A human-computer collaborative intelligence agent system and method based on interactive feedback

Abstract

The application discloses a kind of human-computer collaborative intelligence agent systems and methods based on interaction feedback, it is related to artificial intelligence relevant field, wherein the method includes: receiving the intelligence analysis demand of user end input, through intention analysis, extraction analysis intention element;Generation task execution tree, through interactive interface and user collaborative adjustment task execution tree;Two-stage matching routing decision based on node task characteristics is carried out in agent cluster, determine intelligence agent task cluster and carry out parallel tree task analysis, obtain task analysis result and carry out integration, determine structured intelligence report and build evidence traceability chain;Structural intelligence report and evidence traceability chain are displayed to terminal interface.This application solves the technical problem of imbalance between efficiency and accuracy existing in the prior art intelligence analysis field, achieves intelligence analysis efficiency improvement, analysis result accuracy and reliability enhancement, realizes the technical effect of dynamic balance between efficiency and accuracy.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a human-machine collaborative intelligence agent system and method based on interactive feedback. Background Technology

[0002] Intelligence analysis, as a core component supporting critical decision-making in fields such as science and technology, directly impacts decision quality and response timeliness due to its accuracy and efficiency. It plays an irreplaceable role in reducing decision-making risks and seizing development opportunities. Current mainstream technologies for solving intelligence analysis problems include traditional manual analysis systems, basic AI-assisted systems, direct application of large language models, and intelligent agent systems. These manifest as relying on analysts manually retrieving and processing data, automating some preprocessing work through natural language processing technology, responding to natural language queries based on general-purpose large language models, and autonomously decomposing and executing tasks based on intelligent agent architectures. However, existing methods suffer from inherent limitations, such as an inability to effectively balance efficiency and quality, a lack of interpretability and error correction mechanisms in AI models, difficulty in accumulating tacit knowledge from human experts, and insufficient ability to decompose and collaborate on complex tasks. These limitations lead to problems such as low efficiency in traditional manual analysis, the potential for factual errors in pure AI solutions, the inability to effectively transfer expert experience, and the difficulty in deeply decomposing complex intelligence tasks.

[0003] Currently, there is a technical problem in the field of intelligence analysis where efficiency and accuracy are at odds. Summary of the Invention

[0004] This application provides a human-machine collaborative intelligence agent system and method based on interactive feedback. It receives user-input intelligence analysis requests and extracts core elements through intent parsing. A task decomposition component generates a task execution tree containing main and sub-task nodes. This tree is then adjusted and optimized collaboratively with the user through an interactive interface. An intelligent routing component determines a suitable agent task cluster from the agent cluster based on node task characteristics through two-order matching routing. Parallel tree task analysis is performed, and the results are integrated to generate a structured intelligence report with confidence scores and a corresponding evidence tracing chain. This technology, including presenting the report and evidence tracing chain to the user through a terminal interface, solves the technical problem of efficiency and accuracy imbalance in existing intelligence analysis fields. It achieves the technical effect of improving intelligence analysis efficiency, enhancing the accuracy and credibility of analysis results, and realizing a dynamic balance between efficiency and accuracy.

[0005] This application provides a human-machine collaborative intelligence agent system based on interactive feedback, comprising: an intent parsing module for receiving intelligence analysis requests input by a user terminal and extracting analysis intent elements through intent parsing; a task execution tree generation module for a task decomposition component to generate a task execution tree based on the analysis intent elements, wherein the task execution tree includes a main task node and multiple sub-task nodes, and the task execution tree is adjusted collaboratively with the user through an interactive interface; a parallel tree task analysis module for an intelligent routing component to receive the task execution tree, perform two-order matching routing decisions based on node task characteristics in the agent cluster, determine the agent task cluster and perform parallel tree task analysis, obtain task analysis results and integrate them, determine a structured intelligence report and construct an evidence tracing chain, wherein the task analysis results are marked with a confidence score; and a terminal interface display module for displaying the structured intelligence report and evidence tracing chain on a terminal interface.

[0006] In a possible implementation, the task execution tree generation module includes: a target template matching module, used to match a target template from a preset analysis framework template library based on the analysis intent elements by identifying the analysis object, analysis dimension, and analysis target; an initial task tree generation module, used to allocate main task nodes and several sub-task nodes to the analysis intent elements according to the target template to generate an initial task tree, wherein each task node is identified with the required data source, analysis method, expected output format, and confidence level requirement; and an adjustment instruction receiving module, used to display the initial task tree through a visual interface and receive adjustment instructions from the user side, wherein the adjustment dimension includes at least one of node structure, execution order, and analysis parameters.

[0007] In a possible implementation, the parallel tree task analysis module includes: an agent cluster building module for building an agent cluster, wherein the agent cluster is composed of functional agents based on different intelligence analysis subdomains; an intelligent routing component deployment module for deploying intelligent routing components according to the agent cluster, wherein first-order matching routing and second-order cooperative routing are used as the underlying logic; a task execution tree matching module for matching the task execution tree based on node task features according to the intelligent routing component to determine the agent task cluster; and a task execution tree processing module for calling the agent task cluster to process the task execution tree and determine the task analysis result.

[0008] In a possible implementation, the task execution tree matching module includes: a node task feature determination module, used to determine the node task features of each task node based on the task execution tree, wherein the node task features include at least data type, analysis method, and domain features; a first-order matching routing module, used to traverse the node task features of each task node, perform first-order matching routing, and determine task agents, wherein each task node corresponds to at least one task agent, and each task agent is associated with at least one task node; and a second-order routing coordination module, used to perform second-order routing coordination between task agents based on information transmission, collaborative reasoning, and iterative optimization for the node task features and the task agents, and determine the agent task cluster.

[0009] In a possible implementation, the parallel tree task analysis module includes: a multi-level confidence scoring system establishment module, used to establish a multi-level confidence scoring system, wherein the multi-level confidence scoring system at least covers the data source layer, inference path layer, agent output layer, and conclusion layer; a confidence model establishment module, used to establish a confidence model for dynamic confidence calibration based on the multi-level confidence scoring system as a framework and using Bayesian update rules; and a confidence score analysis module, used to analyze the stage-by-stage confidence scores based on the confidence model during the task analysis process.

[0010] In a possible implementation, the confidence score analysis module includes a manual intervention process triggering module, which is used to trigger a manual intervention process during the task analysis process if at least one of the following conditions is met: the confidence score is lower than a preset threshold or an analysis conflict is detected. The module receives user-end adjudication data and corrects the task execution tree.

[0011] In a possible implementation, the parallel tree task analysis module includes: a data fingerprint generation module, used to generate a data fingerprint for each piece of original data, wherein the data fingerprint includes at least a source identifier, a timestamp, and a content hash value; an evidence tracing chain construction module, used to construct the evidence tracing chain by starting with the original data, determining multiple intermediate quantities through reasoning steps, and ending with the conclusion through directed graph association; and an interactive evidence aiming point setting module, used to locate key conclusions in the structured intelligence report and set interactive evidence aiming points, wherein the structured intelligence report and the evidence tracing chain support forward and reverse tracing, and the conclusion is traced back in reverse based on the interactive evidence aiming points.

