SaaS financial software artificial intelligence voucher auditing method based on expert database

Abstract

The invention discloses a SaaS financial software artificial intelligence voucher auditing method based on an expert database, and relates to the field of voucher auditing, and the method comprises the steps: receiving to-be-audited financial voucher data uploaded by a user through SaaS financial software; integrity verification and logic conflict detection are performed on the structured data, text information in the unstructured data is extracted, and authenticity feature identification is performed on the bill image; constructing a uniform feature vector; obtaining an auditing rule set, calculating a matching degree between the feature vector and the auditing rule set, and outputting an initial auditing confidence coefficient; if the confidence coefficient is higher than a preset threshold value, adopting, and if the confidence coefficient is lower than the preset threshold value, triggering multiple rounds of iterative verification; and updating the expert database, and returning a structured auditing result to the client through the API gateway. The method has the advantages that the AI technology and the expert database are integrated, efficient and accurate financial voucher auditing is achieved, the rules are dynamically updated, manual intervention is reduced, and automation and applicability are improved.

Description

Technical Field

[0001] This invention relates to the field of voucher verification, and in particular to an AI-based voucher verification method for SaaS financial software based on an expert database. Background Technology

[0002] AI-powered voucher verification technology in SaaS financial software, based on an expert database, leverages cloud computing and artificial intelligence to improve the efficiency and accuracy of financial management. The SaaS model, deployed in the cloud, enables flexibility and scalability of the financial software, allowing users to access it in real-time without local installation. The expert database integrates professional financial knowledge, industry standards, and historical data to form a structured rule base and model library, supporting intelligent analysis and decision-making by AI algorithms.

[0003] Current financial voucher verification methods on the market rely on simple rule engines and lack the ability to deeply process unstructured data, such as verifying the authenticity of voucher images and extracting complex text, resulting in low verification efficiency and a high risk of errors. Secondly, existing methods typically do not integrate expert databases, and their verification rules are static and outdated, making them difficult to adapt to diverse or new voucher types and limiting their applicability. Furthermore, existing methods lack a unified feature vector construction mechanism, failing to effectively integrate semantic, numerical, and rule features, resulting in insufficient confidence quantification of verification results and low objectivity and reliability. Especially for low-confidence vouchers, existing methods often lack multi-round iterative verification mechanisms, relying solely on manual review, increasing time and labor costs. Simultaneously, traditional methods provide insufficiently structured data feedback, leading to a poor user experience and failing to meet the needs of large-scale, complex financial scenarios. Summary of the Invention

[0004] To improve existing methods, this paper proposes an AI-based voucher review method for SaaS financial software based on an expert database. This method integrates AI technology with an expert database to achieve efficient and accurate financial voucher review, dynamically updates rules to reduce human intervention, and improves automation and applicability.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] AI-powered voucher verification methods for SaaS financial software based on expert databases include:

[0007] Receive pending financial voucher data uploaded by users through SaaS financial software;

[0008] Preprocessing is performed on the acquired financial voucher data to be audited, integrity verification and logical conflict detection are performed on the structured data, text information is extracted from the unstructured data, and authenticity features are identified on the voucher images based on convolutional neural networks.

[0009] Based on the feature data of the preprocessed financial voucher data to be audited, a unified feature vector is constructed. The feature vector includes a semantic embedding vector, a numerical feature matrix, and a rule matching label.

[0010] Based on the feature vector, the audit rule set associated with the current voucher type is obtained from the expert database, and the matching degree between the feature vector and the audit rule set is calculated. The initial audit confidence is output to quantify the audit result.

[0011] Based on the initial review confidence level obtained, if the confidence level is higher than the preset threshold, it is adopted; if the confidence level is lower than the preset threshold, multiple rounds of iterative verification are triggered.

[0012] The expert database is updated based on the review results, and structured review results are returned to the client via the API gateway.

[0013] Preferably, receiving the financial voucher data to be reviewed uploaded by the user through the SaaS financial software specifically includes:

[0014] The financial voucher data includes structured data and unstructured data;

[0015] The structured data includes at least the subject code, amount, date, and counterparty information;

[0016] The unstructured data includes at least scanned images of the invoices and PDF attachments.

