Intelligent auditing system and method

Abstract

The invention provides an intelligent auditing system and method, and the system is characterized in that the system comprises a configuration module which is used for defining and storing a material template, a form object and an auditing rule according to different auditing business scenes; and the auditing module is connected with the configuration module and is used for extracting key information in the document to be audited based on the material template defined by the configuration module and performing matching and logical judgment on the extracted key information based on the auditing rule and the form object so as to generate an auditing report. The technical scheme provided by the invention can improve the examination and approval timeliness and reduce examination and approval disputes.

Description

Technical Field

[0001] This document relates to the field of document review technology, and in particular to an intelligent review system and method. Background Technology

[0002] There are currently two modes of document review across various industries: one is offline manual review, in which the applicant submits paper materials, conducts initial review at the window, verifies through multiple departments, and then conducts manual review before the paper documents are archived. This relies on manual verification of the authenticity of documents such as property ownership certificates, ID cards, and contracts. The other is basic information-based review, which typically uses a B / S government affairs platform + database + basic OCR recognition as its architecture. The process involves uploading electronic materials to the government affairs system, conducting online manual review, retrieving data from across systems, and notifying the results offline.

[0003] Current systems generally suffer from the following drawbacks: In terms of efficiency, business processes are time-consuming, manual verification is time-consuming, and cross-departmental data silos lead to duplicate verification of information; in terms of risk, manual identification of forged documents has a high error rate, and paper materials are at risk of being tampered with; in terms of user experience, applicants need to supplement materials multiple times, there is no progress feedback, and processing is not possible outside of working hours; in terms of technology, OCR can only extract simple text, lacks semantic understanding and data logic verification capabilities, and lacks a reliable evidence storage mechanism, making it difficult to determine responsibility in case of disputes.

[0004] Therefore, there is an urgent need for an intelligent document review method that integrates intelligent recognition, cross-domain collaboration, and trusted evidence storage to break through the limitations of existing models, achieve efficient verification, controllable risks, optimized user experience, and technological upgrades, and solve the above problems. Summary of the Invention

[0005] This invention provides an intelligent auditing system and method to address the aforementioned problems.

[0006] An intelligent auditing system is provided according to an embodiment of the present invention, comprising: The configuration module is used to define and store material templates, form objects, and audit rules according to different audit business scenarios; The review module, connected to the configuration module, is used to extract key information from the document to be reviewed based on the material template defined by the configuration module, and to match and logically judge the extracted key information based on the review rules and form objects to generate a review report.

[0007] An intelligent auditing method is provided according to an embodiment of the present invention, comprising: Define and store material templates, form objects, and audit rules according to different audit business scenarios; Based on the material template defined by the configuration module, key information is extracted from the document to be reviewed, and the extracted key information is matched and logically judged based on the review rules and form objects to generate an review report.

[0008] By employing the embodiments of this invention, material templates, form objects, and review rules can be flexibly defined through the configuration module. Combined with the review module, a visual big data model and rule engine are used to achieve accurate extraction of key information and intelligent logical judgment, which significantly improves review efficiency and accuracy and reduces the risk of human intervention and misjudgment. At the same time, the introduction of a blockchain evidence storage mechanism ensures the traceability and immutability of the review process, effectively solving the problems of low efficiency, high risk, poor experience, and technical limitations in the traditional review model, and realizing the automation, intelligence, and credibility of the review process. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of the intelligent auditing system according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the configuration module according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the material template configuration unit according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the review module according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the key information extraction unit in an embodiment of the present invention; Figure 6 This is a schematic diagram of the four-layer architecture of the intelligent auditing system according to an embodiment of the present invention; Figure 7 This is a flowchart of the intelligent auditing method according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the configuration stage in an embodiment of the present invention; Figure 9 This is a schematic diagram of the review process according to an embodiment of the present invention. Detailed Implementation

[0011] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0012] System Implementation Examples According to embodiments of the present invention, an intelligent auditing system is provided. Figure 1 This is a schematic diagram of the intelligent review system according to an embodiment of the present invention. Figure 1 As shown, the intelligent auditing system of this invention specifically includes: Configuration module 10 is used to define and store material templates, form objects, and approval rules according to different approval business scenarios. A business scenario can be understood as an abstraction of the approval business process in the business system, with different scenarios representing different approval processes. A created scenario needs to include three parts: material templates, form objects, and approval rules.

