Auditing data analysis method and system based on large language model

Abstract

The invention provides an audit data analysis method based on a large language model. The audit data analysis method comprises the following steps: analyzing a natural language request: judging whether the request is structured data or unstructured data; multi-modal result fusion: an audit data analysis system of the large language model comprises an intelligent analysis module, an SQL assistant module, a document assistant module and a general intelligent module, and data streams among the modules are transmitted through an intranet encryption channel; by constructing an intelligent data analysis system based on a large model, an integrated process of'natural language questioning-gt, automatic SQL generation-gt and data analysis results' is realized, auditing personnel are assisted to efficiently complete data analysis work, the defects in the prior art are overcome, and the method has important research significance and use value.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, in particular to an audit data analysis method and system based on a large language model. BACKGROUND

[0002] The intelligentization of the national governance system and governance capacity puts the most direct requirements on the big data, information technology and intelligent era. The most essential feature of the intelligentization of national governance is to consciously adapt to the general trend of social informatization and integrate big data collection and processing and intelligent decision-making system into the decision-making system and governance system.

[0003] Audit is an important part of the national supervision system and an important force to promote the modernization of the national governance system and governance capacity.

[0004] Currently, there are some pain points in the audit work, such as the difficulty for auditors to establish high-quality risk judgment logic due to unfamiliarity with the business, the difficulty in analyzing unstructured audit evidence, the large amount of standardized and repetitive work, and the lack of accumulation, optimization and application of audit experience. Intelligent means must be adopted to improve the efficiency of audit work and promote the high-quality development of the audit profession. The audit departments at all levels actively carry out the practice of empowering audit work with artificial intelligence. How to simplify the audit process, solve the work pain points and improve the quality and efficiency of audit through intelligent technology has become a hot topic for the practice and academic circles to explore together.

[0005] Currently, large language models show strong general knowledge and reasoning ability. Related research shows that the combination of large language models in audit work not only improves the efficiency and accuracy of audit, but also enhances the depth and breadth of audit, bringing innovation to audit work. The natural language interaction and intelligent question and answer functions of large language models can greatly reduce the threshold for auditors when using intelligent audit systems, improve the coverage and accuracy of audit work. Through text mining technology, large language models can automatically extract key information and data from massive audit-related documents, helping auditors quickly grasp the background information, business characteristics and potential risk points of the audit object. A large amount of data and analysis results generated during the audit process need to be clearly explained and interpreted. Large language models can provide natural language descriptions of data interpretation and analysis results, making audit results easier to understand and communicate to business-related parties. In particular, the code generation capability of large models can generate data analysis SQL scripts for audit items and ideas, assisting in completing data analysis, thereby improving the efficiency of audit data processing and analysis, enhancing the accuracy and reliability of audit work, and providing new ideas and technical paths for national audit.

[0006] With the development of large language model (LLM) technology in natural language understanding, code generation and other fields, its application in audit data analysis is possible. Traditional audit data analysis has high SQL threshold, scattered scenes, and repetitive work, and the specific analysis is as follows: the traditional audit system has three defects:

[0007] High technical threshold: professional personnel is required to write SQL, and the response cycle is long (average > 2 hours);

[0008] Serious data island: on-site audit data is scattered in heterogeneous databases such as Oracle and MySQL;

[0009] Security risk is prominent: original data transmission is easy to cause information leakage (such as CN114580944A scheme), which depends on full data transmission.

[0010] The existing improvement scheme (such as CN116258324A) only supports a single database, and still cannot realize natural language interaction and cross-platform intelligent migration.

[0011] Therefore, an audit data analysis method and system based on a large language model are needed to solve the above problems.

[0012] It should be noted that the above introduction to the technical background is only to facilitate a clear and complete description of the technical solutions of the present application, and to facilitate the understanding of those skilled in the art. The above technical solutions cannot be considered as known to those skilled in the art just because they are described in the technical background part of the present application. SUMMARY

[0013] The purpose of the present application is to provide an audit data analysis method and system based on a large language model, which realizes the integrated process of "natural language question -> automatic SQL generation -> data analysis result" by constructing an intelligent data analysis system based on a large model, and helps auditors to efficiently complete data analysis work, solves the drawbacks in the prior art, and has important research significance and use value.

[0014] To achieve the above purpose, the present application provides the following technical solutions:

[0015] The present application provides an audit data analysis method based on a large language model, comprising the following steps:

[0016] S1: natural language request analysis: through the problem classifier of the intelligent analysis module, receiving the natural language query request transmitted by the client through VPN encryption, identifying the audit intention and data type, and classifying into data query, document retrieval or general task;

[0017] S2: judging the data type involved in the request is structured data or unstructured data;

[0018] S21: If the request involves structured data, trigger the structured query engine to generate abstract query logic, call the SQL assistant module to perform cross-database adaptation, including:

[0019] Identify the target database type through the SQL syntax adapter, including Oracle, MySQL or SQL Server;

[0020] Verify the syntax security of the generated script through the pre-trained model, prevent scripts containing high-risk operations from being executed, including unauthorized table deletion, permission escape query;

[0021] Use the script conversion engine to convert the abstract query logic into the target database dialect script, and perform equivalent reconstruction on the paging query and window function;

[0022] Execute the query and feed back the results in natural language form through the model explanation unit;

[0023] S22: If the request involves unstructured data, start the document assistant module, including:

[0024] Extract text, table and image content in the user's private knowledge base through the multi-format document parser;

[0025] Match the associated clauses in the policy and regulation common library based on the hybrid retrieval engine, adopt a dynamic weighted scoring strategy, and prefer semantic matching results when the semantic similarity is >0.7;

[0026] Mark the data source and confidence through the citation tracing unit;

[0027] Also includes: Construct a joint vector index of the documents in the user's private knowledge base and the policy and regulation common library, and realize cross-library associated retrieval through semantic similarity calculation;

[0028] S3: Multi-modal result fusion:

[0029] The general intelligent module performs joint analysis on structured query results and unstructured retrieval results, including:

[0030] S31: Call the domestic large model to perform text generation and risk pattern recognition;

[0031] S32: Perform OCR recognition and key field extraction on image data;

[0032] S33: Generate an audit analysis report containing data traceability markers;

[0033] The multi-modal result fusion also includes automatically generating an abnormality label and triggering a manual review process when it is detected that the difference between the structured data and the unstructured data in describing the same entity exceeds a preset threshold (numerical difference >= 5% or text description confidence difference >= 0.3);

[0034] The manual review interface supports difference highlighting, comment version tracking and multi-person collaborative review, and the review result is stored through a blockchain.

[0035] The application also provides an audit data analysis system based on a large language model, which comprises an intelligent analysis module, an SQL assistant module, a document assistant module and a general intelligent module, and the data flow between the modules is transmitted through an intranet encryption channel.

[0036] The intelligent analysis module is used for realizing natural language driven audit data query and on-site analysis, and comprises a question classifier, a structured query engine and an abstract generator, and supports interaction with structured metadata through a VPN secure channel; the intelligent analysis function module comprises a data query module and a data analysis module, the data query module is configured to: in response to a user operation on a data query user interface element, receive a data query request input in an associated input box; in response to a user operation on a sending user interface element, acquire relatively complete information corresponding to the data query request based on an industry data resource knowledge base; the data analysis module is configured to: in response to a user operation on a data analysis user interface element, present an industry data resource directory user interface element; wherein the industry data resource directory user interface element comprises a plurality of information table options and associated selection controls; in response to a user operation on the selection control, perform data analysis based on the selected information table and generate an analysis result; the industry data resource directory user interface element is further configured to present preview content of the corresponding information table in response to a user operation on an information table associated preview control;

[0037] The SQL assistant module provides cross-database script migration services, including an SQL syntax adapter, a script conversion engine and a model interpretation unit, and supports Oracle, MySQL and SQL Server multi-dialect environments;

[0038] The document assistant module: builds a user private knowledge base and a policy and regulation common library, integrates a multi-format document parser, a hybrid retrieval engine and a reference tracing unit;

[0039] The general intelligent module: through intranet private deployment of domestic large models, provides general AI capabilities such as text generation, image recognition and multi-language translation, and the domestic large models include DeepSeek, Qwen and GLM.

[0040] The intelligent analysis module includes a data security interaction unit and a temporary table management unit. The client only transmits a database table structure or a desensitized metadata field to the audit data analysis system, and original data is retained for local execution of a query. The temporary table management unit supports storage of a query result of a sensitive field, such as an ID card number or the name of a deceased person, in the form of an encrypted temporary table and sets an automatic cleaning strategy.

[0041] The SQL assistant module includes:

[0042] A dialect adapter: automatic conversion of PL / SQL and T-SQL based on syntax tree analysis, and correction of data type differences;

[0043] An execution plan visualization unit: decomposition of a SQL execution process into natural language interpretation steps and labeling of key performance bottleneck points;

[0044] The document assistant module supports:

[0045] Multi-modal retrieval: paragraph-level vectorization of PDF and DOCX documents, combined keyword matching and semantic similarity calculation to achieve hybrid search;

[0046] Dynamic update mechanism: automatic capture of policy files, generation of a structured knowledge graph, and association with a question and answer system.

