Method and apparatus for providing job prompt management service based on ai model cross-validation

Abstract

A server providing a job prompt management service based on AI model cross-validation can obtain a first job prompt containing job-related input data from a user terminal. Based on the content and context of the first job prompt and the user's job information, the server can determine a first AI agent for generating a response from among a plurality of AI agents. The server can obtain a first response to the first job prompt from the first AI agent. The server can obtain a second response to the first job prompt from a group of AI agents including at least one AI agent other than the first AI agent. The server can verify the suitability of the first response by comparing the first response with the second response. Based on the verification results, the server can correct the selection criteria for the first AI agent.

Description

Technology Field

[0001] The present specification relates to a method and apparatus for providing a job prompt management service based on artificial intelligence model cross-validation. More specifically, it relates to a method and apparatus for providing an artificial intelligence service that can improve job performance efficiency by automatically determining a suitable artificial intelligence agent based on a user's job type and input prompt, evaluating the accuracy of responses through cross-validation among multiple artificial intelligence agents, and dynamically correcting agent selection criteria by reflecting the evaluation results. Background Technology

[0002] Due to the recent advancements in artificial intelligence technology, conversational AI agents are being utilized in various work environments to improve work efficiency. For example, AI language agents are being used to perform diverse tasks such as document creation, report review, marketing content generation, code writing, and technical review; however, the response quality of different AI agents can vary significantly depending on the specific characteristics of each job function.

[0003] However, conventional AI-based job support systems typically involved manually selecting a specific agent based on prompts entered directly by the user or relying on a single agent to generate results. In this scenario, there was a problem in that it was difficult for a single AI agent to provide sufficiently accurate and consistent results when the user's job type or the purpose of the prompt varied.

[0004] For example, if the same AI agent responds in the same way when the user inputs a request for business planning analysis versus a request to draft a research and development document, differences in quality may occur. Furthermore, when the job prompts entered by the user are ambiguous or incomplete, there is a high probability that the agent will generate output different from the intent.

[0005] To address this, approaches utilizing multiple AI agents simultaneously are being attempted; however, existing methods are limited to simply comparing the outputs of various agents in parallel or selecting responses via majority voting. Consequently, there is a lack of a systematic management mechanism to cross-verify the responses of each agent and progressively correct and learn the agent selection criteria based on the results.

[0006] In addition, in a job performance environment, users perform continuous prompt input (e.g., business conversation flow), and at this time, it is necessary to dynamically change an appropriate AI agent according to the job context. However, conventional technology does not provide a function to replace or correct the agent in real time in response to changes in the user's job or conversation context, so users have the inconvenience of having to manually select an agent suitable for each situation or create a separate session.

[0007] Therefore, there is a need for technology that can improve job performance efficiency and response reliability by automatically analyzing user task-specific prompt inputs to select the optimal AI agent, and evaluating and correcting response quality through cross-validation among multiple agents. The problem to be solved

[0008] The purpose of this specification to solve the aforementioned problems is to improve the reliability of the artificial intelligence service and the efficiency of job performance by automatically selecting a suitable agent among a plurality of artificial intelligence agents based on the user's job-specific prompt input, and by performing cross-validation among the multiple agents to evaluate the accuracy and consistency of the response.

[0009] Another objective of this specification to solve the aforementioned problems is to provide a prompt managing service capable of continuous performance improvement by reflecting verification results in agent selection criteria and progressively correcting the agent selection algorithm for identical or similar job inputs.

[0010] Another objective of this specification to address the aforementioned problems is to provide a high level of response quality and adaptability even in complex work environments by switching or readjusting agents in real time according to changes in the user's conversation flow or job context. means of solving the problem

[0011] A method for providing a job prompt management service based on artificial intelligence model cross-validation according to an embodiment of the present specification for achieving the above objective may include: a step of obtaining a first job prompt including job-related input data from a user terminal; a step of determining a first artificial intelligence agent for generating a response among a plurality of artificial intelligence agents based on the content, context, and user job information of the first job prompt; a step of obtaining a first response to the first job prompt from the first artificial intelligence agent; a step of obtaining a second response to the first job prompt from an artificial intelligence agent group including at least one artificial intelligence agent other than the first artificial intelligence agent; a step of verifying the suitability of the first response by comparing the first response and the second response; and a step of correcting the selection criteria of the first artificial intelligence agent based on the verification result.

[0012] Here, the method may further include a step of dynamically determining an artificial intelligence agent corresponding to a second job prompt that is subsequently entered based on corrected selection criteria.

[0013] Here, the method may further include the step of obtaining job classification information including the user's job type, position, and affiliated department information from a user terminal before obtaining a first job prompt; and the step of utilizing said job classification information as an input variable in the decision step of the first artificial intelligence agent.

[0014] Here, character data corresponding to each of a plurality of artificial intelligence agents is stored, and said character data includes job characteristics, response style, and expertise area information of each artificial intelligence agent, and the step of determining the first artificial intelligence agent may include the step of selecting the artificial intelligence agent with the highest job suitability by matching said job classification information and the first job prompt with said character data.

[0015] Here, the method may include the step of creating a group of artificial intelligence agents including a second artificial intelligence agent and a third artificial intelligence agent in the step of obtaining a second response; the step of controlling each artificial intelligence agent to exchange partial responses, grounds, and counterarguments to a first job prompt in the conversation session; and the step of confirming the response integrally derived through the conversation within the agent group as the second response.

[0016] Here, the step of determining the first artificial intelligence agent may include: providing input data to a distribution agent; obtaining an analysis result of the input data from the distribution agent; identifying a plurality of artificial intelligence agent candidates corresponding to the subject, job category, and response type of the first job prompt based on the analysis result; calculating job suitability for each of the plurality of artificial intelligence agents by synthesizing factors including job category match, language style similarity, and past response accuracy; and setting the artificial intelligence agent with the highest job suitability among the candidates to be determined as the first artificial intelligence agent.

[0017] Here, the step of verifying the suitability of the first response may include: a step of controlling the evaluation agent to evaluate the logical structure, factual basis, and responsiveness to the job purpose of each of the first response and the second response; a step of calculating a verification score based on the evaluation result; a step of terminating the verification procedure if the verification score is greater than or equal to a preset standard; a step of correcting the selection criteria of the first artificial intelligence agent if the verification score is less than the preset standard; a step of redetermining the first artificial intelligence agent based on the corrected selection criteria; and a step of setting the step of verifying the suitability of the first response again by obtaining a new response through the redetermined first artificial intelligence agent.

[0018] Here, the method may include the step of storing a communication tone profile corresponding to a job category for each artificial intelligence agent, wherein the communication tone profile is defined as a multidimensional tone parameter vector including politeness, formality, assertiveness, empathy, detail, technical term ratio, and level of structuring; the step of determining a first artificial intelligence agent and then calculating the tone parameter vector by a mapping function that takes job classification information, a target tag of a first job prompt, and a past user feedback indicator as inputs; and the step of controlling the tone parameter vector to be applied as a response generation constraint of the first artificial intelligence agent.