[0013] Figure 2 This is a schematic diagram of the configuration module according to an embodiment of the present invention. Figure 2 The configuration module 10 shown specifically includes: The material template configuration unit 20 is used to select review points in the material template and configure associated prompt word templates to guide information extraction. The material template can be understood as an abstraction of the attachment materials that need to be uploaded in the business process. The content of a well-defined material template should include all review points that need to be reviewed in such materials. Based on the powerful semantic understanding capabilities of the large model, the format of such materials uploaded in actual business does not need to strictly adhere to the template format; it only needs to include the necessary review points. Figure 3 This is a schematic diagram of the material template configuration unit according to an embodiment of the present invention. Figure 3 The material template configuration unit 20 shown specifically includes: The review point annotation submodule 201 is used for interactively selecting regions on the material template image and defining the extraction rule type for the selected content. Specifically, this includes: the system automatically identifying the selected content and adjusting it as needed until the content is confirmed to be correct. Selecting the extraction rule type, including text and stamps, will trigger different strategies for content extraction by the model. It is worth noting that when selecting the stamp type, it is usually necessary to set the base text, auxiliary text, and a shortcut strategy to assist in locating the stamp's ownership. These configurations play a crucial role in scenarios containing multiple stamps.

[0014] The auxiliary positioning strategy submodule 202 is used to configure the base text, auxiliary text and offset strategy when the extraction rule type is seal, so as to help determine the seal ownership. The custom prompt submodule 203 provides a default prompt template and supports overriding and testing custom prompts for specific review points. The system provides a default prompt template; skipping this step is considered a convenient and feasible solution when the material template is relatively simple. When the material template is relatively complex, the prompt template can be overridden for specific needs, including but not limited to declarations of the overall extraction logic and extraction rules for individual review points. After the prompts are written, the model's processing effect can be tested online.

[0015] Form object configuration unit 22 is used to define and provide access interfaces for the standard data required for auditing; A form object is a broad concept, its practical significance being to provide data standards for the entire review process. For the review system, this manifests as a RESTful interface. There are three ways to obtain data standards for the review process in the system, with the form object being the primary method. Other methods include extracting data from other documents within the same scenario and some custom system parameters. In simpler terms, a form object can be understood as a reliable dataset obtained and integrated by the business system through different channels (database, third-party systems, API interfaces). This dataset contains most of the data standards required for the current review process. This data, along with the data standards from the other two methods described above, forms a complete data standard, which is then provided to the rule engine to execute the review logic. Configuring a form object is relatively simple; refer to API tools such as Postman for guidance.

[0016] The rule configuration unit 24 is used to construct the audit logic execution unit that associates extracted data, standard data and comparison rules.

[0017] The review rules are the core of the review system and the sole reference for whether a review process passes. There are three roles in the review rules: extracted data, data standards, and comparison rules. Extracted data includes key data extracted by the AI ​​model engine from each attached document based on the review points defined in the document template. Comparison rules are matching algorithms that determine whether the extracted data matches the corresponding data standards. Configuring the review rules involves sequentially associating these three roles in each document template to form a rule engine execution unit. Once the rule engine has executed all execution units in all documents, the review process is considered complete.

[0018] The audit module 12, connected to the configuration module 10, is used to extract key information from the document to be audited based on the material template defined by the configuration module 10, and to match and logically judge the extracted key information based on the audit rules and form objects to generate an audit report.

[0019] Figure 4 This is a schematic diagram of the review module according to an embodiment of the present invention. Figure 4 The audit module 12 shown includes: The document preprocessing unit 40 is used to perform invalid page filtering, image orientation correction, and image quality enhancement on the received document. At the start of the review process, the business system transmits the document to the review system via the standard API provided by the review system. The review system receives the document from the business system and categorizes the review materials by document template identifiers; each template corresponds to one or more images. Specifically, the document preprocessing unit 40 is used for: By calculating the pixel information coverage of the image, invalid pages below a set threshold are filtered out. For invalid page filtering, the pixel color determination algorithm of the CV engine determines the information coverage of the current page. If it is below the threshold, the page is determined to be invalid and filtered out, thereby reducing the pressure on the large model.

