Multi-agent cooperation report generation method and device, equipment and medium

Abstract

The invention relates to the technical field of artificial intelligence, can be applied to business scenes of financial science and technology, medical health and the like, and discloses a multi-agent cooperation report generation method, device, equipment and medium, and the method comprises the steps: obtaining a compliance requirement file and a task description text; analyzing the task description text through an intention recognition agent to generate a task intention recognition result; processing the compliance requirement file and the task intention recognition result through a key point extraction agent to generate compliance requirement key points; executing the compliance requirement key points through the data processing agent to generate data visual representation; and synthesizing compliance requirement key points and data visual representation through the report generation agent to generate report content. According to the method, task analysis, key point extraction, data processing and report synthesis links are sequentially connected and automatically executed in a multi-agent division cooperation mode, so that compliance files and task description can be quickly converted into structured key points, and business data can be efficiently processed and visually presented.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, and medium for generating multi-agent collaborative reports. Background Technology

[0002] In the fintech sector, with the rapid development of heavily regulated industries such as finance and insurance, grassroots institutions are required to regularly submit various compliance reports to regulatory authorities, such as compliance self-inspection reports, risk assessment reports, and rectification progress reports. These reports typically contain numerous regulatory clauses, compliance requirements, and complex financial and operational indicators. However, existing technologies generally rely on manual methods for completion. While some institutions have deployed document management systems or compliance knowledge bases, several shortcomings remain. On the one hand, the reports are difficult to understand. The legal provisions and professional terminology involved in regulatory documents are complex, and ordinary business personnel often lack legal and document backgrounds, making it difficult to accurately extract regulatory points, leading to reports that easily deviate from regulatory requirements. On the other hand, data extraction efficiency is low. The operational data and financial indicators required for compliance reports usually need to be manually queried in multiple business systems and then copied and pasted into the report, which is prone to errors and extremely inefficient. At the same time, most existing systems only provide full-text search or keyword highlighting functions, failing to achieve key point extraction and automatic outline generation, and lacking semantically driven data retrieval capabilities. Furthermore, most existing solutions are mainly single-point tools, lacking a multi-stage collaborative mechanism covering key point extraction, data extraction, and content organization.

[0003] In the healthcare sector, hospitals, clinics, and relevant regulatory authorities also need to regularly generate regulatory compliance reports, such as medical compliance inspection reports, drug use risk reports, and rectification implementation reports. These reports are complex in structure, including item-by-item inspection requirements of compliance clauses and explanations based on actual medical business data. However, existing technologies face challenges similar to those in the financial sector. First, compliance notices or guidance documents often contain professional medical terminology and compliance requirements. Healthcare professionals, lacking systematic document preparation skills, are prone to misunderstandings of regulatory points, leading to incomplete or non-compliant report content. Second, medical compliance reports require the inclusion of substantial patient service data, drug use data, and operational records. Current methods largely rely on manual searching through hospital information systems and then manually integrating these into report templates, which is not only time-consuming and labor-intensive but also prone to omissions and errors. Furthermore, while current general-purpose language models have some application in medical document generation, their lack of deep adaptation to the medical regulatory context often makes it difficult to generate professional reports that conform to the compliance context, resulting in a lack of relevance and accuracy in the output. Summary of the Invention

[0004] The main objective of this invention is to provide a method, apparatus, device, and storage medium for generating multi-agent collaborative reports, aiming to solve the technical problems of low efficiency and poor accuracy in report generation caused by the lack of automatic extraction of key compliance requirements, semantic-driven invocation of business data, and intelligent synthesis of report content in existing technologies.

[0005] To achieve the above objectives, the present invention provides a method for generating multi-agent collaborative reports, comprising:

[0006] Obtain compliance requirement documents and task description text;

[0007] The task description text is parsed by the intent recognition agent to generate task intent recognition results;

[0008] The intelligent agent extracts key points by processing the compliance requirement documents and task intent recognition results to generate compliance requirement key points.

[0009] The compliance requirements are executed by a data processing intelligent agent, generating a data visualization representation.

[0010] The report generation agent synthesizes the key compliance requirements and data visualizations to generate report content.

[0011] Furthermore, to achieve the above objectives, the present invention provides a multi-agent collaborative report generation device, comprising:

[0012] The input acquisition module is used to obtain compliance requirement documents and task description text;

[0013] The intent recognition module is used to parse the task description text through an intent recognition agent and generate task intent recognition results;

[0014] The key point extraction module is used to process the compliance requirement document and task intent recognition results through the key point extraction intelligent agent to generate compliance requirement key points;

[0015] The data processing module is used to execute the compliance requirements through a data processing intelligent agent and generate a data visualization representation;

[0016] The report generation module is used to synthesize the key points of the compliance requirements and data visualization representation through a report generation agent to generate report content.

[0017] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and a multi-agent cooperative report generation program stored in the memory and executable on the processor, wherein when the multi-agent cooperative report generation program is executed by the processor, it implements the steps of the multi-agent cooperative report generation method as described above.

[0018] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a multi-agent cooperative report generation program, which, when executed by a processor, implements the steps of the multi-agent cooperative report generation method as described above.

[0019] Beneficial Effects: This invention relates to the field of artificial intelligence technology and can be applied to business scenarios such as fintech and healthcare. It discloses a multi-agent collaborative report generation method, apparatus, device, and medium, comprising: acquiring compliance requirement documents and task description text; generating task intent recognition results by parsing the task description text through an intent recognition agent; generating compliance requirement key points by processing the compliance requirement documents and task intent recognition results through a key point extraction agent; generating data visualization representations by executing the compliance requirement key points through a data processing agent; and generating report content by synthesizing the compliance requirement key points and data visualization representations through a report generation agent. This invention, through multi-agent collaborative work, sequentially and automatically executes the task parsing, key point extraction, data processing, and report synthesis stages, enabling compliance documents and task descriptions to be quickly transformed into structured key points, allowing business data to be processed efficiently and presented intuitively, ultimately automatically generating report content based on this. Attached Figure Description

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0021] Figure 1 This is a schematic diagram of an application environment for a multi-agent collaborative report generation method according to an embodiment of the present invention;

[0022] Figure 2 This is a flowchart illustrating an embodiment of the multi-agent collaborative report generation method of the present invention;

[0023] Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the multi-agent collaborative report generation device of the present invention;

[0024] Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention;

[0025] Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0026] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0027] The multi-agent cooperative report generation method provided in this embodiment of the invention can be applied to, for example... Figure 1In this application environment, the user terminal communicates with the server via a network. The server can obtain compliance requirement documents and task description text from the user terminal; generate task intent recognition results by parsing the task description text through an intent recognition agent; generate compliance requirement key points by processing the compliance requirement documents and task intent recognition results through a key point extraction agent; generate data visualization representations by executing the compliance requirement key points through a data processing agent; and generate report content by synthesizing the compliance requirement key points and data visualization representations through a report generation agent. This invention, through multi-agent collaboration, sequentially connects and automates the task parsing, key point extraction, data processing, and report synthesis stages, enabling compliance documents and task descriptions to be quickly transformed into structured key points, business data to be efficiently processed and presented intuitively, and ultimately automatically generating report content. The user terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will be described in detail below through specific embodiments.

[0028] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the multi-agent collaborative report generation method provided by the present invention. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0029] like Figure 2 As shown, the multi-agent cooperative report generation method proposed in this invention includes the following steps:

[0030] S10, Obtain compliance requirement documents and task description text;

[0031] In this embodiment, obtaining compliance requirement documents and task description text is the starting point, involving the collection and standardization of input sources. Compliance requirement documents typically originate from notices, normative provisions, or business guidance materials issued by regulatory agencies. These documents are often stored in the organization's document library, file server, or online system. In practice, a configuration file monitoring mechanism is needed to listen for file change events at preset paths, identify newly added files, and extract file identifiers. Subsequently, the captured files need to be parsed, including character encoding recognition, layout structure analysis, and paragraph and clause extraction. The extracted original compliance text often contains noisy characters or formatting anomalies; therefore, it is necessary to define normative feature markers, such as clause numbers, key legal terms, and binding verbs (e.g., should, must). The completeness of these markers must be verified by setting validation parameters, such as checking the continuity of clause numbers and keyword coverage. Only text that passes verification can be converted into a compliance requirement document.

[0032] In parallel with compliance requirement documents, task description text also needs to be received. This text may be submitted by business personnel through task input channels, in the form of message queues, text streams, or forms. The input text data stream often contains irrelevant information, so it needs to be cleaned up through filtering mechanisms. Filtering methods include keyword matching, regular expression extraction, and context vector discrimination to exclude sentences and segments irrelevant to the task. After filtering, the remaining valid information is encapsulated into task description text to ensure that only semantic content related to task requirements is transmitted in subsequent processing.

[0033] In one implementation, the file monitoring mechanism can employ real-time monitoring technology based on operating system file system events, supporting the capture of add, modify, and delete operations under a specified path, and passing the identifier of the newly added file to the parsing module. The parsing module can use natural language processing tools to segment the document, extract clauses, and identify specification feature markers. The verification process can be based on a rule engine, comparing whether the clause structure is complete and detecting the coverage of markers.

[0034] In another implementation, the task input channel can be a message bus integrated into the organization's business system, supporting real-time reception of text streams. The filtering operation can utilize a semantic classification model to categorize input segments using a trained vector space, retaining only the parts relevant to the task. For example, text concerning the report scope, timeframes, and departmental responsibilities is retained, while other descriptive or incidental information is discarded. The filtered text is then encapsulated to generate a data format with a unified structure for transmission to subsequent modules.

[0035] Alternatively, a distributed architecture can be used, deploying file acquisition services and task text filtering services on distributed nodes and interacting through message queues to improve the overall system's throughput and fault tolerance. For different production environments, the parameters of the file parsing algorithm can be adjusted. For example, in scenarios with complex legal provisions, a deeper syntax analysis model can be enabled; in simpler business environments, regular expressions alone are sufficient.