[0019] Here, to learn a communication tone profile, a feedback score is calculated from user feedback, wherein the feedback score includes explicit response information and implicit response information, wherein the explicit response information includes user evaluation input and whether a re-request was made, and the implicit response information includes response modification rate, dwell time, scroll pattern, and subsequent usage rate; a step of updating a tone parameter vector based on the feedback score using a weighted moving average or gradient-based update rule; and a step of distinguishing between a user-specific personalized vector and an organization-common vector and setting them to be fused and applied according to time weights may be included.

[0020] Here, the method may include the step of correcting a tone parameter vector according to a context, wherein the context includes a recipient role, a delivery channel type, language and region setting information, and a response purpose, wherein the recipient role includes at least one of an executive, a practitioner, and an external customer, wherein the delivery channel type includes at least one of email, chat, and presentation, and wherein the response purpose includes at least one of a summary, report, and proposal; and the step of performing response generation after adjusting the tone parameter vector by applying a correction matrix corresponding to the context.

[0021] Here, the method may include the steps of obtaining a style guardrail rulebook according to organizational regulations, wherein the rulebook includes forbidden word rules, upper and lower limits of expression intensity, minimum formality standards, and length constraints; checking compliance with response candidates generated from tone parameter vectors using a rule-based post-processor; and setting up the rewriting of tone parameters and stylistic templates corresponding to the violations by adjusting them if violations exist.

[0022] An apparatus for providing an artificial intelligence model cross-validation-based job prompt management service according to another embodiment of the present specification for achieving the above objective includes a processor and a memory for storing at least one command executed by the processor, wherein the at least one command may be configured to obtain a first job prompt including job-related input data from a user terminal, determine a first artificial intelligence agent for generating a response among a plurality of artificial intelligence agents based on the content, context, and user job information of the first job prompt, obtain a first response to the first job prompt from the first artificial intelligence agent, obtain a second response to the first job prompt from an artificial intelligence agent group including at least one artificial intelligence agent other than the first artificial intelligence agent, compare the first response and the second response to verify the suitability of the first response, and correct the selection criteria of the first artificial intelligence agent based on the verification result.

[0023] Here, the above at least one command may be configured to dynamically determine an artificial intelligence agent corresponding to a second job prompt subsequently entered based on a corrected selection criterion.

[0024] Here, the at least one command may be configured to obtain job classification information including the user's job type, position, and affiliated department information from the user terminal before obtaining the first job prompt, and to use the job classification information as an input variable when determining the first artificial intelligence agent.

[0025] Here, the at least one command stores character data corresponding to each of a plurality of artificial intelligence agents, and the character data includes job characteristics, response style, and expertise area information of each artificial intelligence agent, and when determining the first artificial intelligence agent, the command may be configured to select the artificial intelligence agent with the highest job suitability by matching the job classification information and the first job prompt with the character data.

[0026] Here, the at least one command may be configured to create an AI agent group including a second AI agent and a third AI agent, control each AI agent to exchange partial responses, grounds, and counterarguments to a first job prompt in the conversation session, and determine a response integrally derived through the conversation within the agent group as the second response.

[0027] Here, the at least one command may be configured to provide input data to a distribution agent, obtain an analysis result for the input data from the distribution agent, identify a plurality of AI agent candidates corresponding to the subject, job category, and response type of the first job prompt based on the analysis result, calculate a job suitability calculated by synthesizing factors including job category match, language style similarity, and past response accuracy for each of the plurality of AI agents, and determine the AI ​​agent with the highest job suitability among the candidates as the first AI agent.

[0028] Here, the at least one command may be configured to evaluate the logical structure, factual basis, and responsiveness to the job purpose of each of the first response and the second response through an evaluation agent, calculate a verification score based on the evaluation result, terminate the verification procedure if the verification score is greater than or equal to a preset standard, correct the selection criteria of the first artificial intelligence agent if the verification score is less than the standard, redefine the first artificial intelligence agent based on the corrected selection criteria, and obtain a new response through the redefine first artificial intelligence agent to verify the suitability of the first response.

[0029] Here, the at least one command stores a communication tone profile corresponding to a job category for each artificial intelligence agent, and the communication tone profile is defined as a multidimensional tone parameter vector including politeness, formality, assertiveness, empathy, detail, technicality ratio, and level of structuring; after determining the first artificial intelligence agent, the tone parameter vector is calculated by a mapping function that takes job classification information, the purpose tag of the first job prompt, and past user feedback indicators as inputs, and the tone parameter vector can be set to be applied as a response generation constraint of the first artificial intelligence agent.

[0030] Here, the at least one command may be configured to calculate a feedback score including explicit response information and implicit response information from user feedback, update a tone parameter vector based on the feedback score using a weighted moving average or gradient-based update rule, and distinguish between a user-specific personalized vector and an organization-common vector to apply fusion according to time weights.

[0031] Here, the at least one command may be configured to correct a tone parameter vector according to a context, wherein the context includes a receiver role, a transmission channel type, language and region setting information, and a response purpose, and to adjust the tone parameter vector by applying a correction matrix corresponding to the context, and to generate a response based on the adjusted tone parameter vector.

[0032] Here, the at least one command may be configured to store a style guardrail rulebook according to organizational regulations, wherein the rulebook includes forbidden word rules, upper and lower limits of expression intensity, minimum formality standards, and length constraints, and to check compliance with response candidates generated as tone parameter vectors using a rule-based postprocessor, and if violations exist, to rewrite by adjusting the tone parameters and stylistic templates corresponding to the violations. Effects of the invention

[0033] According to one embodiment of the present specification, by cross-validating a plurality of artificial intelligence agents for a user's task-specific prompt input and reflecting the results, the reliability and accuracy of the artificial intelligence response during the task performance process can be improved.

[0034] According to one embodiment of the present specification, an artificial intelligence agent corresponding to a subsequent input prompt can be dynamically determined based on corrected selection criteria, so that the selection of the agent is automatically adjusted according to changes in the user's work context, and an artificial intelligence response of consistent quality can be provided even in an environment where job switching is frequent.

[0035] According to one embodiment of the present specification, by utilizing character data and job classification information of each artificial intelligence agent to prioritize the selection of an agent with high job suitability, optimized answers can be provided for each job type, and unnecessary model selection errors can be minimized. Brief explanation of the drawing

[0036] FIG. 1 is a system diagram including a server providing a job prompt managing service based on artificial intelligence model cross-validation according to one embodiment of the present specification. FIG. 2 is a block diagram showing the configuration of a server providing an artificial intelligence model cross-validation-based job prompt managing service according to one embodiment of the present specification. FIG. 3 is a flowchart illustrating a method for providing a job prompt managing service based on artificial intelligence model cross-validation according to one embodiment of the present specification. Specific details for implementing the invention

[0037] As the present specification is susceptible to various modifications and may have various embodiments, specific embodiments are illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the present specification to specific embodiments, and it should be understood that it includes all modifications, equivalents, and substitutions that fall within the spirit and scope of the present specification. Similar reference numerals have been used for similar components in the description of each drawing.

[0038] Terms such as first, second, A, B, etc., may be used to describe various components, but said components should not be limited by said terms. These terms are used solely for the purpose of distinguishing one component from another. For example, without departing from the scope of the rights of this specification, the first component may be named the second component, and similarly, the second component may be named the first component. The term "and / or" includes a combination of a plurality of related described items or any of a plurality of related described items.