[0012] In a possible implementation, the parallel tree task analysis module includes: a confidence score adjustment module, used to receive user feedback data, perform hierarchical matching and confidence score adjustment in the multi-layer confidence score system, and determine confidence correction data; and an intelligence confidence graph update module, used to adjust the confidence of relevant conclusions in the structured intelligence report along the dependency graph based on the evidence chain, and update the intelligence confidence graph, wherein the intelligence confidence graph is dynamically updated.

[0013] In a possible implementation, the parallel tree task analysis module includes: a report generation feedback receiving module, used to receive report generation feedback from the user, wherein the report generation feedback includes direct evaluation, content modification, supplementary evidence, or error type labeling; an implicit behavior data mining module, used to mine implicit behavior data from the user, wherein the implicit behavior data includes browsing duration, frequency of repeated visits, attention to specific intelligence categories, and modification patterns during task execution; and an update learning module, used to update and learn the multi-layer confidence scoring system and the agent cluster based on the report generation feedback and implicit behavior data.

[0014] This application also provides a human-machine collaborative intelligence agent method based on interactive feedback, comprising: receiving intelligence analysis requirements input by a user terminal, extracting analysis intent elements through intent parsing; a task decomposition component generating a task execution tree based on the analysis intent elements, wherein the task execution tree includes a main task node and multiple sub-task nodes, and adjusting the task execution tree collaboratively with the user through an interactive interface; an intelligent routing component receiving the task execution tree, performing two-order matching routing decisions based on node task characteristics in the agent cluster, determining the agent task cluster and performing parallel tree task analysis, obtaining and integrating the task analysis results, determining a structured intelligence report and constructing an evidence tracing chain, wherein the task analysis results are marked with a confidence score; and displaying the structured intelligence report and evidence tracing chain on a terminal interface.

[0015] This application proposes a human-machine collaborative intelligence agent system and method based on interactive feedback. First, it receives intelligence analysis requests from a user. Through intent parsing, it extracts analysis intent elements. Next, a task decomposition component generates a task execution tree based on these elements. This tree includes a main task node and multiple sub-task nodes. The user and the system collaboratively adjust the task execution tree through an interactive interface. Then, an intelligent routing component receives the tree and performs two-order matching routing decisions based on node task characteristics within the agent cluster. This determines the agent task cluster and performs parallel tree task analysis, obtaining and integrating the analysis results to determine a structured intelligence report and construct an evidence tracing chain. The analysis results are marked with a confidence score. Finally, the structured intelligence report and evidence tracing chain are displayed on a terminal interface. Through this process, the system and method proposed in this application achieve the technical effects of improved intelligence analysis efficiency, enhanced accuracy and credibility of analysis results, and a dynamic balance between efficiency and accuracy. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 This is a schematic diagram of the structure of a human-machine collaborative intelligence system based on interactive feedback, provided in an embodiment of this application.

[0018] Figure 2 This is a flowchart illustrating a human-machine collaborative intelligence agent method based on interactive feedback, provided in an embodiment of this application.

[0019] Figure labeling: Intent parsing module 10, task execution tree generation module 20, parallel tree task analysis module 30, terminal interface display module 40. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] This application provides a human-machine collaborative intelligence agent system based on interactive feedback, such as... Figure 1 As shown, the system includes: The intent parsing module 10 is used to receive intelligence analysis requests input from the user terminal and extract analysis intent elements through intent parsing.

[0022] Specifically, the Natural Language Intent Understanding component performs preprocessing operations such as word segmentation, stop word removal, and part-of-speech tagging on the intelligence analysis requests input by the user. For example, using the Jieba word segmentation tool, the phrase "analyze the technological development trend of a certain country in the semiconductor field and its impact on my country" is segmented to obtain words such as "analysis," "a certain country," "semiconductor field," "technological development trend," "my country," and "impact." A BERT-based fine-tuning model is used for semantic understanding. The input of this model is the preprocessed text sequence, and the output is a structured semantic vector. During training, it is fine-tuned using annotated corpus from the intelligence domain, containing different types of intelligence analysis requests and corresponding element annotations. The named entity recognition module identifies intent elements such as the analysis object, analysis dimension, and analysis goal. For example, from the above request, the analysis object is extracted as "a certain country," "semiconductor field," and "my country," the analysis dimension is "technological development trend," and the analysis goal is "to judge the impact on my country." Finally, the extracted elements are validated using a domain rule base to ensure their completeness and accuracy. For example, if no analysis dimension is extracted, the user is prompted to provide relevant information.

[0023] The task execution tree generation module 20 is used by the task decomposition component to generate a task execution tree based on the analysis intent elements. The task execution tree includes a main task node and multiple sub-task nodes. The task execution tree can be adjusted collaboratively with the user through an interactive interface.

[0024] Specifically, an intelligence analysis framework template library is constructed, including various common templates such as technology trend assessment and risk warning. Each template defines the hierarchical relationship between main tasks and sub-tasks, the data source requirements for each node, the analysis method, the expected output format, and the confidence threshold. For example, the technology trend assessment template includes main task nodes such as "Technology Status Analysis," "Core Enterprise Layout," and "Future Development Forecast," and each main task includes sub-task nodes such as "Patent Application Statistics" and "Technology Route Analysis." A template matching algorithm is used to calculate the similarity between the extracted analytical intent elements and each template. The similarity calculation uses the cosine similarity formula, comparing the semantic vector of the analytical intent element with the template feature vector. A matching threshold is set, and when the similarity is higher than this threshold, the template with the highest similarity is selected as the target template. Based on the target template, a tree structure generation algorithm is used to assign analytical intent elements to main task nodes and sub-task nodes. Each node is identified with the required data source, analytical method, expected output format, and confidence level requirement, generating an initial task tree. Data sources include patent databases, academic databases, social media, news websites, and internal intelligence databases. Analytical methods include statistical analysis, time series forecasting, and relation extraction. Expected output formats include tables, charts, and text summaries. Confidence level requirements include minimum confidence levels. Finally, the initial task tree is displayed to the user through a web-based visualization component, such as the ECharts tree diagram plugin. Users can adjust the structure by dragging and dropping nodes, and modify the execution order and analysis parameters by clicking nodes. The system receives adjustment commands by listening for click and drag events on the front end, updates the task tree structure data in real time, and re-renders the display.