[0047] Optionally, the data analysis module is further configured to:

[0048] Below the industry data resource directory user interface element, an audit project data list user interface element is presented;

[0049] The audit project data list user interface element includes a database selection control. In response to user operation of the database selection control, a list of selectable databases is presented and the user is allowed to select a target database;

[0050] The audit project data list user interface element presents, for the selected target database, a plurality of data entry options contained therein and associated selection controls. In response to user operation of the data entry selection controls, data analysis is performed based on the selected data entry and an analysis result is generated;

[0051] The audit project data list user interface element is further configured to, for the data entry, present a detailed information preview of the corresponding data entry in response to user operation of an associated entry preview control;

[0052] The audit project data list user interface element includes a file upload control;

[0053] Configured to:

[0054] in response to a user operation on the file upload control, presenting a file upload dialog box;

[0055] receiving user input through the file upload dialog box, including: database name, table name;

[0056] receiving a data file uploaded by a user through the file upload dialog box, the data file being in a predetermined format;

[0057] in response to a save operation, storing the data file into the audit project data;

[0058] the audit project data list user interface element comprises a database connection adding control;

[0059] configured to:

[0060] in response to a user operation on the database connection adding control, presenting a database type selection dialog box;

[0061] in response to selecting a database type in the database type selection dialog box, presenting a connection configuration dialog box;

[0062] receiving connection configuration information input by a user through the connection configuration dialog box;

[0063] in response to a test connection operation, performing a database connection test based on the connection configuration information;

[0064] after the connection test passes, in response to a save operation, adding the new database connection into the audit project data.

[0065] Optionally, the predetermined format is CSV format.

[0066] Optionally, the system is configured to limit the size of a single uploaded file to be no more than a predetermined threshold.

[0067] The audit data analysis method based on the large language model provided by the present application is a more advanced and comprehensive audit data analysis method based on the large language model. By constructing an intelligent data analysis system based on the large model, an integrated process of "natural language questioning -> automatic SQL generation -> data analysis result" is realized, which helps auditors to efficiently complete data analysis work, solves the disadvantages in the prior art, and better solves the difficulties in analyzing unstructured audit evidence, the problems of standardized and repetitive work, the lack of continuity, optimization and application ability of audit experience accumulation, and other pain points, and has important research significance and use value.

[0068] In addition, the combination of large language models in auditing work not only improves the efficiency and accuracy of auditing, but also enhances the depth and breadth of auditing, bringing innovation to auditing work. The natural language interaction and intelligent question and answer functions of large language models can greatly reduce the threshold for auditors when using intelligent auditing systems, improving the coverage and accuracy of auditing work. For example, through text mining technology, large language models can automatically extract key information and data from massive amounts of auditing-related documents, helping auditors quickly grasp the background information, business characteristics, and potential risk points of the audit object. A large amount of data and analysis results generated during the auditing process need to be clearly explained and interpreted. Large language models can provide natural language descriptions of data interpretation and analysis results for auditors, making the audit results easier to understand and communicate to business-related parties. In particular, the code generation capability of large models can generate data analysis SQL scripts for audit items and ideas, assisting in completing data analysis, thereby improving the efficiency of auditing data processing and analysis, and enhancing the accuracy and reliability of auditing work, providing new ideas and technical paths for auditing. BRIEF DESCRIPTION OF DRAWINGS

[0069] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, a brief introduction of the drawings needed to be used in the embodiments or prior art description will be given below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0070] Figure 1 is a flowchart of the auditing data analysis method based on large language models of the embodiments of the present application.

[0071] Figure 2 is a connection diagram of the auditing data analysis system based on large language models of the embodiments of the present application.

[0072] Figure 3 is GLM pre-training of the embodiments of the present application.

[0073] Figure 4 is a data flow diagram of the model built by Tensorflow of the embodiments of the present application.

[0074] Figure 5 is the main implementation logic of the embodiments of the present application.

[0075] Figure 6 is a general framework diagram of the auditing data analysis system based on large language models of the embodiments of the present application.

[0076] Figure 7 is a data query module diagram of the auditing data analysis system based on large language models of the embodiments of the present application.

[0077] Figure 8 is a query interface diagram of an embodiment of the present application.

[0078] Figure 9 is a query result diagram of an embodiment of the present application.

[0079] Figure 10 is an audit site data analysis diagram of an embodiment of the present application.

[0080] Figure 11 is an audit site data analysis result diagram of an embodiment of the present application.

[0081] Figure 12 is an SQL helper intelligent answering diagram of an embodiment of the present application.

[0082] Figure 13 is an SQL helper script rewriting diagram of an embodiment of the present application.

[0083] Figure 14 is an SQL helper model explanation diagram of an embodiment of the present application.

[0084] Figure 15 is an intelligent analysis running interface diagram of an embodiment of the present application.

[0085] Figure 16 is an SQL helper intelligent answering running interface diagram of an embodiment of the present application.

[0086] Figure 17 is an SQL helper intelligent rewriting running interface diagram of an embodiment of the present application.

[0087] Figure 18 is an SQL helper model explanation running interface diagram of an embodiment of the present application.

[0088] Figure 19 is a document helper running interface diagram of an embodiment of the present application.

[0089] Figure 20 is a general intelligent running interface diagram of an embodiment of the present application.

[0090] Figure 21 is a general intelligent welcome page of an embodiment of the present application.

[0091] Figure 22 is a queuing situation of an embodiment of the present application.

[0092] Figure 23 is a general intelligent function of uploading a file of an embodiment of the present application.

[0093] Figure 24 is a general intelligent function of uploading a picture of an embodiment of the present application.

[0094] Figure 25 is a document helper function of an embodiment of the present application.

[0095] Figure 26 Embodiment of the application document assistant function - reference source.

[0096] Figure 27 Embodiment of the application document assistant - library.

[0097] Figure 28 Embodiment of the application document assistant - personal files.

[0098] Figure 29 Embodiment of the application data analysis - industry data resource directory.

[0099] Figure 30 Embodiment of the application data analysis - audit project data.

[0100] Figure 31 Embodiment of the application audit project data - upload file.

[0101] Figure 32 Embodiment of the application audit project data - add connection.

[0102] In the figure, 1, intelligent analysis module; 2, SQL assistant module; 3, document assistant module; 4, general intelligent module. DETAILED DESCRIPTION

[0103] In order to make the purpose, technical scheme and advantages of the present application clearer, the technical scheme of the present application will be described in detail below. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor belong to the scope of protection of the present application.

[0104] Audit is an important part of the national supervision system and an important force to promote the modernization of the national governance system and governance capacity.

[0105] Currently, there are some problems in audit work, such as the difficulty for auditors to establish high-quality risk judgment logic due to unfamiliarity with business, the difficulty in analyzing unstructured audit evidence, the large amount of standardized and repetitive work, the lack of accumulation, optimization and application ability of audit experience continuity, and other pain points. Intelligent means must be used to improve the efficiency of audit work and promote the high-quality development of the audit profession. The audit departments at all levels actively carry out the positive practice of artificial intelligence empowerment of audit work, and how to simplify the audit process, solve the work pain points and improve the audit quality and efficiency through intelligent technology has become a hot topic for the practical and academic circles to explore together.

[0106] At present, large language models exhibit strong general knowledge and reasoning capabilities. Related research shows that the combination of large language models in auditing not only improves auditing efficiency and accuracy, but also enhances the depth and breadth of auditing, bringing innovation to auditing. The natural language interaction and intelligent question answering capabilities of large language models can significantly reduce the threshold for auditors when using intelligent auditing systems, improving the coverage and accuracy of auditing. For example, through text mining technology, large language models can automatically extract key information and data from massive amounts of audit-related documents, helping auditors quickly grasp the background information, business characteristics, and potential risk points of the audit object. The large amount of data and analysis results generated during the auditing process need to be clearly explained and interpreted. Large language models can provide natural language descriptions of data interpretation and analysis results, making the audit results easier to understand and communicate to business-related parties. In particular, the code generation capabilities of large models can generate data analysis SQL scripts for audit items and ideas, assisting in data analysis, thereby improving the efficiency of auditing data processing and analysis, enhancing the accuracy and reliability of auditing, and providing new ideas and technical paths for national auditing.

[0107] With the development of large language model (LLM) technology in natural language understanding and code generation, its application in auditing data analysis becomes possible. Traditional auditing data analysis has high SQL threshold, scattered scenarios, and repetitive work, as follows:

[0108] High technical threshold: requires professional personnel to write SQL, with long response cycle (average > 2 hours);

[0109] Serious data silos: on-site audit data is scattered in heterogeneous databases such as Oracle and MySQL;

[0110] Security risks are prominent: original data transmission is prone to information leakage (such as CN114580944A solution), which relies on full data transmission.

[0111] Existing improvement solutions (such as CN116258324A) only support a single database and still cannot achieve natural language interaction and cross-platform intelligent migration.

