Invoice intelligent auditing system based on multi-agent cooperation

Abstract

The application relates to the technical field of artificial intelligence application, and provides an invoice intelligent auditing system based on multi-agent cooperation, which comprises an invoice access and analysis module, a multi-agent auditing module, a rule management module, a result aggregation and early warning module and a visual archiving module, wherein the multi-agent auditing module is used for performing authenticity checking, duplication checking, compliance checking and accounting operations on the structured invoice data through multiple parallel agents, and generating respective agent auditing sub-results; the system simultaneously performs authenticity checking, duplication checking, compliance checking and accounting operations on invoice data through multiple parallel agents, parallelizes multiple key auditing links which originally have to be executed in series, and improves enterprise financial processing efficiency.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence application technology, and in particular to an intelligent invoice verification system based on multi-agent collaboration. Background Technology

[0002] In the field of corporate finance and tax management, verifying the accuracy and compliance of invoices is a crucial process for preventing risks and ensuring the safety of funds. Currently, the traditional verification methods commonly used in the industry rely on manual operations by finance personnel. Finance staff must perform multiple independent verification steps for each invoice, including accessing external platforms to verify authenticity, reviewing historical records to check for duplicates, comparing against standard verification documents to determine compliance, and manually verifying the amount calculation.

[0003] This sequential and highly manual review process creates a bottleneck in processing efficiency that is difficult to overcome. Because multiple steps, such as verification, deduplication, compliance checks, and amount calculation, must be completed sequentially by the same person or team, the overall review of a single invoice is time-consuming. Simply increasing manpower cannot linearly improve processing capacity. Therefore, when faced with sudden, large-scale invoice processing demands, systemic processing delays and task backlogs become particularly prominent, severely restricting the efficiency of corporate financial processing. Summary of the Invention

[0004] This invention provides an intelligent invoice verification system based on multi-agent collaboration to improve the efficiency of enterprise financial processing.

[0005] This invention provides an intelligent invoice verification system based on multi-agent collaboration, comprising: The invoice access and parsing module is used to perform optical character recognition processing on the received invoice files in order to extract key fields and generate structured invoice data; The multi-agent audit module is used to perform verification, deduplication, compliance and accounting operations on the structured invoice data through multiple parallel agents, and generate audit sub-results for each agent; The rule management module is used to configure review rules through a graphical interface to generate a set of currently effective review rules and provide the set of currently effective review rules to the multi-agent review module; The result aggregation and early warning module is used to aggregate the review sub-results of each intelligent agent to generate the target review result, and classify the risk points in the target review result into levels to generate an early warning notification. The visualization archiving module is used to display the target audit results and the warning notification, and to archive the structured invoice data and the target audit results and synchronize them to the related business system.

[0006] This invention provides an intelligent invoice review system based on multi-agent collaboration. On one hand, by setting up multiple parallel agents to simultaneously perform verification, deduplication, compliance, and accounting operations on invoice data, the system parallelizes multiple key review steps that were originally required to be executed sequentially, thus breaking the efficiency bottleneck caused by step dependencies in the traditional review process. On the other hand, by decoupling review rules from the program code and having them visualized and dynamically published by an independent rule management module, adjustments to the review logic can be made without interrupting system operation or modifying the program, thereby achieving flexible and real-time updates to the review strategy. Furthermore, by integrating invoice parsing, parallel review, result aggregation, and visual archiving into an automated chain, the system eliminates manual handover and data transfer between steps, thus constructing an end-to-end automated processing closed loop from invoice entry to result output and reuse. Attached Figure Description

[0007] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0008] Figure 1 This is a module structure diagram of an intelligent invoice verification system based on multi-agent collaboration provided in an embodiment of this application. Detailed Implementation

[0009] The technical solutions of the embodiments of the present invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0010] It should be noted that when a component is described as "fixed to / set on" another component, it can be directly on the other component or it can be in a middle component. When a component is described as "connected to" another component, it can be directly connected to the other component or it may be in a middle component.

[0011] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0012] The following detailed description of some embodiments of the present invention is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other. This invention provides an intelligent invoice verification system based on multi-agent collaboration. Figure 1 A flowchart of an intelligent invoice verification system based on multi-agent collaboration provided in an embodiment of the present invention is shown below. Figure 1 As shown, the intelligent invoice verification system based on multi-agent collaboration provided in this embodiment of the invention includes: The invoice access and parsing module 100 is used to perform optical character recognition processing on the received invoice files in order to extract key fields and generate structured invoice data.

[0013] Specifically, the invoice access and parsing module receives invoice files uploaded by users through the system interface or API. This module supports various invoice file formats, including image files, PDF documents, and spreadsheet files. The module first parses the received files to identify their specific format. For image files and PDF files, which are essentially collections of images, the module uses its built-in optical character recognition engine. This engine preprocesses the images to optimize recognition conditions, then detects and recognizes text regions within the images, matching the recognized text information with preset key fields on the invoice to extract core information such as invoice code, invoice number, invoice date, amount, and tax rate. For spreadsheet files, the module directly reads data from corresponding cells based on predefined table templates and field mapping relationships to complete information extraction. Regardless of the processing path, the module ultimately standardizes and organizes all extracted key field information, encapsulating it into a structured data object—structured invoice data. This data object serves as the unified input for all subsequent review processes.

