Domain-specific collaborative general ai system based on reinforcement learning

The AI system addresses limitations in conventional AI by subdividing queries and integrating responses from multiple domain-specific agents, improving accuracy and consistency in complex queries through reinforcement learning.

KR103043036B1Active Publication Date: 2026-11-06PERSONA AI CO LTD
View PDF 7 Cites 0 Cited by

Patent Information

Application Number
KR1020250142819
Authority / Receiving Office
KR · KR
Patent Type
Patents
Current Assignee / Owner
Filing Date
2025-09-30
Publication Date
2026-11-06
Estimated Expiration
2045-09-30

AI Technical Summary

Technical Problem

Conventional generative AI systems face limitations in handling complex queries requiring multi-level logical reasoning and domain knowledge, leading to biased, incomplete, and inconsistent responses, and lack effective methods for conflict resolution during agent collaboration.

Method used

A general-purpose AI system subdivides user queries into semantic units, matches each unit to a domain-specific agent, and integrates responses through reinforcement learning-based collaboration among multiple agents to optimize accuracy and consistency.

Benefits of technology

This approach enhances response accuracy and expertise, prevents conflicts, and satisfies various output constraints by optimizing agent policies, resulting in high-quality, customized responses for complex queries.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 112025111737315-PAT00001_ABST
    Figure 112025111737315-PAT00001_ABST
Patent Text Reader

Abstract

A method of operation of an electronic device is disclosed. The method of operation of at least one electronic device according to the present disclosure comprises the steps of: obtaining a plurality of subtexts by subdividing a question text obtained according to user input into a plurality of request units; selecting at least one domain that matches each of the plurality of subtexts; selecting at least one target agent corresponding to the selected domain among a plurality of agents adaptively learned based on Retrieval-Augmented Generation (RAG) for each of the plurality of domains; and inputting each of the plurality of subtexts into the matching target agent to obtain a plurality of unit responses for each of the plurality of subtexts.
Need to check novelty before this filing date? Find Prior Art

Description

Technology Field

[0001] The present disclosure relates to the field of artificial intelligence technology, and more specifically, to a general-purpose artificial intelligence system that processes user requests through collaboration between agents adaptively learned based on Retrieval-Augmented Generation (RAG) in multiple domains, and a method of operation thereof. Background Technology

[0002] Conventional generative AI systems typically produced responses to given queries based on a single model. While this approach enabled responses specialized for specific domains, it had limitations for questions requiring complex domain knowledge or multi-level logical reasoning.

[0003] For example, in the case of complex queries requiring both medical facts and legal interpretations, single AI models faced problems such as providing responses biased toward specific domain knowledge or generating incomplete and inconsistent answers. Furthermore, methods that generate only a single response were limited in satisfying various output constraints required by users, such as response formats, summary conditions, and multilingual conversion.

[0004] Furthermore, existing multi-agent-based research was limited to simple division of labor or serial processing among agents, and systematic governance and reinforcement learning-based optimization techniques to resolve response conflicts, duplication, and inconsistencies arising during the collaboration process were not sufficiently reflected. Prior art literature

[0005] Registered Patent Publication No. 10-2823037 The problem to be solved

[0006] The present invention aims to solve the above-mentioned problems by providing a general-purpose artificial intelligence system capable of subdividing user queries by request unit, matching each subdivided request unit to a suitable domain and domain-specific agent to generate a unit response, and then producing a complemented optimal response through interaction and collaboration among multiple agents.

[0007] Furthermore, the purpose is to provide a technology that can continuously improve response quality by applying a reinforcement learning-based reward function during the collaboration process of multiple agents to reflect the convergence speed and accuracy of the response policy.

[0008] The purposes of the present disclosure are not limited to those mentioned above, and other purposes and advantages of the present disclosure not mentioned may be understood from the following description and will be more clearly understood from the embodiments of the present disclosure. Furthermore, it will be readily apparent that the purposes and advantages of the present disclosure can be realized by the means and combinations thereof set forth in the claims. means of solving the problem

[0009] A method of operation of at least one electronic device according to the present embodiment may include the steps of: obtaining a plurality of subtexts by subdividing a question text obtained according to user input into a plurality of request units; selecting at least one domain that matches each of the plurality of subtexts; selecting at least one target agent corresponding to the selected domain among a plurality of agents adaptively learned based on RAG (Retrieval-Augmented Generation) for each of the plurality of domains; and inputting each of the plurality of subtexts into the matching target agent to obtain a plurality of unit responses for each of the plurality of subtexts.

[0010] In this case, the step of obtaining the plurality of subtexts may include the step of calculating a plurality of semantic vectors for the question text, and the step of comparing the similarity between the plurality of semantic vectors to detect a topic transition point according to the change in similarity and obtaining a subtext for each request unit.

[0011] Additionally, the step of selecting the domain may be performed by comparing the similarity between the semantic vectors for the plurality of subtexts and the predefined profile vectors for each of the plurality of domains.

[0012] Alternatively, the step of obtaining the plurality of unit responses may include the step of generating a first unit response by inputting a first request unit and a first subtext into a first target agent, and the step of generating a second unit response by inputting a second request unit and a second subtext into a second target agent.

[0013] At this time, the first target agent may further include the step of transmitting the first request unit and the first unit response to the second target agent, and the second target agent may further include the step of transmitting the second request unit and the second unit response to the first target agent.

[0014] In this case, the first target agent may further include the step of generating a third unit response using the first request unit, the received second target request unit, and the second unit response as inputs.

[0015] Meanwhile, the first target agent and the second target agent may further include a mutual evaluation step of assigning evaluation scores to each of the first unit response and the second unit response based on at least one criterion among accuracy, consistency, and domain suitability.

[0016] At this time, based on the evaluation score, the method may further include the step of generating an integrated response candidate by removing duplicate content within the plurality of unit responses or reflecting the complementary relationship between the plurality of unit responses.

[0017] In this case, the method may further include a step of determining the final response by applying a reinforcement learning-based reward function to the aforementioned integrated response candidates to calculate a reward value that reflects the convergence speed and accuracy of the response policy.

[0018] Meanwhile, the method of operating at least one electronic device may further include the step of identifying an overload state of the target agent and, depending on the identified overload state, deactivating the target agent or controlling the generation of a unit response of the target agent. Effects of the invention

[0019] According to the present invention, by subdividing user queries into semantic units and generating responses by selecting a domain and a domain agent corresponding to each request unit, accuracy and expertise can be significantly improved compared to a single model-based response.

