An AI large model-based data asset intelligent report generation method and system

Abstract

The application belongs to the technical field of data asset management and artificial intelligence text generation, and discloses a data asset intelligent report generation method and system based on an AI large model. The method receives a report generation request, generates a report chapter-index-evidence constraint table, extracts asset evidence units from a data asset directory, metadata, blood relationship, quality detection, classification and grading, authorization, calling, cost accounting records, income estimation records and data service settlement records, filters report fact packages according to permissions and sensitivity levels, inputs the report fact packages and constraints into an AI large model to generate candidate report content, reads and checks values, sensitivity levels and effective intervals according to evidence indexes, executes restricted rewriting on sentences and paragraphs that do not pass, and finally outputs a report and generates a report version chain. The method can improve report generation efficiency and enhance the consistency, traceability and permission security of report content.

Description

Technical Field

[0002] This invention relates to the fields of data asset management and artificial intelligence text generation technology, specifically to a method and system for generating intelligent data asset reports based on an AI large model. It is applicable to the automatic generation and evidence verification of reports in scenarios such as data asset inventory, data quality analysis, data asset operation, data asset compliance review, and data asset entry assistance. Background Technology

[0004] With the continuous advancement of digital transformation in enterprises, government agencies, and industry platforms, data resources such as data tables, data interfaces, data products, and indicator datasets are gradually being incorporated into the scope of data asset management. During data asset operation, it is typically necessary to regularly generate data asset inventory reports, quality reports, classification and grading reports, operational reports, value analysis reports, and compliance reports. These reports usually involve multiple types of data, including data asset catalogs, metadata, lineage information, quality inspection results, classification and grading information, authorization records, call records, cost accounting records, revenue calculation records, and data service settlement records. Their content includes statistical indicators such as asset quantity, number of interface calls, quality scores, and authorization status, as well as natural language descriptions such as risk assessments, disposal recommendations, and operational analyses. Therefore, improving report generation efficiency while ensuring accurate data sources, clear permission boundaries, and traceable report conclusions has become a crucial issue in data asset management.

[0005] In existing technologies, there are solutions for report generation based on report templates and data model asset catalogs. For example, prior art document CN119597820B discloses a report generation method and system based on a data model asset catalog. This method constructs a data model asset catalog, selects report data models, report generation rules, and report templates, configures report parameters and permissions based on business permissions, and then imports the associated information from the business data into the report template to obtain the target report, thereby lowering the threshold for report generation and improving its efficiency. This type of solution can solve the problem of template-based generation of some structured reports, but its focus is on data model selection, template configuration, and business data import, and it is generally more suitable for report output with fixed fields, fixed indicators, and fixed formats.

[0006] With the development of large-scale AI models, using them for report writing, summary generation, and text polishing has become a viable approach. Large-scale AI models can generate more natural-looking paragraphs from input materials, reducing the workload of manual writing. However, in data asset reporting scenarios, directly inputting statistical results, asset catalogs, or partial governance data into AI models can easily lead to several problems: First, the model may generate values, proportions, rankings, risk levels, or conclusive statements that are not supported by data evidence based on context, resulting in inconsistencies between the report content and the actual data asset records. Second, data asset reports contain a large number of sensitive fields, important data markers, authorization scopes, and user permission information; direct input or output may result in unauthorized exposure of sensitive fields. Third, existing automated report or text generation methods typically lack sentence-level evidence indexes, making it difficult to trace specific values, conclusions, and risk judgments in the report back to the corresponding data asset evidence. Fourth, when metadata, lineage, quality inspection, authorization, or call records are updated, the report needs to be re-verified of relevant sections; existing technologies struggle to accurately locate affected sentences and sections, leading to repeated generation and a large amount of manual review. Fifth, for incorrectly generated sentences, ordinary rewriting methods usually still rely on the model's free generation and cannot limit modifications to be made solely based on the corresponding evidence units.

[0007] Therefore, existing data asset report generation technologies still need a technical solution that can establish constraints on report chapters, indicators, and evidence, perform permission filtering on input facts before generating the large model, add evidence indexes to key sentences and segments during generation, perform readback verification of values ​​and conclusions after generation, and perform restricted rewriting when verification fails, in order to reduce the problems of report content distortion, leakage of sensitive information, and untraceable versions. Summary of the Invention

[0009] The purpose of this invention is to provide a method and system for generating intelligent data asset reports based on AI large models, so as to improve the efficiency of data asset report generation, reduce numerical distortion, conclusions without evidence and erroneous output of sensitive fields, and improve the traceability of report content and the efficiency of version review.

[0010] To achieve the above objectives, the present invention provides the following technical solution.

[0011] In a first aspect, the present invention provides a method for generating intelligent reports on data assets based on an AI large model, comprising the following steps:

[0012] S1. Receive the report generation request, parse the report purpose, report target, report user and use case, and call the chapter template library to generate the report chapter-indicator-evidence constraint table;

[0013] S2. Extract asset evidence units from the data asset catalog, metadata, lineage, quality inspection, classification and grading, authorization, access, cost accounting records, revenue calculation records and data service settlement records. Each asset evidence unit is bound to the asset identifier, field name, indicator value, source location, timestamp, lineage path, sensitivity level, valid range and evidence summary.

[0014] S3. Match asset evidence units according to the report chapter - indicator - evidence constraint table, and filter them according to the report user's permissions and sensitivity level to generate a report fact package containing fields that can be displayed, evidence index, indicator calculation results and valid ranges;

[0015] S4. Input the report fact package, chapter writing constraints, evidence citation format and prohibition of extrapolation constraints into the AI ​​big model to generate candidate report content. The numerical fields, conclusion fields and risk level fields in the candidate report content are all marked with evidence index tags.

[0016] S5. Analyze the candidate report content, and read back the corresponding indicator calculation results, sensitivity level and valid range from the report fact package based on the evidence index carried by the sentence segment. When the values ​​are inconsistent, the evidence index is missing, the evidence exceeds the valid range or the sensitivity level does not meet the permissions of the report user, the corresponding sentence segment is marked as an unoutputable sentence segment, and a rewrite instruction containing the error type and the corresponding asset evidence unit is generated and returned to the AI ​​big model for rewriting.

[0017] S6. Output a verified data asset report and generate a report version chain, which records the report fact package summary, prompt word summary, model version summary, chapter summary, and evidence index.

[0018] Specifically, in step S1, the report generation request includes the report type, the scope of the report recipients, the report users, the report generation cycle, and the report output format;

[0019] The report types include one or more of the following: data asset inventory report, data asset quality report, data asset classification and grading report, data asset value analysis report, data asset operation report, data asset compliance report, and data asset entry auxiliary report;

[0020] The report chapter-indicator-evidence constraint table includes chapter identifier, chapter title, required indicators, optional indicators, indicator calculation rules, required evidence type, minimum credibility level of evidence, sensitive field processing rules, evidence citation format, and chapter output format.