[0112] The present application provides an auditing data analysis method based on a large language model, which is a more advanced and comprehensive auditing data analysis method based on a large language model. By constructing an intelligent data analysis system based on a large model, an integrated process of "natural language questioning -> automatic SQL generation -> data analysis results" is realized, helping auditors efficiently complete data analysis work, solving the drawbacks of existing technologies, and having important research significance and use value. The specific content is as follows:

[0113] For example,Figures 1-32 As shown, comprising the following steps:

[0114] S1: Natural language request analysis: through the question classifier of the intelligent analysis module, receiving the natural language query request transmitted by the client through VPN encryption, identifying the audit intention and data type, and classifying into data query, document retrieval or general task;

[0115] S2: Determine whether the data type involved in the request is structured data or unstructured data;

[0116] S21: If the request involves structured data, trigger the structured query engine to generate abstract query logic, and call the SQL assistant module to execute cross-database adaptation, including:

[0117] Identify the target database type through the SQL syntax adapter, including Oracle, MySQL or SQL Server;

[0118] Verify the syntax security of the generated script through the pre-trained model, and prevent scripts containing high-risk operations from being executed, including unauthorized table deletion, permission escape query;

[0119] Use the script conversion engine to convert the abstract query logic into the target database dialect script, and perform equivalent reconstruction on the paging query and window function;

[0120] Execute the query and feed back the result in natural language form through the model explanation unit;

[0121] S22: If the request involves unstructured data, start the document assistant module, including:

[0122] Extract text, table and image content in the user's private knowledge base through the multi-format document parser;

[0123] Match the associated clauses in the policy and regulation common library based on the hybrid retrieval engine, and adopt a dynamic weighted scoring strategy, when the semantic similarity is >0.7, prefer to use the semantic matching result;

[0124] Mark the data source and confidence through the reference tracing unit;

[0125] Also includes: constructing a joint vector index of the documents in the user's private knowledge base and the policy and regulation common library, and realizing cross-library associated retrieval through semantic similarity calculation;

[0126] S3: Multi-modal result fusion:

[0127] The general intelligent module performs joint analysis on the structured query result and the unstructured retrieval result, including:

[0128] S31: Call domestic large model to perform text generation and risk pattern identification;

[0129] S32: Perform OCR recognition and key field extraction on image type data;

[0130] S33: Generate an audit analysis report containing data traceability markers;

[0131] Multi-modal result fusion also includes automatically generating an exception marker and triggering a manual review process when detecting that the difference between structured data and unstructured data in describing the same entity exceeds a preset threshold (numerical difference ≥ 5% or text description confidence difference ≥ 0.3);

[0132] The manual review interface supports difference highlighting, comment version tracking, and multi-person collaborative review, and the review results are stored through blockchain.

[0133] At present, large language models exhibit strong general knowledge and reasoning ability. Related research shows that the combination of large language models in auditing not only improves auditing efficiency and accuracy, but also enhances the depth and breadth of auditing, bringing innovation to auditing. The natural language interaction and intelligent question and answer functions of large language models can greatly reduce the threshold for auditors when using intelligent auditing systems, improving the coverage and accuracy of auditing. For example, through text mining technology, large language models can automatically extract key information and data from massive audit-related documents, helping auditors quickly grasp the background information, business characteristics, and potential risk points of the audit object. A large amount of data and analysis results generated during the auditing process need to be clearly explained and interpreted. Large language models can provide natural language descriptions of data interpretation and analysis results for auditors, making audit results easier to understand and communicate to business-related parties. In particular, the code generation capability of large models can generate data analysis SQL scripts for audit items and ideas, assisting in completing data analysis, thereby improving the efficiency of auditing data processing and analysis, and enhancing the accuracy and reliability of auditing, providing new ideas and technical paths for national auditing. However, the current auditing work faces many drawbacks, such as: previous audit projects mainly rely on auditors to set rules for sample anomaly screening, and the audit effect is often closely related to the experience level of auditors. The present invention can expand the coverage of various original audit data models by using large language model intelligent data analysis to work, providing another data-based risk insight and identification path for auditors. In addition, given the demand for big data intelligent analysis: according to the job scene process to establish query judgment logic and form the corresponding code. Furthermore, using artificial intelligence technology to partially realize automated and intelligent auditing work, realizing the saving of auditing labor cost and the rational allocation of auditing resources, forming the focusing ability of auditing resources, helping auditors to complete external data mining, data collection, data comparison and analysis, system security check, and audit working papers writing and other repetitive transactional work, which can efficiently improve the automatic processing ability, solve the problem of having ideas but lacking SQL skills. The auditing data analysis method based on large language model provided by the present invention is a more advanced and comprehensive auditing data analysis method based on large language model, which realizes the integration process of "natural language questioning -> automatic SQL generation -> data analysis results" by constructing an intelligent data analysis system based on large models, helps auditors to efficiently complete data analysis work, solves the drawbacks in the prior art, and has important research significance and use value.

[0134] It needs to be explained that big data technology has played an increasingly important role, but data analysis needs to be deep and thorough, and auditors need to have high data analysis technical ability, and technical "small details" often become "roadblocks" and restrict the quality and effect of data analysis. In order to reduce the data analysis technical difficulty of auditors in work, the technical scheme of the present application purchases intelligent data analysis research service of large model, researches and proves the feasibility of intelligent data analysis of large model technology, and realizes the technology application on the basis of unit intranet environment and hardware conditions, uses large model, Text2SQL and RAG and other technical methods, so that auditors can meet the actual needs of intelligent data analysis and other audits through the question and answer mode, and no longer need to learn to write SQL code, and improve the auditing efficiency.

[0135] The present application also provides an audit data analysis system based on a large language model, which comprises an intelligent analysis module 1, an SQL assistant module 2, a document assistant module 3 and a general intelligent module 4. The data flow between the modules is transmitted through an intranet encryption channel.

[0136] The intelligent analysis module 1 is used to realize natural language driven audit data query and on-site analysis, and comprises a question classifier, a structured query engine and an abstract generator, and supports interaction with the client through a VPN secure channel and structured metadata; the intelligent analysis function module 1 comprises a data query module and a data analysis module, the data query module is configured to: in response to a user operation on a data query user interface element, receive a data query request input in an associated input box; in response to a user operation on a sending user interface element, acquire relatively complete information corresponding to the data query request based on an industry data resource knowledge base; the data analysis module is configured to: in response to a user operation on a data analysis user interface element, present an industry data resource directory user interface element; wherein the industry data resource directory user interface element comprises a plurality of information table options and associated selection controls; in response to a user operation on the selection control, based on the selected information table, execute data analysis and generate an analysis result; the industry data resource directory user interface element is further configured to present preview content of the corresponding information table in response to a user operation on an information table associated preview control;

[0137] The SQL assistant module 2 provides cross-database script migration services, including a SQL syntax adapter, a script conversion engine and a model interpretation unit, and supports Oracle, MySQL and SQL Server multi-dialect environments;

[0138] The document assistant module 3 constructs a user private knowledge base and a policy and regulation common library, integrates a multi-format document parser, a hybrid retrieval engine and a reference tracing unit;

[0139] General intelligent module 4: deploy domestic large models through intranet privatization, provide general AI capabilities such as text generation, image recognition, and multilingual translation, and domestic large models include DeepSeek, Qwen, and GLM;

[0140] The intelligent analysis module 1 includes a data security interaction unit and a temporary table management unit. The client only transmits the database table structure or the desensitized metadata field to the audit data analysis system. The original data is retained for local query execution. The temporary table management unit supports storing the query results of sensitive fields such as ID numbers and the names of the deceased in encrypted temporary tables and setting automatic cleaning strategies.

[0141] The SQL assistant module 2 includes:

[0142] Dialect adapter: based on syntax tree parsing, PL / SQL and T-SQL are automatically converted, and data type differences are corrected;

[0143] Execution plan visualization unit: the SQL execution process is decomposed into natural language interpretation steps, and key performance bottleneck points are labeled;

[0144] The document assistant module 3 supports:

[0145] Multi-modal retrieval: paragraph-level vectorization is performed on PDF and DOCX documents, and hybrid search is implemented by combining keyword matching and semantic similarity calculation;

[0146] Dynamic update mechanism: automatically capture policy documents, generate structured knowledge graphs, and associate them with the question and answer system.

[0147] In specific embodiments of the present application, the data analysis module of the audit data analysis system based on a large language model is further configured to:

[0148] Below the industry data resource directory user interface element, an audit project data list user interface element is presented;

[0149] The audit project data list user interface element includes a database selection control. In response to user operation of the database selection control, a list of selectable databases is presented and the user is allowed to select a target database;

[0150] The audit project data list user interface element presents, for the selected target database, a plurality of data entry options contained therein and associated selection controls. In response to user operation of the data entry selection control, based on the selected data entry, data analysis is performed and analysis results are generated;

[0151] The audit project data list user interface element is further configured to, for a data entry, present a detailed information preview of the corresponding data entry in response to a user operation on an associated entry preview control;

[0152] The audit project data list user interface element comprises a file upload control;

[0153] configured to:

[0154] present a file upload dialog in response to a user operation on the file upload control;

[0155] receive user input through the file upload dialog, including: a database name, a table name;

[0156] receive a data file uploaded by the user through the file upload dialog, the data file being in a predetermined format;

[0157] store the data file into the audit project data in response to a save operation;

[0158] The audit project data list user interface element comprises a database connection adding control;

[0159] configured to:

[0160] present a database type selection dialog in response to a user operation on the database connection adding control;

[0161] present a connection configuration dialog in response to selection of a database type in the database type selection dialog;

[0162] receive connection configuration information input by the user through the connection configuration dialog;

[0163] perform a database connection test based on the connection configuration information in response to a test connection operation;

[0164] add a new database connection to the audit project data in response to a save operation after the connection test is passed.

[0165] In an embodiment of the present application, the predetermined format in the audit data analysis system based on the large language model is CSV format.

[0166] In an embodiment of the present application, the system in the audit data analysis system based on the large language model is configured to limit the size of a single uploaded file to be no more than a predetermined threshold.

[0167] It should be noted that:

[0168] (1) Regarding the research and evaluation of large model capabilities in the specific solutions of the present application: the large language model-based audit data analysis system in the specific embodiments of the present application provides the service capability of researching and evaluating the data analysis capabilities of mainstream large models. According to the requirements of audit work, the original data analysis capabilities of large models are evaluated, and the improvement of the data analysis capabilities of large models is fine-tuned. A large model service framework with good compatibility and stability can be built in an internal network environment or locally, supporting at least three mainstream large models at home and abroad, and allowing users to freely switch between different large models according to their needs. When switching models, users can easily operate without restarting the system or adjusting other settings, and the switching process is smooth without affecting other functions of the system. In addition, the framework should also have an open program-level interface to facilitate calling large models for various operations such as loading models, executing reasoning, and obtaining state information, etc.

