AI model (210). The security information detection unit (160) preprocesses a portion corresponding to security information within the input message received from the user. The security information detection unit (160) transmits the preprocessed input message to the generation AI interface unit (150). The security information detection unit (160) transmits an output message received from the generation AI interface unit (150) to the message processing unit (120). Unlike as illustrated in FIG. 8, the output message received from the generation AI model (210) may be delivered to the message processing unit (120) by bypassing the security information detection unit (160). The security information detection unit (160) may record and manage usage logs regarding the generation AI model (210) in a database (140).

[0078] For example, preset security information may include, but is not limited to, sensitive information such as personal information or corporate trade secrets. "Personal information" refers to information about a living individual that can identify the individual, such as name, resident registration number, or image. Personal information includes information that cannot identify a specific individual on its own, but can be easily combined with other information to identify the individual. Personal information includes objective factual information such as name, address, contact information, income, education, grades, occupation, email address, image, call content, credit, debt, and Internet access IP, as well as subjective information such as third-party opinions or evaluations of the individual. "Corporate trade secrets" refer to production methods, sales methods, and other technical or business information useful for business activities that are not publicly known and have independent economic value and are managed as secrets. The security information detection unit (160) can detect security information using any method well known in the technical field to which the present invention pertains. The security information detection unit (160) may also utilize the similarity AI model (130) in the process of detecting security information.

[0079] The generation AI interface unit (150) plays a role in transmitting data, messages, information, etc. between the security information detection unit (160) and the generation AI model (210).

[0080] Fig. 9 is a flowchart schematically illustrating a method for providing AI generation according to another embodiment of the present invention, and Fig. 10 is a conceptual diagram schematically illustrating a security information filtering process by a method for providing AI generation according to another embodiment of the present invention. The method for providing AI generation illustrated in Fig. 9 can be performed by a device for providing AI generation (100`).

[0081] In Fig. 9, detailed steps of step S510 of Fig. 4 are illustrated as another embodiment of the present invention.

[0082] Referring to FIGS. 9 and 10, in step S511, the generation AI providing device (100`) detects preset security information from a first input message received from a user.

[0083] Next, in step S512, the generation AI providing device (100`) preprocesses the portion corresponding to the security information in the first input message received from the user by one or more of blocking, masking, encryption, and de-identification methods.

[0084] Next, in step S513, the generation AI providing device (100`) inputs the preprocessed first input message to the generation AI model (210) of the external server (200). Unlike as illustrated in FIG. 9, if the preprocessing method in step S512 is blocking, the generation AI providing device (100`) may return the first input message to the user without inputting it to the generation AI model (210) of the external server (200).

[0085] Next, in step S514, the generation AI providing device (100`) records and manages a usage log regarding the generation AI model (210).

[0086] Referring to FIG. 10, the generation AI providing device (100`) not only detects and filters security information, but also records and manages usage logs of the generation AI model. The generation AI providing device (100`) does not input the input message (10) received from the user as is to the generation AI model (210), but detects a portion corresponding to security information within the input message (10) and can identify a portion (11) corresponding to non-security information and a portion (12) corresponding to security information. In addition, the generation AI providing device (100`) can block the portion (12) corresponding to security information from being transmitted to the generation AI model (210). Alternatively, the generation AI providing device (100`) preprocesses the portion (12) corresponding to security information by a method such as de-identification, anonymization, pseudonymization, or encryption.

[0087] In addition, the generation AI providing device (100`) records a usage log of the generation AI model (210). The usage log management is to track, manage, and audit the use of the generation AI model (210) by a user or a user group. For example, the usage log may include, but is not limited to, one or more of the name of the generated AI model used, the date or time of use, the input message received from the user, the part corresponding to the security information, the presence or absence of the preprocessing, the type of the preprocessing, the preprocessed input message, and the output message received from the generated AI model.

[0088] FIG. 11 is a block diagram schematically illustrating a generation AI providing device according to another embodiment of the present invention.

[0089] Referring to FIG. 11, a generation AI providing device (100``) according to another embodiment of the present invention includes a user interface unit (110), a message processing unit (120), a similarity AI model (130), a database unit (140), a generation AI interface unit (150), a security information detection unit (160), and a fine-tuning unit (170). Compared to the generation AI providing device (100`) described with reference to FIG. 8, the generation AI providing device (100``) further includes a fine-tuning unit (170).