[0036] Example Explanation: In the fintech business, when a local regulatory authority issues a quarterly compliance inspection notice, the system captures the notice and extracts the document identifier through a document monitoring mechanism. After parsing, a compliance requirement document containing terms and requirements is generated. Simultaneously, branch office operations personnel submit a task description text through the task input channel, stating that a quarterly risk assessment report needs to be generated. The system filters this text, removing descriptive language irrelevant to the task, retaining only content related to the time frame and data indicators, ultimately forming the task description text.

[0037] In the healthcare sector, when local health regulatory authorities issue data quality inspection guidelines, the system similarly captures the document through monitoring paths and parses it to form a compliance requirement document. Simultaneously, a primary healthcare institution inputs text describing the need to generate a drug usage report through the task input channel. The filtering module identifies statements related to drug categories, time periods, and patient coverage, eliminating redundant supplementary explanations to obtain the task description text. This process ensures effective alignment between the compliance document and task requirements, providing accurate input for subsequent data retrieval and report generation.

[0038] This embodiment, by acquiring compliance requirement documents and task description text, ensures that regulatory notifications and business requirements simultaneously enter the processing chain, achieving standardization and uniformity of document and task input. This provides a stable and accurate input foundation for subsequent intent recognition, key point extraction, and data processing. This mechanism reduces manual intervention and repetitive operations, avoiding risks caused by missing text information or formatting errors.

[0039] S20, the task description text is parsed by the intent recognition agent to generate a task intent recognition result;

[0040] In this embodiment, the process of parsing the task description text and generating the task intent recognition result through an intent recognition agent involves converting the input natural language text into a machine-understandable task semantic expression. First, a syntactic structure analysis of the task description text is required. This syntactic structure analysis is not simply word segmentation, but rather a hierarchical parsing of the text, combining syntactic dependency relations and contextual logic, to form basic semantic units. These basic semantic units can be phrases, predicate-object structures, conditional expressions, etc., and these units represent the smallest semantic fragments in the task description.

[0041] After obtaining the basic semantic units, it is necessary to extract a set of domain keywords from them. The construction of the domain keyword set depends on a domain knowledge base or a pre-set dictionary. Keywords may include time markers (quarters, years), object markers (customers, risk points), and operation markers (inspections, reports), etc. Keyword extraction can be achieved by combining statistical weighting, part-of-speech tagging with rule filtering, or by using an embedding model to calculate word vector similarity and include highly relevant words in the set.

[0042] The extracted set of domain keywords needs to be matched against a pre-built intent tag library. This intent tag library is pre-constructed and contains task types and their corresponding tag structures, such as tags like "risk assessment," "compliance self-inspection," and "rectification progress." The matching process can employ classification models, similarity calculations, or rule trees to associate keywords with tags, thereby constructing a task type topology graph. The task type topology graph is a directed graph structure where nodes represent task types or subtasks, and edges represent the logical or dependency relationships between them.

[0043] Finally, the semantic and logical consistency of the task type topology needs to be verified. The core of logical consistency verification is ensuring that the combination of keywords and task tags forms a complete and conflict-free task intent. For example, time nodes must match action nodes, and task objects cannot be bound to incorrect task types. Once the logical verification passes, the task type topology that passes the semantic and logical consistency verification is used as the task intent recognition result.

[0044] In one implementation, syntactic structure analysis can be achieved using a dependency parser, which parses the sentence into a dependency tree and extracts the predicate-object structure as the basic semantic unit based on dependency relations. When extracting keywords, high-weight words in the text can be calculated using the TF-IDF method, and then domain-specific rules can be used to filter out words irrelevant to the task.

[0045] In another implementation, a deep learning-based named entity recognition model can be used to directly identify time, location, object, and operation in the task description, thereby quickly forming a set of domain keywords. When matching the intent tag library, the BERT model can be used to calculate the semantic similarity between keywords and tag descriptions, and the highest-scoring tag is included in the task type topology graph.

[0046] A hybrid approach can also be used, combining rules and models. For example, rules can be used to identify obvious keywords first, and then a semantic matching model can be used to refine the results, thereby improving the accuracy of identification. For constructing task type topology graphs, a graph database can be used to store the relationships between labels and keywords, supporting efficient querying and consistency verification. Logical consistency verification can be performed through a rule engine to ensure that the generated topology graph conforms to the expected task structure.

[0047] Example Explanation: In the fintech business, when a business user inputs the task description text "Generate a risk assessment report for the fourth quarter of 2023," the system first performs grammatical analysis on the sentence, identifying "generate," "fourth quarter of 2023," and "risk assessment report" as different semantic units. It then extracts a set of keywords including "fourth quarter of 2023" and "risk assessment." After matching these keywords with the intent tag library, they correspond to the "risk assessment" task type. Based on this, the system constructs a task type topology graph, with the node being "risk assessment" and the subordinate attribute being "fourth quarter of 2023." Finally, the system outputs this topology graph as the task intent recognition result.

[0048] In the healthcare business field, when staff input the task description text "Write a monthly report on patient medication adherence," the system identifies the semantic units "write," "patient medication adherence check," and "monthly report" through syntactic analysis. The extracted keyword set is "patient medication adherence check" and "monthly report." After matching these keywords with the tag library, they are mapped to the "adherence check" task type, with the time attribute being "monthly." A topology graph is constructed and consistency verification is completed. The final task intent recognition result is "adherence check task, period: monthly."

[0049] This embodiment uses an intent-recognition intelligent agent to parse the task description text and generate task intent recognition results. This transforms natural language expressions into structured task semantic representations, avoiding ambiguity and biases caused by human interpretation of task requirements. The process realizes a complete link from keyword extraction to task topology map construction, enabling subsequent key point extraction and data retrieval based on explicit task semantics, thereby improving the accuracy and consistency of task processing.

[0050] S30, the intelligent agent extracts key points by processing the compliance requirement document and the task intent recognition result to generate compliance requirement key points;

[0051] In this embodiment, the process of generating compliance requirement key points by processing compliance requirement documents and task intent recognition results through a key point extraction agent first requires parsing the semantic content of the compliance requirement documents. Compliance requirement documents often contain legal provisions, regulatory instructions, and compliance details, with complex text structures and high semantic density. The parsing process involves not only segmenting the text into sentences or words, but also combining semantic dependency relationships, legal terminology databases, and logical reasoning rules to transform the raw text into structured text data. This structured text data can be in tabular form, key-value pair form, or fragment form labeled with semantic roles, ensuring that subsequent processing can proceed based on a clear data structure.

[0052] After generating structured text data, it is necessary to identify the set of obligatory clauses. Obligatory clauses refer to compliance requirements that explicitly stipulate "shall," "must," or "need," and these clauses often constitute the core content of regulatory compliance. Identification can be achieved through regularized keyword rule matching or by using a trained text classification model to divide sentences into obligatory and non-obligatory clauses. By identifying the set of obligatory clauses, the most enforceable items can be extracted from the compliance requirement documents.

[0053] The set of obligatory clauses is further refined by extracting the implementing entity, the subject of implementation, and the time frame. The implementing entity is typically an organizational department or position, such as "Risk Management Department" or "Compliance Department"; the subject of implementation is the matter to be inspected or managed, such as "customer transaction records" or "cash flow status"; and the time frame is the execution cycle specified in the clauses, such as "every quarter" or "within three days." The extraction process requires combining named entity recognition, context window analysis, and rule matching to ensure the accuracy of the extracted content.

[0054] After extraction, scenario attributes need to be extracted from the executing entity, the executing object, and the time node. Scenario attributes are feature information directly related to the task scenario. For example, in a financial business scenario, the executing object might include "credit contracts," while in a healthcare business scenario, the executing object might include "patient follow-up records." These scenario attributes help map general compliance requirements to specific task intentions.

[0055] Next, the scenario classification identifiers from the task intent recognition results are associated with the aforementioned scenario attributes to generate a scenario adaptation strategy. The scenario classification identifier is a task type description obtained from the previous intent recognition step, such as "quarterly risk assessment" or "annual compliance inspection." Aligning it with the scenario attributes generates a specific task execution strategy. Scenario adaptation strategies typically include selecting which clauses apply to the current task, which clauses need to be adjusted in wording, and how to organize the execution logic.

[0056] Based on the scenario adaptation strategy, it is necessary to reorganize the implementing entity, the implementing object, and the time node. The purpose of reorganization is to recombine the scattered semantic units of the original clauses to generate a complete set of reorganized elements. For example, one clause may mention "the compliance department shall inspect" separately, while another clause mentions "every quarter." Through reorganization, the complete expression "the compliance department shall inspect every quarter" can be obtained.

[0057] Finally, a triplet structure consisting of the implementing entity, the implementing object, and the time node is constructed based on the reorganized element set, and this triplet structure serves as the key point of compliance requirements. The triplet structure can express compliance requirements in a unified format, providing standardized input for subsequent data processing and report generation.

[0058] In one implementation, structured text data can be processed through a natural language processing pipeline, including word segmentation, part-of-speech tagging, dependency parsing, and semantic role labeling. The generated structured data is stored in tabular form, with each row corresponding to a clause and each column representing a semantic element, such as subject, object, or time.

[0059] In another implementation, a knowledge graph can be used to store and represent the set of obligatory clauses. The subjects, objects, and time nodes in the clauses are used as nodes, and the relationships between different nodes are represented by edges, thus forming a graph structure that facilitates subsequent rule reasoning and scenario matching.

[0060] A hybrid approach can also be used, first quickly identifying typical obligatory clauses through a rules engine, and then supplementing the identification with a deep learning model to improve coverage of complex clauses. Recombination operations can be implemented using graph matching algorithms to combine elements from different clauses into a complete expression. The triplet structure can be stored in JSON or RDF format for easy cross-system access.

[0061] Example Explanation: In the fintech business, when the input compliance requirement document contains the clause "The compliance department shall check the performance of credit contracts every quarter," the system parses the structured text data, identifies the clause as an obligatory clause, and extracts the implementing entity "compliance department," the implementing object "performance of credit contracts," and the time node "every quarter." Combined with the task intent recognition result "quarterly risk assessment," the system generates a scenario adaptation strategy, directly matching the clause to the task objective. The final output triple structure is (compliance department, performance of credit contracts, every quarter), serving as the key points of the compliance requirement.