[0169] (2) Regarding multi-database adaptation: the large language model-based audit data analysis system in the specific embodiments of the present application provides the ability to interface and manage different types of databases using large models, supporting no less than three mainstream database systems, and the adaptation process should be efficient, automated, and have better flexibility and scalability. It can automatically and intelligently capture the underlying structure of the database, including table structure, field definition, index, view, stored procedure, and other key information; it can analyze the relationships between tables, such as primary-foreign key constraints, index associations, and dependency relationships, and automatically generate a relational model of the database; it can ensure the integrity of the database structure, providing a foundation for subsequent data processing, query optimization, and data analysis; it can flexibly adjust the capture and parsing strategies according to the differences in database types and versions, ensuring wide compatibility with various databases; it can intelligently write complex database structure mappings or relational models.

[0170] (3) Database preprocessing: the large language model-based audit data analysis system in the specific embodiments of the present application provides the ability to intelligently filter and clean multiple types of database tables, intelligently identifying and filtering out tables that do not contain key business information, to eliminate redundant, non-business meaningful tables and data, such as temporary tables, log tables, and other unrelated auxiliary tables. By eliminating unnecessary tables and data, the database structure is more concise, reducing the time and cost of processing irrelevant data.

[0171] (4) Large model intelligent data analysis: the application provides a large language model-based audit data analysis system in the specific embodiment, which provides the ability to provide intelligent and efficient data analysis means for auditors using large model technology, allowing auditors to describe audit ideas in their familiar expression, understand audit requirements through intelligent analysis, and automatically and accurately extract relevant data from databases to generate data tables. In this process, auditors do not need to master complex database query languages to perform complex data analysis tasks more quickly and accurately, improving audit efficiency. In addition, the script automatically generated by the large model can be manually modified and optimized to modify filtering conditions, field selection, logical structure, etc. to ensure that the generated analysis results meet the audit objectives. It can support the SQL general standard and be compatible with multiple database SQL dialects, including P-SQL, T-SQL, MySQL, etc. dialects. At the same time, it has the ability to train database question and answer pairs and integrate more explanatory languages to improve intelligent data analysis capabilities. Auditors can obtain more accurate and personalized data analysis results, further improving work efficiency and decision-making quality.

[0172] (5) Analysis idea intelligent recommendation: the application provides the ability to provide intelligent recommended analysis ideas based on the user's use of the large language model-based audit data analysis system in the specific embodiment, which can intelligently compare the similarity of historical analysis ideas and model usage frequency, identify the most suitable high-frequency or similar analysis model for the task requirements, and help auditors quickly locate key data. It can establish a historical analysis library to save query conditions, table structure, field selection, and analysis ideas, accumulate historical data, analyze data similarity, and improve the intelligent level and recommendation effect of the model through machine learning and deep learning algorithms.

[0173] (6) Large model syntax explanation: the application provides the ability to help auditors solve SQL programming problems in the form of questions and answers using large model technology based on the large language model-based audit data analysis system in the specific embodiment. Auditors can automatically obtain accurate SQL query statements and personalized guidance by asking natural language questions, improving their SQL knowledge and practical skills; it can intelligently adapt to the syntax of not less than 3 different databases such as P-SQL, T-SQL, MySQL, etc. to ensure the compatibility and execution effect of the query statement; it can provide best practices, common error troubleshooting and query optimization suggestions to help auditors improve their SQL skills and avoid common programming pitfalls. It can establish a feedback mechanism based on historical problem optimization answers to improve model understanding capabilities.

[0174] (7) The large model explains the audit analysis approach. The audit data analysis system based on the large language model in the specific embodiment of the present invention provides the ability to interpret the existing analysis model into an analytical intent that auditors can understand using the large model. It can intelligently parse the audit model and explain the role of each part of the complex audit model and the underlying analysis logic in a concise way. It helps auditors to clearly understand the business meaning and execution process of each step of the audit model, so as to better grasp the audit approach and the purpose of data analysis. It can support the SQL syntax of multiple databases such as P-SQL, T-SQL, and MySQL, ensuring that the query statements of different databases can be accurately parsed and interpreted. It can interpret complex SQL script content such as complex multi-table joins, nested queries, complex aggregate functions and condition filtering. It can provide processing methods for handling historical high-quality question and answer records and associating creation time, etc., to further improve the model understanding and analysis capabilities.

[0175] (8) Audit Model for Large Model Correction Issues: The audit data analysis system based on a large language model in the specific embodiments of this invention provides the ability to correct erroneous SQL scripts using a large model. It can automatically check and correct syntax errors, logical problems, and potential performance bottlenecks in the analysis model, and provide correction suggestions or automatic repair functions to ensure the correctness and optimized execution of the analysis model. It can support SQL syntax of multiple databases such as P-SQL, T-SQL, and MySQL, and can accurately check and optimize for different database dialects. It can identify and modify, but is not limited to, missing parentheses, incorrect join conditions, inefficient queries, and other problems.

[0176] (9) Large-Model Migration Analysis Model: The audit data analysis system based on a large language model in this invention provides the ability to migrate existing analysis models to other types of databases using a large model. This simplifies the conversion process between analysis models from different databases, automatically understanding the differences between various databases to complete the migration task. It supports SQL syntax for mainstream databases such as P-SQL, T-SQL, and MySQL, automatically converting SQL scripts from one database into compatible scripts for another. It can automatically identify and handle syntax differences between different databases, such as data type conversions, differences in function calls, and adjustments to query structures. During the migration process, the intelligent service adjusts according to the syntax rules of the target database to ensure that the migrated analysis model can execute correctly in the new environment.

[0177] (10) Document intelligent assistant, the application provides the ability to provide various document Q&A required for relevant analysis ideas during data analysis by using a large model, can intelligently analyze and interpret complex policies, guidelines and other documents through natural language input, establish a special knowledge base for audit projects, support quick search and matching of relevant clauses, and give the source of the relevant content. At the same time, it has the ability to collect historical question and answer records for training the model.

[0178] In the embodiments of the present application, in combination with the audit business scenario, the NVIDIA ecosystem may have more advantages. On the one hand, the NVIDIA hardware performance is excellent, the NVIDIA GPU has powerful parallel computing capability, which can quickly process massive data, etc., which speeds up the audit; and in the current wide application of large models, it can well support TensorFlow, PyTorch and other frameworks, accelerate model training and inference, and help accurate audit. On the other hand, the software ecosystem is rich, CUDA is a parallel computing platform and programming model, combined with cuDNN that optimizes common operations of neural networks and improves computing efficiency, and RAPIDS and other tool libraries, which provide a convenient and efficient way for developers to build and optimize deep learning models and process and analyze audit data. At the same time, the hardware and software are deeply integrated, and the hardware performance is fully played through collaborative optimization, the unified programming model reduces the development cost, and improves the business scalability and flexibility. These advantages make the NVIDIA ecosystem stand out, making it the preferred choice for audit business, rather than other ecosystems. Therefore, in the embodiments of the present application, taking the NVIDIA ecosystem as an example, the hardware environment for the large language model intelligent data analysis research service demand in the embodiments of the present application is built as follows:

[0179] (1) NVIDIA driver installation

[0180] Confirm the graphics card model: In the Linux terminal, input the command "lspci | grep -invidia", this command can list all the device information related to NVIDIA in the system, and accurately identify the specific model of the GPU. In addition, you can also use the "nvidia-smi" command (if the basic driver is installed) to view the GPU model and the current driver version details, which provides a strong basis for downloading the appropriate driver later.

[0181] Download the driver: Visit https: / / www.nvidia.com / Download / index.aspx and manually select the appropriate operating system version (such as Ubuntu, CentOS, etc.), graphics card model, and other relevant parameters under the Linux system to ensure that the selected driver is perfectly compatible with the hardware and operating system. After completing the selection, click "Search" to download the latest version of the adapted driver, unless there are specific compatibility issues that require the use of an older version.

[0182] Driver installation:

[0183] For most Debian or Ubuntu-based systems, after downloading, enter the terminal in the directory where the driver is located and execute the command "sudo chmod+x NVIDIA-Linux-x86_64-*.run" (representing the driver version number) to give the driver execute permissions, then run "sudo. / NVIDIA-Linux-x86_64-.run" to start the installation program. During the installation process, the installation wizard will prompt a series of options, and it is also recommended to choose "Custom Installation" to uncheck unnecessary additional components such as NVIDIA promotional software to avoid wasting resources. Follow the prompts to complete the installation operation.

[0184] For Red Hat or CentOS-based systems, the installation process is slightly different. First, install some necessary dependencies such as "yum install kernel-devel-y", etc., and then run the driver installation program in a similar manner as described above and complete the installation.

[0185] Verify driver installation: After installation, enter "nvidia-smi" in the terminal again. If it can normally display the GPU model, driver version, GPU usage, and GPU memory usage, etc. detailed information, it means that the NVIDIA driver has been successfully installed, and the system's NVIDIA GPU is ready for subsequent CUDA installation.

[0186] (2) CUDA installation

[0187] Select the appropriate version: Go to the CUDA Toolkit Archive page of the NVIDIA official website, and according to the specific release version of the Linux system (such as Ubuntu 22.04, CentOS 7, etc.) and the graphics card model, choose the appropriate CUDA version for download. Note to ensure that the selected version is compatible with the installed driver and the software package to be used subsequently.

