Multi-domain task processing method, system and equipment based on knowledge fusion and Agent cooperation and medium

Abstract

The invention provides a multi-domain task processing method, system and device based on knowledge fusion and Agent cooperation and a medium, and belongs to the technical field of cross-domain knowledge fusion. Tasks input by a user are decomposed, and a subtask set is generated; building a specialized Agent set according to a subtask calling skill mapping table and a role template library; executing cross-domain knowledge retrieval and fusion on the Agent set, monitoring output consistency and classifying collaborative conflicts; triggering a corresponding extension mechanism according to different conflicts, and updating the knowledge graph or generating a new Agent; meanwhile, field knowledge updating is monitored, and knowledge weights are adjusted and synchronized to related modules. Precise disassembly and specialized processing of complex tasks are realized, and the cross-domain knowledge fusion efficiency is improved; cooperation conflicts are solved, and system adaptability is enhanced; the knowledge is updated in real time, so that the processing timeliness is ensured, and the task processing quality and efficiency of the multi-Agent system are improved.

Description

Technical Field

[0001] This invention belongs to the field of cross-domain knowledge fusion technology, specifically relating to a multi-domain task processing method, system, device, and medium based on knowledge fusion and agent collaboration. Background Technology

[0002] In today's rapidly evolving digital transformation environment, the aforementioned technologies are particularly suitable for industries that require highly specialized knowledge and multidisciplinary collaboration, such as healthcare, fintech, smart manufacturing, and scientific research and education.

[0003] In related technologies, when facing complex cross-domain tasks, the inability to accurately identify the relevant professional fields often leads to chaotic subtask decomposition, with overlapping or omissions among subtasks, impacting processing efficiency. Currently, although agents are used, there is a mismatch between agent capabilities and subtask requirements. Agent generation often lacks precise association with subtasks, and skills, tool interfaces, and inference strategies do not match subtask requirements, resulting in insufficient professionalism in subtask processing. When multiple agents collaborate, conflicts easily arise due to inconsistent output semantics, but the lack of real-time monitoring and effective classification mechanisms means that conflicts cannot be detected in a timely manner, or are difficult to resolve due to unclear conflict types. Summary of the Invention

[0004] This invention provides a multi-domain task processing method based on knowledge fusion and agent collaboration. The method improves the processing efficiency and accuracy of complex tasks and enhances the system's adaptability to cross-domain and changing tasks.

[0005] The methods include: Step S101: Perform domain identification and subtask decomposition on the user-input task, identify the semantic elements involved in the task, and generate a set of subtasks; Step S102: Based on the set of subtasks, call the skill mapping table to obtain the correspondence between the professional skills and tool interfaces required by the subtasks, and combine the Agent behavior patterns and reasoning strategies in the role template library to generate specialized Agent instances and form an Agent set; Step S103: Perform cross-domain knowledge retrieval and fusion for the Agent set; Step S104: Based on the results of multi-source knowledge fusion, monitor the output consistency of the Agent set, analyze the semantic consistency of the output results of different Agents in real time, and classify the types of collaboration conflicts according to the semantic consistency analysis results; Step S105: If a collaboration conflict is detected and the confidence level of the Agent output result is lower than the first preset threshold, vertical knowledge expansion is triggered to update the knowledge graph of the corresponding domain. When the collaboration conflict rate exceeds the second preset threshold, horizontal agent expansion is triggered to generate new specialized agent instances; When the frequency of new concept recognition exceeds the third preset threshold, the knowledge graph is updated to supplement new concepts and their related relationships. Step S106: Monitor new domain knowledge updates, update the knowledge graph of the corresponding domain, calculate the comprehensive weight of timeliness, relevance and credibility of each knowledge fragment in the knowledge graph, and reload the updated knowledge weights and knowledge graph information.

[0006] Preferably, the method of performing cross-domain knowledge retrieval and fusion in step S103 includes: a first-layer domain knowledge retrieval device retrieves domain-specific information from a professional knowledge base based on the agent's specialized skill requirements; a second-layer collaborative memory retrieval device retrieves solutions for similar tasks from a historical collaborative case database; and a knowledge weighted fusion device uses a comprehensive weighting algorithm of timeliness, relevance, and credibility to fuse the retrieval results of the two layers to generate multi-source knowledge fusion results.

[0007] Preferably, step S102 specifically includes: For each subtask in the subtask set generated in step S101, the text description is parsed using natural language processing technology to extract the subtask's domain labels, functional requirements, and input / output requirements. Configure the association between domain tags and professional skills and tool interfaces in the skill mapping table. By querying the skill mapping table, obtain the set of professional skills and the list of tool interfaces that perfectly match the domain tags of the subtask. In the role template library, the behavioral patterns and inference strategies of the Agent are configured according to the domain. Based on the functional requirements of the sub-task, the closest combination of behavioral patterns and inference strategies is selected through semantic matching algorithm. The acquired skill sets, tool interfaces, and filtered behavioral patterns and inference strategies are encapsulated in a structured manner, and an independent Agent instance is generated for each subtask through an instantiation engine.

[0008] Preferably, step S103 specifically includes: The first-layer domain knowledge retrieval results and the second-layer collaborative memory retrieval results are collected respectively, and the timeliness marker, relevance tag, and credibility identifier of each knowledge are extracted. Based on preset mapping rules, timeliness markers are converted into timeliness quantification values; relevance tags are converted into relevance quantification values; and credibility identifiers are converted into credibility quantification values. For each piece of knowledge, the fusion weight of a single piece of knowledge is calculated according to the method of timeliness quantification value × α + relevance quantification value × β + credibility quantification value × γ, and the weight result is marked; α, β, and γ are predefined weight allocation coefficients, and α + β + γ = 1; Sort the knowledge fusion results from highest to lowest according to their weight, select the knowledge whose weight reaches the threshold, aggregate them according to the sub-task logic, and generate the knowledge fusion results.