[0017] Preferably, the preprocessing of the acquired financial voucher data to be audited, including performing integrity checks and logical conflict detection on structured data, extracting text information from unstructured data, and identifying the authenticity of the voucher images based on convolutional neural networks, specifically includes:

[0018] Based on the structured data in the financial vouchers to be audited, integrity verification is performed by checking the required fields and field formats.

[0019] Logical conflict detection is performed by checking the consistency of monetary logic in structured data, verifying time logic, and detecting duplicate vouchers.

[0020] Based on the unstructured data in the financial vouchers to be audited, text information in the voucher images is extracted using OCR technology, and the text is then processed into a structured format.

[0021] Based on the collected and labeled invoice image dataset, the data is input into a convolutional neural network-based model for recognizing authenticity features to train the model.

[0022] The image of the invoice to be reviewed is input into the trained model to obtain the authenticity recognition result, and the invoices to be reviewed are re-reviewed if the result is fake.

[0023] The verification results of structured data, the logical conflict detection results, the text extraction results of unstructured data, and the results of ticket authenticity recognition are integrated.

[0024] Preferably, the construction of a unified feature vector based on the feature data of the preprocessed financial voucher data to be audited, wherein the feature vector includes a semantic embedding vector, a numerical feature matrix, and rule matching labels, specifically including:

[0025] Based on the preprocessed financial voucher data to be audited, the text data of the structured data is integrated to obtain a complete text dataset, and key text fields are extracted from the unstructured data to construct semantic embedding vectors.

[0026] Key numerical fields are extracted from structured data, quantitative features are extracted from the authenticity identification results, discrete numerical features are uniquely encoded, continuous numerical features are normalized, and all numerical features are integrated into a matrix form to construct a numerical feature matrix.

[0027] Based on the requirements for auditing financial vouchers, an audit rule set is defined, including integrity rules, logical rules, and truth / false rules. The rules are then classified into hard rules and soft rules, and rule matching tags are generated.

[0028] Semantic embedding vector, numerical feature matrix and rule matching label are concatenated in sequence into a unified feature vector.

[0029] Preferably, the step of obtaining the audit rule set associated with the current voucher type from the expert database based on feature vectors, calculating the matching degree between the feature vectors and the audit rule set, and outputting the initial audit confidence level, specifically includes:

[0030] The expert database stores a knowledge graph of historical audit cases annotated by domain experts;

[0031] Based on the voucher type, the corresponding audit rule set is queried from the expert database. If the voucher type is a mixed or complex type, a comprehensive rule set is generated.

[0032] Map the semantic embedding vector, numerical feature matrix, and rule matching label in the feature vector of the current credential to the conditions of the audit rule set one by one;

[0033] For each rule, the matching degree between the feature vector and the rule conditions is calculated to generate the matching degree of a single rule. The overall matching degree is generated by weighted summation of the matching degrees of all rules.

[0034] Based on the overall matching degree, and combining the matching results of hard and soft rules, the initial review confidence level is calculated.

[0035] Preferably, the step of adopting the initial review confidence score based on the obtained confidence score if the confidence score is higher than a preset threshold, and triggering multiple rounds of iterative verification if the confidence score is lower than the preset threshold, specifically includes:

[0036] Based on business needs and historical audit data, set confidence thresholds and classify confidence levels based on the initial audit confidence levels obtained.

[0037] For vouchers with a confidence level higher than the threshold, they are marked as automatically approved; for vouchers with a confidence level lower than the threshold, they are marked as requiring iterative verification.

[0038] Data completion, in-depth feature analysis, and rule adjustment are performed through multiple rounds of verification until the confidence level exceeds the threshold.

[0039] Preferably, updating the expert database based on the review results and returning the structured review results to the client through the API gateway specifically includes:

[0040] Based on the review results with high confidence levels, review result data is collected, including confidence level, pass / fail status, and iterative verification records;

[0041] Add or modify data in the expert database based on the audit results data;

[0042] For each voucher, a structured audit result is generated, including voucher ID, audit status, confidence level, matching rules, and iterative verification records. The audit result is then pushed to the client via the API gateway.