[0188] Download and installation: After downloading, for Ubuntu system, enter the download directory in the terminal and execute the command "sudo dpkg -i cuda-repo-*.deb" (representing the CUDA version-related identifier) to add the CUDA software source, then run "sudo apt-get update" to update the package list, and then execute "sudo apt-get install cuda-y" to complete the installation. Or through the downloaded RUN package, install the corresponding software package to achieve the installation of CUDA. For CentOS system, you need to download the corresponding RPM package first, run "sudo yum localinstall cuda-repo-.rpm" to add the software source, and then follow the similar operation as Ubuntu to complete the installation process through the "yum" command.

[0189] Environment variable configuration: After installation, you need to manually configure the CUDA environment variables. Open the "~ / .bashrc" file (use the command "vi ~ / .bashrc") and add the following lines at the end of the file:

[0190] export PATH= / usr / local / cuda / bin:$PATH

[0191] export LD_LIBRARY_PATH= / usr / local / cuda / lib64:$LD_LIBRARY_PATH

[0192] Save and exit, then execute "source ~ / .bashrc" in the terminal to make the environment variables effective.

[0193] Verify CUDA installation: In the terminal, enter the "nvcc -V" command. If the command line returns CUDA version information, it means that CUDA has been successfully installed.

[0194] (3) cuDNN installation

[0195] Register and log in to NVIDIA account: Since cuDNN is a proprietary software of NVIDIA, you need to access the https: / / developer.nvidia.com / cudnn-downloads download page to register and log in to the NVIDIA account first.

[0196] Download cuDNN: According to the installed CUDA version, carefully select the corresponding cuDNN version, which is usually presented in the form of a compressed package, and make sure that the selected version is accurately matched with the Linux system distribution and CUDA version.

[0197] Decompression and installation: After downloading, decompress the compressed package to any specified directory, such as " / tmp". Then enter the decompressed directory, for Ubuntu systems, execute the commands "sudo cp -r include / * / usr / local / cuda / include / " and "sudo cp -r lib / * / usr / local / cuda / lib64 / " to copy the relevant files to the corresponding folders in the CUDA installation directory; for CentOS systems, the operation is similar, and ensure that the file copying is accurate.

[0198] Verify cuDNN installation: Enter the " / usr / local / cuda / extras / demo_suite" directory and execute the Linux equivalent commands ". / deviceQuery" and ". / bandwidthTest" for "deviceQuery.exe" and "bandwidthTest.exe". If the results show "pass", it indicates that cuDNN has been successfully installed.

[0199] (4) Adapt the python package

[0200] PyTorch: Access the PyTorch website https: / / pytorch.org / get-started / previous-versions / and select the corresponding PyTorch version according to the installed CUDA version, then use the "conda" or "pip" command to install. For example, if CUDA 10.2 is installed, execute "conda install pytorchtorchvision cuda10.2-c pytorch" or "pip install pytorch torchvision cuda10.2-cpytorch" in the terminal. Verify that PyTorch has been successfully installed and can use CUDA. First, start the Python interactive environment in the terminal and input import torch. If there is no error, it indicates that PyTorch has been successfully installed. Then input torch.cuda.is_available(), if it returns True, it means that PyTorch has successfully configured CUDA. Further, you can check the number of available CUDA devices by torch.cuda.device_count() and obtain the CUDA device name by torch.cuda.get_device_name(0) to confirm.

[0201] TensorFlow: Install the TensorFlow GPU version using the "pip" command, such as "pip install tensorflow-gpu". After installation, open the Python environment, import TensorFlow, and use the "tf.test.is_gpu_available()" function to verify. If the return result is true, it means that the GPU has been successfully adapted.

[0202] In the specific embodiments of the present application, the support service is the core support of the technical solution of the present application, including the support of the basic framework, model service and data storage, which will be analyzed as follows:

[0203] (1) Basic framework

[0204] The deep learning framework is a collection of pre-built tools, libraries and modules, providing a convenient, efficient and standardized platform for developers to simplify the entire process of deep learning model design, training and deployment. Its role is crucial, as it encapsulates the underlying complex mathematical operations, computation graph construction, automatic differentiation and hardware acceleration functions, allowing developers to focus on model architecture design and business logic implementation without delving into technical details, thereby greatly reducing the development threshold of deep learning and improving development efficiency and model maintainability.

[0205] Deep learning frameworks have many advantages, such as high modularity and scalability, allowing developers to easily combine and extend different layers and modules to build complex model architectures; efficient performance optimization, taking full advantage of GPU hardware resources for parallel computing to speed up the model training process; at the same time, rich documentation and active community support provide developers with a wealth of learning resources, code examples and problem-solving approaches, promoting technology exchange and sharing, accelerating the development and application of deep learning technology.

[0206] 1) PyTorch

[0207] Among the many deep learning frameworks, PyTorch is favored for its simple and intuitive dynamic computation graph construction method. When building large language models, developers can easily define the model architecture using PyTorch, such as the Encoder-Decoder structure based on the Transformer architecture. Through its flexible tensor operations and automatic differentiation functions, forward and backward propagation calculations are easily performed, and model parameter optimization is achieved. At the same time, the close integration of PyTorch with Python and its rich pre-trained model library provide developers with the convenience of quickly building and fine-tuning large language models, efficiently handling natural language processing tasks such as text generation and machine translation.

[0208] The large model built using PyTorch includes but is not limited to GPT, BERT, ChatGLM, etc. The specific embodiments of the present application take ChatGLM as an example to introduce the ChatGLM large model building principle, as shown in Figure 3 .

[0209] GLM proposes a novel autoregressive blank filling objective, which constructs NLU tasks as fill-in-the-blank questions containing task descriptions, which are answered by autoregressive generation.

[0210] ①The original text is [x1, x2, x3, x4, x5, x6]. Two ranges [x3] and [x5, x6] are sampled.

[0211] ②Replace the sampled ranges in Part A with [M], and shuffle these ranges in Part B.

[0212] ③GLM autoregressively generates Part B. Each range is preceded by [S] as input and followed by [E] as output. Two-dimensional position encoding represents the position across ranges and within ranges.

[0213] ④Self-attention mask. The area marked as X is shielded. The tokens of Part A can focus on themselves (the first box, there are three boxes in the figure, from top to bottom, the first box, the second box and the third box), but cannot focus on B. The tokens of Part B can focus on the previous generation in A and B (the second box and the third box correspond to two ranges respectively).

[0214] The following introduces multi-objective pre-training of different parameters.

[0215] Three autoregressive fill-in-the-blank pre-training objectives are designed for three types of downstream tasks, which are alternately performed:

[0216] ①token-level objective word-level

[0217] The length of the segment is sampled from a Poisson distribution with a mean of 3, and up to 15% of the characters in the original text are masked, and then the positions of the blank segments are randomly arranged in the text. This objective is for natural language understanding tasks.

[0218] ②sentence-level objective sentence-level

[0219] Randomly mask several text segments from the document, and each text segment must be a complete sentence. The number of words masked is 15% of the length of the entire document. This objective is for seq2seq tasks, which often predict entire sentences or paragraphs.

[0220] ③document-level objective document level

[0221] The pre-training process samples segments from a uniform distribution with lengths ranging from 50% to 100% of the original text length. This pre-training objective is to generate unconditional long text.

[0222] The only difference between the three types of pre-training objectives is the number and length of the span (segment).

[0223] The following content compares and contrasts the three types of pre-training objectives with masked languages.

[0224] Autoregressive imputation is somewhat similar to masked language models. It first samples a portion of the input text, replaces it with the [MASK] marker, and then predicts the text segment corresponding to the [MASK] marker. However, unlike masked language models, the prediction process uses an autoregressive approach.

[0225] ① When the length of the masked segment (span) is 1, the space filling task is equivalent to mask language modeling;

[0226] ② When text 1 and text 2 are concatenated together, and then text 2 is completely masked, the space filling task is equivalent to the conditional language generation task.

[0227] ③ When all the text is masked, the space filling task is equivalent to the unconditional language generation task.

[0228] The principle behind building large models with PyTorch is as follows: it primarily relies on its dynamic computation graph and flexible API design. Users can build complex model structures by defining neural network classes and utilize PyTorch's automatic differentiation mechanism (Autograd) for efficient training and optimization. Furthermore, PyTorch supports distributed training and GPU acceleration, making the training of large-scale models even more efficient.

[0229] The steps to build a large model using PyTorch are as follows:

[0230] ① Deploy the PyTorch deep learning environment on the server;

[0231] ② Use PyTorch to build a Transformer algorithm program;

[0232] ③ On the server, use PyTorch tools to train the model on the algorithm and data;

[0233] ④ Fine-tune and train the above basic model to improve its dialogue capabilities.

[0234] 5. Aligning the large model with human values through training;

[0235] 6. Deploying the trained large model for inference, using the Pytorch parallel framework for inference configuration.

[0236] The advantages of using PyTorch for large model development include:

[0237] Dynamic computation graph: PyTorch is known for its dynamic computation graph, which supports flexible model debugging and rapid prototyping, making it ideal for academic research and experimentation.

[0238] Ease of use: Its simple and intuitive API and dynamic graph features make code writing and debugging more convenient, especially for Python developers.

[0239] Flexibility: Supports real-time variable viewing and model structure adjustment, suitable for rapid iteration and research.

[0240] Ecosystem: Although the ecosystem is not as mature as TensorFlow, the community is active, making it suitable for academic and research use. PyTorch has numerous open-source models and tool libraries available, speeding up the development process.