[0009] Preferably, step S104 specifically includes: Collect the output results of each specialized Agent instance in step S103 and convert them into a unified intermediate representation format; For each standardized output, natural language processing techniques are used to extract key semantic elements; Based on the generated semantic vectors, the consistency of outputs from different agents under the same task is calculated using the following method: Calculate the cosine similarity of the semantic vectors output by any two agents; Calculate the proportion of exact matches of key entities in the two outputs to the total number of entities; Determine whether the intent descriptions of the two outputs belong to the same task sub-goal; By combining the scores from the above three dimensions, an overall consistency score is generated for all Agent outputs; A pre-defined conflict classification rule base is established, which categorizes conflicts into data inconsistency, logical inconsistency, and objective inconsistency based on differences in semantic features.

[0010] Preferably, step S105 specifically includes: Collect the conflict type, severity, associated AgentID, and conflict content information output in step S104, convert them into relevant data, and form an inconsistency event database; Based on the conflict event, the preset extended policy rule base is invoked for matching: If the conflict type is data inconsistency and the severity level is high, match the vertical knowledge expansion strategy; If the conflict type is logical inconsistency and the severity level is medium, match the horizontal agent expansion strategy; If the conflict type is goal inconsistency or the frequency of new concept recognition exceeds θ3, match the knowledge graph update strategy; If multiple types of conflicts are met simultaneously, the primary expansion strategy is selected based on priority: data inconsistency > logical inconsistency > objective inconsistency. Based on the matching strategy, perform vertical and horizontal expansion operations; verify the expanded knowledge graph and the newly added Agent instances.

[0011] Preferably, step S106 specifically includes: When a knowledge base is updated, modified, or deleted, the knowledge base pushes a knowledge update event. Based on the monitored updates, the knowledge graph is modified to ensure the atomicity and traceability of the update operations: If the updated content is a new concept that has not been included in the knowledge graph, a new node will be created in the knowledge graph, and the type, definition, source and creation time will be labeled. If the updated content is a new association of an existing concept, then add a bidirectional edge to the knowledge graph and label the association strength and basis. If the update involves adjusting the attributes of an existing node, then modify the attribute values ​​of the corresponding node and retain the historical version record; Based on the updated knowledge graph, the comprehensive weights of timeliness, relevance, and credibility are recalculated for each knowledge fragment; The updated content and weight changes of the knowledge graph are synchronized to the Agent set in step S102 and step S103.

[0012] This application also provides a multi-domain task processing system based on knowledge fusion and agent collaboration, the system comprising: The task recognition module is used to perform domain recognition and subtask decomposition on the task input by the user, identify the semantic elements involved in the task, and generate a set of subtasks. The instantiation module is used to obtain the correspondence between the professional skills and tool interfaces required by the subtasks by calling the skill mapping table based on the subtask set, and generate professional agent instances by combining the agent behavior patterns and inference strategies in the role template library to form an agent set. The cross-domain fusion module is used to perform cross-domain knowledge retrieval and fusion for a set of Agents; The collaboration monitoring module monitors the output consistency of the Agent set based on the results of multi-source knowledge fusion, analyzes the semantic consistency of the output results of different Agents in real time, and classifies the types of collaboration conflicts according to the semantic consistency analysis results. An extended triggering module is used to trigger vertical knowledge expansion to update the knowledge graph of the corresponding domain when a collaboration conflict is detected and the confidence level of the Agent output result is lower than the first preset threshold. When the collaboration conflict rate exceeds the second preset threshold, horizontal agent expansion is triggered to generate new specialized agent instances; When the frequency of new concept recognition exceeds the third preset threshold, the knowledge graph is updated to supplement new concepts and their related relationships. The knowledge graph continuous update module is used to monitor new domain knowledge updates, update the corresponding domain knowledge graph, calculate the comprehensive weight of the timeliness, relevance and credibility of each knowledge fragment in the knowledge graph, and reload the updated knowledge weights and knowledge graph information.

[0013] According to another embodiment of this application, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the multi-domain task processing method based on knowledge fusion and agent collaboration.

[0014] According to another embodiment of this application, a storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the multi-domain task processing method based on knowledge fusion and agent collaboration.

[0015] As can be seen from the above technical solutions, the present invention has the following advantages: This invention provides a multi-domain task processing method based on knowledge fusion and agent collaboration. By identifying the professional domains involved in the task and extracting core semantic elements, it grasps the key aspects of the task. The independent sub-task set generated based on domain ontology and template library reduces the processing difficulty of complex tasks, provides clear objectives for agent allocation, and reduces information omissions. Based on the sub-task's call to the skill mapping table and role template library, the generated specialized agents possess matching skills, tool interfaces, and reasoning strategies, improving the professionalism and accuracy of sub-task processing. The formation of the agent set enables complex tasks to be efficiently advanced through collaboration. Multi-level knowledge retrieval takes into account domain-specific information and historical cases, making knowledge acquisition more comprehensive. The comprehensive weighted fusion mechanism balances the timeliness, relevance, and credibility of knowledge, making the fusion results more reliable and providing high-quality knowledge support for task processing, reducing the impact of one-sided information. Real-time monitoring of the semantic consistency of agent output can promptly detect collaboration conflicts; the classification of conflict types and severity provides a clear direction for subsequent conflict resolution, preventing conflict spread from affecting the overall task progress and quality. For different conflict-triggered expansion mechanisms, vertical knowledge expansion can deepen the knowledge graph, horizontal agent expansion can enhance processing capabilities, and supplementing new concepts and relationships can improve the knowledge system, ensuring that the system can adapt to changing task requirements and improve the accuracy of conflict resolution. Monitoring and updating domain knowledge ensures the timeliness and accuracy of the knowledge graph; recalculating knowledge weights and synchronizing them to relevant modules enables agents and the knowledge routing engine to operate based on the latest information, maintaining long-term efficient system operation and avoiding processing deviations due to knowledge lag. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a multi-domain task processing method based on knowledge fusion and agent collaboration; Figure 2 This is a schematic diagram of a multi-domain task processing system based on knowledge fusion and agent collaboration; Figure 3 This is a schematic diagram of an electronic device. Detailed Implementation

[0018] This invention provides a multi-domain task processing method based on knowledge fusion and agent collaboration, comprising a task adaptive decomposer, a domain-aware agent builder, a multi-layered knowledge routing engine, and a collaborative conflict intelligent mediator. This framework achieves task decomposition and role mapping through semantic parsing and domain recognition technologies, employs a time-weighted two-layer knowledge retrieval mechanism to achieve intelligent fusion of cross-domain knowledge, and resolves collaborative conflicts between agents through real-time collaborative monitoring and expansion mechanisms. Compared to existing technologies, this invention can adapt to complex cross-domain tasks, improving the task processing efficiency and accuracy of multi-agent systems.