[0043] Compared with the prior art, the advantages of the present invention are:

[0044] By integrating structured and unstructured data, and utilizing OCR technology and convolutional neural networks, efficient text extraction and document authenticity recognition are achieved, significantly improving the accuracy and efficiency of the review process. Secondly, a unified feature vector is constructed and combined with an expert database's review rule set, enabling precise matching of document types and rules, outputting quantified confidence levels, and ensuring the objectivity and reliability of the review results. Particularly for complex or mixed documents, the generation of a comprehensive rule set further enhances the method's applicability. Furthermore, low-confidence documents trigger multiple rounds of iterative verification, continuously optimizing review quality through data completion and rule adjustments, reducing the need for manual intervention, and improving automation. Simultaneously, the dynamic update mechanism of the expert database allows the system to continuously learn from the latest review cases, maintaining the timeliness and adaptability of the rules. Structured results are returned through the API gateway, allowing clients to quickly obtain clear review feedback, improving the user experience. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the method proposed in this invention;

[0046] Figure 2 This is a schematic diagram of the financial voucher data to be audited as proposed in this invention;

[0047] Figure 3 This is a schematic diagram of the data preprocessing proposed in this invention;

[0048] Figure 4 This is a schematic diagram illustrating the construction of a unified feature vector proposed in this invention;

[0049] Figure 5 This is a schematic diagram illustrating the matching degree and initial review confidence level proposed in this invention;

[0050] Figure 6 This is a schematic diagram illustrating the audit confidence level determination proposed in this invention;

[0051] Figure 7 This is a schematic diagram of the updated expert database proposed in this invention. Detailed Implementation

[0052] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0053] See Figure 1 As shown, the AI-powered voucher verification method for SaaS financial software based on an expert database includes:

[0054] Step 1: Receive financial voucher data uploaded by users for review via SaaS financial software;

[0055] Step 2: Based on the acquired financial voucher data to be audited, preprocess the data, perform integrity verification and logical conflict detection on the structured data, extract text information from the unstructured data, and perform authenticity feature recognition on the voucher image based on a convolutional neural network.

[0056] Step 3: Based on the feature data of the preprocessed financial voucher data to be audited, construct a unified feature vector, which includes a semantic embedding vector, a numerical feature matrix, and rule matching labels;

[0057] Step 4: Based on the feature vector, obtain the audit rule set associated with the current voucher type from the expert database, calculate the matching degree between the feature vector and the audit rule set, output the initial audit confidence, and quantify the audit result;

[0058] Step 5: Based on the obtained initial review confidence level, if the confidence level is higher than the preset threshold, it is adopted; if the confidence level is lower than the preset threshold, multiple rounds of iterative verification are triggered.

[0059] Step 6: Update the expert database based on the audit results and return the structured audit results to the client through the API gateway.

[0060] See Figure 2 As shown, the specific data of financial vouchers to be reviewed uploaded by users and received through SaaS financial software includes:

[0061] The financial voucher data includes structured data and unstructured data;

[0062] The structured data includes at least the subject code, amount, date, and counterparty information;

[0063] The unstructured data includes at least scanned images of the invoices and PDF attachments.

[0064] See Figure 3 As shown, based on the acquired financial voucher data to be audited, preprocessing is performed, including integrity verification and logical conflict detection of structured data, extraction of text information from unstructured data, and authenticity feature recognition of the voucher images based on convolutional neural networks. Specifically, this includes:

[0065] Based on the structured data in the financial vouchers to be audited, integrity verification is performed by checking the required fields and field formats.

[0066] Logical conflict detection is performed by checking the consistency of monetary logic in structured data, verifying time logic, and detecting duplicate vouchers.

[0067] Based on the unstructured data in the financial vouchers to be audited, text information in the voucher images is extracted using OCR technology, and the text is then processed into a structured format.

[0068] Based on the collected and labeled invoice image dataset, the data is input into a convolutional neural network-based model for recognizing authenticity features to train the model.

[0069] The image of the invoice to be reviewed is input into the trained model to obtain the authenticity recognition result, and the invoices to be reviewed are re-reviewed if the result is fake.

[0070] The verification results of structured data, the logical conflict detection results, the text extraction results of unstructured data, and the results of ticket authenticity recognition are integrated.