[0025] In one possible implementation, the task execution tree generation module 20 includes: a target template matching module, used to match a target template from a preset analysis framework template library based on the analysis intent elements by identifying the analysis object, analysis dimension, and analysis target; an initial task tree generation module, used to allocate main task nodes and several sub-task nodes to the analysis intent elements according to the target template to generate an initial task tree, wherein each task node is identified with the required data source, analysis method, expected output format, and confidence level requirement; and an adjustment instruction receiving module, used to display the initial task tree through a visual interface and receive adjustment instructions from the user side, wherein the adjustment dimension includes at least one of node structure, execution order, and analysis parameters.

[0026] Specifically, the analysis objects, dimensions, and objectives are structured and coded. For example, analysis objects are categorized into countries, organizations, technologies, events, etc., with each category assigned a unique code, such as 001 for countries and 002 for organizations. Analysis dimensions are categorized into technology level, industry chain, threat assessment, trend prediction, etc., coded as 101, 102, etc. Analysis objectives are categorized into impact judgment, risk identification, decision support, etc., coded as 201, 202, etc. Feature vectors are extracted from each template in the analysis framework template library. These feature vectors contain the corresponding analysis object category code, analysis dimension code, analysis objective code, and applicable scenario keywords. For example, the feature vector for a technology trend assessment template is [001 (country), 101 (technology level), 203 (trend prediction), "semiconductor," "patent analysis"]. A weighted similarity calculation model is used, setting the weights for analysis object dimensions, analysis dimensions, analysis objectives, and scenario keywords, to calculate the weighted similarity between the structured features of the analysis intent elements and the feature vectors of each template. Set a matching threshold to filter out templates with similarity higher than the threshold. If there are multiple templates that meet the criteria, select the one with the highest similarity as the target template. If all similarities are lower than the threshold, recommend the three closest templates for the user to choose from.

[0027] The target template's structure file is parsed using an XML parser. This file stores the hierarchical relationship between main tasks and subtasks, as well as the association IDs of parent and child nodes, in XML format. Based on the specific content in the analysis intent elements, information such as the analysis object and scope is mapped to each node. The node attribute configuration rule base is invoked, and the required data source, analysis method, expected output format, and confidence level requirements are configured according to the node name and type. Finally, a tree data structure construction algorithm is used to organize the main task nodes, subtask nodes, and the attribute information of each node into a tree structure data in JSON format, generating the initial task tree.

[0028] The visualization interface is built using the Vue.js framework and the TreeView component. The initial task tree's JSON data is transformed into a component-recognizable format and rendered as a tree structure. Each node includes expand / collapse buttons, a name display, an edit button, and a delete button. JavaScript event listeners monitor user drag events, clicks on the up / down movement buttons, and form submission events after clicking the edit button. The captured operation information is converted into standardized adjustment instructions in JSON format, including instruction type, node ID, and adjusted attributes. Instruction types include structural adjustments, order adjustments, and parameter modifications. Adjusted attributes include the new parent node ID, the new execution order number, and the modified analysis parameter values. Finally, the adjustment instructions are sent to the backend server via an Axios network request library. The backend receives the data, updates the stored task tree data, and returns the updated tree structure data. The frontend then re-renders the visualization interface.

[0029] The parallel tree task analysis module 30 is used by the intelligent routing component to receive the task execution tree, perform two-order matching routing decisions based on node task characteristics in the intelligent agent cluster, determine the intelligent agent task cluster and perform parallel tree task analysis, obtain task analysis results and integrate them, determine a structured intelligence report and construct an evidence tracing chain, wherein the task analysis results are marked with a confidence score.

[0030] Specifically, feature extraction is performed on each node of the task execution tree, including data type, analysis method, domain characteristics, and data volume. The intelligent routing component employs a two-order matching routing algorithm. First-order matching performs an initial match based on node features and agent capability features, filtering agents that meet basic requirements. Second-order matching assesses the collaborative capabilities among agents to determine the final agent task cluster. A distributed parallel computing framework, such as Spark, is used to assign each subtask to the corresponding agent for parallel execution. Each agent generates intermediate results and confidence scores during task execution. The analysis results from each agent are integrated, including data format standardization, conflict detection and handling, and key information extraction. Based on the integrated results, a structured intelligence report is generated according to the standard framework of intelligence reporting. Finally, data fingerprinting and directed graph modeling techniques are used to construct an evidence tracing chain, linking the key conclusions in the report with the original data and reasoning steps.

[0031] In one possible implementation, the parallel tree task analysis module 30 includes: an agent cluster building module for building an agent cluster, wherein the agent cluster is composed of functional agents based on different intelligence analysis subdomains; an intelligent routing component deployment module for deploying intelligent routing components according to the agent cluster, wherein first-order matching routing and second-order cooperative routing are used as the underlying logic; a task execution tree matching module for matching the task execution tree based on node task features according to the intelligent routing components to determine the agent task cluster; and a task execution tree processing module for calling the agent task cluster to process the task execution tree and determine the task analysis result.

[0032] Specifically, intelligence analysis is divided into sub-fields such as open-source intelligence gathering, deep text understanding, time series prediction, technical intelligence analysis, industry intelligence analysis, science and technology intelligence analysis, and image analysis. Functional agents are designed for each subdomain. For example, the open-source intelligence gathering agent has functions such as web crawling, API data scraping, and data deduplication. It is developed using the Scrapy crawling framework, supports multi-threaded crawling, and is configured with a User-Agent pool and a proxy IP pool to deal with anti-crawling mechanisms. The data scraping frequency is controlled at 1-5 times per second, dynamically adjusted according to the pressure of the data source server. The text deep understanding agent has functions such as word segmentation, named entity recognition, relation extraction, and event extraction. It is based on the BERT-large model and fine-tuned. The input is text data, and the output is structured entity, relation, and event information. The model is trained using 1 million intelligence text corpora with a learning rate of 1e-5 and 5 training epochs. The time series prediction agent uses an LSTM neural network model to predict trends based on historical data. The input is time series data, and the output is the predicted value for the next N periods. The model has 3 hidden layers with 128 neurons per layer and a time step of 12. Each agent is developed using a microservice architecture, packaged as a Docker container, and deployed in a cluster using Kubernetes. A load balancing strategy is configured to dynamically allocate computing resources based on the CPU and memory usage of each agent. The agents communicate with each other using the gRPC protocol, defining standardized request and response data formats to ensure efficient and consistent data transmission.

[0033] A capability profile is constructed for each agent, including sub-domain labels, supported data types, preferred analysis methods, historical task success rate, average processing time, and current load rate. Based on historical task allocation data, including task node features, matched agents, and task execution results, a random forest classification algorithm is used to train the routing decision model. The model's input is the task node feature vector, and the output is the agent matching probability. During training, the data is divided into training and test sets in a 7:3 ratio, and 5-fold cross-validation is used to optimize the model parameters. A prediction probability threshold of 0.6 is set; when the prediction probability is higher than this threshold, the agent is considered to have matched the task node. The intelligent routing component is developed using the Spring Cloud microservice framework and consists of a request receiving module, a feature extraction module, a first-order matching module, a second-order collaborative evaluation module, and a decision output module. These modules communicate via a RESTful API. Finally, the routing component is deployed to a Kubernetes cluster and associated with the agent cluster through a service discovery mechanism. After receiving the task execution tree, the routing component calls each module to complete the two-order routing decision and sends the task execution request to the matched agent.