[0021] Specifically, in step S2, the extraction of asset evidence units includes:

[0022] Standardize the mapping of asset names, asset numbers, departments, business themes, data tables, data fields, interface names, and data product names in the data asset catalog to obtain asset identifiers and field names;

[0023] Associate metadata and lineage to generate a lineage path from the source system, processing task, intermediate table, indicator table to the data product;

[0024] Extract completeness, accuracy, consistency, timeliness, and uniqueness indicators from quality inspection;

[0025] Extract data categories, data levels, personal information tags, important data tags, and exit restriction tags from the classification and grading system;

[0026] Extract the authorization scope, number of calls, caller type, construction cost, operation and maintenance cost, revenue calculation amount, service settlement frequency and service settlement cycle from authorization, invocation, cost accounting records, revenue calculation records and data service settlement records;

[0027] Generate an evidence summary based on the asset identifier, field name, indicator value, source location, timestamp, and lineage path.

[0028] Specifically, in step S3, when the same reporting indicator matches multiple asset evidence units, the credibility of the asset evidence units is calculated. :

[0029]

[0030] In the formula, For asset evidence unit The credibility of the evidence; The score is for quality inspection. The source system's trust level; For the integrity of bloodline; Score for time efficiency; This is the conflict penalty value; , , , , These are the weighting coefficients;

[0031] Asset evidence units whose credibility reaches the minimum level of credibility are selected and included in the report fact package; when the difference in indicator values ​​of multiple asset evidence units under the same reporting indicator exceeds the preset conflict threshold, an indicator conflict mark is generated and the corresponding reporting indicator is included in the set of indicators to be reviewed.

[0032] Specifically, in step S3, the generated report fact package includes:

[0033] Write the asset evidence units into the report fact package according to the chapter identifier and indicator name;

[0034] Fields that meet the requirements of report users' permissions, authorization status, and sensitivity level are identified as fields that can be displayed externally;

[0035] Replace fields that do not meet the display conditions with de-identified summaries, range-based values, enumeration levels, or local retention location indexes;

[0036] The evidence index is generated by combining the storage location of the asset evidence unit in the report fact package, the evidence summary, and the chapter identifier;

[0037] The report fact package includes a fact package identifier, a report object identifier, a chapter identifier, an indicator name, an indicator calculation result, fields that can be displayed, an evidence index, an evidence summary, a valid range, indicator conflict markers, and sensitive field replacement results.

[0038] Specifically, in step S4, the chapter writing constraints include chapter title, chapter word count range, narrative order, definition of indicators, glossary, and list of prohibited terms;

[0039] The prohibition on extrapolation includes: the content of candidate reports shall not generate specific values, rankings, proportions, ownership conclusions, compliance conclusions, value conclusions, service settlement conclusions, and risk levels that are not supported by the evidence index;

[0040] When the AI ​​big model generates candidate report content, it adds evidence index tags to sentences containing numerical fields, conclusion fields, or risk level fields, and outputs a mapping table between sentences, field types, and evidence indexes.

[0041] Specifically, in step S5, the parsing of candidate report content includes:

[0042] Identify numerical fields, conclusion fields, risk level fields, and general description fields in the candidate report content;

[0043] Compare the numerical fields with the calculation results of the indicators pointed to by the corresponding evidence index;

[0044] Compare the conclusion field with the evidence type, evidence credibility, and valid range of the asset evidence unit pointed to by the corresponding evidence index;

[0045] Compare the risk level field with the risk level generation rules in the report section—indicators—evidence constraint table;

[0046] If any comparison result does not meet the corresponding rule, the segment containing that field is marked as an unoutput segment.

[0047] Specifically, in step S5, the rewrite instruction includes a segment identifier, error type, field type, corresponding asset evidence unit, allowed indicator calculation results, allowed evidence index, and rewrite restrictions.

[0048] The error types include one or more of the following: inconsistent values, missing evidence index, evidence exceeding the valid range, sensitivity level not meeting the report user's permissions, conclusions lacking evidence support, and risk level not meeting the generation rules.

[0049] After the AI ​​big model regenerates the candidate segments according to the rewriting instructions, it performs the verification in step S5 again until the candidate segments pass the verification or reach the preset number of rewritings. When the preset number of rewritings is reached and the verification still fails, the corresponding segments are transferred to the manual review queue.

[0050] Specifically, in step S6, the report version chain includes a report version number, a summary of the previous report version, a summary of the report fact package, a summary of the prompt words, a summary of the model version, a summary of the chapter, a set of evidence indexes, records of non-output segments, rewrite records, and a set of changed fields;

[0051] When the data asset catalog, metadata, lineage, quality inspection, classification and grading, authorization, access, cost accounting records, revenue calculation records or data service settlement records are updated, the affected asset evidence units are located based on the set of changed fields, and the affected chapters are determined based on the evidence index set.

[0052] Only the affected chapters will have their candidate report content regenerated and validated; unaffected chapters will retain the chapter summary and content from the previous report version.

[0053] Secondly, the present invention also provides a data asset intelligent report generation system based on an AI large-scale model, used to implement the data asset intelligent report generation method based on an AI large-scale model described in the first aspect, the system comprising:

[0054] The request parsing module is used to receive report generation requests, parse the report purpose, report object, report user and use case, and call the chapter template library to generate the report chapter-indicator-evidence constraint table;

[0055] The evidence extraction module is used to extract asset evidence units from the data asset catalog, metadata, lineage, quality inspection, classification and grading, authorization, access, cost accounting records, revenue calculation records and data service settlement records;

[0056] The fact package generation module is used to match asset evidence units according to the report chapter-indicator-evidence constraint table, and filter them according to the report user's permissions and sensitivity level to generate a report fact package. The fact package generation module includes an evidence matching unit, a permission filtering unit, a sensitive field replacement unit, and an evidence index generation unit. The evidence matching unit is used to match asset evidence units according to the report chapter-indicator-evidence constraint table. The permission filtering unit is used to filter fields that can be displayed according to the report user's permissions and authorization status. The sensitive field replacement unit is used to replace fields that do not meet the display conditions with de-identified summaries, interval values, enumeration levels, or local retention location indexes. The evidence index generation unit is used to combine the storage location of the asset evidence unit in the report fact package, the evidence summary, and the chapter identifier to generate an evidence index.

[0057] The large model generation module is used to input the report fact package, chapter writing constraints, evidence citation format and prohibition of extrapolation constraints into the AI ​​large model to generate candidate report content with evidence index tags;

[0058] The verification and rewriting module is used to parse the content of candidate reports, read back the corresponding indicator calculation results, sensitivity level and effective range from the report fact package based on the evidence index carried by the sentence segment, mark the sentence segment that fails the verification as an unoutputable sentence segment, and generate a rewriting instruction to return to the AI ​​big model for rewriting.

[0059] The version management module is used to output verified data asset reports and generate report version chains.

[0060] Thirdly, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the method described in the first aspect.

[0061] This invention establishes a correspondence between report structure, report indicators, and required evidence through a report chapter-indicator-evidence constraint table. Through asset evidence units and report fact packages, it transforms data asset catalogs, metadata, lineage, quality inspection, classification and grading, authorization, access, cost accounting records, revenue calculation records, and data service settlement records into verifiable, indexable, and access-controlled input objects, ensuring that the AI ​​model can only generate candidate report content based on filtered factual data. By attaching evidence indexes to numerical fields, conclusion fields, and risk level fields, and by rereading the corresponding indicator calculation results, sensitivity levels, and valid ranges from the report fact package before output, it reduces inconsistencies between the content generated by the large model and real data. For segments that fail verification, the system generates rewrite instructions containing the error type and the corresponding asset evidence unit, limiting the rewrite process to the scope of evidence. Simultaneously, the report version chain records fact package summaries, prompt word summaries, model version summaries, chapter summaries, and evidence indexes, facilitating subsequent review, incremental updates, and accountability tracking. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of the structure of an AI-based big data asset intelligent report generation system. Figure 2 A flowchart illustrating the method for generating intelligent reports on data assets based on AI large-scale models. Detailed Implementation

[0064] The embodiments of the present invention will be described below with reference to the accompanying drawings and examples. It should be understood that the following embodiments are for illustrating the technical solutions of the present invention and are not intended to limit the scope of protection. Where there is no conflict, the technical features in the following embodiments can be combined with each other.