[0039] When it is stated that one component is 'connected' or 'connected' to another component, it should be understood that while it may be directly connected or connected to that other component, there may also be other components in between. On the other hand, when it is stated that one component is 'directly connected' or 'directly connected' to another component, it should be understood that there are no other components in between.

[0040] The terms used in this application are used merely to describe specific embodiments and are not intended to limit this specification. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this application, terms such as "comprising" or "having" are intended to indicate the presence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0041] Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by those skilled in the art to which this specification pertains. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and should not be interpreted in an ideal or overly formal sense unless explicitly defined in this application.

[0042] Hereinafter, preferred embodiments of the present specification will be described in more detail with reference to the attached drawings. To facilitate overall understanding in describing the present specification, the same reference numerals are used for identical components in the drawings, and redundant descriptions of identical components are omitted.

[0043] FIG. 1 is a system diagram including a server providing a job prompt managing service based on artificial intelligence model cross-validation according to one embodiment of the present specification.

[0044] Referring to FIG. 1, the method for providing a job prompt management service based on artificial intelligence model cross-validation according to an embodiment of the present specification may be performed on a computing device that is equipped with storage space and connected to the Internet, such as a PC (Personal Computer), and is not easily portable, or on a portable terminal such as a smartphone. In this case, the method for providing a job prompt management service based on artificial intelligence model cross-validation may be executed after an application implementing the method is downloaded from an App Store, etc., and installed on the portable terminal.

[0045] In addition, the method for providing a job prompt management service based on the artificial intelligence model cross-validation described above may be executed by inserting it into a computing device such as a PC while it is recorded on a recording medium such as a CD (Compact Disc) or USB (Universal Serial Bus) memory and performing the operation through an access operation of said computing device, or it may be executed by storing it from said recording medium into the storage space of the computing device and then performing the operation through an access operation of said computing device.

[0046] Meanwhile, if the above computing device or portable terminal can access a server connected to the Internet, the method for providing a job prompt managing service based on artificial intelligence model cross-validation can also be executed on the server in response to a request from the computing device or portable terminal.

[0047] In the following, a computing device, portable terminal, or server, etc., on which the above-mentioned method for providing a job prompt management service based on an artificial intelligence model cross-validation is executed may be collectively referred to as an artificial intelligence model cross-validation-based job prompt management service providing device.

[0048] The artificial intelligence model cross-validation-based job prompt managing service providing device described above may have the same configuration as the artificial intelligence model cross-validation-based job prompt managing service providing device illustrated in FIG. 2, and the artificial intelligence model cross-validation-based job prompt managing service providing device may not be limited to the artificial intelligence model cross-validation-based job prompt managing service providing device illustrated in FIG. 1.

[0049] A system according to one embodiment may include a user terminal (110), a user terminal (120), a user terminal (130), and a server (140) providing a job prompt managing service based on artificial intelligence model cross-validation (hereinafter, server (140)).

[0050] The user terminal (110), user terminal (120), and user terminal (130) may be, but are not limited to, a vehicle, a mobility device, a smartphone, a tablet PC, a PC, a mobile phone, a PDA (personal digital assistant), a laptop, a media player, a micro server, a GPS (global positioning system) device, and other mobile or non-mobile computing devices. Additionally, the user terminal (110), user terminal (120), and user terminal (130) may be wearable devices equipped with communication functions and data processing functions. However, they are not limited to.

[0051] The server (140) may be implemented as a computer device or a plurality of computer devices that communicate with the user terminal (110), user terminal (120), and user terminal (130) through a network to provide commands, code, files, content, services, etc.

[0052] For example, the server (140) may provide a file for installing an application to a user terminal (110), a user terminal (120), and a user terminal (130) connected via a network. In this case, the user terminal (110), the user terminal (120), and the user terminal (130) may install the application using the file provided by the server (140).

[0053] Additionally, user terminals (110), user terminals (120), and user terminals (130) can connect to a server (140) under the control of an operating system (OS) and at least one program (e.g., a browser or an installed application) to receive services or content provided by the server (140).

[0054] As another example, the server (140) may establish a communication session for data transmission and reception and route data transmission and reception between user terminal (110), user terminal (120), and user terminal (130) through the established communication session.

[0055] User terminals (110), user terminals (120), user terminals (130), and a server (140) providing job prompt management services based on artificial intelligence model cross-verification can perform communication using a network. For example, the network includes a Local Area Network (LAN), a Wide Area Network (WAN), a Value Added Network (VAN), a mobile radio communication network, a satellite communication network, and combinations thereof, and is a data communication network in a comprehensive sense that enables each network constituent entity shown in FIG. 1 to communicate smoothly with each other, and may include wired internet, wireless internet, and mobile wireless communication networks. In addition, wireless communication may include, for example, wireless LAN (Wi-Fi), Bluetooth, Bluetooth Low Energy, LoRaWAN, Zigbee, WFD (Wi-Fi Direct), UWB (ultra wideband), infrared communication (IrDA, infrared Data Association), NFC (Near Field Communication), etc., but is not limited thereto.

[0056] FIG. 2 is a block diagram showing the configuration of a server providing an artificial intelligence model cross-validation-based job prompt managing service according to one embodiment of the present specification.

[0057] Referring to FIG. 2, a server (200) (hereinafter, server (200)) providing an artificial intelligence model cross-validation-based job prompt managing service may include a communication unit (210), a processor (220), a storage unit (230), and an artificial intelligence model execution unit (240).

[0058] The communication unit (210) can be configured to reliably transmit and receive data between a user terminal and another server inside or outside the data center via wired and wireless networks, and can be controlled to support standard protocol stacks such as Ethernet, optical interface, 5G NR, and Wi-Fi. The communication unit (210) can include a NIC and a MAC / PHY module to establish a TLS-based encryption channel, transmit and receive prompts, responses, agent metadata, feedback indicators, rulebooks, etc. in packet units, and perform packet loss recovery and flow control. For example, the communication unit (210) can be configured to perform low-latency shared memory I / O between nodes within the data center using RDMA, or to receive requests from a user terminal and transmit streaming responses via a gRPC or HTTP / 2-based API gateway. Additionally, the communication unit (210) can apply access control lists in conjunction with a firewall, IDS / IPS, and DDoS mitigation device, and perform token-based authentication and authorization.

[0059] The communication unit (210) may include an interface capable of transmitting and receiving data to and from multiple IoT devices, authentication terminals, external servers, blockchain networks, etc., via a wired or wireless communication network. The communication unit may be configured to support various transmission protocols such as Bluetooth, Wi-Fi, cellular communication, and Ethernet, and may be configured to reliably receive and transmit various types of data, such as BLE signals, environmental sensor data, and operation logs of IoT devices. In addition, the communication unit enables interoperability between heterogeneous systems through various application layer protocols, such as REST API, MQTT, or Web3 interfaces for calling smart contracts.