[0034] For each node in the task execution tree, features such as data type, analysis method, domain characteristics, data volume, and confidence requirements are extracted and standardized into feature vectors. The first-order matching module of the intelligent routing component inputs the node feature vectors into the trained routing decision model, calculates the matching probability between each agent and the node, and selects agents with a matching probability higher than 0.6 as candidate agents, such as technical intelligence analysis agents and patent data processing agents. The second-order collaborative evaluation module calculates the collaborative fit between candidate agents. Collaborative fit is calculated based on indicators such as historical collaboration success rate, data format compatibility, and communication latency. A collaborative fit threshold is set; if the collaborative fit between candidate agents is higher than this threshold, they are grouped into an agent task cluster. A corresponding agent task cluster is assigned to each task node, and the roles and task assignments of each agent in the cluster are recorded.

[0035] Using a distributed task scheduling framework, such as XXL-Job, tasks at each node of the task execution tree are assigned to corresponding agent task clusters. An execution timeout threshold is set for each task, such as 30 minutes for complex tasks and 5 minutes for simple tasks. The agent clusters execute tasks in parallel according to a pre-defined collaborative process. For example, an open-source intelligence gathering agent crawls patent data from a specified data source, a text deep understanding agent performs entity recognition and relationship extraction on patent text, and a technical intelligence analysis agent performs patent quantitative analysis and technology trend judgment. Each agent transmits data and intermediate results in real time through a message queue, such as RabbitMQ. The system monitors the agents' heartbeat signals and task status reports to track task progress in real time. If a task times out or encounters an error, a retry mechanism is triggered, with a maximum of three retries. If a retry fails, the user is notified. A result fusion algorithm is used to integrate the analysis results from each agent, including data format standardization, duplicate data removal, and conflict result handling. Conflict result handling employs a voting mechanism, selecting the result supported by the majority of agents. If there is significant disagreement, manual intervention is triggered. Finally, the integrated results are quality checked to verify data integrity and logical consistency. If the verification passes, the results are accepted as the task analysis results; otherwise, they are returned to the agent cluster for reprocessing.

[0036] In one possible implementation, the task execution tree matching module includes: a node task feature determination module, used to determine the node task features of each task node based on the task execution tree, wherein the node task features include at least data type, analysis method, and domain features; a first-order matching routing module, used to traverse the node task features of each task node, perform first-order matching routing, and determine task agents, wherein each task node corresponds to at least one task agent, and each task agent is associated with at least one task node; and a second-order routing coordination module, used to perform second-order routing coordination between task agents based on information transmission, collaborative reasoning, and iterative optimization for the node task features and the task agents, and determine the agent task cluster.

[0037] Specifically, the process parses the attribute configuration of task nodes, determines the data type from the required data source fields (e.g., if the data source is the USPTO or WIPO patent database, the data type is patent data), and directly obtains the analysis method features from the analysis method field (e.g., patent measurement analysis, technology lifecycle analysis). Keywords are extracted from the node's name and description text using the TF-IDF algorithm combined with a domain keyword library to extract domain features. The data type, analysis method, and domain features are encoded according to preset standardized coding rules. Finally, the extracted features are verified. If the data type is not recognized or the domain features are missing, a prompt message is returned requesting the user to supplement the node attribute information.

[0038] Load the capability profile data of all agents. Each profile includes information such as the data types supported by the agent, the analysis methods it is proficient in, the domain features it covers, and its historical task success rate. It iterates through the feature codes of each task node and compares them with the corresponding codes in the agent's capability profile. A weighted scoring model is used, with weights set for data type matching, analysis method matching, domain feature matching, and historical success rate, to calculate the matching degree between the node and the agent. A matching degree threshold is set, and agents with matching degrees higher than the threshold are selected as task agents. If at least one agent has a matching degree higher than the threshold for a given node, the top three agents with the highest matching degrees are selected as candidate task agents. If no agent has a matching degree higher than the threshold, the user is prompted to add a suitable agent or adjust the node's task features.

[0039] A collaborative capability evaluation index system was established, including data format compatibility, communication protocol consistency, historical collaboration success rate, and task allocation suitability. Data format compatibility checks whether the input and output data formats of the agents are consistent; communication protocol consistency verifies whether all agents support the gRPC protocol; historical collaboration success rate is the success rate of tasks completed collaboratively by agents over the past three months; and task allocation suitability is scored based on the functional complementarity of the agents. Collaboration process simulations were performed on candidate task agents, simulating data transmission, result feedback, and collaborative reasoning, recording indicators such as data transmission latency and error rate during the simulation. A collaborative suitability score was calculated based on the evaluation index system, and a suitability threshold was set. If the collaborative suitability score between candidate agents exceeded this threshold, they were included in the agent task cluster. Assign specific roles to the agents in the cluster. For example, the master agent is responsible for task scheduling and result integration, while the slave agents are responsible for specific data processing and analysis. Define the information transmission rules and collaboration processes between agents. For example, after the master agent receives a task, it sends a data retrieval request to the slave agent. After the slave agent returns the data, the master agent calls the analysis algorithm to process it, generates intermediate results, and feeds them back to the slave agent for verification.

[0040] In one possible implementation, the parallel tree task analysis module 30 includes: a multi-level confidence scoring system establishment module, used to establish a multi-level confidence scoring system, wherein the multi-level confidence scoring system at least covers the data source layer, inference path layer, agent output layer, and conclusion layer; a confidence model establishment module, used to establish a confidence model for dynamic confidence calibration based on the multi-level confidence scoring system as a framework and using Bayesian update rules; and a confidence score analysis module, used to analyze the stage-by-stage confidence scores based on the confidence model during the task analysis process.

[0041] Specifically, four confidence assessment levels are defined: the data source level assesses the credibility of the original data; the inference path level assesses the reliability of the reasoning process; the agent output level assesses the credibility of the analysis results of each agent; and the conclusion level assesses the credibility of the final integrated conclusion. Evaluation indicators are designed for each level. Data source level indicators include data source authority, timeliness, and historical accuracy; inference path level indicators include the rationality of reasoning logic, the completeness of reasoning steps, and the sufficiency of evidence support; agent output level indicators include the agent's historical accuracy, consistency of output results, and self-assessment of confidence; and conclusion level indicators include the consistency of multi-source evidence, the reliability of the reasoning process, and the history of expert feedback. Develop scoring rules for each indicator. For example, regarding data source authority: well-known industry companies receive 0.7-0.8 points, ordinary websites receive 0.5-0.7 points, and social media receives 0.3-0.5 points; for timeliness: data from the past month receives 0.9-1.0 points, data from the past three months receives 0.8-0.9 points, data from the past six months receives 0.7-0.8 points, data from the past year receives 0.5-0.7 points, and data from more than one year receives 0.3-0.5 points. Use the Analytic Hierarchy Process (AHP) to determine the weights of each level, for example, the data source layer has a weight of 0.25, the inference path layer has a weight of 0.3, the agent output layer has a weight of 0.25, and the conclusion layer has a weight of 0.2. Calibrate the system through historical intelligence analysis cases, adjusting indicator weights and scoring rules to ensure that the confidence score is highly correlated with the credibility of the actual results.