[0065] I. Terminology Explanation

[0066] In this embodiment, the data asset catalog refers to the asset list formed after enterprises, government agencies or other organizations register data resources such as data tables, data fields, data interfaces, data products, indicator datasets, etc.

[0067] Metadata refers to data that describes information such as the structure, meaning, format, update cycle, system to which it belongs, field type, field definition, indicator definition, and interface parameters of data assets.

[0068] The lineage refers to the source, processing, and output path of data assets from the source system, collection task, cleaning task, processing task, intermediate table, indicator table to the data product.

[0069] Quality inspection refers to the record formed by inspecting quality indicators such as the integrity, accuracy, consistency, timeliness, and uniqueness of data assets.

[0070] Classification and grading refers to identifying data assets according to rules such as data category, data level, personal information tagging, important data tagging, and exit restriction tagging.

[0071] The Report Chapter-Indicator-Evidence Constraint Table is a structured table used to constrain the data asset report generation process. It includes at least a chapter identifier, chapter title, required indicators, optional indicators, indicator calculation rules, required evidence types, minimum credibility level of evidence, sensitive field processing rules, evidence citation format, and chapter output format.

[0072] An asset evidence unit refers to a verifiable data object that is extracted and structurally encapsulated from data asset catalogs, metadata, lineage records, quality inspection records, classification and grading records, authorization records, access records, cost accounting records, revenue calculation records, and data service settlement records. An asset evidence unit must at least bind an asset identifier, field name, indicator value, source location, timestamp, lineage path, sensitivity level, valid range, and evidence summary.

[0073] A report fact package refers to the set of inputs that can be used in a large model before generating a data asset report, after filtering based on the report's chapters, metrics, evidence constraint tables, user permissions, and sensitivity levels. The report fact package includes publicly available fields, evidence indexes, metric calculation results, and valid ranges, but does not directly expose raw fields that do not meet permission or sensitivity level requirements.

[0074] Candidate report content refers to the text content to be verified generated by the AI ​​big model based on the report fact package, chapter writing constraints, evidence citation format, and prohibition of extrapolation constraints.

[0075] An evidence index is an index identifier used to trace back from a segment in the candidate report content to the corresponding asset evidence unit in the report's fact package. An evidence index can be formed by combining fact package identifiers, chapter identifiers, indicator names, asset evidence unit identifiers, and evidence summaries.

[0076] Unoutputtable segments refer to segments that, after verification, are found to have inconsistencies in values, missing evidence indexes, evidence exceeding the valid range, sensitivity levels not meeting the report user's permissions, conclusions lacking supporting evidence, or risk levels not conforming to the generation rules, and therefore cannot be included in the final data asset report candidate report segments.

[0077] A report version chain refers to a version record chain that tracks the report generation process. It records at least the report version number, the previous report version summary, the report fact package summary, the prompt word summary, the model version summary, the chapter summary, the evidence index set, the non-output segment record, the rewrite record, and the set of changed fields.

[0078] II. System and Software Implementation Methods

[0079] like Figure 1 As shown, the data asset intelligent report generation system of this embodiment can be deployed in enterprise data asset operation platforms, government data resource management platforms, data governance platforms, or data asset entry auxiliary platforms. The system includes a request parsing module, an evidence extraction module, a fact package generation module, a large model generation module, a verification rewriting module, and a version management module.

[0080] The request parsing module receives report generation requests. These requests can be submitted by users on the front-end page or triggered by a scheduled task. Each request includes the report type, target audience, user, generation cycle, and output format. For example, a report type of "Data Asset Operation Report," targeting data assets in the "Population Services," "Enterprise Services," and "City Operations" categories of a municipal government data resource catalog, targeting the data operations department, generating monthly, and outputting as a Word document or webpage report.

[0081] The request parsing module analyzes the report's purpose, target audience, users, and usage scenarios, and then uses the chapter template library to generate a report chapter-indicator-evidence constraint table. The chapter template library can pre-set templates for data asset inventory reports, data asset quality reports, data asset classification and grading reports, data asset value analysis reports, data asset operation reports, data asset compliance reports, and data asset entry auxiliary reports. Different templates correspond to different chapter structures, indicator ranges, evidence requirements, and output formats.

[0082] The evidence extraction module is used to extract asset evidence units from multiple data sources. These data sources include data asset catalogs, metadata, lineage records, quality inspection records, classification and grading records, authorization records, access records, cost accounting records, revenue calculation records, and data service settlement records. These data sources can originate from databases, data lakes, data warehouses, metadata management systems, data quality platforms, access control systems, API gateways, cost accounting systems, and data service settlement systems. The evidence extraction module performs field mapping and format standardization on data from different sources and generates an evidence summary for each asset evidence unit.

[0083] The fact package generation module is used to generate the report fact package. This module includes an evidence matching unit, a permission filtering unit, a sensitive field replacement unit, and an evidence index generation unit. The evidence matching unit matches asset evidence units to their corresponding chapters and indicators according to the report chapter-indicator-evidence constraint table; the permission filtering unit filters fields that can be displayed based on the report user's permissions and authorization status; the sensitive field replacement unit replaces fields that do not meet the display conditions with anonymized summaries, range-based values, enumeration levels, or local retention location indexes; and the evidence index generation unit combines the storage location of the asset evidence units in the report fact package, the evidence summary, and the chapter identifier to generate an evidence index.

[0084] The large model generation module is used to invoke the AI ​​large model. The AI ​​large model does not directly access the data asset catalog, metadata, lineage, classification and grading, authorization system, or data service settlement system; instead, it only receives report fact packages, chapter writing constraints, evidence citation formats, and prohibition on extrapolation constraints. In this way, the AI ​​large model can only use filtered, indexed, and constrained factual data when generating content.

[0085] The verification and rewriting module is used to parse and verify the content of candidate reports. It includes a field parsing unit, a fact package readback unit, a sentence / segment determination unit, and a rewriting control unit. The field parsing unit identifies numerical fields, conclusion fields, risk level fields, and general description fields in the candidate report content; the fact package readback unit reads the corresponding indicator calculation results, sensitivity levels, valid ranges, and asset evidence units from the report's fact package based on the evidence index; the sentence / segment determination unit determines whether a sentence / segment can be output; and the rewriting control unit generates rewriting instructions and controls the AI ​​model to perform restricted rewriting of sentences / segments that cannot be output.

[0086] The version management module is used to output verified data asset reports and generate a report version chain. When the relevant data source is updated, the version management module locates the affected asset evidence units based on the set of changed fields, and then determines the affected chapters based on the evidence index set. Only the affected chapters are regenerated and verified, while unaffected chapters retain the chapter summary and content from the previous report version.