[0060] The processor (220) is a computing device comprising one or more general-purpose CPU cores and optionally accelerators such as a GPU, NPU, or TPU, and can perform data exchange with a communication unit (210), a storage unit (230), and an artificial intelligence model execution unit (240) via a system bus and interconnect. The processor (220) can execute the control loops of an operating system kernel, driver, runtime, agent orchestration logic, distribution agent, and evaluation agent, and can manage thread and process scheduling, memory management, container isolation, and hardware accelerator invocation. The processor (220) can implement the steps of the claim as a control flow to control the collection of a first job prompt, the determination of a first artificial intelligence agent, the creation of an agent group conversation session, the collection of first and second responses, the calculation of a verification score, and the correction and re-determination of selection criteria in a sequential or parallel manner. In addition, the processor (220) can perform key management, secure model calling, and rulebook integrity verification by linking with a security module (TPM, HSM, SGX / SEV, etc.), and synchronize the timing of the audit log with a system time source (NTP, PTP, RTC).

[0061] The storage unit (230) is a non-transient computer-readable medium, such as an NVMe SSD, HDD, NVRAM, distributed object storage (S3 compatible), etc., and may be configured to provide low-latency access using a DRAM-based cache layer. The storage unit (230) may store agent character data (job characteristics, response style, area of ​​expertise), communication tone profiles (multidimensional parameter vectors defined by politeness, formality, assertiveness, empathy, detail, technicality ratio, and level of structure), job classification tables, routing indexes for distribution agents, verification criteria and weights for evaluation agents, user feedback logs (explicit and implicit), style guardrail rulebooks according to organizational regulations, correction matrices, model checkpoints, token dictionaries, and runtime caches. The storage unit (230) may track and roll back selection criteria correction history and tone parameter update history through snapshots and version control, and may ensure consistency with a journaling file system or a WAL-based transaction log. Additionally, the storage unit (230) may apply an access control list and encryption-at-rest, and the key may be configured to be stored in a protected HSM.

[0062] The artificial intelligence model execution unit (240) may be configured to execute large language models, ranking models, tone matching discriminators, job classifiers, decision models, etc., on a hardware accelerator at high performance, including one or more inference runners and runtime engines. The artificial intelligence model execution unit (240) may include a weight loader, a tokenizer, a KV-cache manager, a batch / padding optimizer, a low-precision computation engine (INT8 / FP8 / FP16), a pipeline and tensor parallelization manager, and a streaming token outputter. This execution unit may be responsible for inference calls for generating responses of the first artificial intelligence agent, generating partial responses of individual agents within the artificial intelligence agent group, generating counterarguments and grounds, generating integrated responses, and executing a discrimination model for evaluating logical structure, factual grounds, and job objective correspondence by an evaluation agent. Additionally, the artificial intelligence model execution unit (240) may support the calculation of tone parameter vectors by a mapping function, reflection of response generation constraints, application of context-specific correction matrices, and chain execution with rule-based post-processing. For example, the artificial intelligence model execution unit (240) can insert tone constraints into the system prompt during prompt preprocessing and apply weighted vectors during the logit projection stage during generation to adjust sampling parameters to satisfy upper and lower limits of expression intensity and formality.

[0063] According to one embodiment of the present specification, each function of the server (200) can be realized by program instructions stored in a non-transient computer-readable medium, and said medium may include a semiconductor memory device, a magnetic disk, an optical disk, a flash memory, a ROM, a PROM, an EPROM, an EEPROM, an NVRAM, or a combination thereof. The program instructions may control the operation of a communication unit (210), a processor (220), a storage unit (230), and an artificial intelligence model execution unit (240) in conjunction with an operating system or middleware, call a hardware accelerator, and direct distributed execution over a network. Accordingly, the server (200) according to one embodiment of the present specification is not an abstract concept that performs a simple set of rules, but can be realized by technical means including specific hardware configuration, data flow, signal processing, non-transient storage, encrypted transmission, and accelerator-based inference.

[0064] The processor (220) may be implemented using at least one of ASICs (application specific integrated circuits), DSPs (digital signal processors), DSPDs (digital signal processing devices), PLDs (programmable logic devices), FPGAs (field programmable gate arrays), controllers, microcontrollers, microprocessors, and other electrical units for performing functions.

[0065] The storage unit (230) may be implemented as a non-transient computer-readable medium and may include both volatile memory (RAM) and non-volatile memory (ROM, flash memory, hard disk, etc.). The storage unit (230) may store the operating system, application programs, communication protocol stack, database, and learning agent parameters of the artificial intelligence module of the server device. Additionally, the storage unit (230) may store log data, communication history, information received from an external terminal, etc., which can be utilized for subsequent operations.

[0066] The storage unit (230) may include RAM (random access memory), such as DRAM (dynamic random access memory) or SRAM (static random access memory), ROM (read-only memory), EEPROM (electrically erasable programmable read-only memory), CD-ROM, Blu-ray or other optical disc storage, HDD (hard disk drive), SSD (solid state drive), or flash memory. The storage unit (230) may also be referred to as memory.

[0067] The storage unit (230) may store at least one command executed through the processor (220). The at least one command may be configured to obtain a first job prompt containing job-related input data from a user terminal, determine a first artificial intelligence agent for generating a response among a plurality of artificial intelligence agents based on the content, context, and user job information of the first job prompt, obtain a first response to the first job prompt from the first artificial intelligence agent, obtain a second response to the first job prompt from an artificial intelligence agent group including at least one artificial intelligence agent other than the first artificial intelligence agent, verify the suitability of the first response by comparing the first response and the second response, and correct the selection criteria of the first artificial intelligence agent based on the verification result.

[0068] Here, the above at least one command may be configured to dynamically determine an artificial intelligence agent corresponding to a second job prompt subsequently entered based on a corrected selection criterion.

[0069] Here, the at least one command may be configured to obtain job classification information including the user's job type, position, and affiliated department information from the user terminal before obtaining the first job prompt, and to use the job classification information as an input variable when determining the first artificial intelligence agent.

[0070] Here, the at least one command stores character data corresponding to each of a plurality of artificial intelligence agents, and the character data includes job characteristics, response style, and expertise area information of each artificial intelligence agent, and when determining the first artificial intelligence agent, the command may be configured to select the artificial intelligence agent with the highest job suitability by matching the job classification information and the first job prompt with the character data.

[0071] Here, the at least one command may be configured to create an AI agent group including a second AI agent and a third AI agent, control each AI agent to exchange partial responses, grounds, and counterarguments to a first job prompt in the conversation session, and determine a response integrally derived through the conversation within the agent group as the second response.

[0072] Here, the at least one command may be configured to provide input data to a distribution agent, obtain an analysis result for the input data from the distribution agent, identify a plurality of AI agent candidates corresponding to the subject, job category, and response type of the first job prompt based on the analysis result, calculate a job suitability calculated by synthesizing factors including job category match, language style similarity, and past response accuracy for each of the plurality of AI agents, and determine the AI ​​agent with the highest job suitability among the candidates as the first AI agent.

[0073] Here, the at least one command may be configured to evaluate the logical structure, factual basis, and responsiveness to the job purpose of each of the first response and the second response through an evaluation agent, calculate a verification score based on the evaluation result, terminate the verification procedure if the verification score is greater than or equal to a preset standard, correct the selection criteria of the first artificial intelligence agent if the verification score is less than the standard, redefine the first artificial intelligence agent based on the corrected selection criteria, and obtain a new response through the redefine first artificial intelligence agent to verify the suitability of the first response.