[0019] The following describes in detail the multi-domain task processing method based on knowledge fusion and agent collaboration involved in this application. Specific details, such as particular system architectures and technologies, are presented for illustrative purposes rather than limiting, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details.

[0020] To facilitate a clear description of the technical solution of this application, the terms "first" and "second" are used to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, nor do they necessarily imply that they are different.

[0021] The statements such as "one embodiment" or "some embodiments" described in this application mean that one or more embodiments of this application include the specific features, structures, or characteristics described in that embodiment. Therefore, the statements such as "in one embodiment," "in some embodiments," "in other embodiments," and "in still other embodiments" in this application do not necessarily refer to the same embodiment, but rather mean one or more, but not all, embodiments, unless otherwise specifically emphasized.

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see Figure 1 The diagram shows a flowchart of a multi-domain task processing method based on knowledge fusion and agent collaboration in a specific embodiment. The method includes: Step S101: Perform domain identification and subtask decomposition on the user-input task, identify the semantic elements involved in the task, and generate a set of subtasks.

[0024] In some embodiments, a pre-trained BERT domain classification model is used to perform vector transformation on the text of the user input task, and then matched with feature vector libraries of various domains to identify one or more professional domains involved. The semantic parsing module uses dependency parsing and named entity recognition technology to extract the core entities, actions, and constraints in the task. The domain ontology is the conceptual hierarchy system of that domain. For example, the conceptual hierarchy system can include hardware and software sub-concepts in the computer domain.

[0025] The task template library stores historical decomposition cases of similar tasks. Combining these two resources, the original task is broken down into independent sub-tasks. Sub-tasks must cover all the core elements of the original task and have no overlap.

[0026] This embodiment uses natural language processing technology to parse task text, determines the professional domain through domain feature matching, and then, based on domain knowledge structure and historical task decomposition experience, breaks down complex tasks into independently executable subtasks, ensuring the integrity and independence of the subtasks. This improves the targeting of subsequent processing; subtask decomposition reduces task complexity, facilitates specialized processing, reduces information omissions, and provides clear objectives for agent allocation.

[0027] Step S102: Based on the set of subtasks, call the skill mapping table to obtain the correspondence between the professional skills and tool interfaces required by the subtasks, and combine the Agent behavior patterns and reasoning strategies in the role template library to generate specialized Agent instances and form an Agent set.

[0028] In some embodiments, the skill mapping table stores the mapping relationship between subtask types and required professional skills and tool interfaces in key-value pairs; the role template library contains the behavioral patterns and inference strategies of agents in various domains. Optionally, the inference strategy can be a combination of rule-based inference and probabilistic inference for agents in the medical field.

[0029] Based on the subtask, the matching skills and tool interfaces are obtained from the skill mapping table. The corresponding domain templates in the role template library are called to instantiate and generate an Agent with specific skills, tool call permissions and reasoning ability. Multiple Agents form an Agent set, and each Agent is responsible for handling one or more subtasks.

[0030] As can be seen, based on the skill requirements and tool dependencies of sub-tasks, corresponding resources and templates are matched to generate specialized Agent instances, enabling each Agent to handle specific sub-tasks and forming a collaborative processing system. Agents possess targeted skills and tool invocation capabilities, improving the professionalism and efficiency of sub-task processing; template-based generation ensures that Agent behavior conforms to domain specifications.

[0031] Step S103: Perform cross-domain knowledge retrieval and fusion for the Agent set.

[0032] In some embodiments, cross-domain knowledge retrieval and fusion are performed through a multi-layered knowledge routing engine. The first-layer domain knowledge retrieval engine retrieves domain-specific information from a professional knowledge base based on the agent's specialized skill requirements. The second-layer collaborative memory retrieval engine retrieves solutions to similar tasks from a historical collaborative case library. The knowledge weighted fusion engine uses a comprehensive weighting algorithm based on timeliness, relevance, and credibility to fuse the retrieval results from the two layers, generating a multi-source knowledge fusion result.

[0033] Specifically, the first-layer domain knowledge retrieval unit receives the agent's specialized skill requirements and retrieves domain-specific information from the professional knowledge base using semantic vector retrieval technology. The second-layer collaborative memory retrieval unit uses case attribute matching to retrieve task solutions with similarity ≥ a threshold from the historical collaborative case library. In the knowledge weighted fusion unit, timeliness is calculated by the difference between the information release time and the current time, relevance is calculated by semantic overlap, and credibility is calculated by the authority level of the information source. The three factors are combined to weight and merge the retrieval results from the two layers to generate a fused result.

[0034] It can be seen that by acquiring professional knowledge and historical processing experience in the field through hierarchical retrieval, and then balancing the value of multi-source information through weighted fusion, comprehensive and prioritized knowledge support can be provided to the Agent.

[0035] Step S104: Based on the results of multi-source knowledge fusion, monitor the output consistency of the Agent set, analyze the semantic consistency of the output results of different Agents in real time, and classify the types of collaboration conflicts according to the semantic consistency analysis results.

[0036] In some embodiments, the collaborative conflict intelligent mediator collects the output results of the Agent set in real time and uses a semantic similarity calculation method to analyze the consistency of the results; when the similarity is lower than the threshold, the conflict type is classified by rule base, the severity is determined by combining the criticality of the tasks involved in the conflict, and a conflict analysis report is generated.

[0037] This embodiment uses semantic analysis technology to monitor the consistency of Agent output, identifies the nature and impact of conflicts based on preset rules, promptly detects inconsistencies in Agent collaboration, and avoids erroneous output results.

[0038] Step S105: If a collaboration conflict is detected, trigger the expansion mechanism according to the conflict type and preset threshold. When the confidence of the Agent output result is lower than the preset threshold θ1, trigger vertical knowledge expansion to update the knowledge graph of the corresponding domain. When the collaboration conflict rate exceeds the preset threshold θ2, trigger horizontal Agent expansion to generate new specialized Agent instances. When the frequency of new concept recognition exceeds the preset threshold θ3, trigger knowledge graph update to supplement new concepts and their related relationships.