[0071] Specifically, for the integrity verification of structured data, the required fields are checked to see if there are any empty values ​​in any field for each voucher record, and the field format is checked to verify whether the value of each field conforms to the format rules.

[0072] For unstructured data, optical character recognition is performed on the ticket image to extract text information from the image, noise in the extracted text is removed, valid information is retained, and the extracted text is classified and formatted by field to generate structured data records.

[0073] Collect a dataset of labeled invoice images, including genuine and counterfeit invoice samples, ensuring clear labeling. Use a convolutional neural network model, input invoice images, extract features, and perform supervised learning through labeled data to optimize model parameters.

[0074] Input the image of the invoice to be reviewed into the trained CNN model to obtain the result of authenticity recognition.

[0075] See Figure 4 As shown, based on the feature data of the preprocessed financial voucher data to be audited, a unified feature vector is constructed. This feature vector includes a semantic embedding vector, a numerical feature matrix, and rule matching labels, specifically:

[0076] Based on the preprocessed financial voucher data to be audited, the text data of the structured data is integrated to obtain a complete text dataset, and key text fields are extracted from the unstructured data to construct semantic embedding vectors.

[0077] Key numerical fields are extracted from structured data, quantitative features are extracted from the authenticity identification results, discrete numerical features are uniquely encoded, continuous numerical features are normalized, and all numerical features are integrated into a matrix form to construct a numerical feature matrix.

[0078] Based on the requirements for auditing financial vouchers, an audit rule set is defined, including integrity rules, logical rules, and truth / false rules. The rules are then classified into hard rules and soft rules, and rule matching tags are generated.

[0079] Semantic embedding vector, numerical feature matrix and rule matching label are concatenated in sequence into a unified feature vector.

[0080] Specifically, all text fields are extracted from the preprocessed structured data, duplicate or redundant fields are merged to generate a complete text dataset; field recognition is performed on the unstructured data to extract key fields, and named entity recognition technology is used to perform structured processing on the text to generate a set of key text fields; the complete text dataset and the set of key text fields are then combined into a unified text input.

[0081] Numerical fields are extracted from structured data, quantitative features are extracted from the authenticity identification results of invoices, discrete numerical features (such as transaction type and voucher status) are uniquely encoded, and all discrete features are integrated into a one-hot encoding matrix, which is the sum of the one-hot dimensions of all discrete features. Continuous numerical features are normalized, and all continuous features are integrated into a normalization matrix, which is the number of continuous features. The one-hot encoding matrix, the normalization matrix, and the authenticity quantitative features are concatenated into a numerical feature matrix.

[0082] For each voucher, check whether it satisfies each rule in the rule set, generate a label vector, and integrate the label vectors of all vouchers into a label matrix;

[0083] The semantic embedding vector, numerical feature matrix, and label matrix are concatenated in sequence to form a unified feature vector.

[0084] See Figure 5 As shown, the system retrieves the set of audit rules associated with the current voucher type from the expert database based on feature vectors, calculates the matching degree between the feature vectors and the audit rule set, outputs the initial audit confidence level, and quantifies the audit results, specifically including:

[0085] The expert database stores a knowledge graph of historical audit cases annotated by domain experts;

[0086] Based on the voucher type, the corresponding audit rule set is queried from the expert database. If the voucher type is a mixed or complex type, a comprehensive rule set is generated.

[0087] Map the semantic embedding vector, numerical feature matrix, and rule matching label in the feature vector of the current credential to the conditions of the audit rule set one by one;

[0088] For each rule, the matching degree between the feature vector and the rule conditions is calculated to generate the matching degree of a single rule. The overall matching degree is generated by weighted summation of the matching degrees of all rules.

[0089] Based on the overall matching degree, and combining the matching results of hard and soft rules, the initial review confidence level is calculated.

[0090] Specifically, the database of historical audit case knowledge graphs annotated by storage experts is accessed, which includes the relationship between voucher types, audit rules, and historical audit results, and the voucher type field is extracted from the structured data of the financial vouchers to be audited;

[0091] Obtain the unified feature vector of the current credential, decompose each rule in the audit rule set into a set of conditions, match the relevant part of the feature vector of each rule with the conditions, and record the matching result of each condition;

[0092] For each rule, the matching degree of a single rule is calculated based on the matching results of its conditions. For all conditions of a rule, the matching degree of a single rule is calculated by weighted average.

