Method for realizing generative bi-decision closed loop based on multi-agent cooperation

Abstract

The application relates to the technical field of artificial intelligence, and particularly discloses a generative BI decision closed loop implementation method based on multi-agent cooperation, which is applied to a generative BI system, the generative BI system comprising a central dispatching agent, a demand analysis agent cluster, a data management agent cluster, a generative analysis agent cluster, a decision simulation agent cluster and a landing tracking agent cluster. A multi-agent division and cooperation architecture is adopted to divide a decision whole process into five independent subtasks of demand analysis, data management, generative analysis, decision simulation and landing tracking, which are respectively executed by corresponding functional agent clusters, each agent focuses on single-link capacity optimization, the data analysis threshold of business personnel is reduced, and the processing professionalism and efficiency of each link are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a generative BI decision-making closed-loop implementation method based on multi-agent collaboration. Background Technology

[0002] As enterprises continue to advance their digital transformation, the scale of operational data accumulated by enterprise business systems is growing rapidly, and business intelligence (BI) has become a core tool for enterprises to support business decision-making. Traditional BI systems are mostly report-based or self-service architectures, relying on technical personnel to pre-complete data modeling, indicator configuration, and report development. Business personnel need to obtain analysis conclusions through fixed reports, which has problems such as long demand response cycles, high usage thresholds, and insufficient flexibility, making it difficult to adapt to rapidly changing business decision-making scenarios.

[0003] In recent years, the development of generative large model technology has driven the emergence of generative BI, which supports the direct generation of analytical conclusions and visualizations through natural language input, effectively lowering the barrier to entry for data analysis. However, existing generative BI solutions still have many shortcomings: First, the analysis process has low transparency, and the data caliber and attribution logic of the generated content lack multi-stage cross-validation, making it prone to data referencing errors, logical illusions, and other problems, and the reliability of the analysis results is difficult to guarantee; Second, the coupling between different parts of the system is insufficient, and most only cover a one-way process from requirement parsing to analysis output, failing to form a complete closed loop from decision implementation to effect feedback. Decision biases cannot be reversed to optimize the system model, and similar errors are prone to recurrence; Third, when facing cross-business domain decision-making needs, existing solutions often require centralized analysis of data from various domains, which can easily lead to data security and compliance risks. If each domain analyzes independently, it is difficult to form a globally optimal decision solution; Fourth, the system resource scheduling method is relatively crude, unable to dynamically match and process resources according to task complexity and value, which can easily lead to delays in the response of high-priority tasks and wasted computing power for low-complexity tasks. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a generative BI decision-making closed-loop implementation method based on multi-agent collaboration, which solves the problems mentioned in the background technology.

[0005] To achieve the above objectives, this invention provides the following technical solution: a generative BI decision-making closed-loop implementation method based on multi-agent collaboration, applied to a generative BI system. The generative BI system includes a central scheduling agent, and clusters of demand analysis agents, data governance agents, generative analysis agents, decision simulation agents, and implementation tracking agents. The method comprises the following steps: The S1 central scheduling agent receives natural language decision-making requirements submitted by the business end, or automatic decision-making requirements triggered by abnormal fluctuations in core business indicators. After initially classifying the requirements, it breaks down the decision-making requirements into five sub-tasks based on preset task decomposition rules: requirement analysis, data governance, generative analysis, decision simulation, and implementation tracking. The S2 central scheduling agent matches the target agent in the corresponding functional agent cluster for each type of subtask, and sends the subtask and context information to the corresponding target agent. S3's target agents collaborate to complete sub-tasks in the following order: First, the requirement parsing agent transforms natural language decision requirements into structured BI task instructions. Then, the data governance agent cleans and integrates multi-source business data based on the BI task instructions. Next, the generative analysis agent generates analysis results containing attribution conclusions and visual dashboards based on the governed data. Finally, the decision simulation agent generates at least two feasible decision schemes based on the analysis results and completes multi-stage cross-validation, outputting the optimal decision scheme and pushing it to the business end. The S4 landing tracking agent collects real-time operational data of the actual business operations after the optimal decision-making scheme is implemented, calculates the actual achievement value of the decision effect, compares the actual achievement value with the expected value in the decision simulation stage to obtain deviation data, and feeds the deviation data back to the central scheduling agent, generative analysis agent, and decision simulation agent to update the model parameters and task matching rules of each agent, forming a closed-loop decision-making process.

[0006] Preferably, the specific method for the central scheduling agent to match the target agent for the subtask in step S2 is as follows: S21 For each subtask, iterate through all candidate agents in the corresponding functional agent cluster and calculate the task matching degree of each candidate agent for the current subtask. The calculation formula is: ; In the formula, The preset weighting coefficients satisfy... ,and ; For the first The capability matching score of each candidate agent, with a value range of [value range missing]. It is obtained by pre-labeling the historical business domain adaptability and task processing accuracy of the intelligent agent; For the first The historical task completion scores of the candidate agents are given, with values ​​ranging from [value range]. This is obtained by accumulating feedback from the agent's historical performance in handling similar subtasks. For the first The current load idle time of each candidate agent, with a value range of [value missing]. It is calculated from the proportion of the agent's current idle computing power to the total computing power; S22 Select Task Matching Degree The highest-ranking candidate agent is selected as the target agent for the current subtask; if at least two subtasks of the same type apply for the same candidate agent at the same time, agent resources are allocated according to the order of task submission time. If the queue of S23 subtasks exceeds the preset waiting time, the backup intelligent agent cluster scheduling will be automatically triggered to re-match the task from the backup cluster. The intelligent agent performs the task.

[0007] Preferably, the specific processing flow of the demand parsing agent in step S3 is as follows: S311 receives natural language decision-making requirements from the central scheduling agent, extracts four types of entity information from the requirements—time dimension, business domain, core performance indicators, and expected goals—and generates a set of requirement tags. ; S312 calculates the semantic similarity between the current requirement and all historical requirements in the historical structured requirement database. The calculation formula is: ; In the formula, This is the vector output function for the semantic embedding model of the business domain. A collection of tags for historical needs; S313 if Then, the decomposition template corresponding to the historical requirements can be directly reused to generate structured BI task instructions. Then, based on the historical template, a structured BI task instruction is generated. Then, based on the generative large model, a brand-new structured BI task instruction is generated. The structured BI task instruction includes three types of information: data query scope, indicator calculation method, and analysis dimension requirements. S314 Computational Requirements Decomposition Adaptability The calculation formula is: ; In the formula, To generate historical confirmation accuracy for similar requirements corresponding to historical templates or generative large models, the value range is [value missing]. ; S315 will generate structured BI task instructions and Synchronize with the business side for confirmation, if It can trigger direct approval without confirmation. If the approval is granted, the requirement parsing subtask will be completed. If the approval is denied, the entity information will be corrected based on feedback from the business side, and a new structured BI task instruction will be generated. After the S316 requirement analysis subtask is completed, the business side confirms the accuracy rate based on this task. Update the historical task completion scores of the analytical agent for this requirement. The updated formula is: ; In the formula, Score based on completing historical missions before the update. This represents the accuracy rate of the task confirmation. A value of 1 indicates that the task was successfully confirmed on the first attempt by the business side, while a value of 0 indicates that the task ultimately failed to pass confirmation. This corresponds to the situation after the correction is passed.