[0014] The multi-agent audit module 200 is used to perform verification, deduplication, compliance and accounting operations on the structured invoice data through multiple parallel agents, and generate audit sub-results for each agent.

[0015] Specifically, the multi-agent auditing module receives structured invoice data from the invoice access and parsing module. After startup, the module first creates an auditing task for the invoice data. Subsequently, multiple independent agents within the module are activated in parallel, each responsible for a specific auditing task. The verification agent obtains information such as the invoice code and number from the structured invoice data, automatically initiates a query request by connecting to the verification interface, and receives the returned invoice status information to complete the authenticity verification. The deduplication agent uses the unique identifier information of the same invoice to query the system's built-in database that records all processed invoices, and can also query external authoritative data sources as needed to determine whether the invoice has been processed. The compliance agent dynamically obtains the currently effective set of auditing rules from the system's rule management module and applies these rules one by one to the structured invoice data to check whether it complies with tax audit standards, reimbursement systems, and other regulations. The accounting agent focuses on numerical logic, verifying the accuracy of the tax calculation, total price including tax, and other numerical information recorded on the invoice. After completing their respective independent audit analysis, all agents generate a sub-result containing their audit conclusions.

[0016] The rule management module 300 is used to configure audit rules through a graphical interface to generate a set of currently effective audit rules and provide the set of currently effective audit rules to the multi-agent audit module.

[0017] Specifically, the rule management module provides the system with an independent environment for defining and maintaining audit rules. At its core is a visual rule configuration interface. Through this interface, business administrators can construct complete audit rules without writing code, using graphical operations such as dragging and dropping logic components, selecting data fields, and setting condition values. For example, a rule can be created stipulating that the tax rate for a certain type of invoice from a specific region must fall within a certain set. Rules created or modified by users are saved in the system's rule knowledge base, which supports version management, enabling, and disabling of rules. Once a rule is enabled and published, the rule management module compiles and integrates it into a unified set of rules currently in use by the system—the currently effective audit rule set. This rule set is provided in real-time to the compliance agent in the multi-agent audit module as the immediate basis for its compliance judgments.

[0018] The result aggregation and early warning module 400 is used to aggregate the review sub-results of each intelligent agent to generate a target review result, and classify the risk points in the target review result into levels to generate an early warning notification.

[0019] Specifically, the result aggregation and early warning module receives multiple agent review sub-results from the multi-agent review module. Internally, the module first aggregates and comprehensively analyzes these scattered sub-results. The module follows a pre-defined decision logic; for example, when the verification agent determines an invoice is fake, regardless of other results, the module will determine that the invoice review fails. Based on this logic, after integrating all sub-results, the module generates a target review result, which includes the final review status of the invoice and various review findings. Subsequently, based on the issues identified in the target review result, the module calls a pre-defined risk assessment model to assign a risk level label to the invoice, such as high risk, medium risk, or low risk. For invoices marked as medium or high risk, the module automatically triggers an early warning process, generating an early warning notification containing core invoice information and a risk description, and sends the notification to designated recipients through pre-configured channels such as system messages and emails.

[0020] The visualization archiving module 500 is used to display the target audit results and the warning notification, and to archive the structured invoice data and the target audit results and synchronize them to the related business system.

[0021] Specifically, the visualization archiving module receives target audit results and alert notifications from the result aggregation and early warning module. The module presents this information to the user through an interactive dashboard interface, displaying it in charts, lists, and other formats. Users can filter and query data on this interface and view the complete audit trajectory of any invoice from upload to the generation of the final result. Users can manipulate the audit result data; for example, they can create a target invoice dataset based on filtering criteria and export it as a standard format file. For data requiring long-term storage, the module is responsible for archiving the structured invoice data along with the final target audit results. Archiving operations include organizing information according to a predetermined data model and synchronizing the archived data package to other business systems designated by the enterprise, such as the enterprise's resource planning system or contract management system, via an application programming interface (API).

[0022] The method provided in this embodiment, on the one hand, sets up multiple parallel intelligent agents to simultaneously perform verification, deduplication, compliance, and accounting operations on invoice data, thereby parallelizing multiple key review steps that originally had to be executed serially, thus breaking the efficiency bottleneck caused by step dependence in the traditional review process; on the other hand, by decoupling the review rules from the program code and having them visualized and dynamically published by an independent rule management module, the adjustment of the review logic does not require interrupting system operation or modifying the program, thus achieving flexible and real-time updates of the review strategy; furthermore, by integrating invoice parsing, parallel review, result aggregation, and visual archiving into an automated link, manual handover and data transfer between each step are eliminated, thus constructing an end-to-end automated processing closed loop from invoice entry to result output and reuse.