[0020] Furthermore, through interaction and collaborative governance among multiple agents, response conflicts can be prevented, redundancy eliminated, and integrated response candidates reflecting complementary relationships can be generated. This enables the improvement of response quality for complex domain queries.

[0021] Furthermore, by applying a reinforcement learning-based reward function, the agents' policies can be continuously optimized, and the convergence speed and accuracy of response policies can be improved. This enables the system to enhance its ability to provide customized responses and satisfy various response constraints. Brief explanation of the drawing

[0022] Aspects, features, and advantages of specific embodiments of the present disclosure will become more apparent from the following description with reference to the accompanying drawings. FIG. 1 is a block diagram for illustrating an electronic device according to one embodiment of the present disclosure. FIG. 2 is a flowchart illustrating a series of operations in which an electronic device acquires a plurality of subtexts for user input and acquires a unit response through a target agent, according to one embodiment of the present disclosure. FIG. 3 is a sequence diagram illustrating a series of operations for generating a unit response through a plurality of target agents according to one embodiment of the present disclosure. FIG. 4 is a flowchart illustrating the operation of an electronic device determining a final response based on a plurality of unit responses according to one embodiment of the present disclosure. FIG. 5 is a flowchart illustrating an algorithm in which an electronic device identifies an overload state of a target agent and controls the generation of a unit response according to one embodiment of the present disclosure. FIG. 6 is a block diagram illustrating an additional configuration of an electronic device according to one embodiment of the present disclosure. Specific details for implementing the invention

[0023] Before specifically describing the present disclosure, the method of description in the specification and drawings is described.

[0024] First, the terms used in this specification and claims have been selected based on general terms considering their functions in the various embodiments of this disclosure. However, these terms may vary depending on the intent of those skilled in the art, legal or technical interpretations, and the emergence of new technologies. Additionally, some terms have been arbitrarily selected by the applicant. Such terms may be interpreted according to the meanings defined in this specification; in the absence of specific definitions, they may be interpreted based on the overall content of this specification and common technical knowledge in the relevant field.

[0025] In addition, the same reference numbers or symbols described in each drawing attached to this specification represent parts or components that perform substantially the same function. For convenience of explanation and understanding, the same reference numbers or symbols are used to describe different embodiments. That is, even if components having the same reference number are all depicted in multiple drawings, the multiple drawings do not imply a single embodiment.

[0026] Additionally, in this specification and claims, terms including ordinal numbers, such as "first," "second," etc., may be used to distinguish between components. These ordinal numbers are used to distinguish identical or similar components from one another, and the meaning of the terms should not be limited by the use of such ordinal numbers. For example, the order of use or arrangement of components combined with such ordinal numbers should not be restricted by the number. If necessary, each ordinal number may be used interchangeably.

[0027] In this specification, singular expressions include plural expressions unless the context clearly indicates otherwise. In this application, terms such as "comprising" or "consisting of" are intended to specify the existence 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.

[0028] In the embodiments of the present disclosure, terms such as "module," "unit," "part," etc. are used to refer to a component that performs at least one function or operation, and such component may be implemented in hardware or software, or in a combination of hardware and software. Additionally, a plurality of "modules," "units," "parts," etc. may be integrated into at least one module or chip and implemented as at least one processor, except where each needs to be implemented in specific individual hardware.

[0029] Furthermore, in the embodiments of the present disclosure, when a part is described as being connected to another part, this includes not only a direct connection but also an indirect connection through another medium. Additionally, the meaning that a part includes a certain component implies that, unless specifically stated otherwise, it does not exclude other components but may include additional components.

[0030] In the present disclosure, the expression “at least one of a, b, or c” may refer to “a”, “b”, “c”, “a and b”, “a and c”, “b and c”, “a, b, and c all”, or variations thereof.

[0031] The terms used in this disclosure have been selected to be as widely used and general as possible, taking into account their functions within this disclosure; however, these terms may vary depending on the intent of those skilled in the art, case law, the emergence of new technologies, etc. Additionally, in specific cases, terms have been selected at the applicant's discretion, and in such cases, their meanings will be described in detail in the relevant explanatory sections. Therefore, terms used in this disclosure should be defined not merely by their names, but based on their meanings and the overall content of this disclosure.

[0032] Singular expressions may include plural expressions unless the context clearly indicates otherwise. Terms used herein, including technical or scientific terms, may have the same meaning as generally understood by those skilled in the art as described in this specification. Additionally, terms including ordinal numbers, such as "first" or "second," used in this specification may be used to describe various components, but said components should not be limited by said terms. Such terms are used solely for the purpose of distinguishing one component from another.

[0033] When a part of a specification is described as "comprising" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components. Furthermore, terms such as "part" or "module" as used in the specification refer to a unit that processes at least one function or operation, and this may be implemented in hardware or software, or as a combination of hardware and software.

[0034] Embodiments of the present disclosure are described below with reference to the attached drawings so that those skilled in the art can easily implement the invention. However, the present disclosure may be embodied in various different forms and is not limited to the embodiments described herein. Furthermore, in order to clearly explain the present disclosure in the drawings, parts unrelated to the explanation have been omitted, and similar parts throughout the specification are denoted by similar reference numerals.

[0035] FIG. 1 is a block diagram for explaining an electronic device (100) according to one embodiment of the present disclosure.

[0036] The electronic device (100) may be a device that performs functions for processing user input, subdividing request units, selecting a domain corresponding to the subdivided request units, and generating and integrating unit responses through multiple agents.

[0037] The electronic device (100) may be implemented as a cloud server, an on-premises server, an edge server, or a similar computing device, and may be implemented software-wise as a server environment including a platform, an operating system, and middleware for executing an artificial intelligence model. Additionally, the electronic device (100) may be implemented by a user terminal, a gateway device, or other network-connected device as needed.

[0038] The electronic device (100) can perform one or more functions and operations, and may be implemented as a single device or distributed among multiple devices so that each function and operation is implemented.

[0039] Referring to FIG. 1, the electronic device (100) may include a memory (110), a communication interface (120), and a processor (130).

[0040] The memory (110) is configured to store at least one instruction or data related to an operating system (OS) for controlling the overall operation of the components of the electronic device (100) and the components of the electronic device (100).

[0041] The memory (110) can store program code required for the operation of the electronic device (100), learned parameters, domain profile vectors, and intermediate data for generating user input and response.

[0042] The memory (110) may include non-volatile memory such as ROM or flash memory, and may include volatile memory such as DRAM. Additionally, the memory (110) may include auxiliary storage devices such as a hard disk or an SSD (Solid State Drive).