[0008] Preferably, the specific processing flow of the data governance intelligent agent in step S3 is as follows: S321 receives the structured BI task instructions and requirements matching output by the demand parsing agent. ,based on Dynamically adjust data quality thresholds The calculation formula is: ; In the formula, The range of values ​​is ; Based on the data query scope specified in the instruction, S322 retrieves raw data from three types of data sources: business database, log database, and third-party database, and calculates the trust weight of each data source to be used. The calculation formula is: ; In the formula, For this data source recently The historical quality score of each participation in governance tasks. The range of values ​​is ; S323 is preferred. The system pulls raw data from the data source. If the coverage of the high-reliability data source is less than 80%, the central scheduling agent is triggered to remind the data source to supplement. S324 preprocesses the retrieved raw data by performing four types of operations: missing value filling, outlier removal, cross-source data primary key alignment, and unified indicator definitions, to obtain standardized governance data. S325 calculates the raw quality score of standardized governance data. The calculation formula is: ; In the formula, The preset weighting coefficients have a range of values. ; For data integrity, the value range is: , which is the proportion of non-missing data to the total required data volume; For data consistency, the value range is: , which is the proportion of conflict-free data after cross-source alignment to the total data volume; S326 combines the data source trust weights to obtain the final data quality score. ,like This will involve standardizing governance data and data quality scores. Synchronize to the generative analytical agent, if Then, the corresponding missing or conflicting data source is retrieved again for secondary processing. After three consecutive processing iterations... If the target is still not met, an abnormal alert will be triggered by the central scheduling agent.

[0009] Preferably, the specific processing flow of the generative analysis agent in step S3 is as follows: S331 receives standardized governance data and data quality scores output by the data governance agent. And the set of requirement tags output by the requirement parsing agent. ; Based on the analytical dimension requirements in the structured BI task instructions, S332 calls a generative large model to complete three types of analytical operations: multi-dimensional attribution analysis, trend prediction, and anomaly localization. The generative large model outputs the model confidence score of the analysis results. , The range of values ​​is The score is obtained by the self-verification score of the analysis logic by the large model. S333 calculates the overall reliability of the analysis results. The calculation formula is: ; S334 matches the visualization template library, selects the visualization template with the highest matching degree with the requirement tags, and generates three types of output content: visualization dashboard, natural language analysis report, and core conclusion summary; S335 will consider overall credibility The analysis output is synchronized to the decision simulation intelligent agent.

[0010] Preferably, the specific processing flow of the decision simulation agent in step S3 is as follows: S341 receives the analysis content and overall reliability of the generative analytical agent's output. Based on the influencing factors and trend prediction results in the analysis conclusions, a simulation environment matching the current business scenario is constructed. The simulation environment includes three types of parameters: business rule constraints, external environmental variables, and internal resource thresholds. S342 generates at least two differentiated decision-making schemes based on the simulation environment, and simulates for each decision-making scheme. Calculate the expected returns for each decision-making scheme under different scenario fluctuation conditions. The calculation formula is: ; In the formula, For the first The probability of occurrence of various scenario fluctuations, with a value range of [value missing]. ,all The sum of is 1; For the first The simulated return value of this decision-making scheme under various scenario fluctuations; S343 Compare the expected returns of all decision options Select The highest-scoring decision is considered the optimal decision, while also taking into account the expected returns of each option. The probability of scenario fluctuations and risk values ​​are simultaneously pushed to the business side for reference.

[0011] Preferably, the process of each target agent executing a subtask in step S3 further includes a multi-agent cross-validation process, specifically: After the S371 generative analysis agent completes the generation of analysis content, it sends the data dimensions and statistical standards referenced in the analysis content to the data governance agent for source data verification. If the data is completely consistent, it passes the verification and enters the logical verification stage. If there are local deviations in standards, it is sent back for correction. If there are core data conflicts, it is rolled back to the data governance stage for re-verification. After the S372 decision simulation agent generates a decision scheme, it sends the corresponding attribution logic and input parameters to the generative analysis agent for analysis logic verification. If the logic is completely matched, the decision scheme is verified and pushed to the business end. If there is a local logic deviation, it is sent back for correction. If there is a core logic conflict, it is rolled back to the generative analysis stage for re-verification. If S373 fails the verification twice in a row, the central scheduling agent will be triggered to re-match the agent in the corresponding stage; if it fails the verification three times in a row, manual intervention will be triggered to remind the user, and the user will review the information to determine whether to continue the iteration or terminate the task.

[0012] Preferably, the specific processing flow of step S4 is as follows: The S41 landing tracking agent continuously collects corresponding actual business operation data according to the evaluation indicators of the optimal decision-making scheme and the preset sampling period. S42 calculates the actual benefit of the decision-making effect after the preset tracking period ends. ; S43 calculates the actual profit and the expected profit of the optimal decision-making plan. deviation rate The calculation formula is: ; S44 If The preset deviation threshold is used to determine whether the current decision-making loop has met the target. If the deviation threshold is not met, the decision-making loop is deemed to have failed to meet the target, triggering the parameter update process. S45 will determine the deviation rate Actual business operation data is synchronized to the central scheduling agent, generative analysis agent, and decision simulation agent as training data for updating model parameters.

[0013] Preferably, the parameter update process in step S4 is as follows: S441 generative analytical agent based on deviation rate Based on actual business operation data, optimize the attribution analysis logic and trend prediction algorithm, and update the model confidence level. The calculation rules, if The case analyzed in this study will then be added to the negative case library. S442 Decision Simulation Agent Based on Deviation Rate Correct the parameter configuration of the simulation environment and the probability of scene fluctuations. The calculation rules, if The simulation scenario will then be added to the risk verification library, and the simulation weight of extreme scenarios will be automatically increased for subsequent similar scenarios. The S443 central scheduling agent is based on deviation rate. Update the capability matching scores of all target agents participating in this task. Score for completing historical missions ,in: The formula for updating the ability matching score is: In the formula This is a preset adjustment coefficient, with a value range of [value range missing]. ; The formula for updating the score for completing historical tasks is: ; S444 If this loop is closed Then, the requirement breakdown template, data governance rules, analysis logic, and simulation scheme for this entire process will be added to the positive case library to improve the template matching priority for similar requirements.