[0020] By calculating the rotation angle of the image and applying a rotation algorithm, the image orientation is corrected page by page. In real-world scenarios, batch-scanned PDF documents or images often have abnormal text orientation, especially for multi-page PDF documents, where there may even be issues with A / B orientation. By calculating the image rotation angle using a CV engine and then correcting the image page by page using a rotation algorithm, this operation can significantly improve the image recognition accuracy.

[0021] Image visual enhancement and size optimization are performed through grayscale conversion, histogram equalization, filtering, and lossy compression. Since there is an upper limit to the length of token sequences that a large model can process in a single inference iteration, appropriately compressing the image size can improve the throughput of the large model. The review system, through repeated testing, determined the image compression ratio and DPI threshold while ensuring recognition accuracy. Subsequently, the CV engine performs grayscale conversion, histogram equalization, and filtering on the image to improve its visual effect and make it more suitable for subsequent recognition operations.

[0022] Key information extraction unit 42 calls the material template and prompt words defined in the configuration module, and extracts text and seal content from the preprocessed document through visual big model and OCR technology; Figure 5 This is a schematic diagram of the key information extraction unit in an embodiment of the present invention. Figure 5 The key information extraction unit 42 shown specifically includes: The form object content extraction submodule 421 is used to call the interface configured in the form object during the review process to obtain and integrate standard data. By using the interface standard configured in the form object during the configuration phase, the review system calls this interface at runtime to obtain form object data, performs deduplication and other operations, and saves it in the system memory for use by the rule engine.

[0023] The key content extraction submodule 422 utilizes a large visual model to extract key textual information from documents based on prompts configured in the material template. After preprocessing, the document images are sent to the AI ​​model engine in batches. The model engine dynamically retrieves the prompts corresponding to the document template and initiates the key content extraction task. The extracted raw data is stored in the database, and an asynchronous uplink is used to save the raw extracted data. The data stored in the database is assembled into standard data according to field mapping rules and stored in system memory for use by the rule engine.

[0024] The seal extraction module 423 uses an object detection model to locate the seal and determines its ownership through a two-layer localization strategy based on baseline text and auxiliary text. Then, it uses a text recognition model to identify the seal content. Because seals may have positional deviations, shape differences, text overlays, or other issues causing blurred content, a simple recognition task cannot meet the requirement of accurate seal content extraction. Therefore, the review system innovatively adopts a method of recognition plus screenshot processing for seals. The configuration method for seal extraction capabilities is described in the material template configuration section of the configuration phase. At the start of the recognition task, all seals in the image are first located using the object detection model, and then the ownership of the seal is determined through a two-layer localization method. The principle of two-layer localization is to detect baseline text around the seal target as the origin. If a baseline text is detected and the confidence level exceeds a threshold, this baseline text is considered the seal's ownership. If no baseline text is detected, auxiliary text is detected using the seal target as the origin combined with an offset strategy. If auxiliary text is detected and the confidence level exceeds a threshold, this baseline text is considered the seal's ownership. Two-layer localization can effectively solve more than 80% of seal ownership problems. Finally, the text recognition model is used to identify the content of the seal. The identified content is used as a parameter for the rule engine, and the seal screenshot is used as a reference for manual review, thus doubly assisting the review process.

[0025] Rule engine unit 44 loads and executes the audit rules defined in the configuration module, comparing the extracted key information with the standard data obtained through the form object. The execution logic of rule engine unit 44 is as follows: the audit rules are organized into multiple rule groups, with rules within a rule group having a logical OR relationship, and rules between different rule groups having a logical AND relationship, and these rules are executed sequentially according to this logical relationship. After the audit system completes the above data acquisition work, it initiates the rule engine execution process. The section on audit rule configuration in the configuration phase describes the configuration method of the execution unit. In addition, the rule engine also includes the concept of a rule group. A rule group can contain multiple rule execution units with an OR (||) relationship, and multiple rule groups are connected through an AND (&&) relationship, thereby forming a relatively flexible execution chain. The execution granularity of the rule engine is document—audit point—rule group—execution unit. When all execution units in all documents have been executed, the execution process of the rule engine is considered complete.