[0039] In some embodiments, preset thresholds θ1 (e.g., 0.6), θ2 (e.g., 30%), and θ3 (e.g., 5 times / hour) are used. When the confidence level of the Agent output is lower than θ1, the vertical knowledge expansion module calls the domain knowledge crawler to retrieve the latest literature from authoritative databases and extracts the deep information of the knowledge graph to update the knowledge. When the collaboration conflict rate exceeds θ2, the horizontal Agent expansion module generates new Agent instances in the same domain based on the role template library to supplement skill gaps. When the frequency of new concept recognition exceeds θ3, the knowledge graph update module parses the definition and context of the new concept, adds new nodes and their associations with existing nodes.

[0040] As can be seen, by comparing the confidence level, conflict rate, and frequency of new concepts with preset thresholds, corresponding expansion strategies are triggered. System resources are adjusted from three dimensions—knowledge depth, number of agents, and knowledge breadth—to resolve collaboration conflicts. This avoids overexpansion and resource waste; by adjusting knowledge and agents, the system's ability to handle complex tasks is improved, enhancing its adaptability.

[0041] Step S106: Monitor new domain knowledge updates, update the knowledge graph of the corresponding domain, calculate the comprehensive weight of timeliness, relevance and credibility of each knowledge fragment in the knowledge graph, and reload the updated knowledge weights and knowledge graph information.

[0042] In some embodiments, new domain knowledge is monitored through RSS subscriptions, API interface listening, etc.; entities and relationships in the new knowledge are added to the knowledge graph using entity linking and relationship extraction techniques, and the attributes of the original nodes are updated; by traversing each knowledge fragment in the knowledge graph, its timeliness, relevance, and credibility are recalculated, and a new weight value is obtained by combining the three; knowledge update notifications are sent to the Agent set through a message queue. After receiving the notification, the Agent reloads the knowledge weights and graph information, and sends an update instruction to the knowledge routing engine to adapt its retrieval strategy to the new knowledge weights.

[0043] This embodiment maintains the timeliness and accuracy of the knowledge graph, avoiding errors caused by the system relying on outdated knowledge; the adjustment of knowledge weights prioritizes the use of key information, improves the processing quality of the Agent and the knowledge routing engine, and enhances the long-term effectiveness of the system.

[0044] In one embodiment of the present invention, based on step S152, the following will provide a possible embodiment and its specific implementation will be described in a non-limiting manner. Step S102 specifically includes: Step S1021: For each subtask in the subtask set generated in step S101, parse the text description using natural language processing technology to extract the subtask's domain labels, functional requirements, and input / output requirements.

[0045] It should be noted that, to ensure accurate matching between the skill set and the requirements of the sub-tasks, an ontology-based semantic similarity algorithm is used to calculate the similarity S between the sub-task domain labels and the domain labels in the skill mapping table. S = |C_task|C_skill| |C_task|C_skill|.

[0046] Here, Task C is the set of domain concepts involved in the subtask, and Skill C is the set of domain concepts in the skill mapping table. A match is considered successful when S ≥ θ4.

[0047] Step S1022: Configure the association between domain tags and professional skills and tool interfaces in the skill mapping table. By querying the skill mapping table, obtain the set of professional skills and the list of tool interfaces that completely match the domain tags of the subtask.

[0048] Step S1023: Configure the Agent's behavior patterns and inference strategies by domain in the role template library, and select the closest combination of behavior patterns and inference strategies based on the functional requirements of the sub-tasks using a semantic matching algorithm.

[0049] Step S1024: The skill set and tool interface obtained in step S1022 are structurally encapsulated with the behavior patterns and inference strategies selected in step S1023. An independent Agent instance is generated for each subtask through the instantiation engine. Each instance includes: a skill execution module, a tool call interface, a behavior control module, and an inference engine.

[0050] Below is a specific example of an annual development plan for a company that includes market analysis, financial budgeting, and legal advice.

[0051] Step S1021: For the market analysis subtask decomposed from S101, natural language processing technology parses its text description and extracts the domain label "marketing." The functional requirements are to analyze the target market size, competitor situation, and consumer preferences. Input requirements include the company's existing market data and industry reports, and output requirements are a market share forecast table and competitive strategy suggestions. Then, the similarity between the domain label of this subtask and the domain label in the skill mapping table is calculated. Assuming task C is {market size, competitors, consumer preferences} and skill C is {market size, competitors, consumer behavior, market share}, then S = 0.75. If θ4 is set to 0.6, since 0.75 ≥ 0.6, the match is considered successful.

[0052] Step S1022: The professional skill set associated with the marketing domain tag in the skill mapping table includes market research skills, data statistical analysis skills, and competitive landscape analysis skills. The tool interface list includes market data query interfaces, statistical analysis software interfaces, and industry report database interfaces. By querying this table, the above-mentioned professional skill set and tool interface list matching the market analysis sub-task can be obtained.

[0053] Step S1023: The Agent behavior patterns in the marketing field of the role template library include regularly collecting market data, analyzing data trends, generating preliminary reports, and adjusting based on feedback. The reasoning strategies include trend prediction based on historical data and comparative analysis reasoning. Based on the functional requirements of analyzing target market size and providing strategic suggestions in the market analysis sub-task, the closest combination of behavior patterns and reasoning strategies is selected using a semantic matching algorithm; that is, the aforementioned behavior patterns and two reasoning strategies.

[0054] Step S1024: The market research skills and market data query interfaces obtained in step S1022 are structurally encapsulated with the behavioral patterns and inference strategies selected in step S1023, and the instantiation engine generates a market analysis agent instance. In this instance, the skill execution module is responsible for processing data using market research skills, the tool call interface connects to market data query tools, the behavior control module operates according to the pattern of collecting data - analyzing trends - generating reports - adjusting, and the inference engine uses trend prediction and comparative analysis inference methods for analysis.

[0055] By using precise semantic parsing and similarity matching, the accuracy of matching subtasks with professional skills and tool interfaces is ensured, avoiding the problem of mismatch between skills and task requirements. Based on the role template library, suitable behavioral patterns and reasoning strategies are selected, enabling the generated Agent instances to have working methods that conform to the characteristics of the domain, thereby improving the professionalism and efficiency of task processing.