[0241] Efficient computing performance: PyTorch significantly improves the training speed of large-scale models through GPU acceleration and distributed training functions.

[0242] Rich community support and resources: PyTorch has an active community and a large number of open-source projects, making it easy for developers to obtain help and support.

[0243] Easy fine-tuning and deployment: PyTorch provides convenient model fine-tuning and deployment tools, making it easy to apply pre-trained models to practical tasks.

[0244] Given the advantages of PyTorch, some current applications based on large model development use the PyTorch framework. This includes the environment required for large model inference and the environment required for large model fine-tuning.

[0245] 2) TensorFlow

[0246] TensorFlow stands out with its powerful visualization tools and distributed training capabilities. In the construction of large language models, TensorFlow's static computational graph mechanism excels in large-scale model training and deployment. By pre-defining and optimizing the computational graph, TensorFlow can efficiently train models in a distributed environment, fully utilizing cluster computing resources to accelerate model convergence. Its high scalability allows it to easily handle large-scale datasets and complex models, such as language models with massive parameters, effectively managing the training process and resource allocation to ensure model training efficiency and stability. TensorFlow is widely used in large language model development in fields such as speech recognition and intelligent customer service.

[0247] The principle of building a large model using TensorFlow is as follows: TensorFlow represents the model through a data flow graph (Computational Graph), and developers can define operations and variables to build neural network models on the data flow graph. The data flow graph diagram of TensorFlow model building is shown in Figure 4 .

[0248] The steps for building a large model using TensorFlow are as follows:

[0249] ① Deploy TensorFlow deep learning environment on the server;

[0250] ② Use TensorFlow to build the Transformer algorithm program;

[0251] ③ On the server, use TensorFlow tools to train the algorithm and data model;

[0252] ④ Fine-tune the above basic large model to train the dialogue ability of the large model;

[0253] ⑤ Align the training of the large model with human values;

[0254] ⑥ Deploy the trained large model for inference, and use TensorFlow parallel framework for inference configuration.

[0255] The advantages of using PyTorch to build a large model include:

[0256] Static computational graph: TensorFlow uses a static computational graph, which requires defining the computational graph before execution, making it suitable for large-scale production deployment.

[0257] Strong ecosystem: TensorFlow has a rich API and tool library (such as Keras, TensorBoard, TFX), supports multiple hardware platforms, and is suitable for training and deployment of large-scale datasets and complex models.

[0258] Community Support: Supported by Google, with extensive community resources and documentation, suitable for industrial applications.

[0259] Performance Optimization: Excellent performance in distributed training and high-performance computing, suitable for training large-scale models.

[0260] Cross-platform support: TensorFlow supports multiple hardware and platforms, including CPU, GPU and TPU, allowing large models to be trained and deployed on different devices.

[0261] 3) PaddlePaddle

[0262] PaddlePaddle, as a domestic deep learning framework, has good support for Chinese natural language processing tasks and localization advantages. When building large language models, PaddlePaddle provides rich Chinese pre-training models and tools, making it easy for developers to train and optimize models for Chinese text data. Its simple and easy-to-use API and efficient kernel implementation allow developers to quickly get started and build high-performance large language models, such as text classification, sentiment analysis, etc. PaddlePaddle also focuses on cooperation with the domestic industry, providing customized deep learning solutions for enterprises, promoting the widespread application and development of deep learning technology in China, and playing an important role in intelligent writing, knowledge graph construction, etc. It provides strong support for the development and application of large language models.

[0263] The principle of building a large model with Paddlepaddle is as follows: relying on its powerful deep learning framework, supporting dynamic and static graph programming modes, providing rich APIs and tools such as PaddleSlim, PaddleDetection, etc., for model optimization, compression and distributed training.

[0264] The steps for building a large model using Paddlepaddle are as follows:

[0265] ① Deploy Paddlepaddle deep learning environment on the server;

[0266] ② Use Paddlepaddle to build a Transformer algorithm program;

[0267] ③ On the server, use Paddlepaddle tools to train the algorithm and data model;

[0268] ④ Fine-tune the above basic large model to train the dialogue ability of the large model;

[0269] 5. Align the large model with human values through training;

[0270] 6. Deploy the trained large model for inference, using the Paddlepaddle parallel framework for inference configuration.

[0271] The steps to build a large model using Paddlepaddle are as follows:

[0272] Ease of use and high performance: PaddlePaddle provides both dynamic and static modes, with clear Chinese documentation, making it easy to get started, and fast running on CPU and GPU.

[0273] Industrial applications: Focus on industrial applications, provide rich pre-training model library and industry solution, suitable for enterprise deployment. Provides a comprehensive deep learning platform, including framework, tools and services, supporting the whole process from model development to industrial application.

[0274] Ecosystem scalability: PaddlePaddle provides high-performance inference engines, supporting large-scale distributed training and model compression, etc. Advanced functions suitable for large-scale production environments.

[0275] In the specific embodiments of the present application, the specific connection logic is:

[0276] 1. Intelligent analysis module → SQL assistant module

[0277] (1) Method: RESTful API (HTTPS two-way authentication)

[0278] (2) Data flow:

[0279] Question classifier identifies SQL query requirements → structured query engine generates abstract syntax tree (AST)

[0280] Send AST to SQL assistant module → script conversion engine adapts to target database dialect (Oracle / MySQL, etc.)

[0281] (3) Return: Encrypted data set (returned to the client through the VPN channel)

[0282] 2. Intelligent analysis module → document assistant module

[0283] (1) Method: gRPC long connection (supports large file streaming)

[0284] (2) Data flow:

[0285] Abstract generator triggers policy retrieval → calls document assistant hybrid retrieval engine

[0286] Document Parser extracts PDF / Word key fields (e.g. contract amount, terms)

[0287] (3) Return: Structured text snippet + reference trace location (for audit evidence chain)

[0288] 3. Intelligent Analysis Module → General Intelligent Module

[0289] (1) Method: Private SDK direct connection (memory-level interaction)

[0290] (2) Capability Invocation:

[0291] Text Generation: Automatically convert SQL results into audit reports

[0292] Image Recognition: Process scanned receipts (in conjunction with Document Assistant)

[0293] Multilingual Translation: Support for cross-border audit scenarios

[0294] 4. Cross-module Collaboration

[0295] (1) Document Assistant → General Intelligence:

[0296] Private Knowledge Base Vector Indexing → Semantic Retrieval through Large Models (DeepSeek / Qwen Embedding)

[0297] (2) SQL Assistant → General Intelligence:

[0298] Model Explanation Unit translates SQL execution plans into natural language (GLM model fine-tuning)

[0299] Intelligent Analysis Module as the dispatch center, General Intelligent Module as the AI capability base, SQL / Document Assistant as vertical capability support, forming a closed-loop audit workflow.

[0300] Document Assistant Module Support:

[0301] Multi-modal Retrieval: Paragraph-level vectorization of PDF, DOCX documents, combined with keyword matching and semantic similarity calculation for hybrid search

[0302] Dynamic Update Mechanism: Automatically scrape policy documents from the Audit Office and Ministry of Finance websites, generate structured knowledge graphs, and associate them with the Q&A system.

[0303] SQL Assistant Module Includes:

[0304] Dialect Adapter: Automatic conversion of PL / SQL and T-SQL based on syntax tree parsing, correcting data type differences (e.g. Oracle's NUMBER to SQL Server's DECIMAL)

[0305] Execution plan visualization unit: decompose SQL execution process into natural language explanation steps, and mark key performance bottleneck points.

[0306] Intelligent analysis module includes:

[0307] Data security interaction unit: the client only transmits the database table structure (DDL) or the desensitized metadata field (work order number, service address) to the platform, and the original data is retained in the local to execute the query;

[0308] Temporary table management unit: support storing the query result of sensitive fields (ID number, dead name) in the form of encrypted temporary table, and setting automatic cleaning strategy.

[0309] As Figures 21-24 shown:

[0310] 1. Click the "Start General Intelligence" button at the bottom right of the page. This function can be used without logging in to display the

General Intelligence

[0311] 2. Enter text in the input box below;

[0312] 3. Click the "Send" button to the right of the input box to start a conversation with the audit data analysis system. During the answering process, you can click the "Stop" button to the right of the input box at any time to stop answering;

[0313] According to the current network situation and the number of users, if the "Queue situation" dialog box pops up, you can continue to wait until the audit data analysis system returns the answer to the conversation. If you don't want to wait, you can click "Leave the queue";

[0314] 4. You can select different large models below the input box. Currently, the audit data analysis system accesses large models deepseek R1 and Tongyi Qianwen 2.5;

[0315] 5. You can click "AI reading" above the input box to upload files or pictures, providing more rich, accurate and dynamic knowledge sources to make the question and answer effect better;

[0316] About: Upload files

[0317] Click the "Upload file" icon. The audit data analysis system supports file formats PDF / DOCX / PPT, etc. The size of a single file is within 20M;

[0318] Pop up a dialog box, browse local files, and click Open;

[0319] The file is displayed in the input box as "Parsing";

[0320] After parsing is completed, the file is successfully uploaded;

[0321] About uploading pictures:

[0322] Click the "Upload Picture" icon, and the audit data analysis system supports picture formats such as JPEG / JPG / PNG, and the size of a single picture is within 20M;

[0323] A pop-up dialog box appears, browse local pictures, and click open;

[0324] The picture "is parsing" is displayed in the input box;

[0325] After parsing is completed, the picture is uploaded successfully.