[0042] Prior confidence levels are set for each level. The prior confidence level of the data source layer is set based on historical statistical data of the data source type. For example, the prior confidence level of the WIPO patent database is 0.9, and the prior confidence level of ordinary news websites is 0.6. The prior confidence level of the reasoning path layer is set based on the reliability of the reasoning method. For example, the prior confidence level of deductive reasoning is 0.85, and the prior confidence level of inductive reasoning is 0.7. The prior confidence level of the agent output layer is set based on the historical performance of the agent. For example, the prior confidence level of the technical intelligence analysis agent is 0.88. The prior confidence level of the conclusion layer is calculated by integrating the prior confidence levels of the first three levels, using a weighted summation formula: Prior confidence level of conclusion layer = Prior confidence level of data source layer × Weight of data source layer + Prior confidence level of reasoning path layer × Weight of reasoning path layer + Prior confidence level of agent output layer × Weight of agent output layer. A likelihood function is constructed, representing the conditional probability of confidence at each level given new feedback data, such as user verification of results or new evidence. According to the Bayesian update rule, posterior confidence ∝ prior confidence × likelihood function. The posterior confidence at each level is calculated; for example, the posterior confidence at the data source layer = (prior confidence × likelihood function value) / normalization constant. Finally, the model iteration update step size is set to update once after each feedback. After each update, the confidence at each level is recalculated, and the impact is propagated to the next level, achieving dynamic calibration.

[0043] Data is collected in real-time from each stage of the task analysis via a data acquisition interface. The data source layer collects information such as the type of data source, release time, and historical accuracy. The inference path layer collects data such as inference steps, algorithms used, and the amount of evidence. The agent output layer collects the agent's output results, self-assessed confidence level, and historical performance data. The collected data is input into a confidence calculation engine. Following the confidence model's calculation rules, the engine sequentially calculates confidence scores for the data source layer, inference path layer, and agent output layer, then integrates these scores to obtain the stage-wise confidence score for the conclusion layer. Confidence thresholds are set for each level. The calculated scores are compared to the corresponding thresholds. If a score at a certain level falls below the threshold, it is marked as an anomaly, and the anomaly level, current score, and threshold are recorded to provide a basis for triggering manual intervention.

[0044] In one possible implementation, the confidence score analysis module includes a manual intervention process triggering module, which is used to trigger a manual intervention process during the task analysis process if at least one of the following conditions is met: the confidence score is lower than a preset threshold or an analysis conflict is detected. The module receives user-end adjudication data and corrects the task execution tree.

[0045] Specifically, the system monitors the confidence scores and agent analysis results at each level in real time through a background monitoring thread. Confidence monitoring employs a timed query mechanism, such as querying every 5 seconds. Analysis conflict monitoring compares the outputs of different agents; if contradictory conclusions exist (e.g., agent A predicts a technology will mature within 3 years, while agent B predicts it will mature within 5 years), an analysis conflict is identified. Trigger conditions are set: if the confidence score at any level falls below a preset threshold, or if an analysis conflict is detected, a manual intervention process is triggered. The system displays an intervention prompt window on the front-end interface, showing the anomaly information and providing a decision-making interface. Users input decision data through the decision-making interface, including their opinion on handling confidence anomalies, the decision result for the analysis conflict, and the reasons for the decision. Decision data is submitted in JSON format. After receiving the decision data, the system corrects the task execution tree based on the decision opinion. For example, if a new data source is added, it is added to the required data source in the corresponding node; if an agent is reselected, the agent task cluster of that node is adjusted, and the analysis task of that node is restarted after correction.

[0046] In one possible implementation, the parallel tree task analysis module 30 includes: a data fingerprint generation module, used to generate a data fingerprint for each piece of original data, wherein the data fingerprint includes at least a source identifier, a timestamp, and a content hash value; an evidence tracing chain construction module, used to construct the evidence tracing chain by starting with the original data, determining multiple intermediate quantities through reasoning steps, and ending with the conclusion through directed graph association; and an interactive evidence aiming point setting module, used to locate key conclusions in the structured intelligence report and set interactive evidence aiming points, wherein the structured intelligence report and the evidence tracing chain support forward and reverse tracing, and the conclusion is traced back in reverse based on the interactive evidence aiming points.

[0047] Specifically, the source identifier for each piece of raw data is extracted. The source identifier is the unique code of the data source plus the unique ID of the data within the data source. The data collection timestamp is extracted, in the format YYYY-MM-DD HH:MM:SS:mmm. The content of the raw data is standardized by removing irrelevant formatting characters such as HTML tags, spaces, and line breaks. The text is uniformly converted to UTF-8 encoding, and the SHA-256 hash algorithm is used to calculate a 256-bit content hash value. Finally, the source identifier, timestamp, and content hash value are connected using the separator "|" to generate a data fingerprint, in the format "source identifier|timestamp|content hash value". The generated data fingerprint is associated with the raw data and stored in the database, with an index created to support fast querying.

[0048] This document outlines the entire intelligence analysis process, clarifying the flow path from raw data to preprocessing, intermediate analysis results, and final conclusions. Intermediate data includes entity identification results, relation extraction results, statistical analysis data, and trend prediction data. In a graph database such as Neo4j, node types are defined, including raw data nodes (attributes: data fingerprint, data content, source, timestamp), preprocessing nodes (attributes: processing step ID, processing method, processing result), intermediate data nodes (attributes: intermediate data ID, type, content, confidence level), and conclusion nodes (attributes: conclusion ID, content, confidence level, generation time). Edge relationship types are defined, including raw data → preprocessing (indicating that the raw data underwent this preprocessing step), preprocessing → intermediate data (indicating that the preprocessing result generated this intermediate data), intermediate data → intermediate data (indicating that one intermediate data generated another intermediate data through reasoning), and intermediate data → conclusion (indicating that this intermediate data supports the conclusion). Each edge records the relationship type, reasoning method, and confidence level transmission value. Based on the actual analysis process, corresponding nodes and edges are created in the graph database, constructing a directed graph structure for evidence tracing. Finally, the graph structure data of the evidence tracing chain is associated with the structured intelligence report, and each conclusion node in the report is labeled with the corresponding graph database node ID, supporting both forward and reverse tracing.