[0043] At least one agent module may be stored in the memory (110). For example, the memory (110) may include a first agent module (111) and a second agent module (113). Each of these modules corresponds to a functional unit module that can be implemented in software or hardware.

[0044] The agent modules (111, 113) correspond to functional unit modules for executing generative artificial intelligence (AI) models adapted to Retrieval-Augmented Generation (RAG) to correspond to specific domains. The agent modules may be modules that receive a user request unit as input and generate a unit response based on the domain knowledge.

[0045] The agent module can be implemented as a pre-trained generative AI model that takes subtext as input and generates a unit response corresponding to the input. The specific operation thereof will be examined in detail in Fig. 2.

[0046] The agent module performs searches on databases, literature repositories, or knowledge graphs corresponding to the domain, and can improve the accuracy of the response by combining the retrieved external knowledge with its own model.

[0047] Agent modules can interact with other agent modules to share or evaluate responses, and in this process, they can contribute to the collaborative governance process by determining conflicts, complementarity, and domain suitability between responses. Additionally, response policies can be optimized by applying a reinforcement learning reward function based on scores assigned through peer evaluation.

[0048] The agent module may be implemented in software as program code stored in memory and executed by a processor (130). In addition, it may be implemented in hardware as a dedicated artificial intelligence chipset, accelerator, or computing device such as an FPGA or ASIC.

[0049] Additionally, the agent module can monitor the usage status of computational resources on its own to report an overload status, or receive a disable or execution delay command from the control module of the electronic device (100) and control the operation accordingly.

[0050] Meanwhile, unlike in Fig. 1, the agent module (111, 113) may be stored on an AI server capable of communicating with the electronic device (100) rather than in the memory (110) of the electronic device (100).

[0051] The communication interface (120) may include a circuit, module, chip, etc. for communicating with at least one external device (AI server, database management server, user terminal, etc.) using various wired or wireless communication methods.

[0052] The communication interface (120) can be connected to an external device through various networks.

[0053] Depending on the area or scale, a network may be a Personal Area Network (PAN), Local Area Network (LAN), Wide Area Network (WAN), etc., and depending on the openness of the network, it may be an Intranet, Extranet, or Internet, etc.

[0054] The communication interface (120) can be connected to external devices through various wireless communication methods such as LTE (long-term evolution), LTE-A (LTE Advance), 5G (5th Generation) mobile communication, CDMA (code division multiple access), WCDMA (wideband CDMA), UMTS (universal mobile telecommunications system), WiBro (Wireless Broadband), GSM (Global System for Mobile Communications), DMA (Time Division Multiple Access), WiFi (Wi-Fi), WiFi Direct, Bluetooth, NFC (near field communication), Zigbee.

[0055] The communication interface (120) may be connected to external devices through various wired communication methods such as Ethernet, the Internet, USB, HDMI, and Thunderbolt.

[0056] The electronic device (100) can be connected to various servers through a communication interface (120), and can also communicate with user terminals, AI servers, user input devices (e.g., keyboard devices, mouse devices, etc.).

[0057] The processor (130) is configured to control the electronic device (100) overall. Specifically, the processor (130) can perform operations according to various embodiments of the present disclosure by being connected to the memory (110) and executing at least one instruction stored in the memory (110).

[0058] The processor (130) may include a general-purpose processor such as a CPU, AP, DSP (Digital Signal Processor), a graphics-dedicated processor such as a GPU, VPU (Vision Processing Unit), or an artificial intelligence-dedicated processor such as an NPU. An artificial intelligence-dedicated processor may be designed with a hardware structure specialized for training or utilizing a specific artificial intelligence model.

[0059] Referring to FIG. 1, the processor (130) can control a text acquisition module (131), a domain selection module (133), an agent selection module (135), and a response generation module (137), etc. Each of these modules corresponds to a functional unit module that can be implemented in software or hardware.

[0060] The text acquisition module (131) can acquire question text from user input and perform preprocessing to subdivide the acquired question text into request units.

[0061] The text acquisition module (131) can convert the question text into multiple semantic vectors and analyze the change in similarity between the semantic vectors to detect the topic transition point, thereby producing multiple subtexts.

[0062] Subtext can refer to partial text obtained by subdividing the question text entered by the user into semantic units.

[0063] Subtext is not limited to simple grammatical division units such as sentences, clauses, and phrases, but may include semantic segments corresponding to changes in the intent of a question or shifts in topic.

[0064] For example, if a user enters the question “Explain cases of side effects from anticancer treatment and inform me of the doctor’s legal liability in such cases,” the subtext can be divided into “Explanation of cases of side effects from anticancer treatment” and “Doctor’s legal liability.”

[0065] Subtext can be generated by converting the question text into multiple semantic vectors and detecting topic transition points based on changes in similarity between the semantic vectors. Therefore, subtext can be defined as a section within the question text that contains different request objectives.

[0066] In addition, the subtext acquisition process may be performed through a pre-trained AI model capable of outputting subtext subdivided into semantic units from the text.

[0067] As such, subtext is not merely the result of dividing text into fixed lengths, but rather partial text generated based on the semantic context and intent of the question, which can be utilized as an input unit for the domain selection and agent matching stages in subsequent steps.

[0068] The domain selection module (133) can perform the function of selecting at least one domain for each subtext produced by the text acquisition module (131).

[0069] A domain refers to a set of knowledge or an area of ​​expertise in a specific field. Domains can include specific subject areas such as medicine, law, finance, technology, education, and entertainment.

[0070] Each domain can be linked to a database, literature set, or knowledge graph specialized for that field, and multiple agents can learn adaptively from this domain-specific data to generate responses optimized for a specific domain.

[0071] A domain can be represented by a predefined profile vector, and the profile vector can contain semantic features representing the domain.

[0072] Profile vectors can be obtained from domain-related datasets through pre-training or fine-tuning processes, and can be compared with the semantic vectors of subtext to be used as criteria for domain selection.

[0073] In addition, domains can be dynamically expanded or new domains can be added based on user requests. For example, if data regarding a specific industry or academic field that was not previously defined is obtained, a new profile vector based on that data can be generated, and a new domain corresponding to it can be registered within the system.

[0074] Therefore, a domain is not merely limited to a classification system but can function as a practical processing unit linked to knowledge resources for response generation.

[0075] The domain selection module (133) can compare the similarity between the semantic vector of the subtext and the predefined profile vector for each domain and match the domain with the highest similarity to the subtext.

[0076] The agent selection module (135) can perform the function of selecting at least one target agent among a plurality of agents corresponding to the domain selected by the domain selection module (133).

[0077] The agent selection module (135) can select the agent with the highest suitability by analyzing features such as keywords, query types, and expected output formats included in the request unit.