[0023] In some embodiments, the invoice access parsing module 100 includes: The multi-format file receiving and classification unit receives original invoice files in various formats and classifies and labels them according to their formats, generating format-marked original invoice files. Specifically, the multi-format file receiving and classification unit provides a file upload interface. When users upload invoice files in batches or the system receives files from other channels, this unit first obtains these original invoice files. The unit analyzes each received file, determining its specific format based on the file extension and file header information, such as image format, document format, or spreadsheet format. Based on the identified format, the unit adds a format mark to each file, thereby converting the original invoice file into a format-marked original invoice file with a clear processing type label. This classification action lays the foundation for subsequently selecting the correct information extraction method.

[0024] The content extraction unit is used to invoke corresponding image recognition or template extraction methods based on the format tags of the original invoice file with format tags, in order to extract key invoice field information from the file and generate a raw field set. Specifically, the content extraction unit receives the original invoice file with format tags from the upstream unit. The unit routes the file to different processing pipelines according to the format tags carried by the file. For files marked as images or documents, the unit invokes the image recognition processing pipeline, first preprocessing the file to improve image quality, and then extracting all text information from the image using optical character recognition technology. For files marked as spreadsheets, the unit invokes the template extraction pipeline, directly reading data from the corresponding positions according to a preset table structure template. Regardless of which pipeline it passes through, the unit ultimately extracts all relevant key invoice field information from the file and aggregates this information in an initial, unverified structure to form a raw field set.

[0025] The structured data encapsulation and output unit is used to perform standardized verification and encapsulation on the original field set, generate the structured invoice data, and output the structured invoice data to the multi-agent audit module.

[0026] Specifically, the structured data encapsulation and output unit receives the raw field set generated by the content extraction unit. The unit first performs standardized validation on these fields, such as checking the existence of required fields and the validity of key numerical field formats. After successful validation, the unit reorganizes and encapsulates the raw field set according to a predefined, unified data model within the system. The encapsulation process constructs the scattered field information into a data object with a clear structure, field names, and type definitions—that is, structured invoice data. Once this data object is assembled, the unit uses it as the final output of this module and passes it to the downstream multi-agent review module for subsequent review.

[0027] The method provided in this embodiment has several advantages. First, by setting up independent units to automatically identify and classify the original files, the system can adaptively process invoice files from different sources and in different formats, thus eliminating the need for manual pre-classification and format conversion. Second, by calling corresponding image recognition or template extraction methods based on file format tags, the system adopts the optimal information extraction strategy for the characteristics of different file formats, thereby ensuring the accuracy and reliability of information extraction from various types of files. Third, by standardizing and validating the extracted original fields and encapsulating them into unified structured data, the system provides consistent and reliable data input to all downstream audit modules, thus laying the foundation for the automated processing of the entire system.

[0028] In some embodiments, the multi-agent auditing module 200 includes: The audit task scheduling and distribution unit receives the structured invoice data, parses and schedules the data, generates a scheduled audit task, and creates a task context and result container. Specifically, the unit receives structured invoice data from the invoice access and parsing module. It parses the data to identify its basic attributes as an invoice awaiting audit. Then, it creates an audit task instance for the invoice and assigns a unique context and result storage container to this instance. This container is used to associate the original data, intermediate states, and final result of the invoice throughout the audit process. Based on the defined agent type in the system, the unit synchronously sends the audit task distribution signal to the corresponding agent subunit, and associates a reference or copy of the structured invoice data with the task context, thereby generating a scheduled audit task and activating the subsequent parallel audit process.

[0029] A multi-agent parallel audit unit executes the scheduled audit tasks. It performs specialized audits on the structured invoice data through multiple parallel intelligent agent sub-units and writes the generated independent intelligent agent audit sub-results into the task context and result container. Specifically, the multi-agent parallel audit unit executes the scheduled audit tasks generated by the scheduling unit. The unit contains multiple intelligent agent sub-units that start simultaneously and work in parallel. Each intelligent agent sub-unit obtains the required invoice data portion from the shared task context and result container. For example, the verification sub-unit obtains the invoice code and number, and the calculation sub-unit obtains the amount and tax rate. Each sub-unit independently executes its own audit logic: the verification sub-unit calls external interfaces, the deduplication sub-unit queries the database, the compliance sub-unit applies rule sets, and the calculation sub-unit performs calculations and verifications. After completing the audit, each sub-unit writes its conclusions, i.e., the independent intelligent agent audit sub-results, into the reserved space in the task context and result container.

[0030] The review sub-result aggregation unit is used to extract all the independent agent review sub-results from the task context and result container, integrate them, and generate the review sub-results for each agent.

[0031] Specifically, the audit sub-result aggregation unit is launched after all parallel audit agent sub-units have completed their work. The unit accesses the task context and result container, extracting all written independent agent audit sub-results. These results may include verification results, deduplication results, compliance audit results, and accounting results. The unit integrates these sub-results, not simply collecting them, but making a comprehensive judgment according to predetermined logic (e.g., overall failure if verification fails). After integration, the unit outputs a unified view representing the audit conclusions of each agent, generating a set of agent audit sub-results for downstream modules to use.