[0087] In some embodiments, the present invention also provides a computer program product, which includes a computer program or computer-executable instructions. The computer program or computer-executable instructions are stored in a non-transitory computer-readable storage medium or provided to an electronic device via a wired network, wireless network, or removable storage medium. The electronic device includes a processor and a memory. When the processor reads and executes the computer program or computer-executable instructions, it implements the aforementioned AI-based big data model-based intelligent report generation method for data assets, including generating a report chapter-indicator-evidence constraint table, extracting asset evidence units, generating a report fact package, calling the AI ​​big model to generate candidate report content carrying evidence index tags, performing readback verification and restricted rewriting of the candidate report content based on the report fact package, and outputting a data asset report and generating a report version chain. The computer program product can be deployed on a single server, server cluster, cloud computing platform, or data asset operation platform; its specific deployment form does not affect the implementation of the above method steps.

[0088] III. Structure, Invocation Method, and Parameter Settings of Large-Scale AI Models

[0089] This implementation involves calling large AI models, but the implementation of this invention does not depend on a specific vendor or a specific name for the large model. The large AI model can be a privately deployed model or a model service that can be called within an enterprise in compliance with regulations. To facilitate implementation by those skilled in the art, the following provides a possible model structure, calling method, and parameter settings.

[0090] The large-scale AI model can employ a Transformer-based autoregressive generative language model. Its structure includes a tokenizer, word embedding layers, multi-layer Transformer decoders, attention layers, feedforward network layers, normalization layers, and an output layer. The model input consists of structured prompts, which include system constraints, report fact packages, chapter writing constraints, evidence citation formats, and prohibition on extrapolation constraints. The model output consists of candidate report content and a mapping table between sentence segments, field types, and evidence indices.

[0091] In one implementation, the context length of the large AI model is no less than 8192 lexical units. If the reported fact package is large, it is segmented according to chapter identifiers, generating only one chapter or several paragraphs under one chapter at a time. The model temperature parameter can be set to 0.1 to 0.4 to reduce the randomness of the generated content; the top-p parameter can be set to 0.7 to 0.95; the maximum output length is set according to the word count range of the chapter, for example, the maximum output length of a single chapter is set to 800 to 1500 words.

[0092] This implementation can directly use a pre-trained general-purpose or industry-specific large-scale language model, without requiring model retraining. If it is necessary to improve the model's adaptability to data asset terminology and report formats, a fine-tuning method can be used. Data used for fine-tuning can include anonymized data asset report samples, chapter templates, indicator descriptions, evidence index annotation samples, and sentence / segment verification samples. The dataset can be divided into training, validation, and test sets in an 8:1:1 ratio. Fine-tuning can employ low-rank adaptation training, with the low-rank matrix rank set to 8 to 16 and the learning rate set to 1×10⁻. 5 Up to 5×10⁻ 5 The number of training rounds is set to 2 to 4, and the batch size is set to 4 to 32 based on the GPU memory capacity. The training data must not contain un-anonymized personal information, important data in plaintext, or unauthorized transaction records. After the model training is completed, the reporting fact package and verification rewriting process of this implementation method is still performed; fact package filtering, evidence index marking, and readback verification are not omitted even though the model has been fine-tuned.

[0093] When calling the model, the system can use the following structured input:

[0094] The first part consists of role and task constraints, which state that the model generates reports solely based on the fact package and must not introduce external knowledge to infer specific values, rankings, amounts, proportions, compliance conclusions, and risk levels.

[0095] The second part outlines the constraints on chapter writing, including chapter titles, word count ranges, narrative order, explanations of indicators, a glossary, and a list of prohibited terms.

[0096] The third part is the evidence citation format, which requires the model to append an evidence index marker to the end of sentences containing numerical fields, conclusion fields, or risk level fields.

[0097] The fourth part is the report fact package, which is organized by chapter identifier, indicator name, indicator calculation result, publicly available fields, valid range, and evidence index.

[0098] Part 5 specifies the output format requirements, which require outputting the content of the candidate reports and a mapping table between sentence segments, field types, and evidence indexes.

[0099] In this way, the generation process of large AI models is jointly restricted by the report fact package, the evidence citation format, and the prohibition on extrapolation.

[0100] IV. Specific Implementation of Method Steps

[0101] like Figure 2 As shown, the data asset intelligent report generation method of this embodiment includes the following steps.

[0102] S1. Receive report generation request and generate report sections—Indicators—Evidence Constraint Table.

[0103] The system first receives a report generation request. This request can be submitted by the user on the front-end page or triggered by a scheduled task. For example, the user selects "Generate April 2026 Data Asset Operation Report," chooses "Population Service Theme Library, Enterprise Service Theme Library, and City Operation Theme Library" for the report target scope, selects "Data Operation Department" for the report user, and selects "Monthly Operation Analysis" for the usage scenario.

[0104] The request parsing module parses the report's purpose, target audience, users, and use case. The report purpose identifies the problem the report aims to solve, such as inventorying existing data assets, analyzing data quality, analyzing data service usage, analyzing data service revenue, or assisting in data asset entry. The target audience defines the scope of the data assets. The users determine permissions and sensitive field handling rules. The use case defines the chapter structure and output style.

[0105] The request parsing module calls the chapter template library to generate a report chapter-indicator-evidence constraint table. For example, a data asset operation report may include the following chapters: Overall Asset Overview, Asset Quality, Asset Usage, Asset Authorization, Asset Service Revenue, and Asset Risk. Each chapter corresponds to several indicators. For example, the Overall Asset Overview chapter corresponds to the total number of assets, the number of data tables, the number of interfaces, the number of data products, and the number of business themes covered; the Asset Quality chapter corresponds to completeness score, accuracy score, consistency score, timeliness score, and uniqueness score; the Asset Usage chapter corresponds to the number of interface calls, the number of data product calls, the number of callers, and the number of active assets; the Asset Authorization chapter corresponds to the number of authorizations, the number of authorized objects, the authorization effectiveness rate, and the number of expired authorizations; the Asset Service Revenue chapter corresponds to the cost accounting amount, the revenue calculation amount, the number of service settlements, and the revenue range; and the Asset Risk chapter corresponds to the number of highly sensitive assets, the number of important data assets, the number of authorization anomalies, the number of expired evidence, and the number of indicator conflicts.

[0106] The report section—Indicators—Evidence Constraints table also records the required evidence type and minimum credibility level for each indicator. For example, the revenue calculation amount needs to come from cost accounting records, revenue calculation records, or data service settlement records, and the credibility of the evidence must not be lower than 0.80; the number of authorized anomalies needs to come from authorization records and call records, and the valid evidence range should cover the reporting period; the risk level needs to simultaneously meet the generation rules of sensitivity level, authorization status, and quality inspection results.

[0107] S2, Asset Evidence Extraction Unit

[0108] The evidence extraction module extracts asset evidence units from the data asset catalog, metadata, lineage, quality inspection, classification and grading, authorization, access, cost accounting records, revenue calculation records, and data service settlement records.

[0109] For the data asset catalog, the system reads the asset name, asset number, department, business theme, data table, data field, interface name, and data product name, and performs standardized mapping. For example, it maps the "Population Basic Table," "Population Basic Information Table," and "Population Basic Data Table" representing the same asset in different systems to the same asset identifier.