[0078] The agent selection module (135) can determine a suitable agent based on the nature of the request unit and processing priority among multiple agents trained on a RAG basis by domain.

[0079] The agent selection module (135) can operate based on a predefined rule or policy table. For example, it is possible to select a summary-specialized agent when a long text summary request is included in a specific domain, and a legal-specialized agent when a legal query is included.

[0080] Alternatively, the agent selection method may operate based on a pre-trained AI model to select an agent that matches the domain. For example, the agent selection module (135) may convert a request unit into an embedding vector and calculate the similarity with the profile vector corresponding to each agent to select the agent with the highest similarity.

[0081] In this way, the agent selection module (135) is not limited to a single criterion and can select a suitable agent using an AI-based method, a rule-based method, or a combination thereof.

[0082] The response generation module (137) can perform the function of inputting subtext to the target agent selected by the agent selection module (135) and obtaining a unit response thereto.

[0083] The response generation module (137) can perform the function of receiving unit responses generated by each of the multiple target agents selected by the agent selection module (135) as input and producing a final response. The response generation module (137) can generate an integrated response candidate by aggregating multiple unit responses and produce a final response by applying a reinforcement learning-based reward function.

[0084] At this time, the response generation module (137) can generate an integrated response candidate by removing overlapping content between multiple unit responses and reflecting mutually complementary content. In this process, the response generation module (137) can calculate the importance of the response based on the evaluation score (accuracy, consistency, domain suitability, etc.) assigned to each unit response.

[0085] Additionally, the response generation module (137) can apply a reinforcement learning-based reward function to multiple integrated response candidates to calculate a reward value that reflects the convergence speed and accuracy of the response policy, and thus determine the final response.

[0086] The response generation module (137) can post-process the final response by reflecting output constraints (e.g., summary length, language, output format, etc.) included in the user request.

[0087] The response generation module (137) can be realized through an AI model, and a pre-trained generative model can be utilized to generate a final response based on multiple unit responses. In this case, the final response can be automatically generated by a reinforcement learning-based policy model, a language generation model, or a combination thereof.

[0088] The response generation module (137) is not limited to simply connecting unit responses, but can collaboratively integrate and optimize multiple unit responses, and can provide a final response that meets user requirements through an AI-based optimization process.

[0089] Here, the AI ​​model included in the electronic device (100) can be constructed by considering the application field of the recognition model, the purpose of learning, the form of the processing data, or the computational performance of the device.

[0090] An artificial intelligence model may be based on a neural network and may include multiple nodes (neurons) that simulate the human neural network. Each node has a weight and is connected to others to perform signal transmission processes similar to synapses.

[0091] Artificial intelligence models can include various structures such as Deep Neural Networks (DNN), Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Restricted Boltzmann Machines (RBM), Deep Belief Networks (DBN), Bidirectional Recurrent Deep Neural Networks (BRDNN), and Reinforcement Learning-based Q-Networks (DQN); recently, hybrid models that combine or modify these structures are also being utilized.

[0092] In addition to supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning, self-supervised learning techniques are being widely applied. Self-supervised learning can be utilized by performing pretraining using large-scale unstructured data, followed by fine-tuning or instruction-based learning to suit a specific domain.

[0093] In particular, Large Language Models (LLMs) and Foundation Models have recently been utilized as the core of artificial intelligence. Representative examples include GPT, BERT, T5, PaLM, LLaMA, Claude, and Gemini; these models are capable of performing advanced Natural Language Processing (NLP) and Natural Language Generation (NLG) by learning from billions to hundreds of billions of parameters.

[0094] Large-scale language models can embed text inputs into vectors in a latent space to analyze semantic similarity and perform various functions such as context understanding, sentence generation, multilingual translation, summarization, question answering (QA), conversational response, and content generation.

[0095] In addition, Multimodal Large-Scale Models (Multimodal LLMs) have recently been developed to simultaneously process various inputs, such as text, images, voice, and video, and to learn the relationships between heterogeneous data to generate richer and more sophisticated responses.

[0096] Therefore, artificial intelligence models are not limited to simple pattern recognition but can provide various advanced functions, such as complex inference, multimodal processing, and reflection of user context, based on large-scale data and foundation model structures.

[0097] FIG. 2 is a flowchart illustrating a series of operations in which an electronic device (100) obtains a plurality of subtexts for user input and obtains a unit response through a target agent, according to one embodiment of the present disclosure.

[0098] The electronic device (100) can obtain multiple subtexts by subdividing the question text obtained according to user input into multiple intention units (S210).

[0099] Subtext includes semantic segments within the question text where the topic shifts or the purpose of the request changes, and can be generated based on the semantic context of the query rather than simple sentence-unit division.

[0100] The electronic device (100) can input a question text into an embedding model to produce multiple semantic vectors. The electronic device (100) can compare the similarity between semantic vectors and detect a point where the change in similarity occurs above a threshold as a topic transition point, and can divide the question text into multiple subtexts based on the topic transition point.

[0101] The electronic device (100) can generate multiple semantic vectors for a question text and can detect a point where the change in similarity exceeds a threshold by comparing the similarity between the generated multiple semantic vectors as a topic transition point.

[0102] The electronic device (100) can divide the question text into request-unit subtexts based on the detected topic transition point, so that each subtext is defined in units of semantic context.

[0103] Additionally, the electronic device (100) can perform candidate segmentation by grammar unit based on syntactic features (phrases, clauses, sentence boundaries) included in the question text, and then combine semantic vector-based analysis to finally determine the subtext corresponding to the actual request intent unit.

[0104] Alternatively, the electronic device (100) can calculate the probability value of the conversion possibility of the semantic unit within the question text using an artificial intelligence classifier that has been pre-trained to extract subtext according to the semantic unit within the text, and can calculate the subtext based on the point where the probability value exceeds a threshold.

[0105] Subtext can be defined not simply as the result of dividing text into fixed lengths, but as a semantic unit produced by reflecting topic shifts or changes in intent within the question text.

[0106] The electronic device (100) can select at least one domain that matches each of the multiple subtexts (S220).

[0107] The electronic device (100) can convert a subtext into an embedding vector and compare the similarity with a profile vector of each predefined domain. The electronic device (100) can determine the domain having the largest similarity value as the matching domain of the corresponding subtext.

[0108] A profile vector refers to a vector that quantifies semantic features representing a specific domain. Profile vectors can be generated through pre-training or fine-tuning processes using datasets, literature, knowledge graphs, or user log data related to the domain.