[0062] In the healthcare business, when a compliance requirement document includes the clause "The Risk Management Department must complete a monthly compliance check of patient follow-up records," the system parses the document and extracts the subject "Risk Management Department," the object "Patient Follow-up Records," and the time node "Monthly." Combining this with the task intent recognition result "Monthly Compliance Check," the system generates an adaptation strategy, aligning the clause to the current task and outputting a triplet structure of (Risk Management Department, Patient Follow-up Records, Monthly) as the key compliance requirement.

[0063] This embodiment extracts key points from the intelligent agent's processing of compliance requirement documents and task intent recognition results. This transforms complex text content into clear, standardized compliance requirement points, ensuring that crucial information is not omitted or misunderstood. This process not only identifies the most core obligatory clauses but also adapts to the scenario by combining task intent, achieving a precise match between compliance requirements and task objectives. This provides a stable semantic foundation for subsequent data retrieval and report generation.

[0064] S40, the compliance requirements are executed by the data processing agent, and a data visualization is generated;

[0065] In this embodiment, the process of generating a data visualization representation by executing compliance requirements through a data processing intelligent agent can be broken down into steps such as data extraction, data access path identification, query compilation and execution, data verification, format conversion, template matching and rendering.

[0066] First, it is necessary to extract business data feature identifiers from the compliance requirements. These identifiers originate from the triple structure formed in the previous step, such as "credit contract performance status" or "patient follow-up records." These identifiers serve as entry conditions for data queries. The extraction process typically involves parsing the execution object portion of the triple, while simultaneously standardizing the feature identifiers using a domain dictionary to ensure accurate matching within the data source.

[0067] Next, it is necessary to identify the data source access paths associated with the business data feature identifiers. An access path refers to the specific location of data within an information system, such as a database table name, view name, or API call address. In fintech scenarios, an access path might correspond to a transaction log or risk event database; in healthcare scenarios, it might correspond to an electronic medical record system or follow-up record database. Identifying access paths relies on a predefined mapping database, which binds feature identifiers to specific data sources.

[0068] After identifying the access path, a structured query instruction is compiled based on the feature identifier and the access path. The structured query instruction can be an SQL statement, a GraphQL query, or a specific API request format. During the compilation process, the feature identifier expressed in natural language needs to be converted into an expression that conforms to the query syntax of the data source, for example, converting "quarterly complaint volume" into an aggregate function in an SQL statement.

[0069] The generated query is executed to obtain the raw business dataset. The raw dataset contains uncleaned or unstandardized business records, such as raw transaction logs, follow-up records, or monitoring logs. This data often contains redundant fields, missing values, or inconsistent formats, thus requiring further processing.

[0070] Data integrity verification must be performed on the original business dataset. Integrity parameters include the number of records, whether fields are missing, the reasonableness of numerical ranges, and the continuity of time series. Verification can be implemented using a validation rule engine, for example, by detecting the proportion of null values ​​in fields, the increasing pattern of timestamps, or the distribution of outliers. Only after successful verification can the data be further processed.

[0071] Data that passes validation needs to be converted to a standardized format. Standardization refers to transforming data from different sources into a unified format, such as merging multiple table structures into a single view or converting various date formats into the ISO standard format. The standardization process ensures that the data can be directly processed by downstream visualization templates, avoiding rendering failures due to format inconsistencies.

[0072] After obtaining standardized formatted data, it is necessary to match the corresponding visualization template type. The template type is usually automatically selected based on data characteristics; for example, time series data matches a line chart template, categorical distribution data matches a bar chart template, and proportional data matches a pie chart template. Template matching rules are predefined by the system or can be automatically recommended through machine learning models.

[0073] Finally, standardized format data is rendered based on the matched template type to generate a data visualization. The rendering process transforms structured data into charts, graphs, or interactive interface elements, and the output can be either a static image or an interactive visualization component.

[0074] In one implementation, business data feature identification is standardized through named entity recognition and keyword mapping, access paths are found by querying a configuration table, query commands are generated using SQL, data integrity verification is completed through predefined rules, standardized data is stored in JSON format, and finally, visualization rendering uses the ECharts engine to output charts.

[0075] In another implementation, the mapping between business data feature identifiers and access paths is achieved through a knowledge graph, query commands are implemented using GraphQL, data integrity verification is automatically completed through statistical analysis functions, standardized data is stored in RDF format, and visualization rendering uses the D3.js library to generate dynamic interactive charts.

[0076] Alternatively, it can be implemented using API calls. When the data source is distributed across multiple heterogeneous systems, the access path is called through a unified API gateway, the query command is encapsulated in the form of request parameters, data integrity verification is performed automatically when the API returns, standardization processing is completed in the middleware layer, and visualization rendering is output through the front-end component library.

[0077] For example, in the fintech business, when the compliance requirement is "the compliance department checks the performance of credit contracts quarterly," the system extracts the business data feature identifier "credit contract performance," maps it to the performance table in the transaction contract database, compiles SQL query statements, and obtains the raw performance dataset. After verifying its completeness, the data is converted into a unified table format and matched with a line chart template adapted for quarterly statistics to render a quarterly performance trend chart, generating a data visualization.

[0078] In the healthcare business, when the compliance requirement is "the risk management department checks patient follow-up records every month", the system extracts the business data feature identifier "patient follow-up records", maps it to the follow-up database path, generates a GraphQL query statement to obtain the raw follow-up data, converts it into standardized format data after integrity verification, matches it with a time series template, renders it as a monthly follow-up compliance rate line chart, and generates a data visualization representation.

[0079] This embodiment utilizes a data processing intelligence agent to generate data visualizations of key compliance requirements. This enables automatic conversion from textual compliance clauses to structured data and then to graphical displays, avoiding manual switching and copying / pasting between multiple systems. This process improves the accuracy and consistency of data acquisition and presentation, ensuring that compliance reports not only contain textual key points but also provide intuitive data charts, facilitating quick understanding and verification of compliance by reviewers.

[0080] S50: The report generation agent synthesizes the key points of the compliance requirements and data visualization representations to generate report content.

[0081] In this embodiment, the process of generating report content by synthesizing compliance requirements and data visualization representations through a report generation agent can be broken down into steps such as key point analysis, content framework construction, reference text retrieval and structure extraction, framework reorganization, visualization embedding, content fusion and language optimization.

[0082] First, the report-generating agent needs to analyze the key compliance requirements and extract the logical relationships between the elements. These elements originate from the tripartite structure generated in the previous step, such as the implementing entity, the implementing object, and the time frame. The purpose of the analysis is to identify the relationships between different elements. For example, when the implementing entity is "Compliance Department," the implementing object is "Credit Contract Performance," and the time frame is "quarter," the logical relationship obtained from the analysis is "The Compliance Department needs to check the contract performance quarterly." This logical relationship serves as the framework for the narrative text.

[0083] Secondly, after establishing the logical relationships, a narrative framework needs to be generated. A narrative framework is an abstract form of text organization, typically consisting of chapter divisions, paragraph layouts, and placeholders. For example, a framework could be generated containing sections such as "Background Information," "Data Analysis," and "Compliance Conclusions," with space reserved in each section to accommodate corresponding compliance points and charts.

[0084] After the content framework is generated, it is necessary to retrieve historical reference templates that match the key points of the current compliance requirements. These reference templates are generally drawn from institutional reporting databases or industry case studies, and their purpose is to provide templates that conform to industry standards and language styles for new reports. By analyzing these reference templates, standard paragraph structures can be extracted, such as legal background descriptions, explanations of data charts, and improvement suggestions.

[0085] After obtaining the standard paragraph structure, the narrative framework needs to be reorganized based on this structure. The reorganization process involves adjusting the original framework to be consistent with the standard paragraph structure, so that the generated report conforms to the established document logic and compliant writing standards.

[0086] Next, the data visualizations need to be embedded into the restructured narrative framework. The embedding locations are typically marked by placeholders, such as the data display area corresponding to "as shown in the chart." As the output of the previous step, the data visualizations include forms such as line charts, bar charts, and pie charts, and can be presented in a combined text and image format within designated sections.

[0087] After the framework and visualization are mapped, content integration is required, which involves combining the textual compliance points and graphical data presentations into a complete draft report. Integration is not just a simple patchwork; it also includes ensuring logical coherence, sentence coherence, and contextual consistency.

[0088] Finally, a compliance language model is applied to optimize the draft report for semantic compliance. The goal of this optimization is to correct parts with non-standard terminology, logical inconsistencies, or unprofessional expression, ensuring that the final text conforms to regulatory terminology requirements and maintains semantic consistency. The optimized text, together with the embedded data visualizations, constitutes the final report content.

[0089] In one implementation, the parsing of compliance requirements is achieved through a rule-based grammar analysis tool, the narrative framework is automatically generated from a predefined template, reference texts are taken from the organization's internal historical document database, paragraph structure extraction is completed through keyword matching, visual embedding is achieved through placeholder replacement, and language optimization is completed by a large language model fine-tuned on financial and legal texts.

[0090] In another implementation, the analysis of compliance requirements is completed through knowledge graph reasoning, the narrative content framework is dynamically generated according to report classification, reference templates are obtained through external industry knowledge base retrieval, paragraph structure is performed through text pattern recognition and extraction using a neural network model, visualization embedding is implemented using dynamic DOM mapping, and language optimization uses a contrastive learning model to improve semantic consistency.

[0091] Alternatively, an interactive configuration approach can be adopted, where users select applicable paragraph structure templates and visualization styles in the interface. The report generation agent adjusts the narrative framework in real time based on user input and performs manual confirmation before optimizing the compliant language model, thereby improving the controllability of the generated results.