[0326] As shown in Figures 25-28 :

[0327] Through the audit document assistant function in the specific embodiments of the present application, the user can upload work plans, operation guidelines, work experience and regulations and systems, etc. to establish his own private knowledge base. Artificial intelligence can provide suggestions and assist in qualifying doubts in new audit projects.

[0328] 1. In the

Document Assistant

Document Assistant

[0329] 2. Different large models can be called below the input box

[0330] 3. If you need to upload reference files, click the "Upload File" button, browse local files, and the size of a single file is within 20M. The uploaded files are saved in the temporary library of "personal files"

[0331] 4. After entering the question in the input box, click the "Send" button to communicate with the audit data analysis system. According to the needs, you can click the "Stop" button at any time to stop answering;

[0332] 5. The shortcut icon in the lower left corner of the page can quickly switch to the corresponding page: home page, new conversation, resource library, personal files and history.

[0333] 6. The resource library contains the internal database that has been uploaded. You can select one or more resource libraries by checking the checkbox in front of the resource library, and query the database. Click the expand button "View File" behind the existing database to pop up the file list of the database. You can search for the file name by entering it in the search bar. Click the page turning button below the file list to turn the page.

[0334] 7. You can also expand the rich built-in resource library and personal files by clicking the arrow of the hidden tab bar on the left side of the page. The calling of resource library and personal files makes the obtained reply faster and better, and the source of the reference literature is annotated in the answer, ensuring that the result is traceable.

[0335] 8. Personal files are personal uploaded personal files stored in the temporary library. One or more personal files can be selected by checking the checkbox in front of the personal file, and the file information can be mounted. The checked personal files will be displayed in the input box below the page. The file can be previewed and deleted by clicking the expand button behind the personal file, or the personal file can be added to the temporary library through the "upload file" button. Currently, docx, ppt and pdf format files are supported.

[0336] In specific embodiments of the present application, the SQL assistant can provide comprehensive SQL learning resources to help users master SQL basic syntax and practical skills. This function can be used without logging in.

[0337] 1. The SQL assistant has three sub-modules: intelligent answering, script rewriting and model explanation.

[0338] 2. Click the "intelligent answering" button, enter your technical problems in the audit work in the input box below, and click the "send" button to get help.

[0339] 3. Click the "script rewriting" button to open a new tab. The left side is the resource database and the right side is the target database. Users can automatically convert one language in the resource database to another language in the target database by clicking the "rewrite" button according to their needs. The blue "rewrite" button will change to a red "stop" button during the rewriting process, which can be stopped at any time by clicking the "stop" button.

[0340] 4. Return to the SQL assistant page and click the "model explanation" button below to open a new tab. The left side is the model content that the user wants to explain, and the right side will appear the corresponding explanation content by clicking the "explain" button. The blue "explain" button will change to a red "stop" button during the explanation process, which can be stopped at any time by clicking the "stop" button.

[0341] Through the audit intelligent analysis assistant, the industry data of social security, business registration, financial support, public accumulation fund and funeral and cremation can be queried, and the user can directly ask questions.

[0342] 1. The intelligent analysis function has two sub-modules: data query and intelligent analysis, as shown in Figures 29-32

[0343] 2. Click the "data query" button and enter the data to be queried in the input box below. Click the "send" button, and the audit data analysis system can obtain relatively complete information corresponding to the query data through industry data resource knowledge.

[0344] ​3. Click the "Data Analysis" button, and the "Industry Data Resource Directory" will pop up on the left side of the page. By checking the checkboxes in front of different information tables in the resource directory, the audit data analysis system can obtain data analysis results more quickly and accurately. By clicking the drop-down list behind the information table, you can also preview the content of the information table.

[0345] 4. Below the "Industry Data Resource Directory" is the "Audit Project Data" list. Click the drop-down box behind it to select different databases. By checking the checkboxes in front of specific rows in the database, the audit data analysis system can obtain data analysis results more quickly and accurately. Click the arrow behind the specific information in the database to preview the detailed information.

[0346] 5. Click "Upload File" behind "Audit Project Data". Currently, csv format databases are supported. In the pop-up dialog box, enter the database name, table name, and upload the csv format file. The size of a single file should be within 20M. Click the "Save" button to upload the file to the "Audit Project Data" in the audit data analysis system.

[0347] 6. Click "Add Connection" behind "Audit Project Data". A dialog box will pop up to select the database type. Select a database type and enter the corresponding information in the pop-up dialog box, including name, environment, host, etc. Click the "Test Connection" button to test the connection. After the test is passed, click the "Save" button to add the new connection to the "Audit Project Data" in the audit data analysis system.

[0348] The traditional audit system has three major defects:

[0349] High technical threshold: requires professional personnel to write SQL, long response period (average > 2 hours);

[0350] Serious data island: on-site audit data is scattered in heterogeneous databases such as Oracle and MySQL;

[0351] Security risk is prominent: the transmission of raw data is easy to cause information leakage (such as CN114580944A scheme).

[0352] Existing improvement schemes (such as CN116258324A) still cannot realize natural language interaction and cross-platform intelligent migration.

[0353] The audit data analysis system and method based on large language model in the specific embodiments of the present application realize audit intelligence through four innovative modules:

[0354] Intelligent analysis module: build a "natural language → SQL → visual report" closed-loop process, and the client only interacts with metadata to ensure security;

[0355] SQL assistant module: solves the problem of cross-database syntax migration (actual conversion accuracy ≥ 96%);

[0356] Document assistant module: realizes semantic-level retrieval of policies and regulations and private documents (recall rate increased by 40%);

[0357] General intelligent module: meets diversified office needs through domestic large model intranet deployment.

[0358] The scheme makes the audit efficiency of non-technical personnel increase by 5-8 times, has been verified in the medical funeral service audit scene, and saves labor cost by more than 70%.

[0359] It should be noted that the data query module is based on user natural language input. First, the intention is identified by the problem classifier, and then it is dispatched to the corresponding SQL generation module by the scene processor. The system generates SQL statements through preset rules or large models and executes database queries, and then passes the results to the summary generator to generate concise result feedback in combination with the scene type. The overall process realizes the automation of the closed loop from semantic recognition to structured query and result summary, improving the query efficiency and user understanding experience.

[0360] The client is deployed on the audit site and can directly connect to multiple types of databases (Oracle, MySQL, SQL Server) to complete data structure extraction and SQL query. Through the VPN security channel, the client initiates an API request to the audit data analysis system of the embodiment of the application and submits DDL or data content. The audit data analysis system completes intelligent processing and returns the result after receiving the request. The overall architecture ensures that data is not directly exported, and only structured information and results are interacted, balancing the synergy of data security, system flexibility and intelligent analysis capability. Regarding intelligent answering: support natural language questioning of SQL related problems, covering syntax correction, optimization suggestions, execution plan analysis, etc.; automatically adapt to different SQL dialects and provide personalized statement generation. Regarding script rewriting, migrate the existing database analysis model to the target database (such as P-SQL to T-SQL), automatically adjust data types, function names, syntax differences, and realize seamless migration. Regarding model explanation, explain the meaning of complex SQL in natural language, and disassemble the logic of the statement step by step to help non-professionals understand the business intent and data processing process. Regarding the document assistant, convert document materials into interactive knowledge, build a user private knowledge base and a policy and regulation common library, and realize an audit material question and answer experience similar to a search engine. Support multi-format file upload and intelligent summary to enhance document information retrieval and reuse efficiency. Multi-document, multi-format support (TXT, PDF, PPTX, DOCX, etc.), build a knowledge retrieval system with extensive content and consistent context.

[0361] In summary, it supports multiple input forms such as text, pictures, etc., and has general intelligent capabilities such as writing generation, translation polishing, image recognition, and document interpretation, meeting the diverse office needs beyond auditing. It can provide a general question and answer platform with the original capabilities of large models, covering writing, translation, reasoning, and image recognition. Private deployment within the network ensures security and provides comprehensive AI support for daily office work and needs beyond auditing. Local private deployment of large models such as DeepSeek, Qwen, and GLM creates a ChatGPT-like question and answer experience, ensuring that data does not leave the network and that sensitive business is controllable and usable. It builds a closed-loop processing flow from natural language analysis to SQL generation to result summary, significantly improving the automation level from asking questions to data insights. Users do not need to master SQL or database structures to complete complex analysis tasks through dialogue. Combined with the context understanding ability and domain knowledge of large models, the user's question is automatically mapped to predefined audit scenarios, dynamically generated through Prompt templates, ensuring that SQL statements are both general and business-oriented, greatly improving the accuracy and practicality of the generated results. It realizes the mixed semantic understanding of structured databases and unstructured documents (such as PDF, DOCX, TXT, etc.). The user's question can trigger both model capabilities and knowledge base retrieval, supporting multi-source knowledge linkage response, and truly realizing the intelligent fusion analysis of data and text.

[0362] Regarding the agent settings:

[0363] 1. One side is agent management, the other side is parameter management;

[0364] 2. You can select a specific scenario in the scene drop-down box above: general intelligence, intelligent answer, model explanation, script rewriting-sql, document assistant, intelligent analysis, script rewriting-explanation, industry data question and answer, different scenarios have different pre-set intelligent agents

[0365] 3. After selecting the scenario, enter a new prompt in the input box below the agent setting on the left side, click "Save", and a new agent can be added

[0366] 4. By controlling the state bar switch on the right side of the agent description, you can turn on or off different agent settings

[0367] 5. In the parameter management on the right side, you can set whether to associate context, maximum number of context dialogues (1-10), MaxTokens (maximum value 131072), Temprature (0-2), Top-p (0-1), Top-K (2-200), FrequencyPenalty (1-2), you can drag the slider or click the right button to increase or decrease the parameters, click "Save" to save the new parameter settings.