[0093] Check the matching results of hard rules: If any hard rule is not met, the confidence level is set to 0, indicating that the review is not passed. If all hard rules are met, the confidence level is calculated based on the overall matching degree and the matching results of soft rules.

[0094] See Figure 6 As shown, based on the obtained initial review confidence level, if the confidence level is higher than the preset threshold, it is adopted; if the confidence level is lower than the preset threshold, multiple rounds of iterative verification are triggered, specifically including:

[0095] Based on business needs and historical audit data, set confidence thresholds and classify confidence levels based on the initial audit confidence levels obtained.

[0096] For vouchers with a confidence level higher than the threshold, they are marked as automatically approved; for vouchers with a confidence level lower than the threshold, they are marked as requiring iterative verification.

[0097] Data completion, in-depth feature analysis, and rule adjustment are performed through multiple rounds of verification until the confidence level exceeds the threshold.

[0098] Specifically, based on business needs and historical audit data, use historical audit cases to statistically analyze the confidence distribution, set reasonable initial audit confidence thresholds, set differentiated thresholds for different voucher types according to business scenarios, and record the basis for threshold setting to ensure traceability;

[0099] The multi-round verification process involves several rounds. The first round completes the data and remaps the audit rule set using the completed feature vectors to calculate a new matching degree. The second round performs deep feature analysis, concatenating the deep features with the original feature vectors to generate an enhanced feature vector. The third round adjusts the rules, relaxing the conditions for soft rules or adding exceptions to hard rules, and using the updated rule set and feature vectors to calculate a new matching degree and confidence level.

[0100] Summarize the classification results, confidence values, and intermediate results of multiple rounds of verification for all vouchers.

[0101] See Figure 7 As shown, updating the expert database based on the review results and returning structured review results to the client through the API gateway specifically includes:

[0102] Based on the review results with high confidence levels, review result data is collected, including confidence level, pass / fail status, and iterative verification records;

[0103] Add or modify data in the expert database based on the audit results data;

[0104] For each voucher, a structured audit result is generated, including voucher ID, audit status, confidence level, matching rules, and iterative verification records. The audit result is then pushed to the client via the API gateway.

[0105] Specifically, based on the voucher ID and voucher type, check if a corresponding record already exists in the expert database. If the record does not exist (i.e., a new voucher), insert the audit result data as a new record into the expert database.

[0106] Analyze high-frequency failure rules or low-confidence credential types, update rule weights or conditions in the knowledge graph, and add association rules or adjust thresholds for low-confidence credential types to optimize subsequent review.

[0107] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0108] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0109] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for AI-powered voucher verification in SaaS financial software based on an expert database, characterized in that: include: Receive pending financial voucher data uploaded by users through SaaS financial software; Preprocessing is performed on the acquired financial voucher data to be audited, integrity verification and logical conflict detection are performed on the structured data, text information is extracted from the unstructured data, and authenticity features are identified on the voucher images based on convolutional neural networks. Based on the feature data of the preprocessed financial voucher data to be audited, a unified feature vector is constructed. The feature vector includes a semantic embedding vector, a numerical feature matrix, and a rule matching label. Based on the feature vector, the audit rule set associated with the current voucher type is obtained from the expert database, and the matching degree between the feature vector and the audit rule set is calculated. The initial audit confidence is output to quantify the audit result. Based on the initial review confidence level obtained, if the confidence level is higher than the preset threshold, it is adopted; if the confidence level is lower than the preset threshold, multiple rounds of iterative verification are triggered. The expert database is updated based on the review results, and structured review results are returned to the client via the API gateway.

2. The AI-powered voucher verification method for SaaS financial software based on an expert database as described in claim 1, characterized in that, The specific details of receiving user-uploaded financial voucher data for review via SaaS financial software include: The financial voucher data includes structured data and unstructured data; The structured data includes at least the subject code, amount, date, and counterparty information; The unstructured data includes at least scanned images of the invoices and PDF attachments.