[0092] Example Explanation: In the fintech business, when the key compliance requirement is "the compliance department needs to check the performance of credit contracts quarterly," the data visualization is represented as a line chart of the quarterly contract performance rate. After the report generation agent parses the logical relationship, it generates a narrative framework of "the compliance department is responsible for quarterly performance checks." It retrieves historical sample documents to extract the standard paragraph structure of "implementation background - data analysis - rectification suggestions." After reorganizing the framework, the line chart is embedded into the "data analysis" paragraph, and the sentences are optimized through a compliance language model. Finally, a report containing performance analysis and improvement suggestions is generated.

[0093] In the healthcare business, when the key compliance requirement is "the Risk Management Department needs to check the completeness of patient follow-up records every month", the data visualization is represented as a bar chart of monthly follow-up completion rate. After the report generation agent parses the logical relationship, it generates a narrative framework of "the Risk Management Department is responsible for follow-up inspection". It retrieves historical sample documents to extract the standard paragraph structure of "explanation of the implementing entity - result display - compliance summary", reorganizes the framework and embeds the bar chart into the "result display" section. It also corrects the terminology through a compliance language model and finally generates a complete patient follow-up compliance analysis report.

[0094] This embodiment uses a report generation agent to synthesize compliance requirements and data visualizations to generate report content, achieving a deep integration of textual key points and data results, thus avoiding the manual process of piecing together data and text. This approach ensures both the compliance and structure of the content while improving the automation level of report writing, making the generated report more complete, intuitive, and easier to review.

[0095] This invention relates to the field of artificial intelligence technology and can be applied to business scenarios such as fintech and healthcare. It discloses a multi-agent collaborative report generation method, apparatus, device, and medium, comprising: acquiring compliance requirement documents and task description text; generating task intent recognition results by parsing the task description text through an intent recognition agent; generating compliance requirement key points by processing the compliance requirement documents and task intent recognition results through a key point extraction agent; generating data visualization representations by executing the compliance requirement key points through a data processing agent; and generating report content by synthesizing the compliance requirement key points and data visualization representations through a report generation agent. This invention, through multi-agent collaborative work, sequentially and automatically executes the task parsing, key point extraction, data processing, and report synthesis stages, enabling compliance documents and task descriptions to be quickly transformed into structured key points, business data to be efficiently processed and intuitively presented, and ultimately automatically generating report content based on this.

[0096] In one embodiment, step S10 above includes:

[0097] S101, Monitor file change events in a preset storage path and capture newly added document identifiers in the file change events;

[0098] S102, parse the file content corresponding to the newly added document identifier to generate the original compliance text;

[0099] S103, Extract the normative feature markers from the original compliance text;

[0100] S104, verify the integrity verification parameters of the specification feature mark, and convert the verified original compliance text into a compliance requirement document;

[0101] S105, Receive text data stream from task input channel;

[0102] S106, filter out non-task description content in the text data stream, and encapsulate the filtered text data stream into task description text.

[0103] In this embodiment, a set of event listening and filtering mechanisms is used to monitor file change events in a preset storage path. The mechanism listens to native event notification interfaces provided by object storage, network file systems, or enterprise document repositories. Event sources can be directory-level additions, object-level uploads, or version updates. The scope of changes is limited by path prefixes and file type whitelists. The newly added document identifier originates from unique location information allocated by the underlying storage. This identifier can be represented as an object key, absolute file path, content hash, or version number. Upon receiving a change event, the listener enqueues the newly added document identifier for subsequent parsing pipeline consumption. During enqueueing, the timestamp, source endpoint, and triggerer information are simultaneously recorded for traceability.

[0104] The parsing engine aims to unify text decoding by analyzing the file content corresponding to newly added document identifiers. It locates the binary payload based on the identifier and selects a decoding strategy. For editable documents, it uses layout structure parsing; for scanned images, it employs layout analysis and character recognition; and for resources with encryption or access restrictions, it performs access verification and controlled decryption. During parsing, it unifies character sets, segments paragraphs, flattens tables, and removes headers and footers. The output is a combination of continuous text and structural fragments as the original compliant text. The source of the original compliant text is the semantic content of the parsed document, which includes multiple text sets such as main text, entries, appendices, and tables. It preserves original paragraph boundaries and basic layout clues to facilitate subsequent extraction.

[0105] The extraction of normative feature tags from the original compliance text is accomplished by the domain text extraction engine. The source of normative feature tags is stable and identifiable language units in compliance writing and legal provisions, covering clause numbers, obligation triggering terms, time constraint words, subject titles, object nouns, citation basis indexes, and attachment prompt indexes. During the extraction process, lexical rules, phrase dictionaries, and named entity recognition are combined to form a multi-granularity tag sequence. At the same time, the correspondence between tags and text intervals and confidence weights are generated, and the appearance order and context window of the tags are preserved to form a labeling result that can be verified and replayed.

[0106] The integrity verification parameters for the specification feature tags correspond to a set of consistency and usability checks. These checks originate from document engineering practices and compliance document quality requirements, covering character set consistency, page sequence continuity, content truncation, empty intervals in image / text parsing, the presence of required fields, whether the proportion of garbled text exceeds a threshold, and whether signatures or serial numbers are missing. The verifier cross-references the tag distribution with the full text, providing pass / fail status and location information. When converting the verified original compliance text into a compliance requirement document, a structured encapsulation process is executed. The original compliance text, specification feature tags, metadata, and version information are written into a standard container, generating a stable external identifier and an immutable content summary. A mapping table is established between newly added document identifiers and compliance requirement documents, ensuring traceability of the source and verifiability of the content.

[0107] For asynchronous text arrival scenarios, a unified entry point is established for receiving text data streams from the task input channel. The task input channel can originate from message queues, API callbacks, enterprise communication robots, or form collection terminals. The text data stream arrives in the form of incremental character fragments or serialized text blocks. The entry module tags each piece of text data with a source label, arrival time, and task session identifier, and provides it to the analysis module for consumption in a streaming buffer. Filtering non-task description content in the text data stream employs multiple filters working collaboratively. Filtering targets include greetings, signatures, disclaimers, duplicate citations, historical forwarding headers, redundant attachment prompts, and desensitized placeholders. The filters combine regular expression templates, a stop phrase library, and a text classifier to perform segment-by-segment discrimination, removing or replacing segments determined to be non-task descriptions, while retaining their identifiers. The encapsulated and filtered text data stream is used as task description text, emphasizing usability and stable interfaces. The encapsulator performs deduplication, sentence segmentation, whitespace normalization, encoding standardization, and lightweight desensitization. The cleaned text and metadata are written into a standard container, generating an external identifier for the task description text and associating it with the task session identifier. At the same time, the timeline relationship between the task description text and the compliance requirement document is recorded to ensure that subsequent processing chains can perform joint retrieval and consistency verification on the shared identifier.

[0108] This embodiment monitors the encapsulated closed-loop input acquisition process, ensuring that compliance requirement documents and task description text undergo source verification, format standardization, tag extraction, and integrity verification before entering subsequent processing. This significantly reduces recognition deviations and the probability of subsequent rework caused by differences in manual uploads, formatting variations, and the inclusion of irrelevant text. It also shortens input preparation time, reduces content offsets caused by cross-system transfers, establishes traceable identifier mappings and timeline records, and provides a stable and consistent data entry point for intent recognition and key point extraction, thereby improving the parsing accuracy and generation efficiency of subsequent links.

[0109] In one embodiment, step S20 above includes:

[0110] S201, The intention recognition agent performs syntactic structure analysis on the task description text to generate basic semantic units;

[0111] S202, Extract the set of domain keywords from the basic semantic unit;

[0112] S203, match the set of domain keywords with the preset intent tag library, and construct a task type topology map based on the matching results;

[0113] S204, verify the semantic and logical consistency of the task type topology graph, and output the task type topology graph that has passed the semantic and logical consistency verification as the task intent recognition result.

[0114] In this embodiment, syntactic structure analysis focuses on the syntactic relations and discourse boundaries of the task description text, aiming to obtain an operable analytical granularity. After processing the linked task description text, character normalization and sentence segmentation are performed first, followed by component segmentation based on dependency relations and phrase hierarchy, outputting segmented results composed of phrases, clauses, and named fragments. To reduce ambiguity caused by differences in expression, synonym mapping and abbreviation expansion are introduced, and common business abbreviations, item numbers, time expressions, and quantifiers are normalized to ensure consistent expression in subsequent extraction. The product of syntactic structure analysis is expressed as structured objects, each carrying the original text index, sentence position, semantic center words, dependency arc information, and confidence level, facilitating subsequent location and backtracking.

[0115] The basic semantic units are formed on the aforementioned structured objects, emphasizing semantic consistency and composability. The formation process aggregates phrases and clauses according to semantic core and modification relationships, distinguishing between action-related expressions, object-related expressions, constraint-related expressions, and contextual conditions, ensuring that each unit fully covers a comprehensible fragment of business intent. To guarantee reusability, the basic semantic units retain the original text fragments, semantic role tags, and cross-sentence referential links, and generate minimally modified representations to reduce noise during subsequent matching.

[0116] The domain keyword set is extracted from basic semantic units, with sources including industry thesaurus, institutional internal control thesaurus, regulatory labels, business process directories, and report item directories. The extraction process employs multi-source dictionary merging and contextual discrimination, prioritizing words at the semantic center and removing generalized words, while supplementing with synonym extensions and derivations to ensure the set both covers industry expressions and maintains discriminative power. Each keyword entry is bound to its location, the underlying semantic unit it belongs to, adjacent qualifiers, and time indicators, subsequently forming a set-level vectorized representation for alignment with the label space.

[0117] The pre-defined intent tag library is a standardized enumeration of business scenarios. Entries can cover categories such as compliance self-inspection, risk assessment, rectification progress, reporting and verification, and thematic analysis, and are organized hierarchically, including tag definitions, typical trigger expressions, key object requirements, time constraint patterns, and exclusion relationships. The tag library is derived from industry standards, internal regulations, and historical archived data, and has undergone multiple rounds of quality verification and refinement, while maintaining scalable metadata to adapt to new scenarios. To improve matching robustness, tag entries maintain both text anchors and vector anchors, with both types of anchors participating in alignment.

