GRC-AI agent based on multi-modal large model and implementation method

Abstract

This invention relates to a GRC-AI intelligent agent based on a multimodal large model and its implementation method, belonging to the field of artificial intelligence technology. Addressing the problems of existing GRC systems, such as difficulties in unstructured data parsing, reliance on manual verification for credential tampering detection, and long rule update cycles, this invention includes: a multimodal data parsing module for semantic understanding and structured extraction of unstructured data such as contracts, credentials, and charts; a risk assessment module for rule comparison and anomaly identification based on a knowledge base; a disposal execution module for automatically triggering business process actions; and a feedback evolution module for iteratively updating the disposal results to the knowledge base and rule base. This invention achieves intelligent parsing of unstructured data through a multimodal large model, credential tampering detection through dual verification of FakeShield and OCR, and zero-code iteration of the rule base through a feedback evolution mechanism. This invention improves the processing efficiency and accuracy of GRC business and can be applied to intelligent management of the entire process of governance, risk, and compliance in industries such as manufacturing.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a GRC-AI intelligent agent system and method based on a multimodal large model, for intelligent management and control in enterprise governance, risk and compliance scenarios, belonging to the field of artificial intelligence and enterprise risk control technology. Background Technology

[0002] The existing GRC digital system has the following technical defects:

[0003] 1. Unstructured data cannot be automatically parsed: Unstructured data such as scanned contracts, expense vouchers, and business charts rely on manual visual inspection and judgment. Taking contract review as an example, a single contract takes an average of 45 minutes, and the audit coverage is generally less than 30%, resulting in a large number of hidden risks going undetected.

[0004] 2. Lack of voucher tampering detection technology: Existing systems cannot automatically identify image-level tampering (such as scanned invoices modified with Photoshop), relying on manual comparison of each voucher. Actual tests show that the accuracy rate of manual detection of tampered vouchers is less than 60%.

[0005] 3. Rule updates rely on manual coding: When compliance rules are adjusted, the existing system requires technical personnel to manually write SQL or code to update the rules, with a response cycle of up to 2-4 weeks.

[0006] 4. Inability to link multi-source data in real time: Data from the financial system, OA system, and contract management system are fragmented, making real-time cross-system correlation analysis impossible. Risk identification has a 3-5 day lag window.

[0007] 5. Lack of self-evolution capability: Expert experience cannot be transformed into system capabilities, making it difficult to continuously optimize the GRC system.

[0008] In existing technologies, most GRC-related digital systems are single-function systems, independent and fragmented, lacking coverage of the entire GRC process, and failing to achieve deep integration of consulting methodologies and AI technology, thus exhibiting the following shortcomings:

[0009] 1. It cannot achieve intelligent parsing of unstructured data (vouchers, contracts, charts, etc.), and has a weak ability to identify hidden risks;

[0010] 2. Rule generation and updates rely on technology development, which has a long cycle and cannot quickly respond to business changes;

[0011] 3. Lacking the ability for autonomous evolution, it is difficult to accumulate expert experience and achieve the transfer of organizational intelligence;

[0012] 4. A complete closed loop of "perception-identification-early warning-response-evolution" has not been formed.

[0013] Therefore, there is an urgent need for a GRC-AI intelligent agent that can integrate GRC consulting methodology with AI technology to achieve full-process autonomous intelligence, evolution, and practicality, thereby addressing the aforementioned shortcomings of existing technologies. Summary of the Invention

[0014] (a) Purpose of the invention

[0015] The purpose of this invention is to provide a GRC-AI intelligent agent and its implementation method based on a multimodal large model, so as to solve the technical problems in the prior art that unstructured data cannot be automatically parsed, credential tampering cannot be automatically detected, and rule updates rely on manual coding.

[0016] (II) Technical Solution

[0017] To achieve the above objectives, this invention adopts a GRC-AI intelligent agent technology solution based on a multimodal large model.

[0018] 1. Scheme Design Concept

[0019] This solution uses the GRC consulting methodology of top consulting firms as its top-level business framework, deeply integrating multimodal large models and industry domain knowledge to construct a GRC-AI intelligent agent that can autonomously evolve and cover "governance-risk-compliance". Its core lies in transforming the experience of top experts into an executable and iterative intelligent closed-loop process, thereby achieving a restructuring of the technical architecture.