[0109] For example, profile vectors corresponding to the medical domain can be trained based on a dataset containing key medical concepts such as disease names, treatment methods, and side effects, and profile vectors corresponding to the legal domain can be trained based on case law, legal terms, and regulatory texts.

[0110] The electronic device (100) calculates the similarity between the profile vector for each of the multiple domains and the semantic vector of the subtext, and can match the domain with the highest similarity with the subtext.

[0111] When multiple domains show equally high similarity, the electronic device (100) can match the corresponding subtext to multiple domains in parallel to perform multi-agent collaboration in a subsequent step.

[0112] According to another embodiment, the electronic device (100) can extract keywords within the subtext and analyze the matching frequency between the extracted keywords and a domain-specific keyword dictionary. If the ratio of keywords corresponding to a specific domain is greater than or equal to a threshold, the electronic device (100) can determine that domain as the matching domain of the subtext.

[0113] Alternatively, the electronic device (100) may calculate the domain classification probability of the subtext using a pre-trained artificial intelligence classification model to select a domain that matches the subtext. In this case, a domain with a probability value greater than or equal to a threshold may be matched with the subtext.

[0114] If the classification probability has similar values ​​for multiple domains, the electronic device (100) can select multiple domains in parallel or make a re-request to the upper control module.

[0115] As such, the domain selection step does not simply rely on predefined rules but can be implemented in various ways, such as embedding similarity, keyword matching, or artificial intelligence classifiers, and parallel selection of multiple domains based on subtext may also be possible.

[0116] Afterwards, the electronic device (100) can select at least one target agent corresponding to the selected domain among a plurality of agents adaptedly learned based on RAG for each of a plurality of domains (S230).

[0117] The electronic device (100) can refer to profile information for each of a plurality of agents belonging to a selected domain and compare the similarity between the semantic vector of the corresponding subtext and the agent profile vector to determine the agent with the highest similarity as the target agent.

[0118] Alternatively, the electronic device (100) may select a target agent according to a predefined policy table by considering the nature of the request unit (e.g., query, summary, translation, commentary, etc.) and the processing priority (e.g., urgent response, accuracy priority, summary priority, etc.).

[0119] Alternatively, the electronic device (100) may perform agent selection using an AI-based selection model. For example, by utilizing a multilayer perceptron or a transformer-based classifier, the probability of a good match between the subtext and the agent may be calculated, and the agent with the highest probability value may be selected as the target agent.

[0120] Meanwhile, the electronic device (100) is not limited to selecting only a single agent, but can perform a collaborative response by selecting multiple agents in parallel. In this case, multiple target agents can each generate a unit response and then contribute to the final response output process through mutual evaluation and collaboration steps. This will be examined in detail in FIG. 3.

[0121] As such, the agent selection step can be implemented in various ways, such as profile-based similarity comparison, policy-based rule selection, AI model-based probability selection, or parallel selection of multiple agents.

[0122] Next, the electronic device (100) inputs each of the multiple subtexts into a matching target agent to obtain multiple unit responses for each of the multiple subtexts (S240).

[0123] The electronic device (100) can input each of a plurality of subtexts to a matching target agent. Each target agent can perform a RAG-style processing process based on the input subtext.

[0124] Specifically, the target agent can convert subtext into a query to perform a search on a domain-specific database, literature repository, or knowledge graph, and generate a unit response by combining the retrieved external knowledge with a self-trained generative artificial intelligence model.

[0125] A unit response may refer to a response generated by a target agent corresponding to a specific subtext and domain. As a partial response that directly corresponds to the request intent of the corresponding subtext, the unit response can serve as foundational data for constituting the final response.

[0126] Unit responses can be generated by combining the output of a generative AI model with results retrieved from external databases, literature repositories, or knowledge graphs based on RAG. In this case, the unit response is not limited to a simple generated result but can function as a verifiable response combined with the search rationale.

[0127] According to one embodiment, the target agent can generate a unit response by considering keywords included in the input subtext, contextual meaning, and constraints of the user request (e.g., summary length, output format, language, etc.).

[0128] If multiple target agents are selected in parallel, the electronic device (100) can obtain multiple unit responses for the same subtext, and the unit responses can be used to generate integrated response candidates through subsequent mutual evaluation and collaboration steps.

[0129] Each unit response generated by multiple target agents can be defined not merely as a mechanical output of the input text, but as a domain-specific response generated through the combination of search-based knowledge and generative artificial intelligence, and can subsequently be utilized as basic data for producing the final response.

[0130] FIG. 3 is a sequence diagram illustrating a series of operations for generating a unit response through a plurality of target agents according to one embodiment of the present disclosure.

[0131] The electronic device (100) can input a plurality of subtexts to each of the matching target agents. In this case, the input subtexts may be the same subtext or different subtexts.

[0132] The first target agent (111) can generate a first unit response based on the input subtext (S310).

[0133] The first unit response may refer to a response generated based on the first subtext input by the first target agent (111). The first unit response may be produced by a RAG-based generative artificial intelligence model adapted to the domain to which the first target agent belongs.

[0134] The first unit response may be a result that directly corresponds to the semantic intent of the corresponding subtext, and may be generated by combining information retrieved from external data resources such as domain databases, literature repositories, and knowledge graphs with the output of the first target agent's trained generative model.

[0135] In addition, the first unit response can be shared during the process of interaction with other target agents and can be used as basic data in the process of comparison and supplementation with the second and third unit responses generated thereafter.

[0136] The second target agent (113) can generate a second unit response based on the input subtext (S320).

[0137] The second unit response can be produced by a generative artificial intelligence model adapted to the RAG method to be specialized for the domain to which the second target agent (113) belongs.

[0138] Specifically, the second target agent (113) can convert the second subtext into a query to perform a search on a domain database, literature repository, or knowledge graph, and generate a response by combining the searched external knowledge with the results of its own model.

[0139] In addition, the second unit response can be produced by reflecting the request type (e.g., description, summary, comparison, commentary, etc.) and output constraints (e.g., volume, language, format, etc.) of the second subtext.

[0140] The second unit response may not belong to the same domain as the first unit response, and in this case, it may reflect a different perspective or knowledge system from the first unit response.

[0141] The second unit response can subsequently form a complementary relationship with the first unit response or be used as an important comparison criterion in the process of detecting response conflicts.

[0142] Afterwards, the first target agent (111) can transmit the first unit response to the second target agent (113) (S330).

[0143] The second target agent (113) can reflect the received first unit response as input to its own response generation process. The second target agent (113) can produce a complementary or interconnected response by referring to the first unit response along with its own domain knowledge.