[0014] Preferably, when the decision-making requirement involves two or more independent business domains, it further includes a cross-domain intelligent agent collaboration process, specifically: The S51 central scheduling agent breaks down the original decision requirements into sub-requirements within each business domain, and matches each business domain with an independent cluster of sub-agents for requirement parsing, data governance, generative analysis, and decision simulation. Each sub-agent cluster can only access the internal data of the corresponding business domain. Each business domain's sub-agent cluster in S52 independently completes sub-tasks within its domain, generating domain-specific decision-making sub-solutions. Each sub-solution includes domain-specific analysis credibility. Expected returns within the region Scheme parameter set ,in Assign a business domain number, , The total number of business domains, and the reliability of the analysis within each domain. Expected returns within the region The definition is consistent with the analytical credibility and expected benefits in a single-domain process; The publicly available parameters of each domain sub-scheme in S53 are synchronized to the cross-domain fusion module of the central scheduling agent. The cross-domain fusion module calculates the global conflict degree of each domain sub-scheme. The calculation formula is: ; In the formula, For the first The number of parameters in a sub-scheme of one domain that conflict with all other sub-schemes of all domains. For the first The total number of parameters for each domain sub-scheme The range of values ​​is ; S54 If Then the sub-solution is sent back to the decision simulation agent of the corresponding business domain to adjust the parameters, until all domain sub-solutions are... ; S55 calculates the global fit of each sub-scheme within the domain. The calculation formula is: ; S56 calculates the fusion weight of each sub-scheme based on global adaptability. The calculation formula is: ; The weighted average yields the global expected return of the globally optimal decision-making scheme. The calculation formula is: ; The S57 globally optimal decision-making solution is distributed to each business domain for implementation, and the implementation tracking agents collect the actual benefits of each business domain. The total actual benefit is obtained by summing the results. Calculate the global deviation rate ,based on Complete the end-to-end parameter update.

[0015] This invention provides a generative BI decision-making closed-loop implementation method based on multi-agent collaboration, which has the following beneficial effects: A multi-agent collaborative architecture is adopted, breaking down the entire decision-making process into five independent sub-tasks: requirement analysis, data governance, generative analysis, decision simulation, and implementation tracking. These sub-tasks are executed by clusters of corresponding functional agents, with each agent focusing on optimizing a single aspect. This lowers the data analysis threshold for business personnel while improving the professionalism and efficiency of each step. Furthermore, dynamic scheduling of agents is achieved based on task matching, and data quality thresholds are dynamically adjusted according to requirement complexity. This ensures the quality of decision output while rationally allocating computing resources and optimizing the overall resource utilization of the system.

[0016] A multi-agent cross-validation mechanism is established, whereby the data governance agent verifies the data caliber of the analysis content, and the generative analysis agent verifies the attribution logic of the decision-making scheme. This forms a two-way verification link between the data layer and the logic layer, which can effectively reduce the illusion problem and data reference bias in generative analysis and improve the accuracy and credibility of the decision-making scheme. A full-link decision-making closed loop is constructed, whereby the landing tracking agent collects actual business data, calculates the expected and actual deviation rate, and updates the model parameters, task matching rules, and positive and negative case libraries of each agent based on the deviation, enabling the system to have continuous self-iterative capabilities.

[0017] For cross-business domain decision-making scenarios, a collaborative model of "data not leaving the domain and solution layer integration" is adopted. The original data of each business domain is kept within the domain for processing, and only publicly available solution parameters are synchronized for global integration. Under the premise of ensuring data security and compliance, global decision optimization of multiple business domains is achieved, which is suitable for multi-department collaborative business scenarios. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the generative BI decision-making closed-loop implementation method based on multi-agent collaboration described in this invention. Figure 2 A block diagram illustrating the principle of a multi-agent system architecture; Figure 3 This is a schematic diagram of the multi-agent cross-validation process; Figure 4 This is a schematic diagram of cross-business domain intelligent agent collaboration. Detailed Implementation

[0019] The technical solution of the present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0020] This embodiment provides a generative BI decision-making closed-loop implementation method based on multi-agent collaboration, applied to a generative BI system deployed in an enterprise private cloud environment. The generative BI system includes one central scheduling agent, as well as clusters of demand analysis agents, data governance agents, generative analysis agents, decision simulation agents, and implementation tracking agents. Each type of functional agent cluster is configured with at least three alternative agents, supporting load balancing and fault tolerance.

[0021] The generative BI multi-agent system architecture of the present invention is as follows: Figure 2 As shown, the central scheduling agent serves as the core of global scheduling, uniformly managing the task allocation, status monitoring, and collaborative interaction of five types of functional agent clusters. Each functional agent cluster focuses on optimizing the capabilities of its corresponding business process, while connecting upwards to the business end to realize demand input and solution output, forming a hierarchical management and collaborative system architecture.

[0022] This method is divided into four core stages: demand reception and decomposition, agent matching and scheduling, subtask collaborative execution, and landing feedback closed loop. The specific implementation process is as follows: The complete execution flow of this method is as follows: Figure 1 As shown, the process of demand decomposition, agent scheduling, multi-stage collaborative execution and implementation tracking are completed sequentially along the forward link. The entire link parameter iteration is achieved through the reverse flow of deviation data, forming a decision-making closed loop of continuous self-optimization.

[0023] The S1 central scheduling agent receives natural language decision-making requirements submitted by the business end, or automatic decision-making requirements triggered by abnormal fluctuations in core business indicators. After initially classifying the requirements, it breaks down the decision-making requirements into five sub-tasks based on preset task decomposition rules: requirement analysis, data governance, generative analysis, decision simulation, and implementation tracking. The S2 central scheduling agent matches the target agent in the corresponding functional agent cluster for each type of subtask, and sends the subtask and context information to the corresponding target agent. S3's target agents collaborate to complete sub-tasks in the following order: First, the requirement parsing agent transforms natural language decision requirements into structured BI task instructions. Then, the data governance agent cleans and integrates multi-source business data based on the BI task instructions. Next, the generative analysis agent generates analysis results containing attribution conclusions and visual dashboards based on the governed data. Finally, the decision simulation agent generates at least two feasible decision schemes based on the analysis results and completes multi-stage cross-validation, outputting the optimal decision scheme and pushing it to the business end. The S4 landing tracking agent collects real-time operational data of the actual business operations after the optimal decision-making scheme is implemented, calculates the actual achievement value of the decision effect, compares the actual achievement value with the expected value in the decision simulation stage to obtain deviation data, and feeds the deviation data back to the central scheduling agent, generative analysis agent, and decision simulation agent to update the model parameters and task matching rules of each agent, forming a closed-loop decision-making process.

[0024] Request reception and task decomposition are executed by the central scheduling agent, corresponding to method step S1: The central scheduling agent supports two types of decision-making inputs: one is natural language decision-making requirements submitted by business personnel through the system front end, such as "analyze the reasons for the decline in sales in East China this quarter and provide optimization solutions"; the other is automatic decision-making requirements triggered by abnormal fluctuations in core business indicators. When the month-on-month fluctuation of core indicators such as sales and profit margin monitored in real time exceeds the preset threshold (set to 20% in this embodiment), the corresponding decision-making requirements are automatically generated.

[0025] After receiving a decision request, the central scheduling agent first performs a preliminary classification according to the business domain and the type of request. Then, based on the preset task decomposition rules, it breaks down the complete decision request into five types of sub-tasks that are dependent on each other: request analysis, data governance, generative analysis, decision simulation, and implementation tracking. Each sub-task carries the corresponding context information and execution time limit requirements.