[0026] The audit report generation module 46 collects and formats the execution process and result data of the rule engine to generate a structured audit report. During the execution of the rule engine, key data on the execution process and results are recorded synchronously. Once the rule engine execution process is complete, the report processing unit retrieves this key data and formats it to form the audit report for the current audit business. The report structure is customizable; only the report interface needs to be implemented, and the corresponding formatting logic needs to be developed. The report is ultimately sent to the business system in JSON format for page rendering and subsequent export operations. The business system can quickly locate problem nodes based on the audit report for manual review and provide handling opinions.

[0027] The system further includes an on-chain evidence storage module, which is used to generate hash values ​​from the original results of key information extraction, rule execution results and corresponding timestamps, and store them in the blockchain.

[0028] Figure 6 This is a schematic diagram of the four-layer architecture of the intelligent review system according to an embodiment of the present invention. Figure 6 It can be seen that the intelligent auditing system of this invention is configured with a four-layer architecture, including: application layer, intelligent engine layer, blockchain evidence storage layer and infrastructure layer; The application layer is used to provide standard API interfaces for interacting with external business systems; The intelligent engine layer integrates a CV engine for image preprocessing, an AI model engine for content understanding, and a rule engine for logical judgment. The CV engine improves image quality by performing operations such as grayscale conversion, filtering, and edge detection on the acquired images. The AI ​​model engine deploys a multimodal large model, supports text / image interaction, and processes material content on demand based on powerful semantic understanding capabilities. It uses dynamic prompt word templates to fill in the key information to be extracted at runtime, accurately extracting document content. The rules engine manages audit rules according to different business processes, dynamically loads them at runtime, supports real-time adjustments, and generates audit reports after execution.

[0029] The blockchain evidence storage layer is used to provide tamper-proof evidence storage services for key data throughout the entire review process; it enables the entire approval process to be on-chain, with material hashes, original extraction results, rule execution results, and timestamps stored in AntChain, providing full traceability of the review process.

[0030] The infrastructure layer employs GPU stacks to integrate computing resources and provide distributed storage to support the dynamic scaling and high-load operation of large model services. Specifically, this includes: using GPU stacks to integrate fragmented computing resources, providing load capacity and dynamic scaling capabilities for large model services; and using distributed storage services to efficiently process template files.

[0031] Approval efficiency is improved, with routine processing time reduced to under 3 minutes based on a one-stop review model (a 75% speedup compared to existing technology). Risk control is strengthened by supporting manual review of results, reducing misjudgments and omissions, improving the accuracy and reliability of reviews, simulating and evaluating review results, identifying and resolving issues in advance, reducing risks and disputes, and ensuring 100% traceability of tampering. Service experience is upgraded with 24 / 7 automated approval availability.

[0032] By employing the embodiments of the present invention, the following beneficial effects are achieved: This system achieves a complete logical closed loop from scenario to report. From an information extraction perspective, the system supports the extraction of key information based on configuration and returns it in key-value format. Compared to simple OCR solutions, it provides a more intelligent and standardized data processing method, greatly simplifying subsequent rule processing logic. Compared to other solutions, the system integrates a dynamic rule engine and a customizable audit report generation module, making up for the shortcomings of complex audit rules that are not configurable or controllable in real business scenarios. At the same time, the flexible audit report generation solution fills a gap in the current field.

[0033] Based on the system's four-layer intelligent audit architecture, it not only completes the closed loop of audit logic at the business level and ensures full data traceability at the security level, but also adopts a computing power integration framework at the infrastructure level, which greatly improves resource utilization.