[0118] The matching process takes a set of domain keywords as input and aligns it with a pre-defined intent tag library using a two-channel approach: one channel compares trigger phrases and qualifiers item by item based on text anchors, while the other channel measures semantic similarity based on vector anchors. The results from both channels converge at the item level and hierarchical level, outputting a candidate tag set and its supporting evidence, including trigger source, matching location, context window, and confidence level. To avoid false triggers, exclusion relationships and sequential dependency constraints are introduced to prune the results, and conflicting tags within the same parent category are selectively retained or merged.

[0119] The task type topology graph is built upon a candidate tag set, employing a directed relational structure to express the inclusion, dependency, parallelism, and mutual exclusion between intents. During construction, tags are treated as nodes, and inclusion edges are established based on the hierarchical relationships provided by the tag library, order edges based on dependency constraints, conflict edges based on mutual exclusion constraints, and parallel edges based on co-occurrence statistics. Simultaneously, key objects, time patterns, and constraints from the domain keyword set are mapped to attributes attached to nodes and edges to support subsequent constraint checks and execution guidance. The topology object provides traversal, pruning, and projection capabilities, facilitating the output of intent views at different granularities.

[0120] Semantic logic consistency verification revolves around topological objects, checking coverage, acyclicity, constraint satisfaction, and evidence consistency. Coverage requires candidate nodes to explain key entries in the domain keyword set; acyclicity requires sequential edges not to form closed loops; constraint satisfaction requires that the attributes on nodes and edges do not conflict, for example, time patterns and object ranges need to be compatible; evidence consistency requires that each retained node can find its source anchor in the task description text. When verification passes, the topological object, after being trimmed and its attributes fixed, is fixed as the task intent recognition result, and a stable identifier, generation time, and source tracing information are generated; if verification fails, the location information of missing anchors, conflicting edges, and uncovered keywords is returned for supplementation by forward processes or manual verification. The task intent recognition result is output as a serializable container, containing the final tag set, relation structure, attribute dictionary, and evidence list, ensuring that downstream processes can directly reference tags, objects, and time patterns to perform key point extraction and data extraction.

[0121] This embodiment decomposes the task description text into basic semantic units and generates a set of domain keywords. Then, it completes dual-channel alignment and structured organization with a preset intent tag library to form a task type topology graph with relational and constraint expressions. After semantic logic consistency verification, the task intent recognition result is output. It can complete scene classification, object confirmation and time pattern fixation before entering the key point extraction and data retrieval, reduce repeated adjustments caused by ambiguous matching and scene deviation, shorten the convergence time of the overall processing chain, and establish directly consumable tags and relational inputs for downstream, improving the accuracy and stability of subsequent extraction and visualization generation.

[0122] In one embodiment, step S30 above includes:

[0123] S301, the intelligent agent extracts key points to parse the semantic content of the compliance requirement document and generates structured text data;

[0124] S302, Identify the set of obligatory clauses in the structured text data;

[0125] S303, Extract the executing entity, executing object, and time node from the set of obligatory clauses;

[0126] S304, Extract scene attributes from the execution subject, execution object and time node;

[0127] S305, Associate the scene classification identifier in the task intent recognition result with the scene attribute to generate a scene adaptation strategy;

[0128] S306, Based on the scenario adaptation strategy, reorganize the execution subject, execution object and time node to generate a set of reorganized elements;

[0129] S307, Based on the set of reorganization elements, construct a triplet structure consisting of the executing entity, the executing object, and the time node, and use the triplet structure as a key point of compliance requirements.

[0130] In this embodiment, when parsing the semantic content of the compliance requirement document, the input document first needs to undergo structured preprocessing, including item number identification, chapter division, format tag removal, and long sentence splitting. The parsing process combines a rule engine and a semantic parsing model to transform natural language expressions into structured text data with hierarchical tags and logical relationships. This structured data retains the position of clauses, binding words, verb predicates, and conditional modifiers to ensure accurate location and retrieval in subsequent steps. In this way, the text content is transformed from unstructured into operable data objects, capable of being indexed and matched within the program.

[0131] When identifying a set of obligatory clauses, it is necessary to search for expressions with normative binding force in structured text data, such as words like "must," "shall," "shall not," and "requires submission," and categorize sentences containing these expressions into the obligatory clause set. This process simultaneously filters out purely descriptive or explanatory statements to prevent clauses without practical binding force from entering the set. The obligatory clause set will form a clear set of requirements to be enforced.

[0132] When extracting the implementing entity, implementing object, and time node from the set of obligatory clauses, entity extraction is required from the sentence component annotation results. The implementing entity is typically an organizational department, position, or role, such as "branch office," "compliance department," or "audit manager." The implementing object is mostly materials, data, actions, or processes that need to be processed, such as "risk assessment report" or "customer complaint list." Time nodes are identified using time parsing tools, identifying explicit or implicit time requirements such as "monthly," "end of quarter," and "within five working days," and standardized into a uniform time expression format. This process ensures that the three core elements of the constraint are accurately extracted.

[0133] When extracting scenario attributes from the implementing entity, the implementing object, and the time node, it is necessary to make judgments based on contextual information and additional qualifiers, such as report type, business scenario, geographical scope, and regulatory agency type. These scenario attributes can help determine the appropriate environment for different provisions in specific implementation. For example, a time node with the "annual" attribute may be associated with an annual compliance inspection scenario, and if the implementing entity is the "finance department," it can be classified as a financial compliance scenario.

[0134] When associating scene classification labels with scene attributes from task intent recognition results, a two-way matching mechanism is used. On one hand, the similarity between the scene labels output from the task intent recognition results and the extracted scene attributes is calculated; on the other hand, logical checks are used to eliminate inconsistent matching results. Finally, a scene adaptation strategy is generated to guide subsequent element reorganization and adjustment. This strategy clarifies the combination of the executing subject, executing object, and time node in different scenarios, avoiding a disconnect between content and the actual task scenario.

[0135] Based on the scenario adaptation strategy, when reorganizing the executing entity, the executing object, and the time node, the three will be recombined according to the priority and dependency relationships set in the strategy. For example, "branch office," "risk investigation report," and "end of quarter" will be combined into a complete requirement, while irrelevant elements will be excluded or delayed. The result of the reorganization forms a logically consistent set of elements, which facilitates unified modeling in the future.

[0136] Finally, when constructing the triplet structure based on the reorganized element set, each triple consists of an executing entity, an executing object, and a time node. The triples are stored in a structured data manner, preserving both the source text index and supporting cross-clause merging and inference. As a key output of compliance requirements, the triplet structure provides clear operational goals and a constraint framework for downstream data retrieval, visualization, and report generation.

[0137] This embodiment performs semantic parsing on compliance requirement documents to generate structured text data, then progressively identifies the set of obligatory clauses and extracts the implementing body, implementing object, and time node, avoiding omissions and misjudgments that may occur during manual interpretation. By combining scenario attribute extraction and task intent recognition results with scenario classification identifiers, the generated scenario adaptation strategy ensures that different clause contents match actual task objectives, avoiding requirement mismatches or confusion regarding the implementing body. The final constructed triplet structure achieves a standardized expression of compliance requirements, facilitating automatic data retrieval and logical verification downstream, thereby improving the accuracy, completeness, and operability of compliance points.

[0138] In one embodiment, step S40 above includes:

[0139] S401, the data processing intelligent agent extracts the business data feature identifiers from the compliance requirements and identifies the data source access path associated with the business data feature identifiers;

[0140] S402, based on the business data feature identifier and data source access path, compile a structured query instruction, and execute the structured query instruction to obtain the original business dataset;

[0141] S403, verify the data integrity parameters of the original business dataset, and when the verification passes, convert the original business dataset into standardized format data;

[0142] S404, Match the visualization template type corresponding to the standardized format data, and render the standardized format data based on the visualization template type to generate a data visualization representation.

[0143] In this embodiment, after the compliance requirements are entered into the data processing agent, the extraction and standardization of business data feature identifiers are first completed. These business data feature identifiers originate from the combined expression of the subject, object, time node, and measurement field in the compliance requirements. Through named entity recognition, phrase segmentation, and term alignment, they form retrieval and compileable identifier units. To avoid semantic ambiguity, the identifier units are bidirectionally mapped to the domain thesaurus, indicator dictionary, and time dictionary. On the one hand, indicators such as "complaint volume," "risk exposure balance," and "number of visits" are mapped to unified indicator codes; on the other hand, time phrases such as "end of quarter," "within the statistical period," and "previous accounting period" are standardized into standard time granularity labels. After extraction, a set of contextual constraints is generated for each business data feature identifier to limit the business scope (e.g., customer base, organizational level, channel dimension) and security domain (e.g., sensitivity level, anonymization form).

[0144] Based on business data feature identifiers, the intelligent agent identifies data source access paths. The determination of access paths employs a joint strategy of directory orchestration and metadata governance: it utilizes data directory services to query the mapping relationships between metrics and data table, view, or data mart subject domains, while simultaneously reading data lineage to confirm the shortest reliable path; when multiple reachable paths exist, they are ranked using quality scores (integrity score, timeliness score, stability score) and access cost scores (cross-domain cost, cross-database cost, computational cost), selecting the path with the highest score. The access path includes not only physical location (database, table, partition, column) but also connection parameters and access control policies (identity, role, key escrow location), and generates a one-time access token to meet auditing and tracing requirements.

[0145] After obtaining the access path, the compiler enters the compilation phase of the structured query instructions. The compiler takes business data feature identifiers as input, combines the schema information and column-level lineage from the access path, and generates an abstract semantic query tree. This tree is then condensed into executable instructions based on the dialect rules of the target execution engine (relational database, distributed engine, or time-series engine). The abstract semantic query tree contains nodes for selection, filtering, aggregation, grouping, sorting, and windowing. Time nodes and organizational levels are instantiated by the connection relationships between time dimension tables and organizational dimension tables. To ensure consistency in indicator definitions, the compiler prioritizes standard aggregation expressions and definition modifiers (e.g., deduplication count, period-end balance, moving average) from the indicator definition dictionary and inserts definition comments during instruction generation for auditing purposes. The generated structured query instructions undergo syntax validation and small-sample probing in a sandbox environment. After successful validation, they are executed in the production engine to obtain the original business dataset.