[0090] The fine-tuning unit (170) can interface directly with the user, similar to the user interface unit (110). The interface for exchanging messages with the user and the interface for fine-tuning the generated AI model can be provided separately. Alternatively, unlike as illustrated in FIG. 11, the fine-tuning unit (170) can interface indirectly with the user through the user interface unit (110).

[0091] The fine-tuning unit (170) is connected to the similarity AI model (130) and the security information detection unit (150). The fine-tuning unit (170) can receive a training data set for fine-tuning from a user. The training data set for fine-tuning can include one or more paired input messages and output messages. The fine-tuning unit (170) can upload the training data set for fine-tuning received from the user to an external server (200) to fine-tune the generated AI model (211). The fine-tuning unit (170) can summarize the fine-tuning result of the generated AI model (211) and output it to the user. The generated AI model (211) represents a model that can be fine-tuned among the generated AI models (210) described above.

[0092] The fine-tuning unit (170) can augment the training data set for fine-tuning received from the user using the generation AI model (212). At this time, the fine-tuning unit (170) can augment the amount of input messages of the training data set using the generation AI model (212). For example, a method of augmenting the amount of input messages of the training data set may be, but is not limited to, inputting the input messages of the training data set to the generation AI model (212) and paraphrasing the input messages, such as summarizing them or changing the tone of the input messages. If the ratio of input messages and output messages of the original training data set received from the user is 1:1, the ratio of input messages and output messages of the training data set augmented by the fine-tuning unit (170) may be N:1 (N is a natural number greater than or equal to 2). In this way, the generation AI providing device (100``) can fine-tune the generation AI model (211) by uploading a training data set for enhanced fine-tuning to an external server (200).

[0093] Alternatively, the ratio of input messages to output messages in the original training data set received from the user may be originally N:1. This is because the user augmented the input messages when generating the training data set for fine-tuning.

[0094] In the generation AI providing device (100``), the security information detection unit (160) is placed between the message processing unit (120) and the generation AI interface unit (150), and between the fine-tuning unit (170) and the generation AI interface unit (150). The security information detection unit (160) is connected to the message processing unit (120), the similarity AI model (130), the database unit (140), the generation AI interface unit (150), and the fine-tuning unit (170).

[0095] The security information detection unit (160) detects preset security information from the training data set for fine-tuning, which the fine-tuning unit (170) receives from the user and uploads to the external server (200), when the fine-tuning unit uploads the training data set for fine-tuning. The security information detection unit (160) preprocesses a portion corresponding to security information within the training data set for fine-tuning received from the user. The security information detection unit (160) transmits the preprocessed training data set to the generation AI interface unit (150). Although not explicitly illustrated, the generation AI interface unit (150) may receive the fine-tuning result of the generation AI model (211) from the external server (200). The security information detection unit (160) may transmit the fine-tuning result received from the generation AI interface unit (150) to the fine-tuning unit (170). Unlike as illustrated in FIG. 11, the fine-tuning result may be transmitted to the fine-tuning unit (170) by bypassing the security information detection unit (160). The security information detection unit (160) can detect and preprocess preset security information from the training data set for fine-tuning received from the user using the generated AI model (212) of the fine-tuning unit (170) during the process of augmenting the training data set for fine-tuning. The security information detection unit (160) can also perform substantially the same role during the process of uploading the augmented training data set for fine-tuning to an external server (200).

[0096] The fine-tuning unit (170) can store the original training data set and the augmented training data set received from the user in the database (140).

[0097] When the message processing unit (120) searches for input messages stored in the database (140) using the similarity AI model (130) to generate an answer to be output to the user, the input messages of the original training data set and the augmented training data set also become search targets.

[0098] The generation AI interface unit (150) plays a role in transmitting data, messages, information, etc. between the fine-tuning unit (170) and the generation AI model (211), between the security information detection unit (160) and the external server (200) (or the generation AI model (211)), or between the security information detection unit (160) and the generation AI model (212).

[0099] The generative AI model (211) that is the target of fine-tuning and the generative AI model (212) used for data augmentation may be the same model or different models depending on the embodiment.

[0100] FIGS. 12 to 14 are flowcharts schematically illustrating a method for providing a generation AI according to another embodiment of the present invention, and FIG. 15 is a conceptual diagram schematically illustrating a process for fine-tuning a generation AI model by a method for providing a generation AI according to another embodiment of the present invention. The generation AI providing methods illustrated in FIGS. 12 to 14 can be performed by a generation AI providing device (100``).