[0032] The method provided in this embodiment, on the one hand, creates an independent task context and result container for each invoice by setting up a scheduling unit, thereby providing an isolated and traceable data environment for subsequent parallel review, and thus ensuring data consistency and independence during multi-task concurrent processing; on the other hand, by enabling multiple intelligent agent sub-units to obtain data from the shared container in parallel and perform review, it achieves true synchronous execution of multiple review dimensions such as verification, deduplication, compliance, and accounting, thereby compressing the review time of a single invoice to mainly depend on the single sub-task with the longest processing time; furthermore, by setting up an independent aggregation unit to extract and integrate all sub-results from the container, it ensures that the scattered review conclusions can be systematically collected and comprehensively evaluated, thereby forming a complete set of review conclusions that can be used for subsequent decision-making.

[0033] In some embodiments, the multi-agent parallel review unit includes: The verification agent is used to query the invoices based on the invoice code and number information in the structured invoice data and generate verification results. Specifically, after activation, the verification agent retrieves structured invoice data from the data associated with the current audit task and extracts the core information necessary for authenticity verification, mainly including the invoice code and invoice number. The agent encapsulates this information according to the format required by the interface and calls the verification interface through a secure network channel. The agent sends a query request and waits for the interface response. Upon receiving the response, the agent parses the response data, determines the registration status of the invoice in the tax system (e.g., normal, non-existent, voided), and generates a clear verification result based on this status.

[0034] The deduplication detection agent is used to query the historical invoice database based on the invoice code and number information in the structured invoice data, compare them, and generate deduplication results. Specifically, after the deduplication detection agent is started, it extracts the unique identifier information of the invoice from the structured invoice data, namely the combination of the invoice code and the invoice number. Using this combination information as a search condition, the agent first queries the historical invoice database maintained internally by the system, which stores all invoice records that have successfully passed the system's review. The agent executes the query to determine whether the unique identifier of the current invoice already exists in the database. In addition, the agent can send query requests to integrated external deduplication data sources according to the configuration. Combining the feedback from internal and external queries, the agent generates deduplication results, indicating whether the invoice is a duplicate.

[0035] The compliance agent is used to perform compliance verification based on the structured invoice data and the currently effective audit rule set, and generate a compliance audit result. Specifically, when the compliance agent is working, in addition to obtaining the structured invoice data of the current invoice, it also needs to obtain the latest currently effective audit rule set provided by the rule management module. The agent loads the rule set and applies the rules to the invoice data one by one. Each rule contains conditions and actions, such as "if the invoice type is X and the amount is greater than Y, then mark it as non-compliant". The agent checks whether the invoice data meets the conditions defined in the rule. If it does, it triggers the corresponding action of the rule, such as marking a risk or recording a violation. After traversing all relevant rules, the agent summarizes the conclusions of all triggered rules and generates an audit result on the compliance of the invoice, i.e., the compliance audit result.

[0036] The accounting agent is used to perform numerical calculations and verifications based on the amount and tax rate information in the structured invoice data, and to generate accounting results.

[0037] Specifically, the accounting agent focuses on the numerical information on the invoice. It extracts key numerical fields such as amount and tax rate from the structured invoice data. Based on basic mathematical calculation logic, the agent recalculates the accrued tax and the total price including tax on the invoice. The agent compares the calculated result with the corresponding field values ​​recorded in the invoice data. Through this comparison, the agent determines whether the original calculation on the invoice is correct and whether there are any calculation errors or human errors. Based on the comparison conclusion, the agent generates the accounting result, indicating the accuracy of the amount calculation.

[0038] The method provided in this embodiment, on the one hand, decouples the potentially time-consuming and unstable review process involving external network requests from other processes by setting up an independent verification agent specifically responsible for calling interfaces to verify authenticity, thereby preventing this process from blocking the entire review process; on the other hand, by setting up an independent compliance agent and dynamically providing it with rule sets, the complex and potentially frequently changing business rule judgments are handled by a dedicated logic unit, thereby separating business logic from program flow and improving the system's responsiveness to changes in review standards; furthermore, by setting up deduplication and calculation agents to handle data comparison and numerical verification respectively, the tasks relying on internal data retrieval and pure logical calculation are professionally divided, thereby improving the execution efficiency and accuracy of various review tasks.

[0039] In some embodiments, the rule management module includes: The visual rule configuration and editing unit receives user input through a graphical interface to configure audit rules and generates a raw rule definition object. Specifically, the unit provides a graphical user interface. Business administrators interact with the system through this interface to configure new audit rules or edit existing ones. Users operate using visual controls within the interface, such as selecting the "Invoice Type" field from a drop-down list, dragging the "equal to" logical operator from the operator panel, and entering "VAT Special Invoice" as a condition value in the input box. Users define a complete rule by combining multiple such conditions with corresponding actions (such as "mark as high risk"). After the user completes the rule definition in the interface, the unit converts this series of interactive operations into a rule logic expression recognizable by the system, generating a raw rule definition object containing all elements of the rule.

[0040] The rule knowledge base storage and management unit is used to store and manage the lifecycle of the original rule definition objects, generating and maintaining managed rule objects. Specifically, the rule knowledge base storage and management unit receives the original rule definition objects from the configuration unit. The unit persistently stores the object in the database, forming a formal rule record. During and after storage, the unit performs full lifecycle management of the rule record. Management operations include creating version history for the rules to track changes; setting the effective status of the rules (such as enabled, disabled, or tested); and batch importing, exporting, or deleting rules. Through these management functions, the original rule definition objects are transformed into managed rule objects with rich metadata under system control.