[0144] At this time, the second target agent (113) may assign an evaluation score to the received first unit response. The evaluation score may be calculated based on at least one of the accuracy of the response, logical consistency between responses, and domain suitability.

[0145] The first unit response can be transmitted in the form of a structured packet that includes not only simple text data but also a confidence score, source document identifier, or search metadata used in the response generation process. Accordingly, the second target agent (113) can derive a more precise response by utilizing the contextual information of the first unit response.

[0146] Likewise, the second target agent (113) can transmit the second unit response to the first target agent (111) (S340).

[0147] The first target agent (111) can generate a subsequent third unit response based on the received second unit response, or review whether there is a conflict and a complementary relationship between the first unit response and the second unit response.

[0148] Likewise, the first target agent (111) can calculate an evaluation score for the second unit response and can calculate the evaluation score based on at least one of accuracy, consistency, and domain suitability.

[0149] The second unit response may be transmitted along with a confidence score, domain suitability indicator, or metadata, just like the first unit response, and the first target agent (111) may refer to this to generate a more collaborative response.

[0150] Afterwards, the first target agent (111) can generate a third unit response (S350).

[0151] The first target agent (111) can generate a third unit response by using the second request unit and the second unit response received from the second target agent (113) as inputs, along with the first request unit and the first unit response.

[0152] The third unit response can be generated by removing the overlapping content of the first and second unit responses and integrating mutually complementary information.

[0153] In this process, the first target agent (111) may consider the evaluation scores for the above responses together. The third unit response can be optimized by prioritizing the reflection of the content of the response that obtained a high evaluation score based on at least one of accuracy, consistency, and domain suitability, and by reducing or excluding the content of the response that has a low evaluation score.

[0154] In this case, the third unit response can be expanded into a more sophisticated response by adjusting for conflicts between the two responses or by reflecting combined knowledge.

[0155] The first target agent (111) can adjust response weights and produce an optimized third unit response by reflecting the accuracy, consistency, reliability score, domain suitability indicator, etc. included in the second unit response.

[0156] The third unit response can contribute to the gradual improvement of the quality of the collaborative response by applying a reinforcement learning-based reward function in subsequent steps.

[0157] The first target agent (111) can transmit the third unit response to the second target agent (113) (S360).

[0158] The second target agent (113) can refer to the received third unit response to supplement additional responses or review consistency with the first response or the second response.

[0159] The third unit response can be input into the evaluation module of the second target agent (113) to receive evaluation scores for accuracy, consistency, and domain suitability, and the results of the mutual evaluation can be used in the final response integration process.

[0160] The first to third unit responses generated by such target agents may be outputs produced by a RAG-based generative artificial intelligence model adapted to the domain to which each target agent (111) belongs.

[0161] The process of generating and exchanging responses among multiple target agents is not limited to a single turn but can be performed repeatedly. By supplementing and refining responses during each iteration, multiple target agents can gradually produce high-quality, collaborative unit responses.

[0162] FIG. 4 is a flowchart illustrating the operation of an electronic device (100) determining a final response based on a plurality of unit responses according to one embodiment of the present disclosure.

[0163] The electronic device (100) can generate at least one integrated response candidate by removing duplicate content within a plurality of unit responses or reflecting a complementary relationship between a plurality of unit responses based on an evaluation score for the unit responses (S410).

[0164] The electronic device (100) can assign an evaluation score to each of a plurality of unit responses generated by at least one target agent. The evaluation score can be calculated based on at least one criterion among the accuracy, consistency, and domain suitability of the unit responses, and can be used as an indicator to quantitatively represent the quality of each unit response.

[0165] An evaluation score can be obtained through a pre-trained artificial intelligence model (hereinafter, the first artificial intelligence model) that takes each unit response as input and calculates an evaluation score based on at least one criterion among the accuracy, consistency, and domain suitability of each unit response. At this time, the electronic device (100) may obtain an evaluation score through the same artificial intelligence model or obtain an evaluation score for each unit response through multiple artificial intelligence models trained separately for each of the multiple criteria.

[0166] The electronic device (100) can generate an integrated response candidate for a question text based on a plurality of unit responses, and the integrated response candidate refers to a response that forms the basis of a final response to the question text obtained according to user input. The electronic device (100) can obtain at least one integrated response candidate based on a plurality of unit responses obtained through an artificial intelligence model.

[0167] The electronic device (100) can generate at least one integrated response candidate by removing duplicate content existing within a plurality of unit responses or reflecting mutually complementary content based on the evaluation score of each unit response, and obtain a final response to the question text based thereon.

[0168] A candidate for an integrated response can be generated through a generative model (hereinafter referred to as the second AI model) that is pre-trained to generate a response using multiple unit responses as input. The generated candidate for an integrated response may be a single response, or multiple candidates for integrated responses may be generated in parallel. The AI ​​model for generating the candidate for an integrated response may be implemented as a separate AI model distinct from the AI ​​model trained to calculate the evaluation score.

[0169] When an integrated response candidate is generated, the electronic device (100) applies a reward function based on reinforcement learning to the integrated response candidate to calculate a reward value that reflects the convergence speed and accuracy of the response policy, and can determine a final response among at least one integrated response candidate based on the reward value (S420).

[0170] Specifically, the electronic device (100) may apply a reinforcement learning-based reward function to at least one generated integrated response candidate. The reward function is a function defined to quantitatively evaluate the quality of multiple unit responses or integrated response candidates, and can produce an indicator that reflects the convergence speed and accuracy of the response policy.

[0171] Here, the compensation function may refer to a mathematical or algorithmic expression that calculates a compensation value for an input unit response by reflecting a response quality criterion based on at least one of the accuracy, consistency, and domain fit of the response.

[0172] According to one embodiment, the compensation function (R) may be defined in the form of a weighted sum based on accuracy (A), consistency (C), and domain fit (D), such as R=(α·A)+(β·C)+(γ·D). α, β, and γ are weights assigned to the accuracy, consistency, and domain fit indicators, respectively, and may be adjusted according to the usage environment, application field, or policy.

[0173] A is an accuracy indicator and may be a value representing how well the generated unit response or integrated response matches a reference knowledge base or external validation data.

[0174] C serves as a consistency index and may be a value that comprehensively reflects the presence of logical contradictions, the degree of context preservation, and conflicts in expression among multiple responses to the same question text. The consistency index can be calculated through correlation analysis of semantic vectors between multiple responses or through mutual evaluation among the responses.

[0175] D is a domain relevance metric that may be a value indicating whether the corresponding response is appropriate for the domain to which a specific request unit belongs. The domain relevance metric can be calculated by measuring the similarity between the domain profile vector of the subtext and the embedding vector of the response, or by calculating the degree of agreement with a predefined set of domain knowledge.