[0026] Agent matching and task scheduling: This step is executed by the central scheduling agent, corresponding to method step S2. For each sub-task obtained from the breakdown, the central scheduling agent matches the optimal target agent from the agent cluster corresponding to the function. The specific process is as follows: Task matching degree calculation: Traverse all candidate agents in the corresponding functional agent cluster and calculate the task matching degree of each candidate agent for the current subtask. The calculation formula is: ; In this embodiment, the weighting coefficient is preset to be 1. , , ,satisfy .in, For the first The capability matching score of each candidate agent, with a value range of [value range missing]. The score is obtained by pre-labeling the agent's historical business domain adaptability and task processing accuracy. Agents that are good at the corresponding business domain score higher. For the first The historical task completion scores of the candidate agents are given, with values ​​ranging from [value range]. It is obtained through cumulative and iterative feedback of the agent's historical performance in handling similar subtasks; For the first The current load idle time of each candidate agent, with a value range of [value missing]. The score is calculated based on the proportion of the agent's current idle computing power to the total computing power; the lower the load, the higher the score.

[0027] This formula comprehensively quantifies the matching degree between the agent and the task from three dimensions: capability adaptability, historical performance, and current load. It takes into account the efficiency of system computing power utilization while ensuring the quality of task processing, and avoids resource imbalance caused by single-dimensional scheduling.

[0028] Numerical calculation example: For the subtask of parsing requirements for a certain sales domain, select candidate agent 1 in the cluster to calculate the matching degree. The matching score of this agent's sales domain capabilities is known. Score for completing historical missions The current percentage of idle computing power corresponds to the load idleness. The calculation process, substituted with the preset weights, is as follows: The calculated task matching degree of this agent is 0.818, which is at a high matching level and can be used as a candidate target agent.

[0029] Agent allocation rules: Under normal circumstances, directly select the task matching degree. The highest-ranking candidate agent will be selected as the target agent for the current subtask. If at least two subtasks of the same type simultaneously request the same candidate agent, agent resources will be allocated according to the order in which the tasks were submitted.

[0030] Timeout scheduling mechanism: If the queuing time of a subtask exceeds a preset threshold (10 minutes in this embodiment), the backup intelligent agent cluster scheduling is automatically triggered, and a new match is made from the backup cluster. The intelligent agent executes tasks, avoiding long-term blocking of high-priority tasks.

[0031] Multi-agent collaborative execution of sub-tasks: This step is completed collaboratively by each target agent in the execution order, corresponding to method step S3. It includes four stages: requirement analysis, data governance, generative analysis, and decision simulation, and is accompanied by a multi-agent cross-validation mechanism, as detailed below: Requirements analysis and processing: This step is executed by the matched requirement parsing agent, which is responsible for converting natural language requirements into machine-executable structured instructions. The specific process is as follows: The requirement parsing agent receives natural language decision requirements from the central scheduling agent and extracts four core entity information categories from the requirements—time dimension, business domain, core performance indicators, and expected goals—using an entity extraction model to generate a set of requirement tags. .

[0032] Calculate the semantic similarity between the current requirement and all historical requirements in the historical structured requirement database. The calculation formula is: ; in, This is the vector output function of the semantic embedding model, which has been fine-tuned using a business domain corpus. It is a set of tags for historical needs, and the cosine similarity is used to measure the degree of matching between the current needs and historical cases.

[0033] The cosine similarity metric is used to quantify the semantic overlap of the requirement label vectors. The closer the value is to 1, the higher the semantic matching degree of the requirement. Based on this, the corresponding requirement decomposition strategy can be selected, which can improve the efficiency and accuracy of requirement parsing.

[0034] Numerical calculation example: After entity extraction, the semantic embedding vector of the current requirement is as follows: The semantic embedding vector of a certain historical demand in the historical demand database is The calculation process is as follows: The first step is to calculate the dot product of the two vectors: The second step is to calculate the magnitudes of the two vectors: The third step is to calculate the cosine similarity: The calculated semantic similarity is 0.965, which satisfies... Under the given conditions, the disassembly templates corresponding to historical requirements can be directly reused.

[0035] Based on semantic similarity matching, a corresponding generation strategy is adopted: If Directly reuse the decomposition template corresponding to historical requirements to generate structured BI task instructions; if Structured BI task instructions are generated after modifications and adjustments based on historical templates; if The generative large model is invoked to generate entirely new structured BI task instructions. The final generated structured BI task instructions contain three core types of information: data query scope, indicator calculation criteria, and analysis dimension requirements.

[0036] The fit of the computational requirements decomposition The reliability of the current disassembly results is quantified by the following formula: ; in, To generate historical confirmation accuracy for similar requirements corresponding to historical templates or generative large models, the value range is [value missing]. .

[0037] The reliability of the demand breakdown results is comprehensively evaluated by combining semantic similarity and historical breakdown accuracy. The higher the fit, the higher the expected accuracy of the breakdown results. This serves as the basis for the determination of the no-confirmation mechanism, which can reduce unnecessary manual review steps.

[0038] Numerical calculation example: Following the semantic similarity calculation results above Historical confirmation accuracy of corresponding historical templates The calculation process is as follows: The calculated requirement breakdown and fit is 0.897, which does not reach the 0.9 threshold for exemption from confirmation. Therefore, manual confirmation is required from the business side.

[0039] Adapting structured BI task instructions to requirements Synchronize with the business side for confirmation: If If the business side confirms the approval, the requirement parsing subtask is completed; if the approval fails, the entity information is corrected based on the feedback from the business side, and a new structured BI task instruction is generated.

[0040] After the requirement analysis subtask is completed, the business side confirms the accuracy based on this task. Update the historical task completion score of the intelligent agent that resolves this requirement. The updated formula is: ; in, Scoring is awarded for completing historical missions prior to the update. This represents the accuracy rate of the task confirmation. A value of 1 indicates that the task was successfully confirmed on the first attempt by the business side, while a value of 0 indicates that the task ultimately failed to pass confirmation. This corresponds to the corrected outcome. Through iterative updates, the capabilities of the agent are dynamically calibrated.

[0041] The historical task completion scores of the agent are iteratively updated using an exponential moving average method, giving higher weight to recent task performance while retaining the reference value of historical performance, thus avoiding excessive fluctuations in the score caused by the results of a single task.

[0042] Numerical calculation example: Historical task completion scores of a certain requirement analysis agent before update. This task has been approved by the business side in one go. The calculation process is as follows: After the update, the agent's historical task completion score is 0.808, which is an improvement over the previous score.

[0043] Data governance processing, this step is executed by the matched data governance intelligent agent, which is responsible for cleaning, fusion and quality control of multi-source data. The specific process is as follows: The data governance agent receives and analyzes the structured BI task instructions output by the agent, and the degree of fit between these instructions and the requirements. Dynamically adjust data quality thresholds based on demand adaptability The calculation formula is: ； The range of values ​​is The lower the demand fit and the higher the uncertainty of decomposition, the higher the corresponding data quality requirements, ensuring the reliability of complex demands from the data source end.

[0044] Establish a dynamic linkage mechanism between demand complexity and data quality requirements. The lower the adaptability of demand decomposition and the higher the uncertainty, the higher the corresponding data quality admission threshold. This will compress the error propagation space from the data source end and ensure the reliability of the output of complex demands.

[0045] Numerical calculation example: Adaptability results based on the above requirements The calculation process is as follows: The calculated data quality threshold for this task is 0.8206. The data quality score after processing must be no lower than this value to proceed to the next stage.