[0041] The rule publishing and supply unit is used to dynamically assemble and publish the currently effective review rule set according to the status of the managed rule object, and send the currently effective review rule set to the multi-agent review module.

[0042] Specifically, the rule publishing and supply unit continuously monitors the status of all managed rule objects in the rule knowledge base. When an administrator changes the status of one or more rules to "enabled" and confirms publication through the management unit, the publishing unit detects this change. The unit extracts the latest rule logic definition from all managed rule objects with the "enabled" status. Subsequently, the unit compiles, links, or assembles these scattered logic definitions into a unified and complete rule package, i.e., the currently effective audit rule set. After this rule set is generated, the unit actively pushes or exposes it to the multi-agent audit modules that need to consume rules, especially the compliance agents, through predefined communication mechanisms or interfaces within the system.

[0043] The method provided in this embodiment, on the one hand, transforms the rule configuration process from coding to graphical operation by providing a visual interface, thereby enabling business personnel without technical backgrounds to directly participate in the formulation and adjustment of review logic, thus reducing the technical threshold and communication costs of system maintenance; on the other hand, by setting up an independent storage management unit to manage the version and status of rules, a systematic and standardized control over review rules is achieved, ensuring the traceability and rollback of rule changes; furthermore, by setting up a publishing unit to dynamically assemble and publish effective rule sets, any rule status change can be reflected in the running review logic in real time, thereby realizing hot updates of review strategies and responding to changes in business needs without interrupting services.

[0044] In some embodiments, receiving user input via a graphical interface to configure audit rules and generating an original rule definition object includes: The system receives configuration operations performed by the user on the graphical interface by dragging and dropping logical operators, selecting predefined invoice data fields, and setting thresholds.

[0045] Specifically, the system displays a rule editing workspace through a graphical interface. Within this workspace, users define rules using the mouse and keyboard. Users drag and drop graphical components representing logical operators (such as "AND," "OR," and "NOT") from the interface's control library into the rule editing area. Users select the desired fields from a predefined list of invoice data fields (such as "Amount" and "Invoice Provider"). Users set specific thresholds or comparison values ​​for the selected fields using input boxes. The system captures these drag-and-drop, selection, and input actions in real time and records them as a set of configuration operation sequences.

[0046] Based on the configuration operation combination, audit rules with conditional and action logic are formed.

[0047] Specifically, after the user completes a series of configuration operations, the system performs logical combinations in the background based on the captured operation sequence. The system translates these discrete operations into a computer-executable logical statement with a clear "condition-action" structure, according to their relationships on the interface. For example, if the user combines the "amount" field, the "greater than" operator, and the value "10000" with the "mark as high risk" action, the system will then form a logical rule: "If the invoice amount is greater than 10000, then mark the invoice as high risk."

[0048] The audit rules were tested and verified using simulated data.

[0049] Specifically, after the rules are combined, the system provides a testing and verification function. Users or the system can trigger this function. The system calls the built-in testing engine, using a pre-prepared set of simulated invoice data covering various typical scenarios. The testing engine applies the newly combined audit rules to this simulated data, simulating a real audit process. The engine records the rule's judgment result for each piece of simulated data and displays these results to the user, or automatically compares them with the expected results. This process is used to verify the correctness of the rule logic and its potential boundary effects.

[0050] Based on the audit rules after testing and verification, the original rule definition object is generated.

[0051] Specifically, after testing and verification confirms that the rule behavior meets expectations, or when the user confirms to skip the test and save directly, the system executes the generation step. The system encapsulates all components of the logically combined audit rule, including its conditional logic, referenced data fields, set thresholds, executed actions, and related metadata (such as the creator), into a structured data entity. This data entity is the original rule definition object, which completely and formally represents a business audit rule defined by the user, and is ready to be submitted to the rule knowledge base for persistent storage and management.

[0052] The method provided in this embodiment, on the one hand, achieves seamless and accurate conversion from business intent to machine instructions by translating complex user graphical operations into computer-executable rule logic in real time, thereby ensuring that the rules configured by the user can be correctly understood and executed by the system; on the other hand, by providing a test and verification stage based on simulated data before the rules take effect, it allows users to discover and correct potential errors or omissions in the rule logic before affecting real business processes, thereby improving the reliability and security of rule configuration; furthermore, by encapsulating the tested and verified rules into standardized original rule definition objects, it provides a unified and standardized data foundation for rule storage, version management, and subsequent dynamic release, thereby ensuring data consistency throughout the rule management process.

[0053] In some embodiments, dynamically assembling and publishing the currently effective audit rule set based on the state of the managed rule object includes: Monitor the status of all managed rule objects.

[0054] Specifically, a status monitoring service runs in the system background. This service periodically, or through an event listener mechanism, continuously scans all managed rule objects stored in the rule knowledge base. The service reads the attribute fields recording the current status (such as enabled, disabled, draft) of each managed rule object. The service maintains a real-time snapshot of the status of all rules and detects whether the status attributes of any rule object have changed.