[0101] Referring to FIGS. 12 and 15, in step S600, the generation AI providing device (100`) receives a training data set for fine-tuning including one or more paired input messages and output messages from a user.

[0102] Next, in step S700, the generation AI providing device (100``) uploads a training data set for fine-tuning received from the user to an external server (200) to fine-tune the generation AI model (211).

[0103] The contents of the remaining steps S300, S400, and S500 are the same as those described with reference to FIG. 3, so redundant descriptions will be omitted. The generation AI providing device (100``) may input the first input message received from the user into the fine-tuned generation AI model (211), and receive a first output message as a response to the first input message from the fine-tuned generation AI model (211). The user may receive an optimized response from the generation AI providing device (100``) according to the fine-tuning.

[0104] In Fig. 13, as another embodiment of the present invention, detailed steps of step S700 of Fig. 12 are illustrated.

[0105] Referring to FIGS. 13 and 15, in step S710, the generation AI providing device (100``) detects preset security information from a training data set for fine-tuning received from a user.

[0106] Next, in step S720, the generation AI providing device (100``) preprocesses a portion corresponding to the security information in the training data set for the fine-tuning received from the user by one or more of blocking, masking, encryption, and de-identification methods.

[0107] Next, in step S730, the generation AI providing device (100``) uploads the preprocessed training data set to an external server (200) to fine-tune the generation AI model. Unlike as illustrated in FIG. 13, if the preprocessing method in step S730 is blocking, the generation AI providing device (100``) may return to the user at least a portion of the training data set without uploading it to the external server (200).

[0108] In Fig. 14, as another embodiment of the present invention, a method for providing a generated AI is illustrated that further includes step S800 compared to Fig. 12.

[0109] Referring to FIGS. 14 and 15, in step S800, the generation AI providing device (100``) increases the amount of input messages of the training data set for fine-tuning using the generation AI model (212).

[0110] Next, in step S700, the generation AI providing device (100``) uploads a training data set for enhanced fine-tuning to an external server (200) to fine-tune the generation AI model (211).

[0111] The contents of the remaining steps S300, S400, S500, and S600 are the same as those described with reference to FIGS. 3 and 12, so duplicate descriptions will be omitted.

[0112] Although not explicitly shown in FIG. 13, the method for providing a generation AI may further include a step in which the generation AI providing device (100``) stores the training data set for the enhanced fine-tuning in a database (140).

[0113] In this case, in step S400, the generation AI providing device (100``) can search for a second input message having a similarity higher than a predetermined standard with a first input message received from a user among one or more messages received from a user stored in a database (140) and one or more input messages of the training data set for the enhanced fine-tuning using a similarity AI model (130).

[0114] Next, in step S500, the generation AI providing device (100``) may extract a second output message paired with the second input message from among one or more output messages received from the generation AI model (211) stored in the database (140) and one or more output messages of the training data set for the enhanced fine-tuning, without inputting the first input message received from the user to the generation AI model (211).

[0115] Referring to FIG. 15, the generation AI providing device (100``) can receive a training data set for fine-tuning from a user. The generation AI providing device (100``) can fine-tune the generation AI model (211) by uploading the original training data set received from the user to an external server (200), or can fine-tune the generation AI model (211) by augmenting the original training data set received from the user using the generation AI model (212) and then uploading the augmented training data set to an external server (200). In addition, the generation AI providing device (100``) can also load the original training data set or the augmented training data set received from the user into an internal database (140). In this way, the generation AI providing device (100``) can not only generate a high-quality output message desired by the user by using the finely tuned generation AI model (211) in the process of outputting an answer to the user, but also can further reduce the cost of the generation AI model (211) because it can use the output message of the training data set in addition to reusing the output message received from the generation AI model (211) in the past.

[0116] Figure 16 is a block diagram schematically illustrating a generation AI providing device according to another embodiment of the present invention.

[0117] Referring to FIG. 16, the generation AI providing device (1000) includes a memory (1100), a processor (1200), and a communication unit (1300).

[0118] The memory (1100) stores various data, programs, or applications for driving and controlling the device (1000). The program or application stored in the memory (1100) includes one or more instructions. The program or application stored in the memory (1100) can be executed by the processor (1200). The memory (1100) may include a random access memory (RAM), a read only memory (ROM), an erasable programmable ROM (EPROM), an electrically erasable programmable ROM (EEPROM), a flash memory, a hard disk, a removable disk, a CD-ROM, or any form of computer-readable recording medium well known in the art to which the present invention pertains.