[0046] Based on the data query scope specified in the command, raw data is retrieved from three types of data sources: enterprise business database, user behavior log database, and third-party external database. Before retrieval, the trust weight of each data source to be used is calculated. The historical reliability of the data source is quantified by the following formula: ; In this embodiment The value is set to 10, which means calculating the average historical quality score of the data source in its most recent 10 participations in governance tasks. The range of values ​​is .

[0047] Quantifying the credibility level of a data source based on its historical quality performance and prioritizing high-credibility data sources with stable historical performance can reduce the impact of single data anomalies on governance results and improve the stability of governance output.

[0048] Numerical calculation example: The historical quality scores of the most recent 10 governance tasks for a company's business database data source are as follows: 0.84, 0.86, 0.82, 0.87, 0.85, 0.83, 0.88, 0.81, 0.85, 0.89. The calculation process is as follows: The first step is to calculate the total historical scores: The second step is to calculate the confidence weights: The calculated confidence weight of this data source is 0.85, which satisfies... The requirement for a highly reliable data source.

[0049] Priority selection If the data coverage of the high-reliability data source is less than 80%, the central scheduling agent will trigger a data source supplement reminder to notify the operation and maintenance personnel to supplement the corresponding data source.

[0050] The raw data is preprocessed to standardize: missing values ​​are filled with mean values ​​of the same dimension, outliers are removed using the 3σ principle, primary key alignment is achieved across source data through primary key mapping, and field mapping and calculation are completed according to a unified indicator standard to finally obtain standardized governance data.

[0051] Calculate the raw quality score of standardized governance data The calculation formula is: ; In this embodiment, the weighting coefficient Take 0.6. Wherein, For data integrity, the value range is: , which is the proportion of non-missing data to the total required data volume; For data consistency, the value range is: This represents the proportion of conflict-free data after cross-source alignment to the total data volume.

[0052] The original quality of the governed data is assessed from two core dimensions: data integrity and consistency. This comprehensively covers the core assessment indicators of data governance and objectively reflects the usability level of the governed data.

[0053] Numerical calculation example: Data integrity after this remediation Data consistency Weighting coefficient The calculation process is as follows: The calculated raw quality score of the treated data was 0.944.

[0054] The final data quality score is obtained by combining the credibility weight of the data source. .like Then, standardized governance data and data quality scores will be combined. Synchronize to the generative analytical agent; if If necessary, the corresponding missing or conflicting data source will be retrieved again for secondary processing; after three consecutive processing iterations... If the target is still not met, an abnormal alert will be triggered by the central scheduling agent, notifying manual intervention.

[0055] Numerical calculation example: Continuing from the above results, the data source credibility weight Original quality score The calculation process is as follows: The final data quality score was calculated to be 0.8024, which is lower than the quality threshold of 0.8206 for this task. Therefore, the corresponding data source needs to be retrieved again for secondary processing.

[0056] Generative analysis processing, this step is performed by the matched generative analysis agent, which is responsible for generating analytical conclusions and visualizations based on the processed data. The specific process is as follows: Generative analytical agents receive standardized governance data and data quality scores output by data governance agents. And the set of requirement tags output by the requirement parsing agent. .

[0057] Based on the analytical dimension requirements in the structured BI task instructions, a business-tuned generative large model is invoked to complete three core analytical operations: multi-dimensional attribution analysis, trend prediction, and anomaly localization. The generative large model outputs the model confidence score of the analysis results. The range of values ​​is It is obtained by converting the self-verification score of the large model's own analysis logic.

[0058] The overall reliability of the analysis results is calculated by combining data quality and model confidence. The calculation formula is: ; By using a data quality weighted model confidence level, the credibility of the analysis results can be traced across the entire chain. The lower the data quality, the lower the upper limit of the credibility of the analysis results.

[0059] By coupling the quality of upstream data with the confidence level of the model itself to calculate the overall credibility, the credibility of the analysis results can be traced back to the whole chain. This mechanism constrains the upper limit of the analysis conclusions corresponding to low-quality data and prevents data errors from being amplified by the model.

[0060] Numerical calculation example: Building upon the data governance results described above, the final data quality score after secondary governance. Model confidence score of generative large model output The calculation process is as follows: The overall confidence level of this analysis result is calculated to be 0.722. This score will be used in conjunction with the weighted correction of the results of subsequent decision simulations.

[0061] The system uses a built-in visualization template library to select the visualization template that best matches the requirements tags, generating three types of output content: visualization dashboards, natural language analysis reports, and core conclusion summaries, to meet the display needs of different scenarios.

[0062] Overall credibility The analysis output is synchronized to the decision simulation agent as input for subsequent scheme simulation.

[0063] The decision simulation process is executed by the matched decision simulation agent, which is responsible for generating feasible decision solutions and simulating their effects. The specific process is as follows: The decision simulation agent receives the analysis content and overall reliability of the output from the generative analysis agent. Based on the influencing factors and trend prediction results in the analysis conclusions, a simulation environment matching the current business scenario is constructed. The simulation environment includes three types of parameters: business rule constraints, external environmental variables, and internal resource thresholds, thus replicating the constraints of real business operations.

[0064] Generate at least two differentiated decision-making schemes based on the simulation environment, and simulate for each decision-making scheme. Different scenarios of fluctuation (in this embodiment) Taking 5 as an example, corresponding to the baseline scenario, optimistic scenario, pessimistic scenario, extreme upside scenario, and extreme downside scenario respectively, calculate the expected return for each decision-making scheme. The calculation formula is: ; in, For the first The probability of occurrence of various scenario fluctuations, with a value range of [value missing]. ,all The sum of is 1; For the first The simulated return value of this decision-making scheme under various scenario fluctuations. This is determined by comprehensive credibility. Weighted simulation results help avoid simulation biases caused by low-confidence analysis.

[0065] The baseline expected return of the decision-making scheme is obtained by multi-scenario probability weighted simulation, and then corrected by combining the comprehensive credibility analysis to avoid overestimation of the return caused by low credibility analysis, so that the simulation results are more in line with the actual business performance.

[0066] Numerical calculation example: This decision simulation sets up five scenarios with varying fluctuations, and the corresponding parameters are as follows: ; Known overall credibility The calculation process is as follows: The first step is to calculate the weighted simulated returns across multiple scenarios: The second step is to calculate the expected return based on the overall credibility: The expected return of this decision-making plan is calculated to be approximately 747,000 yuan.

[0067] Compare the expected benefits of all decision options Select The highest-scoring decision is considered the optimal decision, while also taking into account the expected returns of each option. The probability of scenario fluctuations and risk values ​​are simultaneously pushed to the business side for business personnel to refer to and make decisions.

[0068] A multi-agent cross-validation mechanism is implemented during the execution of each subtask to avoid the illusion of generative content and data referencing errors. The multi-agent cross-validation process set in this embodiment is as follows: Figure 3 As shown, the process includes two levels of verification: source data verification and analysis logic verification. It also includes mechanisms for exception retry and manual intervention. By constraining the generated content from both data scope and logical chain dimensions, it avoids illusions and referencing biases. The specific process is as follows: Source data verification: After the generative analysis agent generates the analysis content, it first sends all data dimensions and statistical definitions referenced in the analysis content to the data governance agent for source data verification. If the data dimensions and statistical definitions are completely consistent with the governed data, the data verification passes and proceeds to the logical verification stage; if there are local discrepancies, the problem points are sent back to the generative analysis agent to correct the reference definitions; if there are core data conflicts, the process directly reverts to the data governance stage to re-verify the data.