[0034] Compared to traditional OCR solutions, this system employs a combination of CV (Visual Character Recognition), OCR, and VLM (Visual Model). Firstly, in the document preprocessing stage, CV capabilities are leveraged to filter invalid pages, correct image orientation, and enhance images, effectively improving visual quality and reducing model load. Simultaneously, image compression enhances batch processing capabilities, ensuring document context integrity and improving the overall analytical capabilities of the large model, resulting in more accurate output. Secondly, in the document content extraction stage, the system employs different extraction strategies based on the content category (text, seal). For text content, a VLM (Visual Model) approach is used, utilizing refined prompts to maximize accurate content extraction. For seals, an OCR method is used. First, an object detection model locates the seal object, then a system-specific localization algorithm determines the seal's ownership, followed by a recognition model to identify the seal content. Finally, the seal content is linked to the seal image. Figure 1 Same as above.

[0035] Method Implementation Examples According to embodiments of the present invention, an intelligent auditing method is provided. Figure 7 This is a flowchart of the intelligent review method according to an embodiment of the present invention. Figure 7 As shown, the intelligent auditing method of this invention specifically includes: S1. Define and store material templates, form objects, and audit rules according to different audit business scenarios; Figure 8 This is a schematic diagram of the configuration stage in an embodiment of the present invention. Figure 8 As shown, S1 in this embodiment of the invention specifically includes: S101. Material Template Configuration: Interactively select the review point area in the material template and define the extraction rule type for the selected content; for review points of the seal type, configure the base text, auxiliary text and offset strategy to help determine the seal ownership; provide a default prompt word template and support overriding and testing custom prompt words for specific review points to guide the information extraction process.

[0036] S102. Form Object Configuration: Defines and provides an access interface for the standard data required for auditing. The form object is provided in the form of a RESTful interface and is used to obtain and integrate standard data from databases, third-party systems or API interfaces during the auditing process.

[0037] S103. Audit rule configuration: Construct an audit logic execution unit that associates extracted data, standard data and comparison rules; organize audit rules into multiple rule groups, with rules within a rule group having a logical OR relationship and different rule groups having a logical AND relationship, forming a flexible audit logic chain.

[0038] S2. Extract key information from the document to be reviewed based on the material template defined by the configuration module, and perform matching and logical judgment on the extracted key information based on the review rules and form objects to generate an review report. Figure 9 This is a schematic diagram of the review process according to an embodiment of the present invention. Figure 9 As shown, S2 in this embodiment of the invention specifically includes: S201. Document Reception and Preprocessing: Receive documents to be reviewed from the business system, classify the review materials by document template identifier; filter invalid pages and correct image orientation page by page for the received documents, and perform visual enhancement and size optimization on the images through grayscale conversion, histogram equalization, filtering and lossy compression operations.

[0039] S202. Key Information Extraction: This involves using the material templates and prompts defined in the configuration phase to extract text and stamp content from the preprocessed document through a large visual model and OCR technology; specifically including: Call the interface configured in the form object to obtain and integrate standard data; Using a large visual model, key text information is extracted from the document based on prompts configured in the material template; The seal is located using an object detection model, and its ownership is determined by a two-layer localization strategy based on baseline text and auxiliary text. Finally, the seal content is identified using a text recognition model.

[0040] S203, Rule Engine Execution: Loads and executes the audit rules defined in the configuration phase, compares the extracted key information with the standard data obtained through the form object; executes in sequence according to the granularity of document, review point, rule group, and execution unit. Rules within a rule group are logical OR relations, and rules between different rule groups are logical AND relations.

[0041] S204. Audit Report Generation and Storage: Collect and format the execution process and result data of the rule engine to generate a structured audit report; at the same time, generate hash values ​​from the original results of key information extraction, rule execution results and corresponding timestamps, and store them in the blockchain to achieve tamper-proof storage of key data throughout the audit process.

[0042] S205. Result Return: The generated audit report is returned to the business system in a structured format for page rendering, export, and subsequent manual review.