[0074] Here, the at least one command stores a communication tone profile corresponding to a job category for each artificial intelligence agent, and the communication tone profile is defined as a multidimensional tone parameter vector including politeness, formality, assertiveness, empathy, detail, technicality ratio, and level of structuring; after determining the first artificial intelligence agent, the tone parameter vector is calculated by a mapping function that takes job classification information, the purpose tag of the first job prompt, and past user feedback indicators as inputs, and the tone parameter vector can be set to be applied as a response generation constraint of the first artificial intelligence agent.

[0075] Here, the at least one command may be configured to calculate a feedback score including explicit response information and implicit response information from user feedback, update a tone parameter vector based on the feedback score using a weighted moving average or gradient-based update rule, and distinguish between a user-specific personalized vector and an organization-common vector to apply fusion according to time weights.

[0076] Here, the at least one command may be configured to correct a tone parameter vector according to a context, wherein the context includes a receiver role, a transmission channel type, language and region setting information, and a response purpose, and to adjust the tone parameter vector by applying a correction matrix corresponding to the context, and to generate a response based on the adjusted tone parameter vector.

[0077] Here, the at least one command may be configured to store a style guardrail rulebook according to organizational regulations, wherein the rulebook includes forbidden word rules, upper and lower limits of expression intensity, minimum formality standards, and length constraints, and to check compliance with response candidates generated as tone parameter vectors using a rule-based postprocessor, and if violations exist, to rewrite by adjusting the tone parameters and stylistic templates corresponding to the violations.

[0078] FIG. 3 is a flowchart illustrating a method for providing a job prompt managing service based on artificial intelligence model cross-validation according to one embodiment of the present specification.

[0079] Referring to FIG. 3, the server can obtain a job prompt (S300). For example, the server can obtain a first job prompt containing job-related input data from a user terminal.

[0080] For example, the first job prompt may consist of request sentences, questions, instructions, etc. generated during the user's job performance process, and may be in the form of natural language that reflects the user's role and work context.

[0081] The server can determine a first artificial intelligence agent (S310). For example, the server can determine a first artificial intelligence agent for generating a response from among a plurality of artificial intelligence agents based on the content, context, and user's job information of the first job prompt.

[0082] The server can determine a first AI agent for generating a response among multiple AI agents by analyzing the content, context, and user job information of the received first job prompt. At this time, the server may refer to character data including attributes such as job characteristics, response style, and area of ​​expertise for each of the predefined multiple AI agents. The server can determine the AI ​​agent with the highest job suitability as the first AI agent by comprehensively considering the linguistic features, keywords, expression structure, and user job type of the prompt.

[0083] The server can obtain a first response (S320). For example, the server can obtain a first response to the first job prompt from the first artificial intelligence agent.

[0084] The server can generate a first response to a first job prompt using a determined first artificial intelligence agent. The first response may be configured in the form of data required for job performance, a draft report, a proposal summary, etc., and may automatically reflect terminology, expression styles, and structures related to the user's job.

[0085] The server may obtain a second response from an artificial intelligence agent group (S330). For example, the server may obtain a second response to the first job prompt from an artificial intelligence agent group that includes at least one artificial intelligence agent other than the first artificial intelligence agent.

[0086] The server may form a group of AI agents including multiple AI agents other than the first AI agent to generate a second response to the same first job prompt. In this case, the server may initiate an interactive verification session between the second AI agent and the third AI agent, and each agent may exchange opinions based on the logical consistency, factual basis, and clarity of expression of the first response. The server may integrate the results of this mutual discussion to finalize the second response.

[0087] The server can verify the suitability of the first response (S340). For example, the server can verify the suitability of the first response by comparing the first response with the second response.

[0088] The server can verify the suitability of the first response by comparing the first response with the second response. At this time, the server can analyze the logical structure, reliability of the evidence, and alignment with the job objective of the two responses through an evaluation algorithm. The server quantifies the comparison results into a verification score, and if the score is above a preset standard, it can confirm the first response as the final result. On the other hand, if the verification score is below the standard, the server can correct the selection criteria of the first AI agent to automatically adjust the system so that a more suitable agent is selected when processing the next prompt.

[0089] The server can correct the selection criteria for the first agent (S350). For example, the server can correct the selection criteria for the first artificial intelligence agent based on the verification results.

[0090] For example, if the server evaluates a specific job prompt as having "low reliability of factual basis," an agent with enhanced basis verification capabilities can be automatically selected for the next similar input. Through this, the server improves its selection criteria through repeated prompt execution and can generate results with higher reliability over time.

[0091] According to one embodiment of the present specification, the series of processes can be automatically controlled within the server, and the generation of a first response, cross-validation of a second response, and correction of selection criteria can be performed cyclically without user intervention. Based on this structure, the server can derive logically valid and fact-based results through mutual verification among multiple artificial intelligence agents, without relying on the results of a single artificial intelligence model.

[0092] According to one embodiment of the present specification, the server can dynamically determine an artificial intelligence agent corresponding to a second job prompt that is subsequently input, using a corrected selection criterion calculated through a verification process for a first job prompt.

[0093] The server can update the selection criteria for AI agents based on validation scores, response quality indicators, job relevance history, and user feedback information obtained during previous prompt execution. The aforementioned selection criteria can be defined by a weighting table that reflects each AI agent's response quality, logical completeness, job relevance, tone suitability, etc. Using these corrected weights, the server can determine the agent optimized for the situation in real time, even when prompts of the same or similar job categories are entered again.

[0094] For example, if an agent with high technical verification capability is evaluated as having demonstrated superior results during the calibration process for the first job prompt, "Summarize the test results of the new part," the server can apply criteria to prioritize the selection of an agent of the same category when the second job prompt, "Organize the analysis results of the heat resistance of the new material," is subsequently input. Conversely, if the first response is evaluated as being excessively theoretical and lacking practical explanation, the server can automatically correct the selection criteria to prioritize a practice-oriented, explanatory agent.

[0095] The server analyzes features such as the context, keywords, topic scope, and sentence structure of the input second job prompt and calculates a score between AI agents by comparing them with the weights of the corrected selection criteria. At this time, the server determines whether the second prompt falls within the same job category as the first prompt, and if the similarity is evaluated as high, it may reflect the previous correction information more strongly. Conversely, if it is classified into a new job category, it may apply an initial value-based selection logic by referencing only a portion of the existing criteria.

[0096] In this way, the server can dynamically determine an AI agent in real-time for every prompt input based on corrected selection criteria, thereby continuously providing responses optimized for each job category. In particular, even when the user's work environment diversifies or the prompt topic changes rapidly, the server can maintain response quality by automatically re-selecting the optimal agent.

[0097] For example, if a user enters an administrative task-related prompt such as "Summarize comments on the personnel management regulations" in the morning and a technical management prompt such as "Propose measures to improve the efficiency of the R&D budget" in the afternoon, the server can switch in real time to the agent most suitable for the context of each task based on corrected selection criteria.

[0098] In addition, according to one embodiment of the present specification, the server may update the corrected selection criteria at regular intervals or immediately reflect them upon the occurrence of specific events (e.g., detection of response discrepancies, negative feedback evaluation). Through this, the server can optimize itself to match changing user work patterns and organizational requirements, and as a result, users can always receive response quality appropriate for the situation without the need for a separate model selection or configuration process.