[0110] For metadata and lineage, the system associates field definitions, field types, update cycles, source systems, processing tasks, intermediate tables, indicator tables, and data products to generate lineage paths. For example, a certain indicator, "monthly call count," could have the following lineage path: Interface gateway call log table — call detail cleaning task — interface call statistics intermediate table — monthly call indicator table — data asset operation report fact package.

[0111] For quality inspection, the system extracts indicators of completeness, accuracy, consistency, timeliness, and uniqueness. For classification and grading, the system extracts data category, data level, personal information tags, important data tags, and outbound restriction tags. For authorization, invocation, cost accounting records, revenue calculation records, and data service settlement records, the system extracts authorization scope, number of invocations, caller type, construction cost, operation and maintenance cost, revenue calculation amount, service settlement frequency, and service settlement cycle, respectively.

[0112] Each asset evidence unit is bound to an asset identifier, field name, indicator value, source location, timestamp, lineage path, sensitivity level, valid range, and evidence digest. The source location can be a database table name, field name, file path, interface address, or record number. The valid range indicates the applicability of the evidence within the reporting period. The evidence digest is obtained by hashing the asset identifier, field name, indicator value, source location, timestamp, and lineage path. The hash algorithm can be SHA-256, or other digest algorithms that meet integrity verification requirements.

[0113] S3, Generate Report Fact Package

[0114] Instead of directly inputting raw data asset records into the AI ​​big model, the system first generates a report fact package.

[0115] First, the evidence matching unit matches asset evidence units according to the report chapter-indicator-evidence constraint table. For example, the "Completeness Score" indicator under the "Asset Quality Status" chapter needs to match the completeness indicator value in the quality inspection record; the "Revenue Calculation Amount" indicator under the "Asset Service Revenue Status" chapter needs to match the cost accounting record, revenue calculation record, or data service settlement record; and the "Number of Highly Sensitive Assets" indicator under the "Asset Risk Status" chapter needs to match the sensitivity level in the classification and grading record.

[0116] When the same reporting metric matches multiple asset evidence units, the system calculates the credibility of the asset evidence units. :

[0117]

[0118] In the formula, For asset evidence unit The credibility of the evidence; The score is for quality inspection. The source system's trust level; For the integrity of bloodline; Score for time efficiency; This is the conflict penalty value; , , , As a positive evaluation weight, and satisfying ; This is the conflict penalty coefficient, with a value ranging from 0 to 1.

[0119] In one implementation, , , , , The values ​​were set to 0.30, 0.20, 0.25, 0.15, and 0.10 respectively. (Quality inspection score) It can be obtained by weighting completeness, accuracy, consistency, timeliness, and uniqueness; source system trust level The degree of lineage can be determined based on whether the data source is a primary data system, an authoritative business system, or a manually uploaded file; The time validity score can be determined based on whether the lineage path covers the source system, processing task, intermediate table, and output object. The conflict penalty value can be determined based on whether the evidence timestamp falls within the reporting period. It can be determined based on the degree of difference between different pieces of evidence under the same indicator.

[0120] Asset evidence units whose credibility reaches the minimum level of credibility are selected for inclusion in the reporting fact package. When the differences in indicator values ​​of multiple asset evidence units under the same reporting indicator exceed a preset conflict threshold, the system generates an indicator conflict flag and adds the corresponding reporting indicator to the set of indicators pending review. For example, if the revenue calculation system records a revenue calculation amount of 382,000 yuan for a certain data product reporting period, while the data service settlement system records a corresponding settlement amount of 369,000 yuan, and the difference exceeds the preset conflict threshold, then an indicator conflict flag is generated for that revenue indicator, preventing the AI ​​large model from directly generating a definitive revenue conclusion.

[0121] Secondly, the permission filtering unit filters fields that can be displayed based on the report user's permissions and authorization status. For example, when the report user is the data operations department, they can view the asset quantity, call count, quality score, and revenue range, but cannot view the plaintext of personal sensitive fields. For fields that do not meet the display conditions, the sensitive field replacement unit replaces them with anonymized summaries, range-based values, enumeration levels, or local retention location indexes. For example, the ID number field is not included in the report fact package, only "Number of Personal Sensitive Fields: 1" is retained; the revenue amount is only displayed as a range for some users; highly sensitive fields only display the sensitivity level and local retention location index.

[0122] Then, the evidence index generation unit combines the storage location of the asset evidence unit in the report fact package, the evidence summary, and the chapter identifier to generate an evidence index. For example, the evidence index can be represented as "FP202604-QA-CH02-M003-E012-H8F2A", where FP202604 represents the fact package identifier, CH02 represents the chapter identifier, M003 represents the indicator identifier, E012 represents the asset evidence unit number, and H8F2A represents the evidence summary truncation value.

[0123] The report fact package includes a fact package identifier, report object identifier, section identifier, indicator name, indicator calculation result, fields that can be displayed, evidence index, evidence summary, valid range, indicator conflict marker, and sensitive field replacement result. The AI ​​large model can only receive the content in the report fact package and cannot bypass the report fact package to access the original data source.

[0124] S4, AI large model generates candidate report content

[0125] The large model generation module inputs the report fact package, chapter writing constraints, evidence citation format, and prohibition of extrapolation constraints into the AI ​​large model to generate candidate report content.

[0126] Chapter writing constraints include chapter titles, word count limits, narrative order, explanations of indicators, glossary, and a list of prohibited terms. For example, the chapter on overall asset overview requires first stating the quantity of assets and the topics covered, followed by information on data products and interfaces; the chapter on asset quality requires first describing the overall quality score, followed by descriptions of completeness, accuracy, consistency, timeliness, and uniqueness, and prohibits the use of absolute statements without supporting evidence.

[0127] The evidence citation format stipulates that each sentence containing a numerical field, a conclusion field, or a risk level field must be accompanied by an evidence index marker. For example, the content of a candidate report could be: During the reporting period, 128 data assets were registered in the Enterprise Service Subject Database, including 86 data tables, 24 data interfaces, and 18 data products. [Evidence: FP202604-CH01-M001-E001].

[0128] The prohibition on extrapolation includes the following constraints: Candidate reports must not generate specific values, rankings, proportions, ownership conclusions, compliance conclusions, value conclusions, service settlement conclusions, or risk levels that are not supported by evidence indexes. If the fact package in the report only provides "the revenue amount range is 300,000 to 400,000 yuan", the AI ​​big model must not generate a sentence or paragraph stating "the revenue amount is 380,000 yuan"; if the fact package in the report contains indicator conflict markers, the AI ​​big model must not generate a definitive conclusion, but can only indicate "this indicator has items pending verification".

[0129] When the AI ​​model generates candidate report content, it adds evidence index tags to sentences containing numerical fields, conclusion fields, or risk level fields, and outputs a mapping table between sentences, field types, and evidence indexes. The mapping table can include sentence identifiers, sentence text, field types, evidence indexes, and chapter identifiers. Field types can be numerical fields, conclusion fields, risk level fields, or general descriptive fields.

[0130] Ordinary description fields do not necessarily require the attachment of evidence indexes, but if ordinary description fields contain specific statistical values, risk assessments, or compliance assessments, they should be re-identified as numerical fields, risk level fields, or conclusion fields.