[0056] In one embodiment of the present invention, based on step S103, the following will provide a possible embodiment and its specific implementation will be described in a non-limiting manner. Step S103 specifically includes: Step S1031: Collect the first-layer domain knowledge retrieval results and the second-layer collaborative memory retrieval results respectively, and extract the timeliness marker, relevance tag, and credibility identifier for each knowledge.

[0057] Step S1032: According to the preset mapping rules, convert the timeliness mark into a timeliness quantification value; convert the relevance label into a relevance quantification value; and convert the credibility identifier into a credibility quantification value.

[0058] Step S1033: For each piece of knowledge, calculate the fusion weight of the single piece of knowledge according to the formula: timeliness quantification value × α + relevance quantification value × β + credibility quantification value × γ, and mark the weight result. α, β, and γ are predefined weight allocation coefficients, and α + β + γ = 1.

[0059] Step S1034: Sort the knowledge fusion weights from high to low, select the knowledge with weights reaching the threshold, aggregate them according to the sub-task logic, and generate the knowledge fusion result.

[0060] This embodiment quantifies and comprehensively calculates the weights of the timeliness, relevance, and credibility of knowledge, enabling a more scientific evaluation of the value of each piece of knowledge and ensuring that the selected knowledge better meets task requirements. Sorting and selecting qualified knowledge according to weight ensures that the final fusion result contains high-value knowledge, improving the effectiveness of the knowledge. Aggregating knowledge according to sub-task logic allows the fused knowledge to closely revolve around the sub-tasks, enhancing the correlation between knowledge and sub-tasks and improving the efficiency and quality of task processing.

[0061] In one embodiment of the present invention, based on step S104, the following will provide a possible embodiment and its specific implementation will be described in a non-limiting manner. Step S104 specifically includes: Step S1041: Collect the output results of each specialized Agent instance in step S103, and convert them into a unified intermediate representation format, including fields such as task ID, Agent ID, output content, timestamp, and associated subtask identifier, to ensure that the outputs from different sources can be compared.

[0062] Step S1042: For each standardized output content, extract key semantic elements using natural language processing technology.

[0063] For example, semantic feature extraction can be achieved by encoding the output content using a pre-trained BERT language model to capture implicit semantic information. For instance, "patient age 30" and "patient age thirty" would be identified as the same entity value; "drug A is suitable for type 1 diabetes" and "drug A is used to treat type 1 diabetes" would be identified as the same intent.

[0064] Step S1043: Based on the semantic vector generated in step S1042, the consistency of outputs from different agents under the same task is calculated using the following method: Calculate the cosine similarity of the semantic vectors output by any two agents; Calculate the proportion of exact matches of key entities in the two outputs to the total number of entities; Determine whether the intent descriptions of the two outputs belong to the same task sub-goal.

[0065] By combining the scores from the three dimensions mentioned above, an overall consistency score is generated for all Agent outputs, ranging from 0 to 1, with higher values ​​indicating lower conflict risk.

[0066] Step S1044: Preset conflict classification rule base, including the following rules: Based on semantic feature differences, they are divided into data inconsistency, logical inconsistency, and goal inconsistency. Severity determination: Based on consistency score and conflict type, a threshold is set: High severity: Consistency score < 0.3, or contains goal inconsistency; Medium severity: 0.3 ≤ consistency score < 0.6, and it is either data inconsistency or logical inconsistency; Low severity: Consistency score ≥ 0.6, or only minor differences in entity values.

[0067] Since different agents use different output formats, a format conversion module is needed to unify them into an intermediate representation to eliminate the interference of format differences on conflict detection. For example, the output of patient data compliance may be presented as GDPR compliance: text or {'gdpr_compliant':True}. After standardization, both are converted into structured objects containing fields such as compliance status and verification basis.

[0068] It should be noted that inconsistency in this embodiment refers to differences or contradictions in semantics, logic, or objectives between information from different sources in the output of multi-agent collaboration or knowledge fusion results. This is an unavoidable phenomenon in the process of multi-source knowledge fusion and a problem that the system needs to monitor and resolve. Data inconsistency points to ambiguity in entity definitions within the knowledge graph; logical inconsistency exposes defects in agent behavior patterns or reasoning strategies; and objective inconsistency reflects deviations in agent behavior patterns within the role template library.

[0069] Step S104 unifies the intermediate representation format, solving the comparison difficulties caused by differences in the output formats of different agents, ensuring the accuracy of conflict detection, extracting key semantic elements to capture deep semantics, such as identifying the same entity or intent in different expressions; multi-dimensional calculation of consistency scores makes conflict assessment more comprehensive and reduces the bias of single-dimensional judgment; classifying conflict types and severity according to the rule base provides a clear basis for subsequent targeted conflict resolution, facilitates rapid problem location and action, and improves the coordination of multi-agent collaboration and the reliability of task processing.

[0070] In one embodiment of the present invention, based on step S105, the following will provide a possible embodiment and its specific implementation will be described in a non-limiting manner. Step S105 specifically includes: Step S1051: Collect the conflict type, severity, associated AgentID, and conflict content information output in step S104, convert them into relevant data, and form an inconsistency event database to facilitate subsequent matching and expansion strategies. The relevant data includes JSON objects containing conflict ID, type label, severity level, associated subtask ID, and conflict description fields.

[0071] Step S1052: Based on the structured conflict events in step S1051, call the preset extended strategy rule base for matching: If the conflict type is data inconsistency and the severity level is high, match the vertical knowledge expansion strategy; If the conflict type is logical inconsistency and the severity level is medium, match the horizontal agent expansion strategy; If the conflict type is goal inconsistency or the frequency of new concept recognition exceeds θ3, match the knowledge graph update strategy; If multiple types of conflicts are met simultaneously, the primary expansion strategy is selected based on priority: data inconsistency > logical inconsistency > objective inconsistency.

[0072] Step S1053: Perform vertical and horizontal expansion operations according to the strategy matched in step S1052.

[0073] Specifically, vertical knowledge expansion addresses data inconsistency by obtaining the latest definitions or revised rules for conflicting entities from authoritative knowledge sources and adding the new rules to the knowledge graph of the corresponding domain. Horizontal agent expansion addresses logical inconsistencies by extracting new skills involved in the conflict from the skill mapping table and combining them with the behavioral patterns of verification agents in the role template library to generate new specialized agent instances. Knowledge graph updates involve identifying new concepts, adding the new concepts and their relationships to the knowledge graph, and supplementing related entities.