[0099] According to one embodiment of the present specification, a server may obtain job classification information including the user's job type, position, and affiliated department information from a user terminal before obtaining a first job prompt. The server may utilize this job classification information as an input variable for an artificial intelligence agent decision step to control the process so that a more precise agent selection is made during prompt processing.

[0100] Job classification information is data used to identify a user's role within an organization and can be provided either automatically collected by the server through integration with user profiles, login accounts, and organizational HR systems, or directly entered by the user. For example, if a user registers information such as 'Department: R&D Division, Position: Senior Researcher, Job: Materials Analysis' upon initial login, the server stores this as job classification information and can refer to it during subsequent prompt analysis processes.

[0101] The server can determine the detailed category to which the user's job belongs by comparing the collected job classification information with a characteristic table for each job category. For example, job information such as "Sales Management Team Manager" can be classified into the "Sales & Marketing" category, while job information such as "Materials Development Center Researcher" can be classified into the "Research & Development" category. These classification results can then be utilized as key judgment criteria during the AI ​​agent decision-making stage.

[0102] The server can adjust weights by including job classification information as an input variable for agent selection logic, rather than simply using it as a classification value. For example, it can assign higher weights to agents that provide technical explanations and evidence-based responses for R&D users, and apply different criteria to prioritize descriptive and persuasive agents for marketing users. In this way, the server can control the direction of the response differently depending on the user's job function, even with the same prompt.

[0103] In addition, according to one embodiment of the present specification, the server checks the consistency between job classification information and prompt content, and can additionally perform a verification procedure if a request outside the user's job scope occurs.

[0104] According to one embodiment of the present specification, a server may refer to character data corresponding to each of a plurality of artificial intelligence agents and match the character data with job classification information to select the artificial intelligence agent with the highest job suitability.

[0105] Character data can consist of metadata representing each AI agent's role, disposition, response style, job expertise, linguistic expression characteristics, etc. The server can manage this character data for each agent in a pre-trained or defined form.

[0106] The server can calculate job fit by comparing the user job type identified in the job classification information with the character data of each agent. Job fit can be defined as a quantitative indicator calculated by the server based on the similarity between the attributes of the job classification information and the attributes of the agent's character. For example, if a user corresponds to a 'Technical Research' position, the server assigns high weights to items such as technical terminology processing ability, logical reasoning ability, and reliability of factual evidence; conversely, if the user corresponds to a 'Sales Management' position, it may apply greater weights to items such as persuasiveness, tone variety, and summarizing ability.

[0107] The server can calculate job suitability scores for all agents by reflecting the weights for each of the above items, and determine the AI ​​agent that obtained the highest score among them as the first AI agent.

[0108] Furthermore, according to one embodiment of the present specification, when calculating job suitability, the server may go beyond simple similarity comparison and incorporate the user's past prompt patterns and response preferences. For example, if a user has a history of providing feedback that a technical question "lacks theoretical basis," the server may increase the job suitability weight of an agent with a strengthened tendency for fact verification in subsequent prompts of similar job categories. Through this, the server can implement dynamic selection logic that adapts to actual user behavior, going beyond simply selecting an agent that matches job classification information.

[0109] Through this structure, the server can automatically select an agent optimized for the job context by matching job classification information with character data in real-time whenever a prompt is entered. This process provides an automated selection mechanism that eliminates the need for humans to manually specify models, and can significantly increase the efficiency of utilizing AI agents suitable for various job types.

[0110] According to one embodiment of the present specification, a server may form an artificial intelligence agent group comprising a plurality of artificial intelligence agents other than a first artificial intelligence agent, and perform an interactive session within the group to derive a second response to a first job prompt.

[0111] First, the server can create a group including a second AI agent and a third AI agent. In this case, the server may consider each agent's area of ​​expertise, conversation style, and response generation method to form the agent group. For example, if the first AI agent is selected as a research and development type agent, the server may form a group by including a verification type agent strong in comparative and critical thinking and an analysis type agent with high descriptive summarization ability.

[0112] The server can initiate conversation sessions within a generated group of AI agents and control each agent to provide partial responses, grounds, or rebuttals to a first job prompt. Conversation sessions are automatically managed by the server, and message exchanges between agents can proceed in stages.

[0113] The server can collect partial responses and logical reasoning submitted by each agent, and generate an integrated conclusion by comparing and analyzing them. During the integration process, the server assigns weights to content with consistent core reasoning among multiple agent responses, and can reconstruct inconsistent items into supplementary sentences or summaries.

[0114] The server can finalize this integrated response as the second response. Since the second response is not limited to the perspective of a single agent but is generated based on the mutual dialogue of multiple AI agents, the logical validity, information reliability, and representational diversity of the response can all be improved.

[0115] Additionally, according to one embodiment of the present specification, the server may control the session to automatically terminate when certain conditions are met within a conversation session in an agent group. For example, if the discussion results between agents achieve a certain level of agreement, or if the quality score of the generated response exceeds a certain threshold, the server may terminate the session and confirm the final response as the second response. Conversely, if the disagreement persists, the server may extend the session by temporarily involving additional verification agents.

[0116] According to one embodiment of the present specification, the server can determine a first artificial intelligence agent corresponding to a first job prompt using a distribution agent.

[0117] When the first job prompt is entered, the server can activate a distribution agent to analyze the content, context, keywords, and job-related terms of the prompt. The distribution agent serves as an intermediate analysis module for selecting an AI agent; it can perform the role of extracting core intent from the prompt and classifying and recommending suitable agent candidates based on this intent.

[0118] Based on the analysis results of the distribution agent, the server can identify multiple AI agent candidates corresponding to the subject, job category, and response type of the first job prompt. At this time, the distribution agent can generate a candidate list by evaluating the correlation between the job domain to which each agent belongs and the subject of the current prompt.

[0119] The server can calculate job suitability for each candidate agent by synthesizing factors such as job category match, language style similarity, and historical response accuracy. Job category match refers to the concordance rate between the prompt category and the agent's area of ​​expertise, while language style similarity refers to the similarity between the prompt's writing style or expression patterns and the agent's language generation characteristics. Historical response accuracy can be calculated based on response quality scores evaluated from the processing history of prompts in the same or similar job categories.

[0120] The server calculates the job suitability of each candidate by calculating the weighted average or composite score of the three elements mentioned above, and can determine the agent that obtained the highest score among them as the first artificial intelligence agent.

[0121] In addition, according to one embodiment of the present specification, the server may not simply use the analysis results produced by the distribution agent as a one-time occurrence, but may continuously update them in conjunction with corrected selection criteria. That is, when the distribution agent performs prompt analysis, by referring to data accumulated from previous verification loops (e.g., verification scores, feedback results), increasingly accurate candidate classification and score calculation become possible.

[0122] According to one embodiment of the present specification, a server may execute an evaluation agent to verify the suitability of a first response. The evaluation agent may be configured to compare and analyze responses generated by a plurality of artificial intelligence agents and to evaluate the quality of the response based on multidimensional criteria such as logical structure, factual basis, and suitability for job purposes.

[0123] The server can perform the evaluation agent's verification process using both the first and second responses as input values. At this time, the evaluation agent can determine the logical sequence of the responses, the causal relationships between key grounds, the clarity of the expressions used, and whether the responses satisfy the job objectives.

[0124] The server can calculate a verification score based on the analysis results of the evaluation agent. The verification score can be defined as a value calculated by weighting the structural completeness of the response, the reliability of factual verification, and the alignment with job objectives.