[0146] After the original business dataset is returned, data integrity parameters are validated. Validation dimensions include record volume threshold, dimension coverage, key column non-null rate, primary key uniqueness, temporal continuity, value range validity, and outlier percentage. Thresholds for each dimension are loaded from a quality rule base, which is bound to the data directory by subject area to adapt to differentiated quality standards in different scenarios in finance and healthcare. The validation process employs a dual-channel comparison: one channel compares trends based on historical snapshots, and the other channel performs cross-validation based on redundant criteria (e.g., comparing differences between independent aggregation results of the same indicator from different fact tables). If any required parameter fails to meet the threshold, an auditable diagnostic report is generated and subsequent processing is halted; otherwise, standardization transformation begins.

[0147] Standardization transforms the original business dataset into standardized data format. The transformation process includes column alias standardization, data type standardization, time granularity homogenization, category dimension encoding consistency, missing value handling, and outlier identification. Time fields are uniformly mapped to a three-element structure of "start of statistical period," "end of statistical period," and "time granularity"; category dimensions are mapped to a dual-track format of stable encoding and display names through dimension tables; numerical fields are subject to unit conversion and decimal precision constraints according to the indicator dictionary; missing values ​​are handled using three strategies based on field attributes: deletion, imputation, or placeholder identification; outliers are retained as labeled columns rather than directly removed to ensure selective visibility in downstream visualizations.

[0148] After standardization, the process moves to visual template type matching. The matching engine reads field roles (measure, dimension, time), data density, category cardinality, presence of hierarchical dimensions, and whether comparative references are included from the standardized data. Combined with display intent tags (trend, comparison, proportion, distribution, correlation), the template type is determined through a rule-based priority and scoring supplementation mechanism. Rule priority is used to clarify the fit relationship; for example, time series with a single measure matches line graphs or areas; high cardinality categories with a single measure match bar graphs; dual measures changing over time match dual-axis line graphs; and geocoding matches regional base map rendering. If multiple templates have similar scores, the template with higher information efficiency and lower cognitive load is selected, and the necessary encoding parameters (axial mapping, scale domain, facet dimension, color channels, annotation strategy) are returned.

[0149] The rendering phase takes standardized format data and visualization template types as input to generate data visualization representations. The rendering pipeline includes five stages: data binding, scale calculation, primitive generation, layout and annotation, and interactive binding. In the data binding stage, dimensions and measures are mapped to positions, lengths, angles, or areas according to the template's channel definitions, ensuring consistency in quantity perception. In the scale calculation stage, linear, logarithmic, or quantile scales are selected based on the data distribution, and appropriate graduation and scale formats are set. In the primitive generation stage, basic primitives such as bars, lines, points, regions, or paths are created, and error bars, threshold lines, or target bands are overlaid as needed. In the layout and annotation stage, graphic areas, legends, and titles are arranged according to container size and facetization strategies, and text annotations or extreme value hints are generated for key data points. In the interactive binding stage, interactive handles such as filtering, hovering, highlighting, and facet switching are provided, with interactive behaviors consistent with data binding to ensure the correctness of quantization encoding is not compromised. The rendering result is output as an embeddable object, along with machine-readable graphic specification descriptions and data summaries, facilitating subsequent position mapping and explanatory text linkage during report generation.

[0150] This embodiment links the business data feature identifiers in the compliance requirements with the data catalog, indicator dictionary, and quality rule base, creating a continuous link between access path selection, instruction compilation, and quality verification. Standardization and rendering are only performed after the data quality reaches a threshold, ensuring that the displayed data has a clear source, consistent definitions, and uniform granularity. The visualization template type is obtained through a dual matching of data features and display intent, ensuring that the generated graphics are accurate in expression and highly interpretable. This achieves automated transmission from semantic requirements to graphical results, reducing errors caused by manual queries, manual splicing, and chart mismatches, improving the efficiency and consistency of graphic generation, and providing stable, auditable, and reusable visualization objects for the automatic synthesis of subsequent report content.

[0151] In one embodiment, step S50 above includes:

[0152] S501, the report generation agent analyzes the element association logic of the key points of the compliance requirements and generates a narrative content framework.

[0153] S502, retrieve historical reference templates that match the key points of the compliance requirements, and extract the standard paragraph structure from the historical reference templates;

[0154] S503, Reorganize the narrative content framework based on the standard paragraph structure;

[0155] S504, map the data visualization representation to the visualization embedding position of the reorganized narrative content framework;

[0156] S505, integrates the reorganized narrative framework with the mapped visual embedding location to generate a draft report content;

[0157] S506, Apply a compliance language model to optimize the semantic compliance of the draft report content, and generate optimized report content.

[0158] In this embodiment, the process of generating the execution report first requires parsing the element association logic of the compliance requirements. The element association logic is constructed from the previously generated triplet structure, including three types of elements: the executing subject, the executing object, and the time node. Through semantic parsing and dependency modeling, the causal, sequential, and constraint relationships between the triples are transformed into narrative structure nodes. For example, the statement "The risk management department completes the summary of customer complaint data at the end of the quarter" is mapped to the subject node "risk management department," the object node "customer complaint data," and the time node "end of the quarter," and a descriptive link is established using logical connectors. This step ensures that the report content is not merely a list of elements, but a narrative framework with inherent logical relationships.

[0159] After generating the narrative framework, it is necessary to retrieve historical reference documents that match the key points of the compliance requirements. The retrieval process combines semantic similarity models and keyword matching. Semantic similarity models capture semantic consistency between different expressions, while keyword matching ensures that key compliance terms and professional expressions are not overlooked. The retrieved historical reference documents are processed through paragraph segmentation, structured parsing, and tag extraction to obtain a standard paragraph structure. This standard paragraph structure typically includes background information, data analysis, risk descriptions, and corrective measures, and retains the organizational relationships between paragraphs to support framework reorganization.

[0160] After acquiring the standard paragraph structure, the system restructures the narrative content framework based on this structure. This restructuring process is not a simple overwrite, but rather maps the triple logic to the skeletal positions of the standard paragraphs, achieving a transition from data elements to a coherent text structure. For example, time-based information is prioritized for mapping to the background description section, the execution object combined with indicator explanations is placed in the data analysis section, and the implementing entity and compliance obligations are embedded in the rectification measures or implementation status paragraphs. This restructuring operation ensures consistency between the content logic and the report format.

[0161] Once the narrative framework is stable, the data visualizations need to be mapped to their designated embedding locations. This mapping process relies on pre-defined visualization placeholders within the narrative framework, which are bound to specific elements. For example, a "complaint volume trend" paragraph might reserve a chart insertion point. The mapping operation embeds the specific bar chart, line chart, or pie chart object into the text flow and generates chart numbers and reference identifiers to ensure consistency between subsequent text descriptions and charts.

[0162] After the visualization is embedded, the framework and charts are merged to generate a draft report. This merging process involves more than just text and image stitching; it also includes consistency checks on data references in the text descriptions. For example, the year-on-year growth rate mentioned in the text needs to match the values ​​in the charts. The system generates consistency labels through cross-checking and automatically revises the text to align with the charts when necessary, ensuring content consistency.

[0163] Finally, the draft report content was submitted to a compliance language model for semantic compliance optimization. Optimization included standardizing technical terminology, refining compliant expressions, and enhancing logical flow. For example, "customer complaints decreased" was automatically changed to "customer complaints decreased month-over-month," and transitional phrases such as "based on this" and "therefore" were inserted between paragraphs to ensure the language conformed to regulatory context and written expression standards. The optimized result formed the optimized report content, serving as the basis for subsequent review and revision.

[0164] This embodiment automatically generates reports with logical rigor, consistent graphics and text, and compliance standards by logically analyzing, structurally reorganizing, visually embedding, and semantically optimizing the key points of compliance requirements and their data visualization. Compared to manual writing or relying on a single template, this method not only reduces the risk of missing key points and data discrepancies but also improves the efficiency and consistency of report generation. This reduces the manpower required for grassroots organizations to write complex reports and enables the generated reports to better meet the actual needs of regulatory requirements and business scenarios.

[0165] In one embodiment, after step S50 above, the method further includes:

[0166] S601, The report content is presented to the interactive review interface;

[0167] S602, Receive a set of revision instructions through the interactive review interface;

[0168] S603, parse the content modification items and format adjustment items in the set of revision instructions;

[0169] S604, Locate the paragraph in the report content that needs to be modified, corresponding to the content modification item;

[0170] S605, Apply the content modification items to the paragraph to be modified to generate an updated content fragment;

[0171] S606, Match the format adjustment item with the predefined style template to generate style matching parameters;

[0172] S607, Integrate the updated content fragment and the style matching parameters into the report content to generate the final report document.

[0173] In this embodiment, after generating the report content, it needs to be presented to an interactive review interface. The interactive review interface is built based on front-end visual interaction technology and back-end report content management mechanism. The front-end interface can display the report content in a segmented and layered manner, and configure operable areas for each paragraph, table, or chart, where users can trigger revision operations. The back-end is responsible for storing the report content in a structured manner and performing two-way data binding with the front-end, so that every interface operation can be transmitted to the structural representation of the report content in real time.

[0174] In the interactive review interface, users input a set of revision instructions. These instructions are received in batches and include various types such as text modification, formatting adjustments, and chart replacement. Each instruction is encapsulated into a structure containing three parts: instruction type, target locator, and modification parameters. For example, an instruction might be "Change the title of paragraph 2 to Compliance Analysis." These instructions are triggered by interface interaction events and sent to the backend, entering a unified revision processing module.

[0175] The revision processing module first parses the set of revision instructions, dividing them into content modification items and formatting adjustment items. Content modification items refer to the semantic content of text paragraphs or chart descriptions, while formatting adjustment items affect fonts, colors, indentation, table lines, or chart styles. The parsing process is completed through pattern matching and parameter extraction, ensuring that different types of instructions can enter the corresponding processing branches.