[0074] Step S1054: Verify the expanded knowledge graph and the newly added Agent instance.

[0075] Specifically, knowledge graph verification involves verifying, through domain experts or historical case libraries, whether the updated knowledge fragments resolve the original conflicts. Agent instance verification is a simulation of task scenarios to test the effectiveness of newly added agents' skills; After successful verification, the skill mapping table update is pushed to the Agent set in step S102, and the knowledge graph version number and newly added Agent information are sent to the knowledge routing engine in step S103 to ensure that subsequent tasks call the latest resources.

[0076] This embodiment structures conflict information into an event database, facilitating rapid querying and matching of extension strategies, thereby improving the efficiency of conflict handling. Based on a rule base and priority matching of extension strategies, it ensures that the most suitable solution is adopted for different conflicts, avoiding strategy misuse. For example, high-severity data inconsistencies are prioritized for resolution through vertical knowledge expansion. Specific extension operations are highly targeted: vertical expansion supplements knowledge to resolve data inconsistencies, horizontal expansion adds agents to resolve logical inconsistencies, and updating the knowledge graph can address target inconsistencies and new concept issues. The verification process ensures the effectiveness of the expanded resources, preventing invalid expansion from affecting system operation. Pushing update information ensures that relevant modules use the new resources synchronously, maintaining the stability of multi-agent collaboration and improving the continuity and accuracy of task processing.

[0077] In one embodiment of the present invention, based on step S106, the following will provide a possible embodiment and its specific implementation will be described in a non-limiting manner. Step S106 specifically includes: Step S1061: When a knowledge base is added, modified, or deleted, the knowledge base pushes a knowledge update event, which includes information such as update type, update content, and update timestamp. Step S1062: Based on the updated content detected in step S1061, modify the knowledge graph to ensure the atomicity and traceability of the update operation: If the updated content is a new concept that has not been included in the knowledge graph, a new node will be created in the knowledge graph, and the type, definition, source and creation time will be labeled. If the updated content is a new association of an existing concept, then add a bidirectional edge to the knowledge graph and label the association strength and basis. If the update involves adjusting the attributes of an existing node, then modify the attribute values ​​of the corresponding node and retain the historical version record.

[0078] Step S1063: Based on the updated knowledge graph, recalculate the comprehensive weights of timeliness, relevance, and credibility for each knowledge fragment to ensure that the weights reflect the current value of the knowledge. Specifically, based on the difference Δt between the last update time and the current time of the knowledge fragment, it is divided into four levels; Scores are calculated based on the degree of matching between knowledge fragments and the current task domain, using predefined domain label matching rules. Based on the authority of the source of the knowledge fragment and the results of historical verification, a weighted average of the two is taken; The scores of the above dimensions are weighted and summed according to timeliness α, relevance β, and credibility γ to obtain the comprehensive weight of the knowledge fragment W=α⋅T+β⋅R+γ⋅C.

[0079] Step S1064: Synchronize the updated content and weight changes of the knowledge graph to the Agent set in step S102 and step S103 in the following manner.

[0080] Specifically, a message queue is used to send knowledge graph update events. The event content includes information such as the knowledge graph version number, the ID of the newly added / modified knowledge fragment, the change value of the comprehensive weight, and the update timestamp. Each Agent instance in the Agent set listens to the message queue. Upon receiving an update event, it updates its local knowledge graph cache to ensure that the latest knowledge is used when subsequent tasks are executed. Upon receiving an update event, refresh its knowledge retrieval index and reload the weights of the latest knowledge fragments to ensure that the retrieval results for new tasks are based on the latest knowledge.

[0081] As can be seen, step S106 receives knowledge base update events in real time, which can promptly capture knowledge changes and avoid the system relying on outdated knowledge; targeted modifications to the knowledge graph ensure that new knowledge is accurately integrated into the existing system, and retain historical versions for easy traceability, such as adding concept creation nodes and adding basis for labeling related edges; recalculating the comprehensive weights ensures that the knowledge value assessment is consistent with the current state, and ensures that important knowledge is used first; and updating information synchronously through the message queue ensures that the Agent set and knowledge routing engine load the latest knowledge and weights in a timely manner, maintains the consistency of knowledge in each module of the system, improves the accuracy of multi-Agent collaboration and the reliability of task processing, and enables the system to continuously adapt to changes in domain knowledge.

[0082] The technical solution of the present invention will be described in detail below with reference to specific embodiments: Example 1: Performing medical compliance consultation.

[0083] Step 1: Task Input and Decomposition User input task: "Develop personalized diabetes treatment plans that meet GDPR requirements for patients in a certain region".

[0084] Task Adaptive Decomposer Workflow: The domain classification module identifies two main domains: "medical diagnostics" and "data compliance". The semantic parsing module extracts key entities: [a certain region, GDPR, diabetes, personalized treatment]; Subtask generation module outputs: T1="Diabetes treatment plan development", T2="Patient data GDPR compliance review".

[0085] Step 2: The agent builder constructs a domain-aware agent based on the subtask requirements: For T1 mapping skills: [Endocrinology expertise, drug interaction analysis, patient data interpretation], generate "Endocrine Therapy Expert Agent"; For T2 mapping skills: [GDPR interpretation, data anonymization technology, cross-border data transfer compliance], generate a "Data Compliance Review Agent".

[0086] Step 3: Knowledge Retrieval and Fusion – Multi-layered Knowledge Routing Engine Execution: First-level search: The endocrine therapy expert agent searches the medical literature database for the latest diabetes treatment guidelines; Second-level retrieval: Retrieve historical cases of "medical data processing in a certain region" from the collaborative memory; Knowledge weighted fusion: Calculate the weight of each knowledge source and prioritize the use of information with high weight to generate a preliminary scheme.

[0087] Step 4: Conflict Detection and Resolution. The collaborative conflict intelligent mediator discovered that the genetic testing data included in the treatment plan conflicted with GDPR sensitive data processing requirements.