[0069] Analysis Logic Verification: After the decision simulation agent generates a decision plan, it first sends the corresponding attribution logic and input parameters to the generative analysis agent for analysis logic verification. If the attribution logic, input parameters, and analysis conclusions match perfectly, the logic verification passes, and the decision plan is officially pushed to the business side. If there are local logic deviations, the problem points are returned to the decision simulation agent to correct the plan. If there are core logic conflicts, the process reverts to the generative analysis stage to re-verify the conclusions.

[0070] Abnormal retry mechanism: If the same subtask fails the verification twice in a row, the central scheduling agent will be triggered to rematch the agent of the corresponding link; if the cross-verification of any link fails three times in a row, manual intervention will be triggered to remind the user, and the user will review the case to determine whether to continue iterating or terminate the task.

[0071] Landing tracking and end-to-end closed-loop updates: This step is executed collaboratively by the landing tracking agent and various functional agents, corresponding to method step S4. It realizes feedback on the decision-making effect and self-iteration of system capabilities. The specific process is as follows: The landing tracking agent continuously collects corresponding actual business operation data according to the evaluation indicators of the optimal decision-making scheme and a preset sampling period. In this embodiment, the default sampling period is 1 day and the tracking period is 7 days.

[0072] After the preset tracking period ends, the actual benefits of the decision-making plan after its implementation will be statistically calculated. .

[0073] Calculate the actual returns and the expected returns of the optimal decision-making scheme. deviation rate The calculation formula is: ; The relative deviation between the expected value of the quantitative decision simulation and the actual value is measured. The higher the deviation rate, the worse the accuracy of the decision simulation. This is used as the basis for triggering the update of the entire link parameters and driving the iteration of system capabilities.

[0074] Numerical calculation example: Based on the above expected returns, the actual returns during the tracking period after the solution is implemented. 10,000 yuan, expected return The calculation process for 10,000 yuan is as follows: The calculated deviation rate of this decision-making loop is approximately 0.062, which is lower than the preset deviation threshold of 0.15, thus the decision-making loop is deemed to have met the standard.

[0075] like If the preset deviation threshold (set to 0.15 in this embodiment) is met, the decision-making loop is deemed to have met the target; if If the deviation threshold is reached, it is determined that the current decision-making loop has not met the standard, triggering a full-link parameter update process.

[0076] Deviation rate Actual operational data is synchronized to the central scheduling agent, generative analysis agent, and decision simulation agent as training data for model parameter updates. The specific rules for parameter updates are as follows: Generative analytical agents based on bias rate Based on actual business operation data, optimize the attribution analysis logic and trend prediction algorithm, and update the model confidence level. Calculation rules; if If the case is not found, it will be added to the negative case library for subsequent model fine-tuning to avoid the recurrence of similar errors.

[0077] Decision-making simulation agents based on bias rate Correct the parameter configuration of the simulation environment and the probability of scene fluctuations. Calculation rules; if If the simulation scenario is not found, it will be added to the risk verification library. In subsequent similar scenarios, the simulation weight of extreme scenarios will be automatically increased to improve the robustness of the simulation.

[0078] Central scheduling agent based on deviation rate Update the capability matching scores of all target agents participating in this task. Score for completing historical missions ,in: The formula for updating the ability matching score is: In this embodiment, the adjustment coefficient Take 0.5; The agent's capability matching score is dynamically adjusted based on the closed-loop deviation rate. The greater the deviation, the greater the reduction in capability score, thereby achieving dynamic calibration of the agent's capability rating and making the task matching results more consistent with the agent's actual performance.

[0079] Numerical calculation example: the ability matching score of an agent participating in this task before the update. Adjustment coefficient The closed-loop deviation rate in this case The calculation process is as follows: Because the deviation rate was low this time, the ability matching score was only slightly reduced, and the updated score is 0.795.

[0080] The formula for updating the score for completing historical tasks is: .

[0081] By combining the historical baseline score with the current closed-loop effect to update the historical task completion score, the smaller the deviation, the better the current task performance, and the greater the positive increase in score, continuously reflecting the changes in the overall performance of the agent.

[0082] Numerical calculation example: the agent's historical task completion score before the update. The closed-loop deviation rate in this case The calculation process is as follows: In practical applications, the score is truncated to an upper limit. The final updated score is 1, which means that the agent performed well in this task and has reached the full score level in history.

[0083] If this loop is closed The requirement breakdown template, data governance rules, analysis logic, and simulation scheme for this entire process will be added to the positive case library to improve the template matching priority for similar requirements and shorten the processing time for subsequent similar requirements.

[0084] Cross-business domain collaboration implementation methods: Intelligent agent collaboration modes in cross-business domain scenarios, such as... Figure 4 As shown, each business domain retains an independent sub-agent cluster and original data, and only the publicly available scheme parameters are synchronized to the cross-domain fusion module of the central scheduling agent for global optimization. This achieves optimal global decision-making without the data leaving the domain. The specific implementation method is as follows: When decision-making requirements involve two or more physically isolated business domains, this method supports cross-domain agent collaboration, achieving global decision optimization without data leaving the domain. The specific implementation method is as follows: The central scheduling agent breaks down the original decision requirements into sub-requirements within each business domain. It then matches each business domain with an independent cluster of sub-agents for requirement analysis, data governance, generative analysis, and decision simulation. Each sub-agent cluster can only access the internal data of its corresponding business domain, ensuring data security and compliance.

[0085] Each business domain's sub-agent cluster independently completes domain-specific sub-tasks and generates domain-specific decision-making sub-solutions. Each sub-solution includes domain-specific analysis credibility. Expected returns within the region Scheme parameter set ,in Assign a business domain number, , Total number of business domains. Domain-specific analysis reliability. Expected returns within the region The calculation rules are completely consistent with the aforementioned single-domain process.

[0086] The publicly available parameters (excluding original business data and sensitive business rules) of each domain sub-scheme are synchronized to the cross-domain fusion module of the central scheduling agent. The cross-domain fusion module calculates the global conflict degree of each domain sub-scheme. The calculation formula is: ; in, For the first The number of parameters in a sub-scheme of one domain that conflict with all other sub-schemes of all domains. For the first The total number of parameters for each domain sub-scheme The range of values ​​is .

[0087] The compatibility of cross-domain sub-solutions is quantified by the proportion of conflict parameters. The higher the conflict degree, the more core contradictions there are between the solutions, and the worse the feasibility after global integration. This is used as the basis for judging whether a solution should be rejected for adjustment.

[0088] Numerical calculation example: This cross-domain requirement involves two independent business domains: sales and operations. The sales domain sub-solution has a total of 10 parameters, with 2 parameters conflicting with the operations domain solution; the operations domain sub-solution has a total of 8 parameters, with 2 parameters conflicting with the sales domain solution. The calculation process is as follows: The conflict degrees of the two sub-schemes are 0.2 and 0.25, respectively, both below the threshold of 0.3. Therefore, no adjustment is needed, and the schemes can proceed to the scheme fusion stage.