[0049] This algorithm employs a text keyword extraction method combined with domain rules to identify key conclusions in structured intelligence reports. Key conclusions typically contain explicit judgments, predictions, and assessments. Text segments with a keyword matching degree higher than 0.8 are identified as key conclusions. The algorithm queries the graph database of the evidence tracing chain to find the conclusion node ID corresponding to each key conclusion, establishing a relationship between the key conclusion and the conclusion node. In the structured intelligence report displayed on the front end, interactive evidence anchors are set for each key conclusion, using icons and underlines. Clicking an icon or underline triggers the tracing function, enabling both forward and reverse tracing. Forward tracing, clicking an anchor displays the complete reasoning path from the original data node to the conclusion node. Path data is obtained through path query statements in the graph database and displayed on the front end in the form of a tree diagram or flowchart. For reverse tracing, users can click on an original data node or intermediate node in the visual interface of the evidence tracing chain to query all conclusions supported by that node. Relevant conclusions are obtained through graph database queries, and the corresponding key conclusions and evidence anchors are highlighted in the report.

[0050] In one possible implementation, the parallel tree task analysis module 30 includes: a confidence score adjustment module, used to receive user feedback data, perform hierarchical matching and confidence score adjustment in the multi-layer confidence score system, and determine confidence correction data; and an intelligence confidence graph update module, used to adjust the confidence of relevant conclusions in the structured intelligence report along the dependency graph based on the evidence chain, and update the intelligence confidence graph, wherein the intelligence confidence graph is dynamically updated.

[0051] Specifically, user feedback data is analyzed to extract feedback type, feedback object, feedback result, and feedback reason. For example, if a user reports a factual error in a piece of information from data source NW023, the feedback type is "data source credibility feedback," the feedback object is data source NW023, and the feedback result is "negative error." Based on the feedback type and feedback object, the corresponding confidence assessment level is matched: data source-related feedback matches the data source layer, inference process-related feedback matches the inference path layer, conclusion-related feedback matches the conclusion layer, and agent output-related feedback matches the agent output layer. A Bayesian update rule is used to adjust the confidence score of the corresponding level. Simultaneously, the confidence adjustment at this level affects related levels. For example, after adjusting the confidence at the data source layer, the confidence scores of the agent output layer and the conclusion layer are recalculated using a weighted summation formula, with an adjustment step size set to 0.1. The confidence change after each feedback adjustment does not exceed 0.1 to avoid excessive score fluctuations. Finally, confidence correction data is generated, including the adjusted confidence scores for each level, the reason for adjustment, and the feedback data ID. This data is stored in the confidence database and synchronized to the confidence model.

[0052] From the dependency graph of the evidence chain, nodes associated with the confidence adjustment level are identified, such as preprocessing nodes, intermediate nodes, and conclusion nodes related to the adjusted data source node. A depth-first search algorithm is used to traverse the dependency paths, and a confidence propagation model is employed to propagate the adjusted confidence along these paths, with propagation weights set based on the importance of the edge relationships. An update threshold is set; if the confidence adjustment of a conclusion exceeds this threshold, it is marked as requiring close attention and highlighted in the intelligence confidence graph. Finally, the intelligence confidence graph is updated, storing the adjusted confidence of each node in the graph database, and the corresponding confidence labels for the conclusions in the structured intelligence report are updated synchronously.

[0053] In one possible implementation, the parallel tree task analysis module 30 includes: a report generation feedback receiving module, used to receive report generation feedback from the user, wherein the report generation feedback includes direct evaluation, content modification, supplementary evidence, or error type labeling; an implicit behavior data mining module, used to mine implicit behavior data from the user, wherein the implicit behavior data includes browsing duration, frequency of repeated visits, attention to specific intelligence categories, and modification patterns during task execution; and an update learning module, used to update and learn the multi-layer confidence scoring system and the agent cluster based on the report generation feedback and implicit behavior data.

[0054] Specifically, the front-end feedback collection interface is designed, including a direct evaluation area, a modification input area, a supplementary evidence upload area, and an error type labeling area. For example, the direct evaluation area uses a 5-point rating system, with 1 being the worst and 5 being the best, covering four dimensions: accuracy, completeness, usefulness, and comprehensibility. The modification input area supports rich text editing, allowing users to directly modify the report content. The supplementary evidence upload area supports uploading evidence in formats such as documents, images, and links, with a maximum of 5 files. The error type labeling area provides a drop-down menu, including five types: factual errors, logical errors, missing information, formatting errors, and other errors, allowing multiple selections and input of error locations and descriptions. User feedback data is received via a front-end form submission event, using FormData format for file data and JSON format for text data. The feedback data is standardized: direct evaluation data is converted to numerical values, modification content is converted to HTML format for storage, supplementary evidence uploads generate a unique file ID and access link, and error type labels are converted to encoding, with text fragments extracted from the error locations. Finally, the standardized feedback data is stored in a database, establishing a link between feedback data and structured intelligence reports through report IDs, providing data support for updated learning.

[0055] Tracking points are placed at key locations on the front-end pages, including report browsing pages, evidence tracing interfaces, and task adjustment interfaces. JavaScript code records user behavior events, such as page load events, page close events, click events, scrolling events, and modification events. The back-end server records user access requests through a logging system, including access time, access IP, accessed report ID, and request type. The collected behavioral data is preprocessed, browsing duration is calculated, duplicate access frequencies are counted after deduplication, and report segments with a browsing duration exceeding 30 seconds are selected as high-attention segments. The operation sequence of task adjustment is analyzed to determine the modification pattern. Statistical analysis algorithms are used to extract features; for example, the attention level to a specific intelligence category is calculated as the sum of browsing durations of reports in that category divided by the sum of browsing durations of all reports. If this ratio exceeds 0.3, it is considered high-attention. Modification patterns are categorized into parameter adjustment, structural adjustment, and content supplementation, and the proportion is statistically analyzed based on the type and frequency of modification operations. Finally, the mined implicit behavioral data is standardized into JSON format, including information such as user ID, report ID, browsing duration, frequency of repeated visits, high-attention intelligence categories, modification patterns and proportions, and stored in the user behavior database.

[0056] Weights are assigned to explicit and implicit feedback, and report generation feedback (explicit feedback) and implicit behavioral data (implicit feedback) are weighted and fused. Based on the fused feedback data, a proximal policy optimization algorithm is used to optimize the parameters of the multi-layer confidence scoring system. Feedback scores are used as reward signals to adjust the indicator weights and scoring rules at each level. The capability profile of the agent cluster is adjusted according to the feedback data. For example, if an agent's output is repeatedly rejected by users, its historical accuracy index is lowered; if the collaborative results of two agents are highly recognized by users, their collaborative fit index is increased. Simultaneously, the collaboration strategy of the agents is optimized, and task allocation rules are adjusted. For example, tasks in high-interest intelligence categories are prioritized for agents that perform well in that domain. Finally, the updated confidence scoring system parameters, agent capability profiles, and collaboration strategies are stored in the system database, and relevant models and components are updated synchronously to achieve continuous learning of the system.