[0088] Conflict type identification: Data privacy versus healthcare needs; Extension: Generate a "Medical Data Ethics Coordination Agent"; Conflict resolution: Redesign the data processing workflow and use differential privacy technology to process genetic data.

[0089] Example 2: System adaptive optimization method.

[0090] When the system detects the release of a new medical guideline during operation, the extended module: Update the medical knowledge graph and add new diagnosis and treatment nodes; Recalculate the timeliness weight of relevant knowledge; Notify the relevant Agents to reload the updated knowledge weights.

[0091] As can be seen, through the task-driven agent generation mechanism, the system can construct specialized agent combinations based on task characteristics, avoiding the problem of fixed roles in traditional systems. Through real-time monitoring and expansion, the system can identify and resolve collaboration conflicts between agents, improving system stability and reliability. It also features knowledge base updates and agent capability expansion capabilities, maintaining continuous system optimization and adaptability.

[0092] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0093] The following are embodiments of a multi-domain task processing system based on knowledge fusion and agent collaboration provided in this disclosure. This system and the multi-domain task processing methods based on knowledge fusion and agent collaboration in the above embodiments belong to the same inventive concept. For details not described in detail in the embodiments of the multi-domain task processing system based on knowledge fusion and agent collaboration, please refer to the embodiments of the multi-domain task processing methods based on knowledge fusion and agent collaboration described above.

[0094] like Figure 2 As shown, the system includes: The task identification module 201 is used to perform domain identification and subtask decomposition on the task input by the user, identify the semantic elements involved in the task, and generate a set of subtasks.

[0095] The instantiation module 202 is used to obtain the correspondence between the professional skills and tool interfaces required by the subtasks by calling the skill mapping table based on the subtask set, and generate professional agent instances by combining the agent behavior patterns and inference strategies in the role template library to form an agent set.

[0096] The cross-domain fusion module 203 is used to perform cross-domain knowledge retrieval and fusion for a set of Agents.

[0097] The collaboration monitoring module 204 monitors the output consistency of the Agent set based on the results of multi-source knowledge fusion, analyzes the semantic consistency of the output results of different Agents in real time, and classifies the types of collaboration conflicts according to the semantic consistency analysis results.

[0098] The extended trigger module 205 is used to trigger vertical knowledge expansion to update the knowledge graph of the corresponding domain when a collaboration conflict is detected and the confidence level of the Agent output result is lower than the first preset threshold.

[0099] When the collaboration conflict rate exceeds the second preset threshold, horizontal agent expansion is triggered to generate new specialized agent instances.

[0100] When the frequency of new concept recognition exceeds the third preset threshold, the knowledge graph is updated to supplement new concepts and their related relationships.

[0101] The knowledge graph continuous update module 206 is used to monitor new domain knowledge updates, update the knowledge graph of the corresponding domain, calculate the comprehensive weight of the timeliness, relevance and credibility of each knowledge fragment in the knowledge graph, and reload the updated knowledge weights and knowledge graph information.

[0102] like Figure 3 As shown, this application also provides an electronic device, including a display module 103, a memory 102, a processor 101, and a computer program stored in the memory and executable on the processor 101. When the processor 101 executes the program, it implements the steps of the method for parsing and loading a GIM model of a power transmission project.

[0103] In embodiments of the present invention, electronic devices include, but are not limited to, laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the embodiments described and / or claimed herein.

[0104] In this embodiment, processor 101 may be implemented using at least one of an application-specific integrated circuit, a programmable logic device, a field-programmable gate array, a processor, a controller, a microcontroller, a microprocessor, or an electronic unit designed to perform the functions described herein. In some cases, such an implementation may be implemented within a controller. For software implementation, implementations such as processes or functions may be implemented with separate software modules that allow the performance of at least one function or operation. Software code may be implemented by a software application (or program) written in any suitable programming language, and the software code may be stored in memory and executed by the controller.

[0105] The display module 103 is used to display information input by the user or information provided to the user. The display module 103 may include a display panel, which may be configured in the form of a liquid crystal display, an organic light-emitting diode, or the like.

[0106] The memory 102 can be used to store software programs and various data. The memory 102 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0107] This application also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the multi-domain task processing method based on knowledge fusion and agent collaboration.

[0108] The storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example,, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0109] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-domain task processing method based on knowledge fusion and agent collaboration, characterized in that, The methods include: S101: Perform domain identification and subtask decomposition on the user-input task, identify the semantic elements involved in the task, and generate a set of subtasks; S102: Based on the set of subtasks, call the skill mapping table to obtain the correspondence between the professional skills and tool interfaces required by the subtasks, and combine the Agent behavior patterns and reasoning strategies in the role template library to generate professional Agent instances and form an Agent set; S103: Perform cross-domain knowledge retrieval and fusion for the Agent set; S104: Based on the results of multi-source knowledge fusion, monitor the output consistency of the Agent set, analyze the semantic consistency of the output results of different Agents in real time, and classify the types of collaboration conflicts according to the semantic consistency analysis results; S105: If a collaboration conflict is detected and the confidence level of the Agent output result is lower than the first preset threshold, vertical knowledge expansion is triggered to update the knowledge graph of the corresponding domain. When the collaboration conflict rate exceeds the second preset threshold, horizontal agent expansion is triggered to generate new specialized agent instances; When the frequency of new concept recognition exceeds the third preset threshold, the knowledge graph is updated to supplement new concepts and their related relationships. S106: Monitor new domain knowledge updates, update the knowledge graph of the corresponding domain, calculate the comprehensive weight of the timeliness, relevance and credibility of each knowledge fragment in the knowledge graph, and reload the updated knowledge weights and knowledge graph information.

2. The multi-domain task processing method based on knowledge fusion and agent collaboration according to claim 1, characterized in that, The execution of cross-domain knowledge retrieval and fusion in step S103 includes: a first-layer domain knowledge retrieval engine retrieves domain-specific information from a professional knowledge base based on the agent's specialized skill requirements; a second-layer collaborative memory retrieval engine retrieves solutions to similar tasks from a historical collaborative case library; and a knowledge weighted fusion engine uses a comprehensive weighting algorithm based on timeliness, relevance, and credibility to fuse the retrieval results from the two layers, generating a multi-source knowledge fusion result.