[0119] In some embodiments, the memory (1100) may store one or more data, programs, or applications that constitute an artificial neural network or AI model. Alternatively, the memory (1100) may store one or more data, programs, or applications that control an artificial neural network or AI model.

[0120] The processor (1200) executes an operating system (OS) stored in the memory (1100) and various programs or applications. The processor (1200) may include one or more processors including single cores, dual cores, triple cores, quad cores, and multiples thereof. The processor (1200) may include a main processor and a sub-processor operating in sleep mode.

[0121] By having the processor (1200) execute one or more instructions stored in the memory (1100), the device (1000) can execute the methods according to the various embodiments described above.

[0122] The communication unit (1300) can transmit and receive any type of signal, data, or information. The communication unit (1300) can be connected to a network including a wired network, a wireless network, or a combination thereof. The communication unit (1300) can communicate with an external device or system.

[0123] The device (1000) can be implemented in the form of a server that provides online services.

[0124] Alternatively, the device (1000) may be implemented as a smartphone, tablet PC, PC, smart TV, mobile phone, PDA (personal digital assistant), laptop, media player, micro server, GPS (global positioning system) device, e-book reader, digital broadcasting terminal, navigation, kiosk, MP3 player, digital camera, home appliance, and other mobile or non-mobile computing devices, but is not limited thereto. In addition, the device (100) may also be implemented as a wearable device such as a watch, glasses, hair band, and ring having communication functions and data processing functions.

[0125] Devices according to various embodiments of the present invention may further include other components, such as an input device, a display unit, etc., depending on the implementation example of the present invention.

[0126] The steps of the methods according to various embodiments of the present invention may be further divided into additional steps or combined into fewer steps, depending on the implementation of the present invention. Furthermore, some steps may be omitted as needed, and the order of the steps may be changed.

[0127] The methods according to the various embodiments of the present invention described above may be implemented as a computer program (or application) to be executed in combination with a computer as hardware and stored in a computer-readable recording medium.

[0128] The above-described program may include codes coded in a computer language, such as C, C++, JAVA, Ruby, or Python machine language, that can be read by the processor (CPU) of the computer through the device interface of the computer, so that the computer reads the program and executes the methods implemented as a program. Such codes may include functional codes related to functions that define functions necessary for executing the methods, and may include control codes related to execution procedures necessary for the processor of the computer to execute the functions according to a predetermined procedure. In addition, such codes may further include memory reference-related codes regarding which location (address address) of the internal or external memory of the computer should reference additional information or media necessary for the processor of the computer to execute the functions. In addition, if the processor of the computer needs to communicate with any other computer or server located remotely in order to execute the functions, the code may further include communication-related code regarding how to communicate with any other computer or server located remotely using the communication module of the computer, and what information or media to send and receive during communication.

[0129] The above storage medium refers to a medium that stores data semi-permanently and can be read by a device, rather than a medium that stores data for a short period of time, such as a register, cache, or memory. Specifically, examples of the storage medium include, but are not limited to, ROM, RAM, CD-ROM, magnetic tape, floppy disk, and optical data storage device. That is, the program can be stored in various recording media on various servers that the computer can access or in various recording media on the user's computer. In addition, the medium can be distributed across network-connected computer systems, so that computer-readable code can be stored in a distributed manner.

[0130] The foregoing description of the present invention is for illustrative purposes only, and those skilled in the art will readily appreciate that the present invention can be readily modified into other specific forms without altering the technical spirit or essential characteristics of the present invention. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. For example, each component described as a single entity may be implemented in a distributed manner, and similarly, components described as distributed may be implemented in a combined manner.

[0131] The scope of the present invention is indicated by the claims described below rather than the detailed description above, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be interpreted as being included in the scope of the present invention.

Claims

1. A method performed by a generation AI providing device, A step of receiving a first input message from a user; A step of searching for a second input message having a similarity higher than a predetermined standard with the first input message received from the user among one or more input messages received from the user stored in a database using a similarity AI model; and Including a step of outputting a reply to the first input message to the user according to the search result, The step of outputting a reply to the first input message to the user according to the search result is: If the above second input message is not found, A step of inputting the first input message received from the user into the generated AI model of an external server, A step of receiving a first output message as a response to the first input message from the above generating AI model; A step of pairing the first input message received from the user and the first output message received from the generated AI model and storing them in a database, and outputting the first output message to the user as a response to the first input message; When the above second input message is retrieved, A step of extracting a second output message paired with the second input message from one or more output messages received from the generating AI model stored in the database without inputting the first input message received from the user to the generating AI model, comprising the step of outputting the second output message as a response to the first input message to the user; Method for providing AI-based fine-tuning generative similarity model.