[0131] S5. Candidate report content verification and restricted rewriting

[0132] The verification and rewriting module parses the candidate report content and reads back the corresponding indicator calculation results, sensitivity level, and effective range from the report fact package based on the evidence index carried by the sentence segment.

[0133] First, the field parsing unit identifies the numerical fields, conclusion fields, risk level fields, and general descriptive fields in the candidate report content. Numerical fields include asset quantity, field quantity, quality score, number of calls, estimated revenue amount, service settlement frequency, ratio, ranking, and time. Conclusion fields include judgmental statements such as "Requirements met," "Anomalies exist," "Can be included in the table," and "Not recommended for open sharing." Risk level fields include risk level descriptions such as "Low risk," "Medium risk," and "High risk."

[0134] Secondly, the fact package readback unit retrieves the corresponding indicator calculation results, sensitivity level, valid range, and asset evidence unit from the report fact package based on the evidence index. For example, if a sentence states "completeness score is 98.2," its evidence index points to the completeness indicator result in the report fact package, and the system compares the 98.2 in the sentence with the 98.2 in the fact package. If the sentence states 98.5, the values ​​are determined to be inconsistent.

[0135] Next, the segment determination unit performs the following judgment:

[0136] When the numerical field does not match the calculation result of the indicator pointed to by the evidence index, the corresponding segment will be marked as an unoutput segment;

[0137] When a segment contains a numerical field, a conclusion field, or a risk level field but does not carry an evidence index, the corresponding segment will be marked as an unoutputable segment.

[0138] When the valid range of the asset evidence unit pointed to by the evidence index does not cover the reporting period, it is determined that the evidence exceeds the valid range, and the corresponding segment is marked as an unoutput segment.

[0139] When the sensitivity level of the asset evidence unit pointed to by the evidence index does not meet the permissions of the report user, the corresponding segment will be marked as an unoutput segment;

[0140] When the type of evidence referenced by the conclusion field does not meet the required evidence type in the report section—indicators—evidence constraint table, the corresponding sentence segment will be marked as an unoutputable sentence segment;

[0141] When the risk level field is inconsistent with the risk level generation rules in the report section—indicators—evidence constraint table, the corresponding segment will be marked as an unoutputable segment.

[0142] For example, if the report is used by an external partner, and the fact package only allows displaying a range of call counts, but the candidate report specifies the exact number of calls, then that segment will be marked as unoutputtable because its sensitivity level does not meet the report user's permissions. Similarly, if a revenue indicator has a conflict flag, but the candidate report concludes with a certain revenue growth, then that segment will be marked as unoutputtable because the conclusion lacks supporting evidence or cites conflicting evidence.

[0143] For segments that cannot be output, the rewrite control unit generates a rewrite instruction. The rewrite instruction includes the segment identifier, error type, field type, corresponding asset evidence unit, allowed indicator calculation results, allowed evidence indexes, and rewrite restrictions. Error types include one or more of the following: inconsistent values, missing evidence index, evidence outside the valid range, sensitivity level not meeting the report user's permissions, lack of evidence to support the conclusion, and risk level not conforming to the generation rules.

[0144] The rewrite instruction does not provide the original sensitive fields to the AI ​​large model, nor does it open additional data sources. It only provides the content that can be explicitly displayed in the corresponding asset evidence unit and the evidence index that can be cited. For example, the rewrite instruction can require: rewrite the sentence S2-05 into a range-based expression, without specifying the number of calls, and only using the range "10,000 to 15,000 times" and the evidence index FP202604-CH03-M002-E015.

[0145] After the AI ​​model regenerates candidate segments based on the rewrite instructions, it executes step S5 again for verification until the candidate segments pass verification or reach the preset number of rewrites. The preset number of rewrites can be set to 3. If the verification still fails after reaching the preset number of rewrites, the corresponding segment is transferred to the manual review queue, and the segment that cannot be output and the rewrite record are recorded in the report version chain.

[0146] S6. Output data asset report and generate report version chain.

[0147] Once all segments in the candidate report pass validation, or if any segments that fail validation have been deleted, replaced, or transferred to the manual review queue, the system outputs a validated data asset report. The data asset report can be output as a Word document, PDF document, web page report, or structured JSON report.

[0148] The version management module generates a report version chain. The report version chain includes the report version number, a summary of the previous report version, a summary of the report fact package, a summary of prompt words, a summary of the model version, a summary of chapters, a set of evidence indexes, records of non-output segments, rewrite records, and a set of changed fields. The summary of the report fact package can be obtained by hashing the entire content of the report fact package; the summary of prompt words can be obtained by hashing system constraints, chapter writing constraints, evidence citation format, and prohibition of extrapolation constraints; the summary of the model version can be generated from the model name, model version number, and calling parameters; the summary of each chapter can be generated from the final text of each chapter; and the set of evidence indexes records all evidence indexes used in the final report.

[0149] When data asset catalogs, metadata, lineage, quality inspection, classification and grading, authorization, access, cost accounting records, revenue calculation records, or data service settlement records are updated, the version management module locates the affected asset evidence units based on the set of changed fields and determines the affected chapters based on the evidence index set. For affected chapters, the system regenerates candidate report content and performs verification; for unaffected chapters, the chapter summary and content from the previous report version are used. This approach avoids a full report rewrite after each data update, facilitating report version comparison and review.

[0150] V. Examples

[0151] This embodiment uses the generation of a monthly data asset operation report from a municipal-level data asset operation platform as an example to illustrate the implementation process of the data asset intelligent report generation method and system of the present invention.

[0152] The platform connects to three business theme databases: population services, enterprise services, and urban operations, with a reporting period from April 1, 2026 to April 30, 2026. The system registers 312 data assets, including 198 data tables, 72 data interfaces, and 42 data products. The platform stores approximately 68,000 metadata records, including table structure, field definitions, interface parameters, indicator definitions, update cycles, tag information, and quality rules; approximately 12,000 table-level, field-level, interface-level, and indicator-level lineage records; approximately 9,000 quality inspection records; approximately 15,000 field-level and asset-level classification and grading records; approximately 2,300 authorization records; approximately 4.2 million interface and data product call records; and approximately 1,800 cost accounting records, revenue calculation records, and data service settlement records.

[0153] In this embodiment, the report generation request is: Generate a data asset operation report for April 2026. The report user is the data operation department, the output format is a Word document, and the usage scenario is monthly operation analysis. The report user has permission to view the number of assets, asset quality score, call statistics, authorization statistics, and data service revenue range, but does not have permission to view plaintext personal sensitive fields, plaintext important data fields, and details of external service objects.

[0154] After receiving the report generation request, the system generates a report chapter-indicator-evidence constraint table comprising 6 chapters and 34 reporting indicators. The 6 chapters are: Overall Asset Overview, Asset Quality, Asset Usage, Asset Authorization, Asset Service Revenue, and Asset Risk. The 34 reporting indicators include: Total Assets, Number of Data Tables, Number of Data Interfaces, Number of Data Products, Completeness Score, Accuracy Score, Consistency Score, Timeliness Score, Uniqueness Score, Number of Interface Calls, Number of Data Product Calls, Number of Authorizations, Authorization Effectiveness, Calculated Revenue Amount, Number of Service Settlements, Number of Highly Sensitive Assets, Number of Important Data Assets, Number of Authorization Anomalies, Number of Expired Evidence, and Number of Indicator Conflicts.