[0125] If the verification score is above a preset threshold, the server may determine the first response to be valid and terminate the verification procedure. Conversely, if the verification score is below the threshold, the server may correct the selection criteria of the first artificial intelligence agent. Correction may be performed differently depending on the cause of the response quality degradation; for example, if there is a lack of logical consistency, the weight of the analytical agent may be increased, and if the reliability of the factual basis is low, the priority of the data verification agent may be strengthened.

[0126] The server can redefine the first AI agent based on the corrected selection criteria and use the redetermined agent to generate a new response for the same first job prompt. Subsequently, the server can verify the newly generated response again through an evaluation agent and repeat the redetermining and re-verification process until the result meets or exceeds the criteria.

[0127] Through this iterative loop structure, the server can implement an autonomous improvement mechanism that checks and corrects response quality on its own. In this process, the verification scores and correction history generated during each iteration cycle are accumulated and managed in the server's internal database, allowing them to be automatically reflected upon subsequent prompt input.

[0128] According to one embodiment of the present specification, the server stores a communication tone profile corresponding to a job category for each artificial intelligence agent and can apply the tone profile as a constraint when generating a response.

[0129] A communication tone profile is a data structure that quantifies the expression style and tone of an AI agent using multidimensional parameters, and may include items such as politeness, formality, assertiveness, empathy, detail, the proportion of technical terms, and the level of structuring. The server can pre-define different tone profiles according to each job category.

[0130] After determining the first artificial intelligence agent, the server can calculate a tone parameter vector to control the expression direction of the response to be generated by the agent. At this time, the server can calculate the tone parameter vector using a mapping function that takes the target tag of the first job prompt, user job classification information, past user feedback indicators, etc., as input variables.

[0131] The server can apply the calculated tone parameter vector as a constraint to the response generation process of the first AI agent. That is, when the agent selects words or constructs sentences, weights corresponding to each parameter are reflected, allowing for control to ensure that a specific tone or writing style is automatically maintained. For example, if the 'assertiveness' parameter is set high, the response may be composed of clear conclusions and instructions, and if the 'empathy' parameter is set high, the response may include a narrative structure that reflects the user's emotions or situation.

[0132] According to one embodiment of the present specification, the server can control the expression style differently depending on the job category or the purpose of the prompt, even for responses with the same content. Additionally, the server can continuously learn a communication tone profile by accumulating feedback data from a specific user or organization. That is, when a user provides positive or negative feedback regarding the tone or format of the response, the server can gradually reflect a personalized expression tendency by using that information to automatically update the tone parameter vector.

[0133] According to one embodiment of the present specification, the server can learn a communication tone profile based on user feedback and periodically update a tone parameter vector.

[0134] The server can calculate a feedback score by analyzing evaluation inputs, modification actions, and re-requests made by the user after receiving a response from an AI agent. This feedback score can be divided into explicit response information and implicit response information. Explicit response information may include numerical evaluations or text comments directly left by the user (e.g., "too formal," "summary is clear"), while implicit response information may include indirect indicators estimated from the user's actual behavior (e.g., number of re-requests, modification rate, dwell time, scroll range, follow-up usage rate, etc.).

[0135] The server can calculate an overall feedback score by combining this explicit and implicit response information. For example, if a user immediately copies the response or quotes it in an external document, it can be considered positive feedback, while if the user enters the "rewrite it" command immediately after the response, it can be interpreted as negative feedback.

[0136] The server can update the tone parameter vector based on the generated feedback scores. This update can be performed using weighted moving averages or gradient-based learning rules. For example, if a specific user's feedback is consistently collected as "the response is too assertive," the server can adjust the vector by gradually decreasing the weight of the assertiveness parameter. Conversely, if feedback that the response is "ambiguous" is repeated, the server can strengthen the clarity and structure level parameters.

[0137] According to one embodiment of the present specification, a server may manage personalized vectors and organizational common vectors separately when updating feedback. Personalized vectors represent communication tone tendencies specialized for individual users, while organizational common vectors may refer to standard communication styles applied to the same department or the entire organization. The server may fuse and apply the two vectors by considering time weights or usage frequency weights. For example, in the case of new users, the organizational common vector may be applied with a high weight, and if a specific user has used the service continuously for a certain period, the influence of the personalized vector may gradually expand.

[0138] The server can repeat this vector update procedure at regular intervals or whenever new feedback accumulates. Through this, the server can control the communication style of the AI ​​agent to gradually adapt to the user's actual response patterns.

[0139] According to one embodiment of the present specification, the server can correct the communication tone parameter vector according to context information and apply a format of expression suitable for the situation when generating an actual response.

[0140] Context information is an element that defines the environmental and situational context in which a response will be used, and may include the recipient's role, the type of transmission channel, language and region settings, the purpose of the response, etc. The server can identify the context by analyzing metadata or user profile information transmitted along with a prompt entered from a user terminal.

[0141] The server can adjust the existing tone parameter vector by applying a correction matrix corresponding to the identified context. The correction matrix includes predefined weight adjustment rules for each context element and can be configured, for example, to increase the weights of formality and politeness items when the recipient role is 'executive', and to enhance the level of structuring and visual expressiveness when the delivery channel is 'presentation'.

[0142] The server can incorporate the new tone parameter vector generated as a result of applying the correction matrix into the response generation process. In this case, the response generation logic can dynamically control sentence length, narrative density, expression intensity, and the usage rate of technical terms based on the adjusted vector values. For example, if the recipient is designated as an external customer, the server can adjust the writing style to reduce the proportion of technical terms in the response and strengthen empathetic expressions and descriptive conjunctions. Conversely, if it is determined to be for internal technical reporting, the system can control the response to increase the level of structuring and prioritize fact-based narratives.

[0143] According to one embodiment of the present specification, the server can control various communication tones even for the same artificial intelligence agent by dynamically applying a correction matrix on a situational basis rather than using a single tone parameter vector as a single standard.

[0144] In addition, the server can apply a weighted integration method for each element of the correction matrix, taking into account cases where context information interacts in a complex manner. For example, if the delivery channel is 'email' and the recipient role is 'senior executive,' the server can simultaneously reinforce the formality weight and the conciseness weight to construct a response that is professional yet not excessive.

[0145] According to one embodiment of the present specification, a server may validate the response of an artificial intelligence agent by applying a style guardrail rulebook to comply with the organization's communication policy and expression norms, and rewrite it if necessary.

[0146] The style guardrail rulebook may consist of a set of rules including the organization's document writing standards, language usage guidelines, prohibited words, restrictions on expression intensity, and stylistic format standards. The server can pre-load the rulebook or integrate with an external management system to automatically reflect the latest version of the policy.

[0147] After response generation is complete, the server can execute a rule-based post-processor on the output controlled by tone parameter vectors. During this post-processing, the server can verify compliance with each item by comparing the response's sentence structure, word choice, format level, length, etc., against a rulebook.

[0148] When a violation is detected, the server can rewrite the response by adjusting tone parameters and style templates corresponding to the violation. For example, if a specific expression in the response is determined to be inappropriate according to organizational norms, the server can automatically replace it with an alternative phrase of the same meaning. Additionally, if a sentence is longer than the standard length, the content can be summarized or the sentence split, and if it is judged to lack formality, the expression can be modified by reinforcing politeness.