[0089] like Then, the sub-solution is returned to the decision simulation agent of the corresponding business domain to adjust the parameters, and the process is repeated iteratively until all domain sub-solutions are found. Eliminate core parameter conflicts.

[0090] Calculate the global fitness of sub-solutions within each domain. The calculation formula is: ; By combining the credibility of the analysis within the domain with the degree of conflict between the solutions, the adaptability value of the sub-solutions to the overall decision-making is comprehensively evaluated. The higher the credibility and the lower the degree of conflict, the higher the reference value in the overall integration.

[0091] Numerical calculation example: Based on the above results, the credibility of the sales domain analysis. Credibility of operational domain analysis The calculation process is as follows: The calculated global fitness of the two domain sub-schemes are 0.568 and 0.57, respectively.

[0092] The fusion weight of each sub-scheme is calculated based on the global adaptability. The calculation formula is: ; The global expected return of the globally optimal decision scheme is obtained through weighted calculation. The calculation formula is: ; The fusion weight of each sub-scheme is calculated based on the global adaptability normalization. The scheme with higher adaptability has a larger proportion in the global decision-making, which ensures the rationality and reliability of the global fusion result.

[0093] Numerical calculation example: Following the global fitness results above, the calculation process is as follows: The first step is to calculate the total global fit: The second step is to calculate the fusion weights of each sub-scheme: The calculated fusion weights for the sales domain and operations domain sub-solutions are approximately 0.499 and 0.501, respectively.

[0094] Example of global expected revenue calculation: Given the normalized expected revenue score of the sales domain sub-solution. Normalized expected return score of the operation domain sub-solution Based on the aforementioned fusion weights, the calculation process is as follows: The normalized expected return score of the globally optimal decision scheme is calculated to be 0.775. This return is the global comprehensive expected value after scheme conflict correction and credibility correction.

[0095] The globally optimal decision-making solution is distributed to each business domain for implementation, and the implementation tracking agents collect the actual benefits of each business domain. The total actual benefit is obtained by summing the results. Calculate the global deviation rate Based on the global deviation rate, the end-to-end parameters are updated according to the aforementioned rules.

[0096] Numerical calculation example: After the solution is implemented, the normalized actual revenue score for the sales domain is 0.79, and for the operations domain it is 0.71. The global actual revenue score is calculated using a weighted average consistent with the expected revenue. The calculation process is as follows: The first step is to calculate the weighted global actual payout score: The second step is to calculate the global deviation rate: The calculated global deviation rate is approximately 0.032, which is relatively small, indicating that the implementation effect of the global decision-making scheme meets expectations.

[0097] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A generative BI decision-making closed-loop implementation method based on multi-agent collaboration, applied to a generative BI system, wherein the generative BI system includes a central scheduling agent, and clusters of demand analysis agents, data governance agents, generative analysis agents, decision simulation agents, and implementation tracking agents, characterized in that... Includes the following steps: The S1 central scheduling agent receives natural language decision-making requirements submitted by the business end, or automatic decision-making requirements triggered by abnormal fluctuations in core business indicators. After initially classifying the requirements, it breaks down the decision-making requirements into five sub-tasks based on preset task decomposition rules: requirement analysis, data governance, generative analysis, decision simulation, and implementation tracking. The S2 central scheduling agent matches the target agent in the corresponding functional agent cluster for each type of subtask, and sends the subtask and context information to the corresponding target agent; S3's target agents collaborate to complete sub-tasks in the following order: First, the requirement parsing agent transforms natural language decision requirements into structured BI task instructions. Then, the data governance agent cleans and integrates multi-source business data based on the BI task instructions. Next, the generative analysis agent generates analysis results containing attribution conclusions and visual dashboards based on the governed data. Finally, the decision simulation agent generates at least two feasible decision schemes based on the analysis results and completes multi-stage cross-validation, outputting the optimal decision scheme and pushing it to the business end. The S4 landing tracking agent collects real-time operational data of the actual business operations after the optimal decision-making scheme is implemented, calculates the actual achievement value of the decision effect, compares the actual achievement value with the expected value in the decision simulation stage to obtain deviation data, and feeds the deviation data back to the central scheduling agent, generative analysis agent, and decision simulation agent to update the model parameters and task matching rules of each agent, forming a closed-loop decision-making process.

2. The generative BI decision-making closed-loop implementation method based on multi-agent collaboration as described in claim 1, characterized in that, The specific method for the central scheduling agent to match the target agent for the subtask in step S2 is as follows: S21 For each subtask, iterate through all candidate agents in the corresponding functional agent cluster and calculate the task matching degree of each candidate agent for the current subtask. The calculation formula is: ； In the formula, The preset weighting coefficients satisfy... ,and ; For the first The capability matching score of each candidate agent, with a value range of [value range missing]. ; For the first The historical task completion scores of the candidate agents are given, with values ​​ranging from [value range]. ; For the first The current load idle time of each candidate agent, with a value range of [value missing]. ; S22 Select Task Matching Degree The highest-ranking candidate agent is chosen as the target agent for the current subtask. If there are at least two subtasks of the same type applying for the same candidate agent at the same time, the agent resources will be allocated according to the order of task submission time. If the queue of S23 subtasks exceeds the preset waiting time, the backup intelligent agent cluster scheduling will be automatically triggered to re-match the task from the backup cluster. The intelligent agent performs the task.

3. The generative BI decision-making closed-loop implementation method based on multi-agent collaboration as described in claim 2, characterized in that, The specific processing flow of the demand parsing agent in step S3 is as follows: S311 receives natural language decision-making requirements from the central scheduling agent, extracts four types of entity information from the requirements—time dimension, business domain, core performance indicators, and expected goals—and generates a set of requirement tags. ; S312 calculates the semantic similarity between the current requirement and all historical requirements in the historical structured requirement database. The calculation formula is: ； In the formula, This is the vector output function for the semantic embedding model of the business domain. A collection of tags for historical needs; S313 if Then, the decomposition template corresponding to the historical requirements can be directly reused to generate structured BI task instructions. Then, based on the historical template, a structured BI task instruction is generated. Then, based on the generative large model, a brand-new structured BI task instruction is generated. The structured BI task instruction includes three types of information: data query scope, indicator calculation method, and analysis dimension requirements. S314 Computational Requirements Decomposition Adaptability The calculation formula is: ； In the formula, To generate historical confirmation accuracy for similar requirements corresponding to historical templates or generative large models, the value range is [value range missing]. ; S315 will generate structured BI task instructions and Synchronize with the business side for confirmation, if It can trigger direct approval without confirmation. If the approval is granted, the requirement parsing subtask will be completed. If the approval is denied, the entity information will be corrected based on feedback from the business side, and a new structured BI task instruction will be generated. After the S316 requirement analysis subtask is completed, the business side confirms the accuracy rate based on this task. Update the historical task completion scores of the analytical agent for this requirement. The updated formula is: ； In the formula, Score based on completing historical missions before the update. This represents the accuracy rate of the task confirmation. A value of 1 indicates that the task was successfully confirmed on the first attempt by the business side, while a value of 0 indicates that the task ultimately failed to pass confirmation. This corresponds to the situation after the correction is passed.