[0155] The evidence extraction module extracts asset evidence units from various data sources. After standardization mapping, lineage association, quality indicator extraction, classification and grading labeling and identification, and authorization call record organization, the system extracts a total of 1486 asset evidence units. Each asset evidence unit is bound to an asset identifier, field name, indicator value, source location, timestamp, lineage path, sensitivity level, valid range, and evidence summary.

[0156] The fact package generation module matches asset evidence units according to the report section—indicators—evidence constraint table and calculates the credibility of the evidence. In this embodiment, , , , , The values ​​are set to 0.30, 0.20, 0.25, 0.15, and 0.10, respectively. For evidence units from authoritative business systems with complete lineage paths and timestamps within the reporting period, their source system credibility level, lineage completeness, and time validity scores are higher. For evidence units with multiple sources under the same indicator and numerical differences exceeding the threshold, the system assigns a conflict penalty value.

[0157] For example, the estimated revenue for a certain data product in the enterprise service subject library is 382,000 yuan in the revenue calculation record, while the corresponding settlement amount in the data service settlement record is 369,000 yuan. The difference exceeds the preset conflict threshold of 3%. The system generates a conflict flag for this revenue indicator and adds it to the set of indicators pending review. In the report fact package, this indicator is not used as a definitive revenue conclusion, but is only included in the asset risk status section as "There is a revenue indicator pending review".

[0158] The permission filtering unit filters fields that can be displayed based on the user's permissions and authorization status. For fields that are allowed to be displayed, such as the number of assets, the number of data tables, the number of interfaces, the number of data products, the quality score, the range of call counts, and the number of authorizations, the system writes them into the report fact package. For fields that do not meet the display conditions, such as ID card numbers, mobile phone numbers, detailed addresses, important plaintext data fields, and details of external settlement objects, the system replaces them with de-identified summaries, range-based values, enumeration levels, or local retention location indexes.

[0159] In this embodiment, the system-generated report fact package contains 621 fields that can be made explicit, 843 evidence indexes, 34 indicator calculation results, 219 sensitive field replacement results, and 7 indicator conflict markers. The AI ​​big model can only receive the report fact package, chapter writing constraints, evidence citation format, and prohibition of extrapolation constraints; it cannot directly access the original data asset catalog, original call details, original personal sensitive fields, or original settlement object details.

[0160] The large-scale model generation module generates candidate report content based on the report fact package. The candidate report content includes 6 chapters, with an initial total of 132 sentences, including 48 numerical field sentences, 39 conclusion field sentences, 12 risk level field sentences, and 33 general description field sentences. The AI ​​large-scale model also outputs a mapping table between sentence segments, field types, and evidence indexes.

[0161] The verification and rewriting module parses and verifies the content of the candidate reports. The first round of verification found 12 unoutputable segments, including: 3 segments with inconsistent values, 2 segments with missing evidence indexes, 4 segments that cited evidence outside the valid range, 2 segments whose sensitivity level did not meet the permissions of the report user, and 1 segment that made a definitive conclusion on the amount of revenue calculation with conflicting indicator markers.

[0162] For the aforementioned non-output segments, the rewrite control unit generates rewrite instructions. These instructions include the segment identifier, error type, field type, corresponding asset evidence unit, allowed indicator calculation results, allowed evidence indexes, and rewrite restrictions. For example, for segments with inconsistent values, the rewrite instruction requires the AI ​​large model to use only the indicator calculation results pointed to by the corresponding evidence index in the report fact package; for segments whose sensitivity level does not meet the permission requirements, the rewrite instruction requires the deletion of plaintext sensitive fields and the use of only anonymized summaries, enumeration levels, or local retention location indexes; for segments with conflicting revenue calculation amounts, the rewrite instruction requires that no definitive revenue conclusions be generated, only indicating that the indicator has items pending review.

[0163] The AI ​​model regenerated candidate segments based on the rewriting instructions. After the first round of rewriting, 10 segments passed the verification; the remaining 2 segments were returned for rewriting again because they still involved definitive statements of conflicting indicators. After the second round of rewriting, 1 segment passed the verification, and the other segment was transferred to the manual review queue, and was presented in the final report as "This indicator has items to be reviewed, and no definitive conclusion is made at this time".

[0164] After all outputtable segments pass verification, the system outputs the final data asset operation report and generates a report version chain. The report version chain records the report version number V202604-01, a summary of the previous report version, a summary of the report fact package, a summary of the prompt words, a summary of the model version, summaries of 6 chapters, 843 evidence indexes, 12 records of non-outputtable segments, 13 rewrite records, and a set of changed fields. The report fact package summary, prompt word summary, model version summary, and chapter summaries in the report version chain can all be generated using hash digests for subsequent review and version comparison.

[0165] To illustrate the effectiveness of this embodiment, three report generation methods were compared using the same batch of data. Method A involved manually compiling the data and then writing the report; Method B involved directly inputting the data asset statistical results into the AI ​​model to generate the report, without setting up a fact package, evidence index readback verification, or restricted rewriting; Method C involved generating the report using the method described in this embodiment. Ten monthly reports were generated for each method, and the average value was taken. The test items included average generation time, number of numerically incorrect sentences, number of sentences with conclusions without evidence, number of erroneous outputs of sensitive fields, and the proportion of traceable sentences. The test results are shown in Table 1.

[0166] Table 1 Results of different report generation methods

[0167] As shown in Table 1, when generating data asset reports using Method C in this embodiment, the report generation time is shorter than that of manual compilation and writing. Compared with the method of directly calling AI large models to generate reports, the number of numerical error segments, unsupported conclusion segments, and erroneous output of sensitive fields are all reduced. Since the numerical fields, conclusion fields, and risk level fields in the final report all carry evidence indexes and have been verified by reading back from the report fact package, the proportion of traceable segments is relatively high.

[0168] In another incremental update test, after generating the April 2026 report, the system updated the call count and authorization status of a certain data product in the enterprise service theme library. The version management module located the corresponding asset evidence unit based on the set of changed fields and determined the affected chapters as asset usage and asset authorization status based on the evidence index set in the report version chain. The system only regenerated candidate report content and performed verification for the above two chapters; the remaining four chapters used the chapter summary and content from the previous report version. This incremental update report generation took 8 minutes, less than the 32 minutes required to regenerate the entire report.

[0169] In this embodiment, the system limits the required evidence for the report through a report chapter-indicator-evidence constraint table, limits the input range of the AI ​​large model through asset evidence units and report fact packages, determines whether the candidate report content can be output through evidence index marking and fact package back-read verification, and records the report generation process through a report version chain. Those skilled in the art can implement the data asset intelligent report generation method and system of the present invention based on the above data structure, processing steps, parameter settings, model calling methods, verification rules, and embodiment data.