[0176] For content modifications, the system needs to locate the paragraph in the report content that corresponds to the target to be modified. The location process relies on multiple mechanisms, including paragraph indexing, semantic tagging, and content hash comparison, to ensure precise targeting of the modification scope without affecting other parts. Once located, the system applies the modification instructions to generate an updated content fragment. This updated content fragment is constructed as an independent text block or table unit and mapped differentially to the original paragraph content, thus preserving modification traces in the report history.

[0177] For formatting adjustments, the system matches predefined style templates and generates style matching parameters. These predefined style templates store common paragraph styles, chart styles, and layout specifications for compliance reports, such as regulatory-specified title fonts or table border standards. Style matching parameters are generated through template selection and attribute inheritance to ensure that the modified report maintains a consistent overall style.

[0178] After generating the updated content snippets and style matching parameters, the system integrates them into the report content to form the updated document structure. The integration process includes content replacement, style overriding, and cross-paragraph consistency correction, ensuring that the modified paragraphs not only semantically conform to the user's instructions but also maintain visual consistency with the overall structure. Finally, the system outputs the integrated report document as the version completed through interactive review and revision.

[0179] Example Description: In health management institutions within the healthcare sector, staff are required to periodically submit compliance reports to meet the requirements of local regulatory authorities regarding chronic disease follow-up management and health data security. The entire process begins with the acquisition of compliance requirement documents and task description text. The system detects a new regulatory guidance document in a preset storage path. This document, issued by the superior health management committee, contains the latest resident health record standards and data security requirements. After the document change event is captured, the system automatically parses the document content and generates the original compliance text, further extracting regulatory feature markers such as "follow-up frequency," "data anonymization," and "information obligation." After integrity parameter verification, the text is converted into a structured compliance requirement document. Simultaneously, community health service personnel submit a task description text through the task input channel. The text explicitly requires the generation of a quarterly compliance report on the follow-up status of diabetic patients. When receiving this data stream, the system filters out non-task descriptive information, retaining only the content directly related to the task and encapsulating it as the task description text.

[0180] Subsequently, the intent recognition agent performs syntactic structure analysis on the task description text, breaking down natural language into basic semantic units, such as "generate report," "quarterly," and "diabetes follow-up." Key domain keywords are extracted from these semantic units and compared with a pre-defined intent tag library to construct a task type topology map, clarifying that the current task belongs to "generation of a chronic disease follow-up compliance report." After semantic and logical consistency verification, this topology map outputs the task intent recognition result, serving as input for subsequent key point extraction.

[0181] Key takeaways: At this stage, the intelligent agent performs semantic parsing of the compliance requirement documents, transforming them into structured text data and identifying a set of obligatory clauses. For example, the document might specify "the follow-up subject is a community general practitioner," "the target is registered diabetes patients," and "the time point is the end of each quarter." After these elements are extracted, the system further identifies the scenario attributes, such as "chronic disease management" and "resident health records." Simultaneously, the scenario classification identifier "quarterly compliance report" included in the task intent recognition results is associated with these scenario attributes to generate a scenario adaptation strategy. Based on this strategy, the system reorganizes the implementing entity, the implementing object, and the time point to form a structured set of elements, ultimately representing the compliance requirement key points in the form of triples.

[0182] After obtaining the key compliance requirements, the data processing agent begins executing data processing tasks. The system first extracts business data feature identifiers from these requirements, such as "number of follow-up visits for diabetic patients" and "glucose monitoring compliance rate." These identifiers are matched against data source access paths in the medical data warehouse to generate corresponding structured query instructions. After executing the query, the system obtains the raw business dataset, which contains data from the community health record system and follow-up record platform. The dataset undergoes integrity verification parameter checks, such as record missing rate and field consistency. Once verification is passed, the data is converted into standardized format data, such as tabular statistical results. The standardized format data is matched with appropriate visualization template types based on its characteristics; a line chart template is selected for blood glucose monitoring trends, and a bar chart template is selected for follow-up coverage. After rendering, the system generates the corresponding data visualization representation.

[0183] Subsequently, the report-generating agent is responsible for synthesizing the key compliance requirements with data visualizations into the report content. This process first parses the logical relationships between elements in the key compliance requirements to generate a narrative framework, such as "Background Information—Data Analysis—Compliance Assessment—Improvement Recommendations." The system retrieves historical reference documents from the report template library, extracts their standard paragraph structures, and uses them to reorganize the narrative framework. Data visualizations are mapped to specified visualization embedding locations within the reorganized framework, such as inserting charts in the "Data Analysis" section. The framework and visualization results are merged to generate a preliminary draft of the report content. Then, a compliance language model is used to semantically optimize the draft, making its expression more consistent with regulatory terminology and professional standards, ultimately outputting the optimized report content.

[0184] Once the report is generated, its content is presented to an interactive review interface for online review by health service department heads. Users can input a set of revision instructions, such as modifying paragraph headings, adding indicator descriptions, or adjusting table styles. The system parses these instructions, applies the content changes to the corresponding paragraphs in the draft report to generate updated segments, and matches formatting adjustments with predefined style templates to generate style matching parameters. Finally, the updated segments and style parameters are integrated back into the report, outputting the final report document. After the entire process is completed, the health management organization receives a efficiently generated, semantically accurate, and regulatory-compliant report.

[0185] In the fintech business, grassroots branches often need to frequently compile various compliance reports, such as fund flow monitoring reports, risk assessment reports, and abnormal transaction investigation reports. The process of generating these reports is extremely complex, requiring both an accurate understanding of the compliance requirements documents issued by regulatory authorities and dynamic analysis of the institution's transaction data, account information, and risk indicators. In practice, the system first monitors for updates to regulatory notices in a preset storage path. Upon detecting a new document, it automatically parses its content to generate the original compliance text and extracts regulatory feature markers with clear instructions, such as "reporting suspicious anti-money laundering transactions" and "verification of single large fund flows." After verifying the integrity parameters of these markers, the document is converted into a structured compliance requirement document. Simultaneously, compliance personnel submit task description text through a task input channel, such as "generate the Q4 2023 Anti-Money Laundering Risk Compliance Report." The system filters out non-task information and encapsulates it into task description text.

[0186] Subsequently, the intent recognition agent parses the task description text, generating basic semantic units through syntactic structure analysis and extracting keywords such as "large transaction," "quarterly analysis," and "abnormal account." These keywords are compared with a pre-set intent tag library to construct a task type topology map, clarifying that the task belongs to the anti-money laundering risk analysis compliance report category. After logical consistency verification, this topology map is input as the task intent recognition result to the next stage. The key point extraction agent parses the compliance requirement document, generating structured text data and identifying key clause sets, such as "the transaction monitoring entity is the compliance department," "the object is an account with an average daily transaction volume exceeding the threshold," and "the time node is the end of the quarter." The system further extracts the executing entity, executing object, and time node, and generates a scenario adaptation strategy based on the scenario classification identifiers in the task intent recognition result. Through this strategy, the system reorganizes these elements into a triplet structure, representing the compliance requirement key points.

[0187] The data processing agent extracts business data feature identifiers, such as "transaction frequency" and "capital inflow / outflow ratio," based on compliance requirements. After matching these identifiers with access paths in the financial transaction database, the system compiles and executes structured query instructions to retrieve the raw business dataset from the data warehouse. Data integrity parameter verification ensures that transaction data is free of missing values ​​and outliers; verified datasets are then converted into standardized data formats. The system matches appropriate visualization template types, generating line charts of quarterly large transaction trends and bar charts of suspicious account distributions. After rendering, a data visualization representation is created.

[0188] Building upon this foundation, the report-generating agent analyzes the logical elements of compliance requirements, generates a narrative framework, retrieves matching historical reference texts, extracts standard paragraph structures, reassembles the framework, and embeds data visualizations into designated locations to form a draft report. A compliance language model semantically optimizes the draft, ensuring professional terminology and compliance with regulatory conventions, generating the optimized report content. The report content is presented in an interactive review interface, where compliance officers can input a set of revision instructions, such as adjusting risk level descriptions or modifying chart annotations. The system parses these revision instructions, integrating content modifications and format adjustments into the report content, ultimately generating the final report document.

[0189] Through this process, fintech institutions can efficiently automate the generation of compliance reports when faced with complex regulatory requirements and massive amounts of business data. This avoids the deviations and risks caused by manual extraction of key points or data duplication, ensuring that the report content meets regulatory expectations and can be delivered quickly.

[0190] This embodiment links the report content with the interactive review interface and supports the parsing and application of revision command sets, enabling parallel processing of content modification and format adjustment. Users no longer need to manually revise repeatedly across multiple file versions; instead, they can complete modifications by directly inputting commands through a centralized interface. The report content automatically inherits predefined style templates during updates, making the modification process both efficient and standardized. The final generated report document not only reflects the user's personalized revision needs but also maintains overall compliance and professional formatting, reducing manual editing errors and improving the efficiency of the review and finalization stages.

[0191] In one embodiment, a multi-agent collaborative report generation apparatus is provided, which corresponds one-to-one with the multi-agent collaborative report generation method described in the above embodiments. (Refer to...) Figure 3 , Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the multi-agent collaborative report generation device of the present invention. The module includes an input acquisition module 10, an intent recognition module 20, a key point extraction module 30, a data processing module 40, and a report generation module 50. Detailed descriptions of each functional module are as follows:

[0192] Input acquisition module 10 is used to acquire compliance requirement documents and task description text;

[0193] The intent recognition module 20 is used to parse the task description text through the intent recognition agent and generate a task intent recognition result;

[0194] The key point extraction module 30 is used to process the compliance requirement document and the task intent recognition result through the key point extraction intelligent agent to generate compliance requirement key points;

[0195] Data processing module 40 is used to execute the compliance requirements through a data processing intelligent agent and generate a data visualization representation;

[0196] The report generation module 50 is used to synthesize the key points of the compliance requirements and data visualization representation through the report generation agent to generate report content.