[0055] In response to the managed rule object being set to the enabled state, the rule logic is extracted from the enabled managed rule object. Specifically, this step is triggered when the status monitoring service detects that the status of a rule object has been updated to "enabled," or when an explicit rule publishing instruction is received. Based on the publishing instruction or the latest status snapshot, the system filters all managed rule objects with the status "enabled" from the rule knowledge base. The system accesses these objects one by one, parsing and extracting the core part of each object, namely the rule logic that defines the audit conditions. Here, the rule logic is a formalized expression describing what action should be performed under what conditions.

[0056] The extracted rule logic is compiled or assembled into a unified set of currently effective audit rules. Specifically, after obtaining the logic fragments of all enabled rules, the system inputs these fragments into a rule processing engine. The engine performs compilation or assembly operations. For rules based on a specific rule language (such as Drools DRL), the engine performs compilation, converting the text logic into executable bytecode or internal representation. For simpler rules, assembly may be performed, merging all rule logic into a single rule list or rule file. Regardless of the technology used, the end result is the generation of a single rule package containing all currently effective rules, which can be directly loaded and interpreted by the rule engine; this is the set of currently effective audit rules.

[0057] The currently effective set of audit rules is sent to the multi-agent audit module through a predefined interface.

[0058] Specifically, once the currently effective audit rule set is generated, the system initiates the transmission process. The system transmits the newly generated, currently effective audit rule set as a data packet through predefined internal communication interfaces, such as application programming interface calls or message queue publishing. The target of this transmission is the compliance audit component within the multi-agent audit module. Upon receiving the new rule set, this component loads it into memory, replacing the old rule set, thus ensuring that the latest audit strategy immediately takes effect in the audit process for all subsequent invoices.

[0059] The method provided in this embodiment, on the one hand, achieves precise and automated control over the timing of rule activation by continuously monitoring the rule status and responding to status change events, thereby ensuring that only explicitly enabled rules are included in the review logic; on the other hand, by extracting logic from enabled rule objects and compiling and assembling it into a unified rule set, the discrete rule definitions stored in the database are transformed into a complete rule execution package that is efficient and usable at runtime, thereby improving the efficiency of rule matching and execution; furthermore, by actively pushing the generated rule set to the review module through a standard interface, the review module can obtain the latest review policies in real time, thereby achieving instant synchronization between business rule changes and online review services, eliminating the delay in policy updates.

[0060] In some embodiments, the result aggregation and early warning module includes: The multi-source audit result aggregation unit receives the audit sub-results from each intelligent agent and integrates them based on preset adjudication logic to generate the target audit result. Specifically, the multi-source audit result aggregation unit receives multiple intelligent agent audit sub-results from the multi-agent audit module. These sub-results represent audit findings on invoices in different dimensions such as verification, deduplication, compliance, and accounting. The unit has a pre-set adjudication logic. First, the unit parses the content of each sub-result; for example, the verification result may be "passed" or "failed." The unit performs a comprehensive analysis of these results according to the preset logic. For example, the adjudication logic stipulates that if the verification sub-result is "failed," then regardless of other results, the overall audit conclusion is "failed." After applying this logic to evaluate all sub-results, the unit generates a comprehensive conclusion, i.e., the target audit result. This result includes the final audit status of the invoice and a summary of the specific audit situation for each dimension.

[0061] A multi-level risk rating engine unit is used to determine the risk level identifier of the target audit result based on a risk assessment model. Specifically, the multi-level risk rating engine unit receives the target audit result generated by the aggregation unit. Internally, the unit is configured with a risk assessment model that defines the risk levels corresponding to different types and severity of audit issues; for example, mapping "fake invoice" to "high risk" and "error in amount calculation" to "medium risk." The engine unit parses the target audit result and identifies all the issues it contains. Then, the engine matches these issues with entries in the risk assessment model, calculating and assigning a risk level identifier, such as "high," "medium," "low," or a specific risk score, to each issue or the entire invoice.

[0062] The real-time early warning generation and distribution unit is used to generate and send a real-time early warning notification when the early warning triggering conditions are met, based on the target review results and the risk level identifier.

[0063] Specifically, the real-time alert generation and distribution unit simultaneously receives the target audit results and their associated risk level identifiers. The unit has an internal alert triggering mechanism that defines the risk level at which an alert should be triggered, for example, when the risk level identifier is "high" or "medium." The unit makes a judgment based on the received risk level identifier. If the triggering condition is met, the unit extracts key invoice information (such as number and amount) and a specific risk description from the target audit results. The unit organizes this information according to a predefined notification template to generate a structured alert message, i.e., a real-time alert notification. After generation, the unit calls the integrated messaging service to send the notification to a pre-specified list of recipients through configured channels (such as internal system messaging, mail servers, or office software robot interfaces).

[0064] The method provided in this embodiment, on the one hand, integrates scattered and potentially one-sided audit findings into a global and authoritative audit conclusion by setting up an aggregation unit and applying preset adjudication logic to comprehensively evaluate the audit sub-results of each dimension, thereby eliminating the potential for conflicting and uncertain conclusions caused by multi-dimensional audits; on the other hand, by setting up an independent risk rating engine and quantifying audit issues into risk levels based on a model, it achieves an objective and standardized assessment of invoice risks, providing a clear decision-making basis for subsequent differentiated processing; furthermore, by setting up an early warning generation and distribution unit to automatically generate and send notifications when risks reach a threshold, it establishes an automated response link from risk identification to risk notification, shortening the time from risk exposure to risk response.