[0149] Furthermore, the server can learn rule violation patterns to progressively improve the application of the rulebook. For example, if a specific prohibited word appears repeatedly in the responses of users within the same department, the server can increase the enforcement intensity of that rule or automatically generate additional filtering rules. Conversely, rules deemed to be unnecessarily excessive restrictions can be relaxed or adjusted to be applied conditionally.

[0150] In this way, the server can automatically reflect the organization's communication norms and verify and correct in real time the responses generated by the AI ​​agent to ensure they comply with internal guidelines. Accordingly, according to one embodiment of this specification, the server can operate as an intelligent compliance engine that performs organizational-level language quality management functions, going beyond a simple language model control system.

[0151] The operation according to the embodiments of this specification may be implemented as a computer-readable program or code on a computer-readable recording medium. A computer-readable recording medium includes all types of recording devices in which data that can be read by a computer system is stored. Additionally, a computer-readable recording medium may be distributed across networked computer systems, allowing computer-readable programs or code to be stored and executed in a distributed manner.

[0152] When the embodiment is implemented in software, the above-described technique may be implemented as a module (process, function, etc.) that performs the above-described function. The module may be stored in memory and executed by a processor. The memory may be located inside or outside the processor and may be connected to the processor by various well-known means.

[0153] In addition, computer-readable recording media may include hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Program instructions may include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc.

[0154] Some aspects of this specification have been described in the context of a device, but may also be described according to a corresponding method, wherein a block or device corresponds to a method step or a feature of a method step. Similarly, aspects described in the context of a method may also be described according to a corresponding block or item or a feature of a corresponding device. Some or all of the method steps may be performed by (or using) a hardware device, such as, for example, a microprocessor, a programmable computer, or an electronic circuit. In some embodiments, one or more of the most important method steps may be performed by such a device.

[0155] In the embodiments, a programmable logic device (e.g., a field-programmable gate array) may be used to perform some or all of the functions of the methods described herein. In the embodiments, the field-programmable gate array may operate with a microprocessor to perform one of the methods described herein. Generally, it is preferable that the methods be performed by some hardware device.

[0156] Although the foregoing has been described with reference to preferred embodiments of this specification, those skilled in the art will understand that various modifications and changes can be made to this specification without departing from the spirit and scope of the specification as set forth in the following claims.

Claims

Claim 1 A method for providing a job prompt management service based on artificial intelligence model cross-validation performed by at least one server, comprising: obtaining a first job prompt including job-related input data from a user terminal; determining a first artificial intelligence agent for generating a response among a plurality of artificial intelligence agents based on the content, context, and user job information of the first job prompt; obtaining a first response to the first job prompt from the first artificial intelligence agent; obtaining a second response to the first job prompt from a group of artificial intelligence agents including at least one artificial intelligence agent other than the first artificial intelligence agent; verifying the suitability of the first response by comparing the first response and the second response; correcting the selection criteria of the first artificial intelligence agent based on the verification result; storing a communication tone profile corresponding to a job category for each artificial intelligence agent, wherein the communication tone profile is defined as a multidimensional tone parameter vector including politeness, formality, assertiveness, empathy, detail, technical term ratio, and level of structuring; and after determining the first artificial intelligence agent, job classification information, the purpose tag of the first job prompt, and past users A step of calculating the tone parameter vector by a mapping function that takes a feedback indicator as input; a step of controlling the tone parameter vector to be applied as a response generation constraint of the first artificial intelligence agent; a step of calculating a feedback score from user feedback to learn the communication tone profile, wherein the feedback score includes explicit response information and implicit response information, the explicit response information includes user evaluation input and whether a re-request was made, and the implicit response information includes a response modification rate, dwell time, scroll pattern, and subsequent usage rate;A step of updating the tone parameter vector using a weighted moving average or gradient-based update rule based on the feedback score; a step of distinguishing between a user-specific personalized vector and an organization-common vector and setting them to be fused according to time weights; a step of correcting the tone parameter vector according to context, wherein the context includes a recipient role, a delivery channel type, language and region setting information, and a response purpose, wherein the recipient role includes at least one of an executive, a practitioner, and an external customer, wherein the delivery channel type includes at least one of email, chat, and presentation, and wherein the response purpose includes at least one of a summary, report, and proposal; a step of adjusting the tone parameter vector by applying a correction matrix corresponding to the context and then generating a response; a step of obtaining a style guardrail rulebook according to organizational regulations, wherein the rulebook includes forbidden word rules, upper and lower limits of expression intensity, minimum formality standards, and length constraints; and a step of checking compliance with response candidates generated from the tone parameter vector using a rule-based post-processor. A method for providing an AI model cross-validation-based job prompt managing service, comprising the step of setting up rewriting by adjusting tone parameters and style templates corresponding to the violation items when violation items exist. Claim 2 A method for providing an AI model cross-validation-based job prompt managing service according to claim 1, further comprising the step of dynamically determining an AI agent corresponding to a second job prompt subsequently input based on the corrected selection criteria. Claim 3 A method for providing an artificial intelligence model cross-validation-based job prompt managing service, wherein, in claim 1, prior to obtaining the first job prompt, the method further comprises: obtaining job classification information including the user's job type, position, and affiliated department information from the user terminal; and utilizing the job classification information as an input variable in the decision step of the first artificial intelligence agent. Claim 4 A method for providing an AI model cross-validation-based job prompt managing service, wherein, in claim 3, character data corresponding to each of a plurality of AI agents is stored, and said character data includes job characteristics, response style, and expertise domain information of each AI agent, and the step of determining said first AI agent includes the step of selecting the AI ​​agent with the highest job suitability by matching said job classification information and said first job prompt with said character data. Claim 5 A method for providing an AI model cross-validation-based job prompt managing service according to claim 1, wherein the step of obtaining the second response comprises: generating an AI agent group including a second AI agent and a third AI agent; controlling each AI agent to exchange partial responses, grounds, and counterarguments to the first job prompt in a conversation session; and determining the response integrally derived through the conversation within the agent group as the second response. Claim 6 A method for providing an AI model cross-validation-based job prompt managing service according to claim 1, wherein the step of determining the first AI agent comprises: providing the input data to a distribution agent; obtaining an analysis result for the input data from the distribution agent; identifying a plurality of AI agent candidates corresponding to the subject, job category, and response type of the first job prompt based on the analysis result; calculating a job suitability calculated by synthesizing factors including job category matching, language style similarity, and past response accuracy for each of the plurality of AI agents; and setting the AI ​​agent with the highest job suitability among the candidates to be determined as the first AI agent. Claim 7 A method for providing an AI model cross-validation-based job prompt managing service according to claim 1, wherein the step of verifying the suitability of the first response comprises: controlling an evaluation agent to evaluate the logical structure, factual basis, and correspondence to job objectives of each of the first response and the second response; calculating a verification score according to the evaluation result; terminating the verification procedure if the verification score is greater than or equal to a preset standard; correcting the selection criteria of the first AI agent if the verification score is less than the preset standard; redetermining the first AI agent based on the corrected selection criteria; and setting the step of verifying the suitability of the first response again by obtaining a new response through the redetermined first AI agent.

Images