3. The AI-powered voucher verification method for SaaS financial software based on an expert database as described in claim 1, characterized in that, The preprocessing of the acquired financial voucher data to be audited, including integrity verification and logical conflict detection of structured data, extraction of text information from unstructured data, and authenticity feature recognition of the voucher images based on convolutional neural networks, specifically includes: Based on the structured data in the financial vouchers to be audited, integrity verification is performed by checking the required fields and field formats. Logical conflict detection is performed by checking the consistency of monetary logic in structured data, verifying time logic, and detecting duplicate vouchers. Based on the unstructured data in the financial vouchers to be audited, text information in the voucher images is extracted using OCR technology, and the text is then processed into a structured format. Based on the collected and labeled invoice image dataset, the data is input into a convolutional neural network-based model for recognizing authenticity features to train the model. The image of the invoice to be reviewed is input into the trained model to obtain the authenticity recognition result, and the invoices to be reviewed are re-reviewed if the result is fake. The verification results of structured data, the logical conflict detection results, the text extraction results of unstructured data, and the results of ticket authenticity recognition are integrated.

4. The AI-powered voucher verification method for SaaS financial software based on an expert database as described in claim 1, characterized in that, Based on the feature data of the preprocessed financial vouchers to be audited, a unified feature vector is constructed. This feature vector includes a semantic embedding vector, a numerical feature matrix, and rule-matching labels, specifically comprising: Based on the preprocessed financial voucher data to be audited, the text data of the structured data is integrated to obtain a complete text dataset, and key text fields are extracted from the unstructured data to construct semantic embedding vectors. Key numerical fields are extracted from structured data, quantitative features are extracted from the authenticity identification results, discrete numerical features are uniquely encoded, continuous numerical features are normalized, and all numerical features are integrated into a matrix form to construct a numerical feature matrix. Based on the requirements for auditing financial vouchers, an audit rule set is defined, including integrity rules, logical rules, and truth / false rules. The rules are then classified into hard rules and soft rules, and rule matching tags are generated. Semantic embedding vector, numerical feature matrix and rule matching label are concatenated in sequence into a unified feature vector.

5. The AI-powered voucher verification method for SaaS financial software based on an expert database according to claim 1, characterized in that, The process of retrieving the audit rule set associated with the current voucher type from the expert database based on feature vectors, calculating the matching degree between the feature vectors and the audit rule set, and outputting the initial audit confidence level, specifically includes: The expert database stores a knowledge graph of historical audit cases annotated by domain experts; Based on the voucher type, the corresponding audit rule set is queried from the expert database. If the voucher type is a mixed or complex type, a comprehensive rule set is generated. Map the semantic embedding vector, numerical feature matrix, and rule matching label in the feature vector of the current credential to the conditions of the audit rule set one by one; For each rule, the matching degree between the feature vector and the rule conditions is calculated to generate the matching degree of a single rule. The overall matching degree is generated by weighted summation of the matching degrees of all rules. Based on the overall matching degree, and combining the matching results of hard and soft rules, the initial review confidence level is calculated.

6. The AI-powered voucher verification method for SaaS financial software based on an expert database as described in claim 1, characterized in that, The process of obtaining an initial review confidence level, and then accepting the application if the confidence level is higher than a preset threshold, and triggering multiple rounds of iterative verification if the confidence level is lower than the preset threshold, specifically includes: Based on business needs and historical audit data, set confidence thresholds and classify confidence levels based on the initial audit confidence levels obtained. For vouchers with a confidence level higher than the threshold, they are marked as automatically approved; for vouchers with a confidence level lower than the threshold, they are marked as requiring iterative verification. Data completion, in-depth feature analysis, and rule adjustment are performed through multiple rounds of verification until the confidence level exceeds the threshold.

7. The AI-powered voucher verification method for SaaS financial software based on an expert database as described in claim 1, characterized in that, The process of updating the expert database based on the review results and returning the structured review results to the client via the API gateway specifically includes: Based on the review results with high confidence levels, review result data is collected, including confidence level, pass / fail status, and iterative verification records; Add or modify data in the expert database based on the audit results data; For each voucher, a structured audit result is generated, including voucher ID, audit status, confidence level, matching rules, and iterative verification records. The audit result is then pushed to the client via the API gateway.

Images