[0065] In some embodiments, generating and sending a real-time early warning notification when the early warning triggering conditions are met, based on the target review result and the risk level identifier, includes: Determine whether the preset warning triggering conditions are met based on the risk level identifier.

[0066] Specifically, after the real-time alert generation and distribution unit is activated, it first reads the risk level identifier calculated by the risk rating engine unit attached to the currently processed target review result. The unit has a pre-set set of alert triggering conditions, which are directly associated with the risk level identifier; for example, the condition might be "risk level identifier is 'high' or 'medium'". The unit compares and matches the current risk level identifier with the pre-set triggering conditions. Through logical judgment, the unit determines whether the alert triggering conditions are met. If the match is successful, the process proceeds to the next step; otherwise, the process terminates.

[0067] When the aforementioned warning triggering conditions are met, key invoice information and risk point descriptions are extracted from the target audit results to generate a structured real-time warning notification. Specifically, the warning generation component is activated after the warning triggering conditions are determined to be met. This component accesses the associated target audit results. The component parses and extracts several key pieces of information from these results, which typically include, but are not limited to, the unique identifier of the invoice (such as code or number), the specific amount involved in the risk, and a description of the type of issue that triggered the risk (such as "double reimbursement" or "non-compliant tax rate"). The component fills the extracted information into the corresponding positions of a predefined warning notification template within the system. The template specifies the title, body structure, and display format of the notification. After completion, the component generates a complete, fixed-format notification document, i.e., a structured real-time warning notification.

[0068] The real-time alert notification will be sent to a pre-configured recipient via at least one of the following channels: system message, email, or external office robot.

[0069] Specifically, after an alert notification is generated, the distribution component begins its operation. The distribution component accesses the system configuration to obtain the list of recipients associated with this type of risk alert and the specified sending channels. Sending channels may include the system's internal message center, an external email system, or an integrated instant messaging tool's bot interface. Based on the configuration, the distribution component encapsulates the generated structured real-time alert notification as message content and sends it through the application programming interfaces of one or more selected channels. The message is successfully delivered to the server of the corresponding channel, and ultimately delivered to the recipient by the server.

[0070] The method provided in this embodiment, on the one hand, automates and objectifies the decision-making process for issuing early warnings by explicitly defining the early warning trigger logic as a matching judgment between risk level identifiers and preset conditions, avoiding reliance on manual judgment and ensuring the timeliness and consistency of risk response; on the other hand, it standardizes and automates the production of early warning content by extracting key information from the review results and applying preset templates to generate structured notifications, ensuring the accuracy and completeness of risk information transmission and avoiding omissions that may occur during manual writing; furthermore, by supporting the sending of notifications to pre-configured recipients through multiple channels, it achieves the diversity and reliability of risk information transmission channels, ensuring that key risks can effectively reach relevant responsible personnel.

[0071] In some implementations, the visual archiving module includes: The multi-dimensional data visualization unit receives the target review results and the early warning notifications, and provides an interactive interface to visualize and trace the entire review process. Specifically, the unit receives the target review results and real-time early warning notifications from the result aggregation and early warning module. The unit presents this data as visual elements through the system's user interface, such as a web dashboard. The interface displays global statistical data charts, such as pass rate, rejection rate, and risk distribution pie charts. It also displays detailed invoice review results and early warning notification entries in a list format. Users can interact with the interface, such as using filters to filter the result list by time, department, or risk level. For any invoice in the list, users can click to view details. The unit retrieves all processing records for that invoice from file upload to the generation of the final result from the system log and organizes them into a coherent review trajectory view for the user.

[0072] An interactive data manipulation and export unit is used to respond to user operations on the interactive interface. The data corresponding to user operations is filtered to form a target invoice data set, and the target invoice data set is exported.

[0073] Specifically, the interactive data manipulation and export unit is tightly integrated with the visual display interface. Users issue commands through the operation controls on the interface; for example, a user sets filter conditions and clicks the "Export" button. The unit captures this user command and, based on the implicit filter conditions in the command (such as "past week, high risk"), queries and filters the entire target audit result data in the background, forming a target invoice data set that matches the user's intent. Subsequently, the unit converts and organizes the data in this data set according to a common data exchange format (such as spreadsheet format), generating a data file that can be downloaded by the user.

[0074] The archiving and cross-system synchronization unit is used to archive the structured invoice data and the target audit results and synchronize them to the associated business system.

[0075] Specifically, the archiving and cross-system synchronization unit performs data persistence and distribution tasks. The source data processed by this unit includes structured invoice data from the invoice access and parsing module and the final target audit results from the result aggregation and early warning module. Following the system-defined archiving data model, the unit correlates, cleans, and reorganizes these two data components, encapsulating them into data packets suitable for long-term storage and cross-system exchange. Then, the unit sends the encapsulated data packets to these target systems by calling their standard data interfaces (such as ERP systems and contract management systems). Upon receiving the data, the target systems import it into their own databases, completing the data synchronization and archiving process.