[0197] In one embodiment, the input acquisition module 10 is specifically used for:

[0198] Monitor file change events in a preset storage path and capture newly added document identifiers in the file change events;

[0199] Parse the file content corresponding to the newly added document identifier to generate the original compliance text;

[0200] Extract the normative feature markers from the original compliance text;

[0201] Verify the integrity verification parameters of the specified feature markers, and convert the verified original compliance text into a compliance requirement document;

[0202] Receive text data stream from the task input channel;

[0203] Filter out non-task description content from the text data stream and encapsulate the filtered text data stream into task description text.

[0204] In one embodiment, the intent recognition module 20 is specifically used for:

[0205] The task description text is analyzed by an intent-recognition intelligent agent to generate basic semantic units.

[0206] Extract the set of domain keywords from the basic semantic units;

[0207] The domain keyword set is matched with a preset intent tag library, and a task type topology map is constructed based on the matching results;

[0208] Verify the semantic and logical consistency of the task type topology graph, and output the task type topology graph that passes the semantic and logical consistency verification as the task intent recognition result.

[0209] In one embodiment, the key point extraction module 30 is specifically used for:

[0210] The intelligent agent extracts key points to parse the semantic content of the compliance requirement document and generates structured text data.

[0211] Identify the set of obligatory clauses in the structured text data;

[0212] Extract the executing entities, executing objects, and time points from the set of obligatory clauses;

[0213] Extract scene attributes from the execution subject, execution object, and time node;

[0214] By associating the scene classification identifier in the task intent recognition result with the scene attribute, a scene adaptation strategy is generated;

[0215] Based on the scenario adaptation strategy, the execution subject, execution object and time node are reorganized to generate a set of reorganized elements;

[0216] Based on the aforementioned set of reorganization elements, a triplet structure consisting of the executing entity, the executing object, and the time node is constructed, and the triplet structure is used as a key point of compliance requirements.

[0217] In one embodiment, the data processing module 40 is specifically used for:

[0218] The data processing intelligent agent extracts the business data feature identifiers from the key points of the compliance requirements and identifies the data source access path associated with the business data feature identifiers.

[0219] Based on the business data feature identifier and data source access path, compile a structured query instruction, and execute the structured query instruction to obtain the original business dataset;

[0220] Verify the data integrity parameters of the original business dataset, and when the verification passes, convert the original business dataset into standardized format data.

[0221] Match the standardized format data with the corresponding visualization template type, and render the standardized format data based on the visualization template type to generate a data visualization representation.

[0222] In one embodiment, the report generation module 50 is specifically used for:

[0223] The report generation agent analyzes the element association logic of the key compliance requirements and generates a narrative framework.

[0224] Retrieve historical reference documents that match the key points of the compliance requirements, and extract the standard paragraph structure from the historical reference documents;

[0225] The narrative framework is reorganized based on the standard paragraph structure.

[0226] Map the data visualization representation to the visualization embedding location of the reorganized narrative content framework;

[0227] By integrating the reorganized narrative framework with the visualized embedding locations of the mappings, a draft report is generated.

[0228] The semantic compliance of the draft report content is optimized using a compliance language model to generate the optimized report content.

[0229] In one embodiment, the report generation module 50 is specifically used for:

[0230] The report content is displayed on the interactive review interface;

[0231] Receive a set of revision instructions through the interactive review interface;

[0232] Analyze the content modification items and format adjustment items in the set of revision instructions;

[0233] Locate the paragraph in the report content that corresponds to the content modification item and needs to be modified;

[0234] Apply the content modification items to the paragraph to be modified to generate an updated content fragment;

[0235] Match the format adjustment items with the predefined style template to generate style matching parameters;

[0236] The updated content fragments and the style matching parameters are integrated into the report content to generate the final report document.

[0237] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external user terminals via a network connection. When executed by the processor, the computer program implements the functions or steps of a multi-agent cooperative report generation method on the server side.

[0238] In one embodiment, a computer device is provided, which may be a user terminal, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements the user-side functions or steps of a multi-agent cooperative report generation method.

[0239] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:

[0240] Obtain compliance requirement documents and task description text;

[0241] The task description text is parsed by the intent recognition agent to generate task intent recognition results;

[0242] The intelligent agent extracts key points by processing the compliance requirement documents and task intent recognition results to generate compliance requirement key points.

[0243] The compliance requirements are executed by a data processing intelligent agent, generating a data visualization representation.

[0244] The report generation agent synthesizes the key compliance requirements and data visualizations to generate report content.

[0245] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0246] Obtain compliance requirement documents and task description text;

[0247] The task description text is parsed by the intent recognition agent to generate task intent recognition results;

[0248] The intelligent agent extracts key points by processing the compliance requirement documents and task intent recognition results to generate compliance requirement key points.

[0249] The compliance requirements are executed by a data processing intelligent agent, generating a data visualization representation.

[0250] The report generation agent synthesizes the key compliance requirements and data visualizations to generate report content.

[0251] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and user side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0252] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0253] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0254] It should be noted that if any software tools or components not belonging to this company appear in the embodiments of this application, they are merely illustrative examples and do not represent actual use. The embodiments described above are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for multi-agent collaborative report generation, the method comprising: Includes the following steps: Obtain compliance requirement documents and task description text; The task description text is parsed by the intent recognition agent to generate task intent recognition results; The intelligent agent extracts key points by processing the compliance requirement documents and task intent recognition results to generate compliance requirement key points. The compliance requirements are executed by a data processing intelligent agent, generating a data visualization representation. The report generation agent synthesizes the key compliance requirements and data visualizations to generate report content.

2. The multi-agent collaborative report generation method of claim 1, wherein, Obtain compliance requirement documents and task description text, including: Monitor file change events in a preset storage path and capture newly added document identifiers in the file change events; Parse the file content corresponding to the newly added document identifier to generate the original compliance text; Extract the normative feature markers from the original compliance text; Verify the integrity verification parameters of the specified feature markers, and convert the verified original compliance text into a compliance requirement document; Receive text data stream from the task input channel; Filter out non-task description content from the text data stream and encapsulate the filtered text data stream into task description text.

3. The multi-agent collaborative report generation method of claim 1, wherein, The task intent recognition agent parses the task description text to generate task intent recognition results, including: The task description text is analyzed by an intent-recognition intelligent agent to generate basic semantic units. Extract the set of domain keywords from the basic semantic units; The domain keyword set is matched with a preset intent tag library, and a task type topology map is constructed based on the matching results; Verify the semantic and logical consistency of the task type topology graph, and output the task type topology graph that passes the semantic and logical consistency verification as the task intent recognition result.

4. The multi-agent cooperative report generation method as described in claim 1, characterized in that, The intelligent agent processes the compliance requirement document and task intent recognition results through key point extraction to generate compliance requirement key points, including: The intelligent agent extracts key points to parse the semantic content of the compliance requirement document and generates structured text data. Identify the set of obligatory clauses in the structured text data; Extract the executing entities, executing objects, and time points from the set of obligatory clauses; Extract scene attributes from the execution subject, execution object, and time node; By associating the scene classification identifier in the task intent recognition result with the scene attribute, a scene adaptation strategy is generated; Based on the scenario adaptation strategy, the execution subject, execution object and time node are reorganized to generate a set of reorganized elements; Based on the aforementioned set of reorganization elements, a triplet structure consisting of the executing entity, the executing object, and the time node is constructed, and the triplet structure is used as a key point of compliance requirements.

5. The multi-agent cooperative report generation method as described in claim 1, characterized in that, The compliance requirements are executed by a data processing intelligent agent, generating a data visualization representation, including: The data processing intelligent agent extracts the business data feature identifiers from the key points of the compliance requirements and identifies the data source access path associated with the business data feature identifiers. Based on the business data feature identifier and data source access path, compile a structured query instruction, and execute the structured query instruction to obtain the original business dataset; Verify the data integrity parameters of the original business dataset, and when the verification passes, convert the original business dataset into standardized format data. Match the standardized format data with the corresponding visualization template type, and render the standardized format data based on the visualization template type to generate a data visualization representation.

6. The multi-agent cooperative report generation method as described in claim 1, characterized in that, The report generation agent synthesizes the key compliance requirements and data visualizations to generate report content, including: The report generation agent analyzes the element association logic of the key compliance requirements and generates a narrative framework. Retrieve historical reference documents that match the key points of the compliance requirements, and extract the standard paragraph structure from the historical reference documents; The narrative framework is reorganized based on the standard paragraph structure. Map the data visualization representation to the visualization embedding location of the reorganized narrative content framework; By integrating the reorganized narrative framework with the visualized embedding locations of the mappings, a draft report is generated. The semantic compliance of the draft report content is optimized using a compliance language model to generate the optimized report content.

7. The multi-agent cooperative report generation method as described in claim 1, characterized in that, After generating the report content by synthesizing the key compliance requirements and data visualizations through a report generation agent, it also includes: The report content is displayed on the interactive review interface; Receive a set of revision instructions through the interactive review interface; Analyze the content modification items and format adjustment items in the set of revision instructions; Locate the paragraph in the report content that corresponds to the content modification item and needs to be modified; Apply the content modification items to the paragraph to be modified to generate an updated content fragment; Match the format adjustment items with the predefined style template to generate style matching parameters; The updated content fragments and the style matching parameters are integrated into the report content to generate the final report document.

8. A multi-agent collaborative report generation device, characterized in that, The multi-agent collaborative report generation device includes: The input acquisition module is used to obtain compliance requirement documents and task description text; The intent recognition module is used to parse the task description text through an intent recognition agent and generate task intent recognition results; The key point extraction module is used to process the compliance requirement document and task intent recognition results through the key point extraction intelligent agent to generate compliance requirement key points; The data processing module is used to execute the compliance requirements through a data processing intelligent agent and generate a data visualization representation; The report generation module is used to synthesize the key compliance requirements and data visualization representations through a report generation agent to generate report content.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a multi-agent cooperative report generation program stored in the memory and executable on the processor. When executed by the processor, the multi-agent cooperative report generation program implements the steps of the multi-agent cooperative report generation method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a multi-agent cooperative report generation program, which, when executed by a processor, implements the steps of the multi-agent cooperative report generation method as described in any one of claims 1-7.

Images