Claims

1. A method for generating intelligent reports on data assets based on an AI large-scale model, characterized in that, include: S1. Receive the report generation request, parse the report purpose, report target, report user and use case, and call the chapter template library to generate the report chapter-indicator-evidence constraint table; S2. Extract asset evidence units from the data asset catalog, metadata, lineage, quality inspection, classification and grading, authorization, access, cost accounting records, revenue calculation records and data service settlement records. Each asset evidence unit is bound to the asset identifier, field name, indicator value, source location, timestamp, lineage path, sensitivity level, valid range and evidence summary. S3. Match asset evidence units according to the report chapter - indicator - evidence constraint table, and filter them according to the report user's permissions and sensitivity level to generate a report fact package containing fields that can be displayed, evidence index, indicator calculation results and valid ranges; S4. Input the report fact package, chapter writing constraints, evidence citation format and prohibition of extrapolation constraints into the AI ​​big model to generate candidate report content. The numerical fields, conclusion fields and risk level fields in the candidate report content are all marked with evidence index tags. S5. Analyze the candidate report content, and read back the corresponding indicator calculation results, sensitivity level and valid range from the report fact package based on the evidence index carried by the sentence segment. When the values ​​are inconsistent, the evidence index is missing, the evidence exceeds the valid range or the sensitivity level does not meet the permissions of the report user, the corresponding sentence segment is marked as an unoutputable sentence segment, and a rewrite instruction containing the error type and the corresponding asset evidence unit is generated and returned to the AI ​​big model for rewriting. S6. Output a verified data asset report and generate a report version chain, which records the report fact package summary, prompt word summary, model version summary, chapter summary, and evidence index.

2. The method for generating intelligent reports on data assets based on AI large-scale models according to claim 1, characterized in that, The report generation request includes the report type, the scope of the report recipients, the report users, the report generation cycle, and the report output format; The report types include one or more of the following: data asset inventory report, data asset quality report, data asset classification and grading report, data asset value analysis report, data asset operation report, data asset compliance report, and data asset entry auxiliary report; The report chapter-indicator-evidence constraint table includes chapter identifier, chapter title, required indicators, optional indicators, indicator calculation rules, required evidence type, minimum credibility level of evidence, sensitive field processing rules, evidence citation format, and chapter output format.

3. The method for generating intelligent reports on data assets based on AI large-scale models according to claim 1, characterized in that, In step S2, the extraction of asset evidence units includes: Standardize the mapping of asset names, asset numbers, departments, business themes, data tables, data fields, interface names, and data product names in the data asset catalog to obtain asset identifiers and field names; Associate metadata and lineage to generate a lineage path from the source system, processing task, intermediate table, indicator table to the data product; Extract completeness, accuracy, consistency, timeliness, and uniqueness indicators from quality inspection; Extract data categories, data levels, personal information tags, important data tags, and exit restriction tags from the classification and grading system; Extract the authorization scope, number of calls, caller type, construction cost, operation and maintenance cost, revenue calculation amount, service settlement frequency and service settlement cycle from authorization, invocation, cost accounting records, revenue calculation records and data service settlement records; Generate an evidence summary based on the asset identifier, field name, indicator value, source location, timestamp, and lineage path.

4. The method for generating intelligent reports on data assets based on AI large-scale models according to claim 1, characterized in that, In step S3, when the same reporting indicator matches multiple asset evidence units, the evidentiary credibility of the asset evidence units is calculated. : In the formula, For asset evidence unit The credibility of the evidence; The score is for quality inspection. The source system's trust level; For the integrity of bloodline; Score for time efficiency; This is the conflict penalty value; , , , , These are the weighting coefficients; Asset evidence units whose credibility reaches the minimum level of credibility are selected and included in the report fact package; when the difference in indicator values ​​of multiple asset evidence units under the same reporting indicator exceeds the preset conflict threshold, an indicator conflict mark is generated and the corresponding reporting indicator is included in the set of indicators to be reviewed.

5. The method for generating intelligent reports on data assets based on AI large-scale models according to claim 1, characterized in that, In step S3, the generated report fact package includes: Write the asset evidence units into the report fact package according to the chapter identifier and indicator name; Fields that meet the requirements of report users' permissions, authorization status, and sensitivity level are identified as fields that can be displayed externally; Replace fields that do not meet the display conditions with de-identified summaries, range-based values, enumeration levels, or local retention location indexes; The evidence index is generated by combining the storage location of the asset evidence unit in the report fact package, the evidence summary, and the chapter identifier; The report fact package includes a fact package identifier, a report object identifier, a chapter identifier, an indicator name, an indicator calculation result, fields that can be displayed, an evidence index, an evidence summary, a valid range, indicator conflict markers, and sensitive field replacement results.

6. The method for generating intelligent reports on data assets based on AI large-scale models according to claim 1, characterized in that, In step S4, The writing constraints for the chapters include the chapter title, the word count range for the chapter, the narrative order, the interpretation of indicators, the glossary, and the list of prohibited words; The prohibition on extrapolation includes: the content of candidate reports must not generate specific values, rankings, proportions, ownership conclusions, compliance conclusions, value conclusions, service settlement conclusions, and risk levels that are not supported by the evidence index; When the AI ​​big model generates candidate report content, it adds evidence index tags to sentences containing numerical fields, conclusion fields, or risk level fields, and outputs a mapping table between sentences, field types, and evidence indexes.

7. The method for generating intelligent reports on data assets based on AI large-scale models according to claim 1, characterized in that, In step S5, parsing the candidate report content includes: Identify numerical fields, conclusion fields, risk level fields, and general description fields in the candidate report content; Compare the numerical fields with the calculation results of the indicators pointed to by the corresponding evidence index; Compare the conclusion field with the evidence type, evidence credibility, and valid range of the asset evidence unit pointed to by the corresponding evidence index; Compare the risk level field with the risk level generation rules in the report section—indicators—evidence constraint table; If any comparison result does not meet the corresponding rule, the segment containing that field is marked as an unoutput segment. And / or, the rewrite instruction includes a segment identifier, error type, field type, corresponding asset evidence unit, allowed indicator calculation results, allowed evidence index, and rewrite restrictions; The error types include one or more of the following: inconsistent values, missing evidence index, evidence exceeding the valid range, sensitivity level not meeting the report user's permissions, conclusions lacking evidence support, and risk level not meeting the generation rules. After the AI ​​big model regenerates candidate segments according to the rewriting instructions, it performs the verification in step S5 again until the candidate segments pass the verification or reach the preset number of rewritings. When the preset number of rewritings is reached but the verification still fails, the corresponding segments are transferred to the manual review queue.

8. The method for generating intelligent reports on data assets based on AI large-scale models according to claim 1, characterized in that, In step S6, the report version chain includes the report version number, the previous report version summary, the report fact package summary, the prompt word summary, the model version summary, the chapter summary, the evidence index set, the non-output segment record, the rewrite record, and the change field set; When the data asset catalog, metadata, lineage, quality inspection, classification and grading, authorization, access, cost accounting records, revenue calculation records or data service settlement records are updated, the affected asset evidence units are located based on the set of changed fields, and the affected chapters are determined based on the evidence index set. Only the affected chapters will have their candidate report content regenerated and validated; unaffected chapters will retain the chapter summary and content from the previous report version.

9. A data asset intelligent report generation system based on an AI large-scale model, characterized in that, The system is used to implement the AI-based big data model-based intelligent report generation method for data assets according to any one of claims 1 to 8, the system comprising: The request parsing module is used to execute step S1; The evidence extraction module is used to execute step S2; The fact package generation module is used to execute step S3; The large model generation module is used to perform step S4; The verification rewrite module is used to execute step S5; The version management module is used to execute step S6.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the method described in any one of claims 1-8.

Images