[0057] The terminal interface display module 40 is used to display the structured intelligence report and evidence tracing chain on the terminal interface.

[0058] Specifically, the terminal interface layout is designed with a top navigation bar, a left-side table of contents, a central content area, and a right-side toolbar. The structured intelligence report is visualized, presented in a chapter structure: executive summary, background introduction, detailed analysis, supporting evidence, trend prediction, uncertainty explanation, and recommended actions. Rich text format is used to render the text content, table data is displayed using Element UI's Table component, and trend charts are drawn using the ECharts component. Key conclusions are labeled with confidence scores and interactive evidence anchors. The evidence tracing chain is triggered by the evidence tracing button in the right-side toolbar, displayed in a new pop-up window or sidebar. A directed graph is drawn using Neo4j's front-end visualization plugin, supporting zooming, dragging, and node clicking to view details. Interactive functions are integrated, including report content search, chapter navigation, evidence tracing interaction, and feedback submission. A responsive layout is adopted, using CSS media queries to adapt to different terminals. For example, the full layout is displayed on the computer, the left-side table of contents is hidden on tablets, and a single-column layout with paginated chapter content is used on mobile devices, ensuring clear viewing and interactive operations on different devices.

[0059] This application embodiment receives intelligence analysis requests input by the user and extracts core elements through intent parsing. The task decomposition component generates a task execution tree containing main and sub-task nodes based on this. After interactive interface and collaborative adjustment and optimization with the user, the intelligent routing component determines the appropriate intelligent agent task cluster from the intelligent agent cluster based on the node task characteristics through two-order matching routing. Parallel tree task analysis is performed and the results are integrated to generate a structured intelligence report with confidence score and a corresponding evidence tracing chain. The report and evidence tracing chain are displayed to the user through the terminal interface. These technical means solve the technical problem of the imbalance between efficiency and accuracy in the existing intelligence analysis field, and achieve the technical effect of improving intelligence analysis efficiency, enhancing the accuracy and credibility of analysis results, and realizing a dynamic balance between efficiency and accuracy.

[0060] In the above text, refer to Figure 1 A human-machine collaborative intelligence agent system based on interactive feedback according to an embodiment of the present invention is described in detail. Next, reference will be made to... Figure 2 This invention describes a human-machine collaborative intelligence agent method based on interactive feedback according to an embodiment of the present invention.

[0061] According to an embodiment of the present invention, a human-machine collaborative intelligence agent method based on interactive feedback is used to solve the technical problem of the imbalance between efficiency and accuracy in the existing intelligence analysis field, achieving the technical effect of improving intelligence analysis efficiency, enhancing the accuracy and credibility of analysis results, and realizing a dynamic balance between efficiency and accuracy. The human-machine collaborative intelligence agent method based on interactive feedback includes: receiving intelligence analysis requests input from a user terminal; extracting analysis intent elements through intent parsing; a task decomposition component generating a task execution tree based on the analysis intent elements, wherein the task execution tree includes a main task node and multiple sub-task nodes; adjusting the task execution tree collaboratively with the user through an interactive interface; an intelligent routing component receiving the task execution tree; performing two-order matching routing decisions based on node task characteristics in the agent cluster; determining the agent task cluster and performing parallel tree task analysis; obtaining and integrating the task analysis results; determining a structured intelligence report and constructing an evidence tracing chain; wherein the task analysis results are marked with a confidence score; and displaying the structured intelligence report and evidence tracing chain on a terminal interface.

[0062] The process of generating a task execution tree may further include: based on the analysis intent elements, matching a target template from a preset analysis framework template library by identifying the analysis object, analysis dimension, and analysis goal; allocating main task nodes and several sub-task nodes to the analysis intent elements according to the target template to generate an initial task tree, wherein each task node is identified with the required data source, analysis method, expected output format, and confidence level requirement; displaying the initial task tree through a visual interface and receiving adjustment instructions from the user, wherein the adjustment dimension includes at least one of node structure, execution order, and analysis parameters.

[0063] The process of obtaining the task analysis results may further include: building an intelligent agent cluster, wherein the intelligent agent cluster is composed of functional intelligent agents based on different intelligence analysis subdomains; deploying an intelligent routing component based on the intelligent agent cluster, wherein first-order matching routing and second-order cooperative routing are used as the underlying logic; matching the task execution tree based on node task features based on the intelligent routing component to determine the intelligent agent task cluster; and calling the intelligent agent task cluster to process the task execution tree to determine the task analysis results.

[0064] The process of matching the task execution tree based on node task features to determine the agent task cluster may further include: determining the node task features of each task node based on the task execution tree, wherein the node task features include at least data type, analysis method, and domain features; traversing the node task features of each task node, performing first-order matching routing to determine the task agents, wherein each task node corresponds to at least one task agent, and each task agent is associated with at least one task node; and performing second-order routing coordination between the task agents based on information transmission, collaborative reasoning, and iterative optimization for the node task features and the task agents to determine the agent task cluster.

[0065] The task analysis results are identified by a confidence score, which may further include: establishing a multi-layer confidence score system, wherein the multi-layer confidence score system covers at least the data source layer, the inference path layer, the agent output layer, and the conclusion layer; based on the multi-layer confidence score system as a framework, using Bayesian update rules, establishing a confidence model for dynamic confidence calibration; and during the task analysis process, analyzing the confidence scores at different stages based on the confidence model.

[0066] This may further include: during the task analysis process, if at least one of the following conditions is met—a confidence level below a preset threshold and an analysis conflict is detected—a manual intervention process is triggered to receive user-end adjudication data and correct the task execution tree.

[0067] The process of identifying structured intelligence reports and constructing an evidence tracing chain can further include: generating a data fingerprint for each piece of original data, wherein the data fingerprint includes at least a source identifier, a timestamp, and a content hash value; constructing the evidence tracing chain by starting with the original data, determining multiple intermediate quantities through reasoning steps, and ending with the conclusion, through directed graph association; locating key conclusions in the structured intelligence report and setting interactive evidence anchor points; wherein the structured intelligence report and the evidence tracing chain support forward and reverse tracing, and the conclusion is traced back in reverse based on the interactive evidence anchor points.

[0068] After identifying the structured intelligence report and constructing the evidence tracing chain, the process may further include: receiving user feedback data, performing hierarchical matching and confidence score adjustment in the multi-layered confidence scoring system, and determining confidence correction data; adjusting the confidence of relevant conclusions in the structured intelligence report along the dependency graph based on the evidence chain, and updating the intelligence confidence graph, wherein the intelligence confidence graph is dynamically updated.