[0020] 1) Comprehensive Perception and Understanding: Deep analysis of multimodal large models, intelligent parsing of unstructured data such as vouchers, contracts, and charts, and understanding of business semantics and GRC (Governance, Risk, and Compliance) context;

[0021] 2) Intelligent identification and analysis: Identify potential risks and compliance deviations, capture anomalies based on domain knowledge base and real-time data, output analysis process and results, and provide a basis for decision-making;

[0022] 3) Autonomous decision-making and handling: Intelligent agents work together to trigger early warnings, link up with BPM processes to execute standardized handling (such as interception), provide strategy suggestions for complex scenarios, and assist human judgment;

[0023] 4) Continuous feedback and evolution: Accumulate data, optimize rules and models, transform responses into knowledge bases and training data, evolve autonomously, continuously empower, and form a GRC closed loop.

[0024] 2. Key technical difficulties solved by this invention

[0025] To address the aforementioned technical deficiencies, this invention focuses on solving the following key technical challenges:

[0026] Challenge 1: Unified representation and alignment of multimodal data

[0027] In existing technologies, images (scanned vouchers), text (contract terms), and structured data (exported from financial systems) are processed using different workflows, making cross-modal correlation analysis impossible. This invention constructs a unified multimodal embedding space, mapping image features (extracting tampered region features via FakeShield), text semantics (embedding via a large model), and structured fields (such as reimbursement amounts and dates) to the same vector space, enabling cross-modal joint retrieval and reasoning.

[0028] Challenge 2: The "Dual Authentication" Mechanism for Token Tampering Detection

[0029] Traditional solutions rely solely on pixel-level tampering detection (such as JPEG artifact analysis, which has a high false positive rate) or solely on OCR logical verification (such as field format checking, which has a high false negative rate). This invention innovatively proposes a dual verification architecture of "pixel-level detection + semantic logic verification":

[0030] 1) First layer: Pixel-level tampering location is achieved through the FakeShield multimodal framework, and the tampered area and tampering type are output;

[0031] 2) Second step: After extracting data through OCR, perform cross-field logical consistency verification (such as the time sequence of "invoice date" and "reimbursement date", and the consistency of "invoice amount" and "reimbursement amount").

[0032] Based on the combined assessment of both methods, the accuracy of tamper detection was increased from approximately 80% with a single method to 92.5% (test data).

[0033] Challenge 3: Zero-code iteration mechanism for the rule base

[0034] Traditional rule updates require technical personnel intervention, which is time-consuming and costly. This invention, through a feedback evolution module, automatically transforms expert-confirmed decisions into structured rules. Specifically:

[0035] 1) When an expert confirms that a reimbursement is for "special business entertainment", the system automatically extracts the characteristics of the scenario (department, amount range, and keywords of the reason).

[0036] 2) Combine features into executable rule expressions (e.g., "IF Department = Sales Department AND Amount > 500 AND Reason includes 'Customer Entertainment' THEN Advance Declaration Required");

[0037] 3) Rules are automatically added to the database, and the rule base can be iterated with zero code through the optimization of prompt word templates.

[0038] Tests show that the rule update cycle has been shortened from 2-4 weeks to hours.

[0039] 2. Implementation Method of the Solution

[0040] like Figure 1 As shown, the overall architecture of the GRC-AI intelligent agent of this invention includes a multimodal data parsing layer, a risk assessment layer, an intelligent handling layer, and a feedback evolution layer. Figure 2 As shown, the method flow of the present invention includes four steps: data perception, risk identification, automatic handling, and feedback evolution.

[0041] This invention employs a multi-agent collaborative agent matrix system to form a fully intelligent closed loop for the GRC process, encompassing "data perception—risk identification—automatic handling—feedback evolution".

[0042] 1) Financial Intelligent Agent: Focuses on voucher analysis, expense reimbursement verification, and voucher forgery detection.

[0043] Based on RAG (Retrieval Enhanced Generation) and large models, it is possible to comprehensively verify corporate expense reimbursement data, including whether vouchers have been tampered with.

[0044] a) Analyze reimbursement data using a large-scale model and output key points for verification;

[0045] b) Based on the key points, retrieve the vector database and output the verification items corresponding to the reimbursement data;

[0046] c) Input the reimbursement data and verification items into the large model. After analysis, the large model will provide the risk analysis process and results.