[0176] The reward function can progressively optimize the response policy by providing feedback on the quality of the response generated by the target agent during the learning process and strengthening collaboration among multiple agents. The reward function can be defined to reflect the convergence speed and accuracy of the response policy.

[0177] According to a related embodiment, the electronic device (100) may calculate a compensation value for each integrated response candidate based on a compensation function and select integrated response candidates whose calculated compensation value is greater than or equal to a preset threshold. If there are multiple selected integrated response candidates, the electronic device (100) may obtain a final response among the multiple integrated response candidates based on domain suitability.

[0178] Specifically, the electronic device (100) inputs each of a plurality of integrated response candidates, each having a reward value greater than or equal to a threshold, into a pre-trained artificial intelligence model (hereinafter referred to as the third artificial intelligence model) to calculate an evaluation score for domain suitability, and can obtain an evaluation score for each integrated response candidate as an output. The evaluation score is a value calculated based on the degree of similarity between the integrated response candidate and the domain, and a higher evaluation score can be calculated as the similarity between the integrated response candidate and the domain increases.

[0179] Here, the third artificial intelligence model can be implemented as a separate artificial intelligence model distinct from the first and second artificial intelligence models, and can be implemented as an artificial intelligence model trained to take as input the domain information of the agent that generated the integrated response candidate and the unit response, and output an evaluation score based on the similarity to the domain of each integrated response candidate.

[0180] The electronic device (100) generates a unit response through a first artificial intelligence model, inputs the generated unit response into a second artificial intelligence model to generate at least one integrated response candidate, inputs the generated integrated response candidate into a third artificial intelligence model to obtain an evaluation score for each integrated response candidate, and can obtain the integrated response candidate with the highest evaluation score as the final response to the question text.

[0181] In this way, the electronic device (100) can provide a final response optimized for the domain context while ensuring the accuracy and reliability of the final response by obtaining a final response that additionally considers domain suitability based on the magnitude of the reward value for a plurality of integrated response candidates.

[0182] Meanwhile, when the reward value for an integrated response candidate approaches a certain threshold, the electronic device (100) can determine the final response by applying additional evaluation criteria (e.g., user feedback, time efficiency, response diversity).

[0183] According to another embodiment, the final response may be a post-processed result reflecting output constraints included in the user request. For example, if a summary constraint is included, the electronic device (100) may reduce the integrated response candidate to a specified amount or less, and if a multilingual support request is included, the electronic device (100) may call a translation module to provide a response converted into a specified language.

[0184] Alternatively, if a specific response format (e.g., table format, list format, report format, etc.) is required, the final response may be restructured to fit the corresponding output format and provided to the user.

[0185] In this way, the electronic device (100) can generate unit responses through a matching target agent for each subtext through text analysis, and by applying a reinforcement learning-based reward function to multiple unit responses and integrated response candidates, it can optimize the convergence speed and accuracy of the response policy and secure the expertise and domain suitability of the response.

[0186] FIG. 5 is a flowchart illustrating an algorithm in which an electronic device (100) identifies an overload state of a target agent and controls the generation of a unit response according to one embodiment of the present disclosure.

[0187] The electronic device (100) can identify the overload status of at least one target agent (S510).

[0188] Here, an overload state may include a state in which the usage status of resources, such as the CPU usage, memory usage, network latency, and number of concurrent requests of the target agent, exceeds a predefined threshold.

[0189] To this end, the electronic device (100) can obtain resource usage indicators for operations performed by the target agent. Resource usage indicators may include CPU usage, GPU computation amount, memory occupancy, network latency, packet loss rate, number of concurrent processing requests, etc.

[0190] The electronic device (100) can compare the indicator value with a reference threshold and identify the overloaded state of the target agent if it exceeds the threshold.

[0191] Alternatively, the electronic device (100) can measure the response time (latency) of the target agent and identify the target agent as being in an overloaded state if the average response time exceeds a predefined threshold value or if the fluctuation range is unstable above a certain level.

[0192] Alternatively, the electronic device (100) may acquire status events provided by log data or a monitoring module and determine that the target agent is in an overloaded state if it exhibits a state such as an increased error rate, an increased frequency of timeouts, or a queue backlog.

[0193] The electronic device (100) can determine whether the state of the target agent corresponds to an overload state (S520).

[0194] If the target agent is in an overloaded state, the electronic device (100) can disable the target agent or control the generation of a unit response of the target agent (S530).

[0195] The electronic device (100) can relieve the load on the agent by temporarily blocking input to the target agent that is in an overloaded state, or by storing the request in a queue and processing it later.

[0196] In this case, the electronic device (100) can select another target agent within the same domain that is not in an overloaded state to handle requests that would otherwise be handled by the overloaded agent.

[0197] The electronic device (100) can acquire status indicators (e.g., CPU usage, memory usage, average response delay time, etc.) for each of a plurality of agents in real time and determine the availability of the agent based on the acquired status indicators.

[0198] The electronic device (100) can reassign requests assigned to an overload agent to another candidate agent within the same domain depending on availability, and when reassigning, it can determine a replacement agent by considering the priority of the request, processing deadline, response quality requirements, etc.

[0199] According to another embodiment, the electronic device (100) can process requests in parallel by distributing them to a plurality of non-overloaded agents, and in this case, the unit responses generated by each agent can be reflected in the final response after undergoing an integration and evaluation process in a subsequent step.

[0200] According to another embodiment, the electronic device (100) may control the response frequency or output length of an agent in an overloaded state, or prioritize the omission of non-essential requests (e.g., requests for supplementary explanations, additional information, etc.).

[0201] Response length limits can be applied based on output constraints included in the user request or criteria defined internally within the system.

[0202] For example, if a summary response is required, the response length can be limited to a predefined number of tokens or less, and even if a detailed response is required, the output length can be reduced to include only key keywords, main sentences, and summarized paragraphs under overload conditions.

[0203] In addition, the output length limit can be determined based on importance scores derived from a reward function. By including only high-scoring items and omitting low-scoring items based on the calculated importance score for each sentence or unit of information within the response, an efficient length limit can be achieved.

[0204] Alternatively, the omission of non-essential requests may be performed based on the priority classification of request units. For example, among user requests, core queries (e.g., definition, cause, solution) are processed as essential, while auxiliary explanations (e.g., background information, additional examples, related cases) or supplementary requests (e.g., references, extended questions, etc.) may be omitted preferentially in an overloaded state.