[0069] After identifying the structured intelligence report and constructing the evidence tracing chain, the process may further include: receiving report generation feedback from the user, wherein the report generation feedback includes direct evaluation, content modification, supplementary evidence, or error type labeling; mining implicit behavioral data from the user, wherein the implicit behavioral data includes browsing duration, frequency of repeated visits, attention to specific intelligence categories, and modification patterns during task execution; and updating and learning the multi-layer confidence scoring system and the intelligent agent cluster based on the report generation feedback and implicit behavioral data.

[0070] The human-machine collaborative intelligence agent system based on interactive feedback provided in this embodiment of the invention can execute the human-machine collaborative intelligence agent method based on interactive feedback provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0071] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A human-machine collaborative intelligence agent system based on interactive feedback, characterized in that, The system includes: The intent parsing module is used to receive intelligence analysis requests input from the user terminal and extract the elements of the analysis intent through intent parsing; The task execution tree generation module is used by the task decomposition component to generate a task execution tree based on the analysis intent elements. The task execution tree includes a main task node and multiple sub-task nodes. The task execution tree can be adjusted collaboratively with the user through an interactive interface. The parallel tree task analysis module is used by the intelligent routing component to receive the task execution tree, perform two-order matching routing decisions based on node task characteristics in the intelligent agent cluster, determine the intelligent agent task cluster and perform parallel tree task analysis, obtain task analysis results and integrate them, determine a structured intelligence report and construct an evidence tracing chain, wherein the task analysis results are marked with a confidence score. The terminal interface display module is used to display the structured intelligence report and evidence tracing chain on the terminal interface; The parallel tree task analysis module includes: An intelligent agent cluster building module is used to build an intelligent agent cluster, wherein the intelligent agent cluster is composed of functional intelligent agents based on different intelligence analysis subdomains; The intelligent routing component deployment module is used to deploy intelligent routing components according to the intelligent agent cluster, wherein the underlying logic is based on first-order matching routing and second-order cooperative routing. The task execution tree matching module is used to match the task execution tree based on node task characteristics according to the intelligent routing component to determine the intelligent agent task cluster. The task execution tree processing module is used to call the intelligent agent task cluster, process the task execution tree, and determine the task analysis results; The task execution tree matching module includes: The node task feature determination module is used to determine the node task features of each task node based on the task execution tree, wherein the node task features include at least data type, analysis method and domain features; The first-order matching routing module is used to traverse the node task features of each task node, perform first-order matching routing, and determine the task agent. Each task node corresponds to at least one task agent, and each task agent is associated with at least one task node. The second-order routing coordination module is used to perform second-order routing coordination between task agents based on information transmission, collaborative reasoning and iterative optimization, according to the task characteristics of the node and the task agent, to determine the task cluster of the agent. The parallel tree task analysis module includes: The data fingerprint generation module is used to generate a data fingerprint for each piece of original data, wherein the data fingerprint includes at least a source identifier, a timestamp, and a content hash value; The evidence tracing chain construction module is used to construct the evidence tracing chain by starting with the original data, determining multiple intermediate quantities through reasoning steps, and ending with the conclusion through directed graph association. An interactive evidence targeting module is used to locate key conclusions in the structured intelligence report and set interactive evidence targeting points. The structured intelligence report and evidence tracing chain support forward and reverse tracing, and the conclusions are traced back in reverse based on the interactive evidence targeting points.

2. The human-machine collaborative intelligence agent system based on interactive feedback as described in claim 1, characterized in that, The task execution tree generation module includes: The target template matching module is used to match target templates from a preset analysis framework template library based on the analysis intent elements by identifying the analysis object, analysis dimension and analysis target; The initial task tree generation module is used to allocate main task nodes and several sub-task nodes to the analysis intent elements according to the target template, and generate an initial task tree, wherein each task node is identified with the required data source, analysis method, expected output format and confidence level requirement. The adjustment instruction receiving module is used to display the initial task tree through a visual interface and receive adjustment instructions from the user side, wherein the adjustment dimension includes at least one of node structure, execution order and analysis parameters.

3. The human-machine collaborative intelligence agent system based on interactive feedback as described in claim 1, characterized in that, The parallel tree task analysis module includes: The multi-level confidence scoring system establishment module is used to establish a multi-level confidence scoring system, which at least covers the data source layer, inference path layer, agent output layer and conclusion layer. The confidence model building module is used to build a confidence model for dynamic confidence calibration based on the multi-level confidence scoring system and using Bayesian update rules. The confidence score analysis module is used to analyze the confidence score at each stage during the task analysis process, based on the confidence model.

4. The human-machine collaborative intelligence agent system based on interactive feedback as described in claim 3, characterized in that, The confidence score analysis module includes: The manual intervention process triggering module is used to trigger the manual intervention process during task analysis if at least one of the following occurs: the confidence level is lower than a preset threshold or an analysis conflict is detected. The module receives adjudication data from the user terminal and corrects the task execution tree.

5. The human-machine collaborative intelligence agent system based on interactive feedback as described in claim 3, characterized in that, The parallel tree task analysis module includes: The confidence score adjustment module is used to receive user feedback data, perform hierarchical matching and confidence score adjustment in the multi-level confidence score system, and determine the confidence correction data. The intelligence confidence graph update module is used to adjust the confidence of relevant conclusions in the structured intelligence report along the dependency graph based on the evidence chain, and update the intelligence confidence graph, wherein the intelligence confidence graph is dynamically updated.

6. The human-machine collaborative intelligence agent system based on interactive feedback as described in claim 3, characterized in that, The parallel tree task analysis module includes: The report generation feedback receiving module is used to receive report generation feedback from the user, wherein the report generation feedback includes direct evaluation, content modification, supplementary evidence or error type annotation; The implicit behavior data mining module is used to mine implicit behavior data from users, including browsing time, frequency of repeated visits, attention to specific intelligence categories, and modification patterns during task execution. The update learning module is used to generate feedback and implicit behavioral data based on the report, and to update the multi-layer confidence scoring system and the agent cluster.

7. A human-machine collaborative intelligence agent method based on interactive feedback, characterized in that, The method is implemented by a human-machine collaborative intelligence agent system based on interactive feedback as described in any one of claims 1-6, and the method includes: Receive intelligence analysis requests from users and extract the elements of the analysis intent through intent parsing; The task decomposition component generates a task execution tree based on the analysis intent elements. The task execution tree includes a main task node and multiple sub-task nodes. The task execution tree can be adjusted collaboratively with the user through an interactive interface. The intelligent routing component receives the task execution tree, performs two-order matching routing decisions based on node task characteristics in the intelligent agent cluster, determines the intelligent agent task cluster and performs parallel tree task analysis, obtains and integrates the task analysis results, determines the structured intelligence report and constructs the evidence tracing chain, wherein the task analysis results are marked with a confidence score. The structured intelligence report and evidence tracing chain are displayed on the terminal interface.

Images