3. The multi-domain task processing method based on knowledge fusion and agent collaboration according to claim 1, characterized in that, Step S102 specifically includes: For each subtask in the subtask set generated in step S101, the text description is parsed using natural language processing technology to extract the subtask's domain labels, functional requirements, and input / output requirements. Configure the association between domain tags and professional skills and tool interfaces in the skill mapping table. By querying the skill mapping table, obtain the set of professional skills and the list of tool interfaces that perfectly match the domain tags of the subtask. In the role template library, the behavioral patterns and inference strategies of the Agent are configured according to the domain. Based on the functional requirements of the sub-task, the closest combination of behavioral patterns and inference strategies is selected through semantic matching algorithm. The acquired skill sets, tool interfaces, and filtered behavioral patterns and inference strategies are encapsulated in a structured manner, and an independent Agent instance is generated for each subtask through an instantiation engine.

4. The multi-domain task processing method based on knowledge fusion and agent collaboration according to claim 2, characterized in that, Step S103 specifically includes: The first-layer domain knowledge retrieval results and the second-layer collaborative memory retrieval results are collected respectively, and the timeliness marker, relevance tag, and credibility identifier of each knowledge are extracted. Based on preset mapping rules, timeliness markers are converted into timeliness quantification values; relevance tags are converted into relevance quantification values; and credibility identifiers are converted into credibility quantification values. For each piece of knowledge, the fusion weight of a single piece of knowledge is calculated according to the method of timeliness quantification value × α + relevance quantification value × β + credibility quantification value × γ, and the weight result is marked; α, β, and γ are predefined weight allocation coefficients, and α + β + γ = 1; Sort the knowledge fusion results from highest to lowest according to their weight, select the knowledge whose weight reaches the threshold, aggregate them according to the sub-task logic, and generate the knowledge fusion results.

5. The multi-domain task processing method based on knowledge fusion and agent collaboration according to claim 1, characterized in that, Step S104 specifically includes: Collect the output results of each specialized Agent instance in step S103 and convert them into a unified intermediate representation format; For each standardized output, natural language processing techniques are used to extract key semantic elements; Based on the generated semantic vectors, the consistency of outputs from different agents under the same task is calculated using the following method: Calculate the cosine similarity of the semantic vectors output by any two agents; Calculate the proportion of exact matches of key entities in the two outputs to the total number of entities; Determine whether the intent descriptions of the two outputs belong to the same task sub-goal; By combining the scores from the above three dimensions, an overall consistency score is generated for all Agent outputs; A pre-defined conflict classification rule base is established, which categorizes conflicts into data inconsistency, logical inconsistency, and objective inconsistency based on differences in semantic features.

6. The multi-domain task processing method based on knowledge fusion and agent collaboration according to claim 5, characterized in that, Step S105 specifically includes: Collect the conflict type, severity, associated AgentID, and conflict content information output in step S104, convert them into relevant data, and form an inconsistency event database; Based on the conflict event, the preset extended policy rule base is invoked for matching: If the conflict type is data inconsistency and the severity level is high, match the vertical knowledge expansion strategy; If the conflict type is logical inconsistency and the severity level is medium, match the horizontal agent expansion strategy; If the conflict type is goal inconsistency or the frequency of new concept recognition exceeds θ3, match the knowledge graph update strategy; If multiple types of conflicts are met simultaneously, the main expansion strategy is selected in descending order of priority: data inconsistency, logical inconsistency, and goal inconsistency. Based on the matching strategy, perform vertical and horizontal expansion operations; verify the expanded knowledge graph and the newly added Agent instances.

7. The multi-domain task processing method based on knowledge fusion and agent collaboration according to claim 1, characterized in that, Step S106 specifically includes: When a knowledge base is updated, modified, or deleted, the knowledge base pushes a knowledge update event. Based on the monitored updates, the knowledge graph is modified to ensure the atomicity and traceability of the update operations: If the updated content is a new concept that has not been included in the knowledge graph, a new node will be created in the knowledge graph, and the type, definition, source and creation time will be labeled. If the updated content is a new association of an existing concept, then add a bidirectional edge to the knowledge graph and label the association strength and basis. If the update involves adjusting the attributes of an existing node, then modify the attribute values ​​of the corresponding node and retain the historical version record; Based on the updated knowledge graph, the comprehensive weights of timeliness, relevance, and credibility are recalculated for each knowledge fragment; The updated content and weight changes of the knowledge graph are synchronized to the Agent set in step S102 and step S103.

8. A multi-domain task processing system based on knowledge fusion and agent collaboration, characterized in that, The system is used to implement the multi-domain task processing method based on knowledge fusion and agent collaboration as described in any one of claims 1 to 7; The system includes: The task recognition module is used to perform domain recognition and subtask decomposition on the task input by the user, identify the semantic elements involved in the task, and generate a set of subtasks. The instantiation module is used to obtain the correspondence between the professional skills and tool interfaces required by the subtasks by calling the skill mapping table based on the subtask set, and generate professional agent instances by combining the agent behavior patterns and inference strategies in the role template library to form an agent set. The cross-domain fusion module is used to perform cross-domain knowledge retrieval and fusion for a set of Agents; The collaboration monitoring module monitors the output consistency of the Agent set based on the results of multi-source knowledge fusion, analyzes the semantic consistency of the output results of different Agents in real time, and classifies the types of collaboration conflicts according to the semantic consistency analysis results. An extended triggering module is used to trigger vertical knowledge expansion to update the knowledge graph of the corresponding domain when a collaboration conflict is detected and the confidence level of the Agent output result is lower than the first preset threshold. When the collaboration conflict rate exceeds the second preset threshold, horizontal agent expansion is triggered to generate new specialized agent instances; When the frequency of new concept recognition exceeds the third preset threshold, the knowledge graph is updated to supplement new concepts and their related relationships. The knowledge graph continuous update module is used to monitor new domain knowledge updates, update the corresponding domain knowledge graph, calculate the comprehensive weight of the timeliness, relevance and credibility of each knowledge fragment in the knowledge graph, and reload the updated knowledge weights and knowledge graph information.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the multi-domain task processing method based on knowledge fusion and agent collaboration as described in any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the multi-domain task processing method based on knowledge fusion and agent collaboration as described in any one of claims 1 to 7.

Images