[0043] The intelligent review method adopted in this invention significantly improves review efficiency and accuracy by flexibly configuring material templates, form objects, and review rules, and combining a visual big data model and rule engine to achieve accurate extraction of key information and intelligent logical judgment. At the same time, the introduction of a blockchain evidence storage mechanism ensures the traceability and immutability of the review process, effectively solving the problems of low efficiency, high risk, poor experience, and technical limitations in the traditional review model, and realizing the automation, intelligence, and credibility of the review process.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent auditing system, characterized in that... include: The configuration module is used to define and store material templates, form objects, and audit rules according to different audit business scenarios; The review module, connected to the configuration module, is used to extract key information from the document to be reviewed based on the material template defined by the configuration module, and to match and logically judge the extracted key information based on the review rules and form objects to generate a review report.

2. The system according to claim 1, characterized in that, The configuration module specifically includes: The material template configuration unit is used to select review points in the material template and configure associated prompt word templates to guide information extraction; The form object configuration unit is used to define and provide access interfaces for the standard data required for auditing; The rule configuration unit is used to build an audit logic execution unit that associates extracted data, standard data, and comparison rules.

3. The system according to claim 2, characterized in that, The material template configuration unit specifically includes: The review point annotation submodule is used to interactively select areas on the material template image and define the extraction rule type for the selected content; The auxiliary positioning strategy submodule is used to configure the base text, auxiliary text, and offset strategy when the extraction rule type is seal, in order to help determine the ownership of the seal; The custom prompts submodule provides a default prompt template and supports overriding and testing custom prompts for specific review points.

4. The system according to claim 1, characterized in that, The audit module includes: The document preprocessing unit is used to perform invalid page filtering, image orientation correction, and image quality enhancement on the received document; The key information extraction unit calls the material template and prompt words defined in the configuration module, and extracts the text and seal content from the preprocessed document through visual big model and OCR technology. The rules engine unit loads and executes the audit rules defined in the configuration module, and compares the extracted key information with the standard data obtained through the form object; The audit report generation module is used to collect and format the execution process and result data of the rule engine to generate a structured audit report.

5. The system according to claim 4, characterized in that, The key information extraction unit specifically includes: The form object content extraction submodule is used to call the interface configured in the form object during the review process to obtain and integrate standard data; The key content extraction submodule uses a large visual model to extract key text information from the document based on the prompts configured in the material template. The seal extraction module uses a target detection model to locate the seal, determines the seal's ownership through a two-layer localization strategy based on baseline and auxiliary text, and then identifies the seal's content through a text recognition model.

6. The system according to claim 4, characterized in that, The document preprocessing unit is specifically used for: Invalid pages below a set threshold are filtered out by calculating the coverage of image pixel information. The image orientation is corrected page by page by calculating the rotation angle of the image and applying a rotation algorithm. Image visual enhancement and size optimization are achieved through grayscale conversion, histogram equalization, filtering, and lossy compression.

7. The system according to claim 4, characterized in that, The execution logic of the rule engine unit is as follows: organize the audit rules into multiple rule groups, with the rules within a rule group having a logical OR relationship and the rules between different rule groups having a logical AND relationship, and execute them sequentially according to this logical relationship.

8. The system according to claim 1, characterized in that, The system further includes an on-chain evidence storage module, which is used to generate hash values ​​from the original results of key information extraction, rule execution results and corresponding timestamps, and store them in the blockchain.

9. The system according to claim 1, characterized in that, The system is configured with a four-layer architecture, including: application layer, intelligent engine layer, blockchain evidence storage layer, and infrastructure layer. The application layer is used to provide standard API interfaces for interacting with external business systems; The intelligent engine layer integrates a CV engine for image preprocessing, an AI model engine for content understanding, and a rule engine for logical judgment. The blockchain evidence storage layer is used to provide tamper-proof evidence storage services for key data throughout the audit process; The infrastructure layer uses GPU stacks to integrate computing resources and provide distributed storage to support the dynamic scaling and high-load operation of large model services.

10. An intelligent auditing method, characterized in that, include: Define and store material templates, form objects, and audit rules according to different audit business scenarios; Based on the material template defined by the configuration module, key information is extracted from the document to be reviewed, and the extracted key information is matched and logically judged based on the review rules and form objects to generate an review report.

Images