4. The generative BI decision-making closed-loop implementation method based on multi-agent collaboration as described in claim 3, characterized in that, The specific processing flow of the data governance intelligent agent in step S3 is as follows: S321 receives the structured BI task instructions and requirements matching output by the demand parsing agent. ,based on Dynamically adjust data quality thresholds The calculation formula is: ； In the formula, The range of values ​​is ; Based on the data query scope specified in the instruction, S322 retrieves raw data from three types of data sources: business database, log database, and third-party database, and calculates the trust weight of each data source to be used. The calculation formula is: ； In the formula, For this data source recently The historical quality score of each participation in governance tasks. The range of values ​​is ; S323 is preferred. The system pulls raw data from the data source. If the coverage of the high-reliability data source is less than 80%, the central scheduling agent is triggered to remind the data source to supplement. S324 preprocesses the retrieved raw data by performing four types of operations: missing value filling, outlier removal, cross-source data primary key alignment, and unified indicator definitions, to obtain standardized governance data. S325 calculates the raw quality score of standardized governance data. The calculation formula is: ； In the formula, The preset weighting coefficients have a range of values. ; For data integrity, the value range is: , which is the proportion of non-missing data to the total required data volume; For data consistency, the value range is: , which is the proportion of conflict-free data after cross-source alignment to the total data volume; S326 combines the data source trust weights to obtain the final data quality score. ,like This will involve standardizing governance data and data quality scores. Synchronize to the generative analytical agent, if Then, the corresponding missing or conflicting data source is retrieved again for secondary processing. After three consecutive processing iterations... If the target is still not met, an abnormal alert will be triggered by the central scheduling agent.

5. The generative BI decision-making closed-loop implementation method based on multi-agent collaboration as described in claim 4, characterized in that, The specific processing flow of the generative analysis agent in step S3 is as follows: S331 receives standardized governance data and data quality scores output by the data governance agent. And the set of requirement tags output by the requirement parsing agent. ; Based on the analytical dimension requirements in the structured BI task instructions, S332 calls a generative large model to complete three types of analytical operations: multi-dimensional attribution analysis, trend prediction, and anomaly localization. The generative large model outputs the model confidence score of the analysis results. , The range of values ​​is The score is obtained by the self-verification score of the analysis logic by the large model. S333 calculates the overall reliability of the analysis results. The calculation formula is: ； S334 matches the visualization template library, selects the visualization template with the highest matching degree with the requirement tags, and generates three types of output content: visualization dashboard, natural language analysis report, and core conclusion summary; S335 will consider overall credibility The analysis output is synchronized to the decision simulation intelligent agent.

6. The generative BI decision-making closed-loop implementation method based on multi-agent collaboration as described in claim 5, characterized in that, The specific processing flow of the decision simulation agent in step S3 is as follows: S341 receives the analysis content and overall reliability of the generative analytical agent's output. Based on the influencing factors and trend prediction results in the analysis conclusions, a simulation environment matching the current business scenario is constructed. The simulation environment includes three types of parameters: business rule constraints, external environmental variables, and internal resource thresholds. S342 generates at least two differentiated decision-making schemes based on the simulation environment, and simulates for each decision-making scheme. Calculate the expected returns for each decision-making scheme under different scenario fluctuation conditions. The calculation formula is: ； In the formula, For the first The probability of occurrence of various scenario fluctuations, with a value range of [value missing]. ,all The sum of is 1; For the first The simulated return value of this decision-making scheme under various scenario fluctuations; S343 Compare the expected returns of all decision options Select The highest-scoring decision is considered the optimal decision, while also taking into account the expected returns of each option. The probability of scenario fluctuations and risk values ​​are simultaneously pushed to the business side for reference.

7. The generative BI decision-making closed-loop implementation method based on multi-agent collaboration as described in claim 6, characterized in that, The process of each target agent executing a subtask in step S3 also includes a multi-agent cross-validation process, specifically: After the S371 generative analysis agent completes the generation of analysis content, it sends the data dimensions and statistical standards referenced in the analysis content to the data governance agent for source data verification. If the data is completely consistent, it passes the verification and enters the logical verification stage. If there are local deviations in standards, it is sent back for correction. If there are core data conflicts, it is rolled back to the data governance stage for re-verification. After the S372 decision simulation agent generates a decision scheme, it sends the corresponding attribution logic and input parameters to the generative analysis agent for analysis logic verification. If the logic is completely matched, the decision scheme is verified and pushed to the business end. If there is a local logic deviation, it is sent back for correction. If there is a core logic conflict, it is rolled back to the generative analysis stage for re-verification. If S373 fails the verification twice in a row, the central scheduling agent will be triggered to re-match the agent in the corresponding stage; if it fails the verification three times in a row, manual intervention will be triggered to remind the user, and the user will review the information to determine whether to continue the iteration or terminate the task.

8. The generative BI decision-making closed-loop implementation method based on multi-agent collaboration as described in claim 7, characterized in that, The specific processing flow of step S4 is as follows: The S41 landing tracking agent continuously collects corresponding actual business operation data according to the evaluation indicators of the optimal decision-making scheme and the preset sampling period. S42 calculates the actual benefit of the decision-making effect after the preset tracking period ends. ; S43 calculates the actual profit and the expected profit of the optimal decision-making scheme. deviation rate The calculation formula is: ； S44 If The preset deviation threshold is used to determine whether the current decision-making loop has met the target. If the deviation threshold is not met, the decision-making loop is determined to be unsuccessful, triggering the parameter update process. S45 will determine the deviation rate Actual business operation data is synchronized to the central scheduling agent, generative analysis agent, and decision simulation agent as training data for updating model parameters.

9. The generative BI decision-making closed-loop implementation method based on multi-agent collaboration as described in claim 8, characterized in that, The parameter update process in step S4 is as follows: S441 generative analytical agent based on deviation rate Based on actual business operation data, optimize the attribution analysis logic and trend prediction algorithm, and update the model confidence level. The calculation rules, if The case analyzed in this study will then be added to the negative case library. S442 Decision Simulation Agent Based on Deviation Rate Correct the parameter configuration of the simulation environment and the probability of scene fluctuations. The calculation rules, if The simulation scenario will then be added to the risk verification library, and the simulation weight of extreme scenarios will be automatically increased for subsequent similar scenarios. The S443 central scheduling agent is based on deviation rate. Update the capability matching scores of all target agents participating in this task. Score based on completion of historical missions ,in: The formula for updating the ability matching score is: In the formula This is a preset adjustment coefficient, with a value range of [value range missing]. ; The formula for updating the score for completing historical tasks is: ; S444 If this loop is closed Then, the requirement breakdown template, data governance rules, analysis logic, and simulation scheme for this entire process will be added to the positive case library to improve the template matching priority for similar requirements.

10. The generative BI decision-making closed-loop implementation method based on multi-agent collaboration as described in claim 9, characterized in that, When the decision-making requirements involve two or more independent business domains, cross-domain intelligent agent collaboration processes are also included.

Images