[0368] It includes:

[0369] a scene selection module, providing a drop-down box containing a plurality of preset scenes, the preset scenes including general intelligence, intelligent answering, model explanation, script rewriting-SQL, document assistant, intelligent analysis, script rewriting-explanation, and industry data question answering;

[0370] an agent management module, in response to a scene selection operation, loading a preset agent configuration of a corresponding scene;

[0371] an agent editing module, providing an agent description input box and a save control, in response to a new prompt word input by a user and a save operation, generating a new agent configuration item;

[0372] an agent state control module, providing an independent switch for each agent configuration item, for enabling or disabling the corresponding agent;

[0373] a parameter management module, providing the following parameter groups that can be interactively adjusted:

[0374] a context association switch;

[0375] a context dialogue round adjuster, with a value range of 1-10 rounds;

[0376] a maximum label number controller, with an upper limit value of 131072;

[0377] a temperature parameter adjuster, with a value interval of [0, 2];

[0378] a top probability cutoff adjuster, with a value interval of [0, 1];

[0379] a top label cutoff adjuster, with a value interval of [2, 200];

[0380] a frequency penalty coefficient adjuster, with a value interval of [1, 2];

[0381] a parameter saving module, in response to a parameter modification operation and a save instruction, persistently storing the parameter configuration.

[0382] It should be noted that the adjuster of the parameter management module adopts a composite interactive component of a slider control and a step button, supporting continuous adjustment by dragging the slider or discrete adjustment by clicking the button by a preset step size.

[0383] In specific embodiments of the present application, the switch state of the agent state control module is dynamically associated with the scene selection module, and when the scene is switched, the preset agent of the scene is automatically enabled and the agents of other scenes are disabled. When the context association switch is disabled, the context dialogue round is forcibly set to 0, regardless of the set value of the context dialogue round adjuster.

[0384] In the specific embodiments of the present application, the parameter saving module performs numerical boundary check when performing the saving operation, and automatically corrects the parameter value that exceeds the set range to the nearest boundary value. The agent configuration item added by the agent editing module is bound to a specific scene, and is only displayed in the agent management module when the drop-down option of the scene is selected.

[0385] Regarding the public library:

[0386] 1. Click the public library, and the page displays the existing libraries of the platform;

[0387] 2. Click the upper "New public library", a dialog box pops up, enter the library name and corresponding description, click "Save", and a new public library can be created;

[0388] 3. Enter the keyword in the upper "Search bar" to search in the existing public library;

[0389] 4. Taking the social security library as an example, click the "Social security library" in the public library, a dialog box will pop up, displaying all the data of the social security library, which can be deleted or deleted in batches. You can also enter a keyword in the upper search bar to quickly search for a certain item in the library. If you need to return to the previous level, you can click the "Back" button;

[0390] 5. Click the "Add" button, a new data dialog box will pop up, you can drag or click to upload new data, supporting multiple file formats.

[0391] It should be noted that the audit data analysis system of the specific embodiments of the present application supports the local private deployment of multiple domestic main models such as DeepSeek, Tongyi Qianwen and GLM, and has dynamic switching capability, which is suitable for different scene tasks. Effectively balance model effect, safety compliance and system flexibility, and ensure stable and reliable operation of the system in the audit intranet environment.

[0392] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A large language model-based audit data analysis method, characterized by, Comprise the following steps: S1: natural language request analysis: through the problem classifier of the intelligent analysis module, receive the natural language query request transmitted by the client through VPN encryption, identify the audit intention and data type, and classify into data query, document retrieval or general task; S2: determine whether the data type involved in the request is structured data or unstructured data; S21: if the request involves structured data, trigger the structured query engine to generate abstract query logic, call the SQL assistant module to execute cross-database adaptation, including: Identify the target database type through the SQL syntax adapter, including Oracle, MySQL or SQLServer; Verify the script generation syntax security through the pre-trained model, prevent scripts containing high-risk operations from being executed, including unauthorized table deletion, permission escape query; Use the script conversion engine to convert the abstract query logic into the target database dialect script, and perform equivalent reconstruction on the paging query and window function; Execute the query and feed back the result in natural language form through the model explanation unit; S22: if the request involves unstructured data, start the document assistant module, including: Extract text, table and image content in the user's private knowledge base through the multi-format document parser; Match the associated clauses in the policy and regulation common library based on the hybrid retrieval engine, and use a dynamic weighted scoring strategy. When the semantic similarity is >0.7, the semantic matching result is preferred; Mark the data source and confidence through the reference tracing unit; Also includes: constructing a joint vector index of the documents in the user's private knowledge base and the policy and regulation common library, and realizing cross-library associated retrieval through semantic similarity calculation; S3: multi-modal result fusion: The general intelligent module performs joint analysis on the structured query result and the unstructured retrieval result, including: S31: call the domestic large model to execute text generation and risk pattern recognition; S32: perform OCR recognition and key field extraction on image data; S33: generate an audit analysis report containing data source markers; Multi-modal result fusion also includes automatically generating an exception marker and triggering a manual review process when the difference between the descriptions of the same entity by structured data and unstructured data exceeds the preset threshold (numerical difference ≥5% or text description confidence difference ≥0.3); The manual review interface supports difference highlighting, comment version tracking and multi-person collaborative review, and the review result is stored through blockchain.

2. An audit data analysis system based on a large language model, characterized by, The audit data analysis system of the large language model comprises an intelligent analysis module (1), an SQL assistant module (2), a document assistant module (3) and a general intelligent module (4), and the data flow between the modules is transmitted through an internal network encrypted channel; The intelligent analysis module (1) is used to realize natural language driven audit data query and on-site analysis, and contains a problem classifier, a structured query engine and an abstract generator, which supports interaction with the client through a VPN secure channel to obtain structured metadata; The intelligent analysis function module (1) comprises a data query module and a data analysis module, the data query module is configured to: in response to a user operation on a data query user interface element, receive a data query request input in an associated input box; in response to a user operation on a sending user interface element, based on an industry data resource knowledge base, obtain relatively complete information corresponding to the data query request; the data analysis module is configured to: in response to a user operation on a data analysis user interface element, present an industry data resource directory user interface element; wherein the industry data resource directory user interface element contains a plurality of information table options and associated selection controls; in response to a user operation on the selection control, based on the selected information table, perform data analysis and generate an analysis result; the industry data resource directory user interface element is further configured to present preview content of the corresponding information table in response to a user operation on an information table associated preview control; The SQL assistant module (2) provides cross-database script migration services, including SQL syntax adapters, script conversion engines, and model interpretation units, supporting Oracle, MySQL, and SQL Server multi-dialect environments; The document assistant module (3) builds a user private knowledge base and a policy and regulation common library, integrating multi-format document parsers, hybrid search engines, and citation tracing units; The general intelligent module (4) provides general AI capabilities such as text generation, image recognition, and multi-language translation through internal network private deployment of domestic large models, including DeepSeek, Qwen, and GLM; The intelligent analysis module (1) includes a data security interaction unit and a temporary table management unit, the client only transmits database table structures or desensitized metadata fields to the audit data analysis system, and the original data is retained for local query execution, the temporary table management unit supports storing query results of sensitive fields in encrypted temporary tables and setting automatic cleaning strategies, sensitive fields include ID numbers and the names of the deceased; The SQL assistant module (2) includes: Dialect adapter: based on syntax tree parsing, automatically converts PL / SQL and T-SQL, and corrects data type differences; Execution plan visualization unit: decomposes SQL execution process into natural language interpretation steps and labels key performance bottleneck points; The document assistant module (3) supports: Multi-modal retrieval: paragraph-level vectorization of PDF and DOCX documents, combined with keyword matching and semantic similarity calculation for hybrid search; Dynamic update mechanism: automatically captures web policy files, generates structured knowledge graphs, and associates them with the question and answer system. 3.The big language model-based audit data analysis system of claim 1, wherein, The data analysis module is further configured to: Below the industry data resource directory user interface element, present an audit project data list user interface element; The audit project data list user interface element contains a database selection control; in response to a user operation on the database selection control, present a list of selectable databases and allow the user to select a target database; The audit project data list user interface element presents, for a selected target database, a plurality of data entry options and associated selection controls contained therein; in response to user operation of the data entry selection controls, performs data analysis based on selected data entries and generates analysis results; The audit project data list user interface element is further configured to, for the data entries, present detailed information preview of a respective data entry in response to user operation of an associated entry preview control; The audit project data list user interface element contains a file upload control; configured to: present a file upload dialog in response to user operation of the file upload control; receive user input through the file upload dialog, including: database name, table name; receive a data file uploaded by the user through the file upload dialog, which conforms to a predetermined format; store the data file into the audit project data in response to a save operation; The audit project data list user interface element contains a database connection adding control; configured to: present a database type selection dialog in response to user operation of the database connection adding control; present a connection configuration dialog in response to selection of a database type in the database type selection dialog; receive connection configuration information input by the user through the connection configuration dialog; perform database connection test based on the connection configuration information in response to a test connection operation; add the new database connection into the audit project data in response to a save operation after the connection test is passed.

4. The large language model-based audit data analysis system of claim 3, wherein, The predetermined format is CSV format.

5. The large language model-based audit data analysis system of claim 4, wherein, The system is configured to limit the size of a single uploaded file to be no more than a predetermined threshold.

Images