[0205] In this way, the electronic device (100) can prevent response delays or quality degradation caused by exceeding the computational resources of a specific agent by controlling the generation of unit responses differently according to the overload status of the target agent. Accordingly, stable response quality can be maintained even in an environment where multiple agents operate simultaneously, and the processing efficiency and reliability of the entire system can be improved.

[0206] FIG. 6 is a block diagram illustrating an additional configuration of an electronic device (100) according to one embodiment of the present disclosure.

[0207] Referring to FIG. 6, the electronic device (100) may further include at least one of a user interface (140) and an output interface (150) in addition to a memory (110), a communication interface (120), and a processor (130).

[0208] The user interface (140) is configured to receive user input related to user commands or information. The user interface (140) may include, but is not limited to, buttons, keypads, touchpads, touchscreens, cameras, microphones, etc.

[0209] Meanwhile, the electronic device (100) may also receive user input from a separate user input device (e.g., keyboard device, mouse device, remote controller, etc.) connected via a communication interface (120).

[0210] The output interface (150) may include a display, a speaker, etc. as a configuration for outputting various information. The electronic device (100) may provide parameter adjustment information for parameter adjustment performed in real time through the output interface (150), and may also provide the aforementioned prediction information regarding changes in performance evaluation information.

[0211] Meanwhile, the electronic device (100) may also provide information through a separate output device (e.g., display device, speaker device, etc.) connected via a communication interface (120).

[0212] Meanwhile, the various embodiments described above may be implemented by combining two or more embodiments, provided that they do not conflict or contradict each other.

[0213] Meanwhile, the various embodiments described above may be implemented in a recording medium readable by a computer or a similar device using software, hardware, or a combination thereof.

[0214] According to hardware implementation, the embodiments described in this disclosure 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), processors, controllers, microcontrollers, microprocessors, and other electrical units for performing functions.

[0215] In some cases, the embodiments described herein may be implemented as the processor itself. In a software implementation, embodiments such as the procedures and functions described herein may be implemented as separate software modules. Each of the aforementioned software modules may perform one or more functions and operations described herein.

[0216] Meanwhile, computer instructions or computer programs for performing processing operations in an electronic device (100) according to various embodiments of the present disclosure described above may be stored in a non-transitory computer-readable medium. When such computer instructions or computer programs stored in the non-transitory computer-readable medium are executed by a processor of a specific device, the specific device described above performs processing operations in the electronic device (100) according to various embodiments described above.

[0217] A non-transient computer-readable medium refers to a medium that stores data semi-permanently and can be read by a device, unlike media that store data for a short period of time such as registers, caches, and memory. Specific examples of non-transient computer-readable media include CDs, DVDs, hard disks, Blu-ray discs, USBs, memory cards, and ROMs.

[0218] Although preferred embodiments of the present disclosure have been illustrated and described above, the present disclosure is not limited to the specific embodiments described above. It is understood that various modifications can be made by those skilled in the art without departing from the essence of the present disclosure as claimed in the claims, and such modifications should not be understood individually from the technical spirit or perspective of the present disclosure. Explanation of the symbols

[0219] 100: Electronic device 110: Memory 111: 1st Agent Module 113: Second Agent Module 120: Communication interface 130: Processor 131: Text Acquisition Module 133: Domain Selection Module 135: Agent Selection Module 137: Response generation module 140: User Interface 150: Output interface

Claims

Claim 1 A method of operating at least one electronic device comprises: a step of subdividing a question text obtained according to user input into a plurality of request units to obtain a plurality of subtexts corresponding to semantic segment units corresponding to a change in request intent; a step of selecting at least one domain that matches each of the plurality of subtexts; a step of selecting at least one target agent corresponding to the selected domain among a plurality of agents adaptively learned based on RAG (Retrieval-Augmented Generation) for each of the plurality of domains; and a step of inputting each of the plurality of subtexts into the matching target agent to obtain a plurality of unit responses for each of the plurality of subtexts; wherein the step of obtaining the plurality of unit responses comprises: a step of inputting a first request unit and a first subtext into a first target agent to generate a first unit response; and a step of inputting a second request unit and a second subtext into a second target agent to generate a second unit response; and a mutual evaluation step in which the first target agent and the second target agent assign evaluation scores to each of the first unit response and the second unit response based on at least one criterion among accuracy, consistency, and domain suitability. A method of operating an electronic device, further comprising the step of: the first target agent and the second target agent removing duplicate content within the plurality of unit responses based on the evaluation score or generating an integrated response candidate by reflecting the complementary relationship between the plurality of unit responses. Claim 2 A method of operation of an electronic device according to claim 1, wherein the step of acquiring a plurality of subtexts comprises: a step of calculating a plurality of semantic vectors for the question text; and a step of acquiring a subtext for each request unit by comparing similarity between the plurality of semantic vectors and detecting a topic switching point according to a change in similarity. Claim 3 A method of operation of an electronic device according to claim 1, wherein the step of selecting the domain is performed by comparing the similarity between the semantic vector for the plurality of subtexts and the profile vector defined for each of the plurality of domains. Claim 4 delete Claim 5 A method of operation of an electronic device according to claim 1, further comprising: the first target agent transmitting the first request unit and the first unit response to the second target agent; and the second target agent transmitting the second request unit and the second unit response to the first target agent. Claim 6 A method of operation of an electronic device according to claim 5, further comprising the step of the first target agent generating a third unit response by taking the first request unit, the received second target request unit, and the second unit response as inputs. Claim 7 delete Claim 8 A method of operation of an electronic device according to claim 1, further comprising the step of generating an integrated response candidate by removing duplicate content within the plurality of unit responses or reflecting the complementary relationship between the plurality of unit responses based on the evaluation score. Claim 9 A method of operation of an electronic device, further comprising the step of determining a final response by applying a reward function based on reinforcement learning to the integrated response candidate to calculate a reward value that reflects the convergence speed and accuracy of the response policy in claim 8. Claim 10 A method of operation of an electronic device according to claim 1, further comprising the step of identifying an overload state of the target agent and, according to the identified overload state, deactivating the target agent or controlling the generation of a unit response of the target agent.

Citation Information

Patent Citations

  • Question-and-answer system, server and operation method for generating answer to user question based on rag technique

    KR102694794B1

  • Multilingual AI DOC System Building Method for Hindi English And Korean

    KR102800075B1

  • Method, apparatus, and recording medium for generating customized robot model using artificial intelligence

    KR102841575B1

  • Document Processing and Response Generation System

    KR1020230025714A

  • System and method for request processing based on multi-agent collaboration

    KR1020250136185A