2. In paragraph 1, The step of searching the second input message using the above similarity AI model is: When multiple input messages having a similarity level higher than a predetermined standard with the first input message received from the user are searched among one or more input messages received from the user stored in the database, Among the plurality of input messages having a similarity higher than a predetermined standard with the first input message, the input message having the highest similarity is determined as the second input message. Method for providing AI-based fine-tuning generative similarity model.

3. In paragraph 1, The step of inputting the first input message received from the user to the generated AI model of the external server is as follows: A step of detecting preset security information from the first input message received from the user; A step of preprocessing a portion corresponding to the security information in the first input message received from the user by one or more of blocking, masking, encryption, and de-identification methods; A step of inputting the preprocessed first input message to the generated AI model of the external server, Method for providing AI-based fine-tuning generative similarity model.

4. In paragraph 3, Further comprising a step of recording and managing usage logs regarding the generated AI model of the external server. Method for providing AI-based fine-tuning generative similarity model.

5. In paragraph 4, The above usage log includes at least one of the generated AI model name used, the date or time of use, the input message received from the user, the part corresponding to the security information, the presence or absence of the preprocessing, the type of the preprocessing, the preprocessed input message, and the output message received from the generated AI model. Method for providing AI-based fine-tuning generative similarity model.

6. In paragraph 1, receiving a training data set for fine-tuning comprising one or more paired input messages and output messages from said user; and Further comprising a step of uploading a training data set for fine-tuning received from the user to the external server to fine-tune the generated AI model. Method for providing AI-based fine-tuning generative similarity model.

7. In paragraph 6, The step of uploading the training data set for the above fine-tuning to the external server and fine-tuning the generated AI model is as follows. A step of detecting preset security information from a training data set for the above fine-tuning received from the above user, A step of preprocessing a portion corresponding to the security information in the training data set for the fine-tuning received from the user by one or more of blocking, masking, encryption, and de-identification methods; Comprising a step of uploading the preprocessed training data set to the external server and fine-tuning the generated AI model. Method for providing AI-based fine-tuning generative similarity model.

8. In paragraph 6, Further comprising a step of increasing the amount of input messages of the training data set for the fine-tuning by using the generated AI model of the external server, The step of uploading the training data set for the above fine-tuning to the external server and fine-tuning the generated AI model is as follows. Uploading the training data set for the above-mentioned enhanced fine-tuning to the external server to fine-tune the above-mentioned generated AI model. Method for providing AI-based fine-tuning generative similarity model.

9. In paragraph 8, Further comprising the step of storing the training data set for the above-mentioned enhanced fine-tuning in the database; The step of searching the second input message is: Using the above similarity AI model, a second input message having a similarity higher than a predetermined threshold with the first input message received from the user is searched for among one or more messages received from the user stored in the database and one or more input messages of the training data set for the augmented fine-tuning, The step of extracting the second output message is: Extracting a second output message paired with the second input message from one or more output messages received from the generated AI model stored in the database and one or more output messages of the training data set for the augmented fine-tuning; Method for providing AI-based fine-tuning generative similarity model.

10. Memory for storing one or more instructions; and comprising one or more processors executing one or more instructions stored in said memory; The one or more processors execute the one or more instructions, Receive a first input message from the user, Using a similarity AI model, a second input message having a similarity higher than a predetermined standard with the first input message received from the user is searched for among one or more input messages received from the user stored in the database, and Outputting a response to the first input message to the user based on the search results, Outputting a reply to the first input message to the user based on the search results, If the above second input message is not found, Input the first input message received from the user to the generated AI model of the external server, Receive the first output message as a response to the first input message from the above generating AI model, Pairing the first input message received from the user and the first output message received from the generated AI model and storing them in a database, and outputting the first output message to the user as a response to the first input message, When the above second input message is retrieved, Extracting a second output message paired with the second input message from one or more output messages received from the generated AI model stored in the database, Outputting the second output message as a response to the first input message to the user; A device that provides AI-based fine-tuning generative similarity model.

Images