[0076] The method provided in this embodiment, on the one hand, visualizes the audit results and warning information in the form of multi-dimensional charts and interactive lists, thereby providing users with the ability to have a global grasp and trace details, making the situational awareness and problem location of massive audit data intuitive and efficient; on the other hand, by responding to user interaction, it directly performs an export operation on the filtered result set, thereby integrating data query, filtering, and formatted output into a fast and coherent action, meeting users' needs for offline analysis and reporting of audit result data; furthermore, by archiving the original data of invoices and audit conclusions according to a standard model and synchronizing them to external business systems, it opens up the data flow between the intelligent audit process and the enterprise's core business system, realizing the value transfer and reuse of audit data in enterprise-level applications.

[0077] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An intelligent invoice verification system based on multi-agent collaboration, characterized in that, include: The invoice access and parsing module is used to perform optical character recognition processing on the received invoice files in order to extract key fields and generate structured invoice data; The invoice access and parsing module includes: a multi-format file receiving and classification unit, used to receive original invoice files in various formats, classify and mark the original invoice files according to their formats, and generate original invoice files with format markings; a content extraction unit, used to call corresponding image recognition or template extraction methods according to the format markings of the original invoice files with format markings, in order to extract key invoice field information from the files and generate an original field set; and a structured data encapsulation and output unit, used to perform standardized verification and encapsulation on the original field set, generate the structured invoice data, and output the structured invoice data to the multi-agent review module. A multi-agent auditing module is used to perform verification, deduplication, compliance, and accounting operations on the structured invoice data through multiple parallel agents, generating audit sub-results for each agent. The multi-agent auditing module includes: an audit task scheduling and distribution unit, used to receive the structured invoice data, parse and schedule the structured invoice data, generate scheduled audit tasks, and create task contexts and result containers; and a multi-agent parallel auditing unit, used to execute the scheduled audit tasks, perform specialized audits on the structured invoice data through multiple parallel agent sub-units, and write the generated independent agent audit sub-results into the task context and result containers. The multi-agent parallel auditing unit includes: a verification agent, which... Based on the invoice code and number information in the structured invoice data, the system calls the verification interface to query and generate verification results; the deduplication agent queries the historical invoice database for comparison based on the invoice code and number information in the structured invoice data and generates deduplication results; the compliance agent performs compliance verification based on the structured invoice data and the currently effective audit rule set and generates compliance audit results; the accounting agent performs numerical calculation verification based on the amount and tax rate information in the structured invoice data and generates accounting results; and the audit sub-result aggregation unit extracts all the independent agent audit sub-results from the task context and result container, integrates them, and generates the audit sub-results for each agent. The rule management module is used to configure audit rules through a graphical interface to generate a currently effective audit rule set and provide the currently effective audit rule set to the multi-agent audit module. The rule management module includes: a visual rule configuration and editing unit, used to receive configuration operations performed by the user on the graphical interface by dragging and dropping logical operators, selecting predefined invoice data fields, and setting thresholds; forming audit rules with conditional and action logic based on the configuration operations; testing and verifying the audit rules using simulated data; generating original rule definition objects based on the tested and verified audit rules; a rule knowledge base storage and management unit, used to store and manage the lifecycle of the original rule definition objects, and generate and maintain managed rule objects; and a rule publishing and supply unit, used to monitor the status of all managed rule objects; extracting rule logic from the managed rule objects in the enabled state in response to the managed rule objects being set to the enabled state; compiling or assembling the extracted rule logic into a unified currently effective audit rule set; and sending the currently effective audit rule set to the multi-agent audit module through a predefined interface. The result aggregation and early warning module is used to aggregate the review sub-results of each intelligent agent to generate a target review result, and classify the risk points in the target review result into risk levels to generate an early warning notification. The result aggregation and early warning module includes: a multi-source review result aggregation unit, used to receive the review sub-results of each intelligent agent and integrate them based on preset adjudication logic to generate a target review result; a multi-level risk rating engine unit, used to determine the risk level identifier of the target review result according to a risk assessment model; and a real-time early warning generation and distribution unit, used to determine whether preset early warning triggering conditions are met based on the risk level identifier; when the early warning triggering conditions are met, extracting key invoice information and risk point descriptions from the target review result to generate a structured real-time early warning notification; and sending the real-time early warning notification to a pre-configured recipient through at least one of the following channels: system message, email, or external office robot. The visualization archiving module is used to display the target audit results and the warning notification, and to archive the structured invoice data and the target audit results and synchronize them to the associated business system. The visualization archiving module includes: a multi-dimensional data visualization display unit, used to receive the target audit results and the warning notification, and to provide an interactive interface for visually displaying the target audit results and the warning notification and tracing the entire audit process; an interactive data operation and export unit, used to respond to user operations on the interactive interface, filter the data corresponding to the user operations to form a target invoice data set, and export the target invoice data set; and an archiving and cross-system synchronization unit, used to archive the structured invoice data and the target audit results and synchronize them to the associated business system.

Images