[0047] If there are supporting documents, the system will first call a dedicated large model to determine whether the images have been tampered with at the pixel level. Then, it will use OCR to recognize all the data in the supporting documents. The large model will then perform logical relationship verification on the data recognized by OCR and determine whether it has been tampered with again.

[0048] d) The intelligent agent drives the corresponding response process;

[0049] e) The audit specialist completes the closed-loop process;

[0050] f) The data generated by the closed loop is continuously fed back to RAG for knowledge optimization.

[0051] 2) Legal / Contract Intelligent Agent: Responsible for contract review, compliance clause comparison, and legal risk alerts.

[0052] a) Parse PDF electronic contracts using a large model and output key contract terms, including basic information, business terms, financial terms, and legal terms;

[0053] b) Based on the agent-specific prompts and the parsed contract content, the large model analysis provides the analysis process and risk recommendations;

[0054] c) Estimate potential gains and losses based on the terms and conditions.

[0055] 3) Supply Chain Intelligent Agent: Responsible for comprehensive monitoring and real-time interception of suppliers and distributors.

[0056] a) Obtain publicly available information about companies through APIs, including business registration information;

[0057] b) Analyze enterprise information using a large model and generate a report on enterprise relationships;

[0058] c) Continue to use the large model to further analyze the comprehensive risks of the supply chain based on the correlation, including analysis of bid rigging and other issues;

[0059] d) Standardize supply chain risk analysis into API interfaces and provide them to external systems.

[0060] 4) Chart / Data Intelligence: Generate business charts through Natural Language Question Answering (NL2SQL).

[0061] a) User input natural language query

[0062] b) After the system receives the input question, it calls the large model to analyze the user's intent and converts it into a database operation description;

[0063] c) The system uses the converted database description and database tables to infer and identify the tables and table structures related to the problem by calling the large model.

[0064] d) The system uses the comprehensive database operation description, combined with relevant tables and table structures, to call the large model to construct SQL statements;

[0065] e) The system executes and validates the SQL syntax;

[0066] f) The system calls the large model to generate visualization charts based on the SQL execution results;

[0067] g) Finally, the system returns the SQL execution results to the user in tabular form.

[0068] 5) Knowledge Base:

[0069] Stored procedures:

[0070] a) Upload files (txt / pdf / word);

[0071] b) Read file content: use PyPDF2 to parse PDF files and python-docx to parse Word files;

[0072] c) Use LangChain's RecursiveCharacterTextSplitter to perform semantic segmentation on the text content, breaking long texts into text chunks of controllable size and complete semantics.

[0073] d) Use the vector embedding model "bge-base-zh-v1.5" to convert the segmented data into high-dimensional vectors;

[0074] e) Store the vectorized data and metadata in the vector database ChromaDB.

[0075] Search process:

[0076] a) Specify the data set ID to be retrieved, the similarity index, the maximum number of results returned, and the question to be retrieved;

[0077] b) Use the vector embedding model "bge-base-zh-v1.5" to convert the query into a high-dimensional vector;

[0078] c) Retrieve relevant content from the ChromaDB vector database based on the vectorized question.

[0079] 6) Integration tools: Provides a unified rule base, model library, and dialogue support.

[0080] a) Rule base: Structure, standardize, and make enforceable governance, risk control, and compliance requirements to form a unified rule engine that supports automated review, verification, and compliance determination.

[0081] b) Model Library: A collection of AI prediction and risk measurement models built based on business data and risk scenarios, used for risk identification, anomaly detection, and trend prediction to enhance the intelligence and foresight of GRC.

[0082] c) Dialogue:

[0083] i) The user enters content in the input box and clicks send to trigger a request;

[0084] ii) Instantiate an EventSource object and send a long-lived connection request to the backend;

[0085] iii) Bind the onmessage event to capture data fragments pushed by the server in real time;

[0086] iv) Parse the data fragments and append them to Vue's ref or reactive state;

[0087] v) Vue detects data changes and automatically updates the DOM to achieve a streaming display effect;

[0088] vi) Once data reception is complete, call close() to disconnect and reset the loading state.

[0089] 3. Solution Technology Selection

[0090] 1) Fine-tuning of multimodal large model

[0091] The multimodal large model fine-tuning strategy is based on a general large model (such as Qwen2.5-14B-Instruct) and combined with GRC domain knowledge for fine-tuning, thereby improving the accuracy of GRC business semantic understanding and risk assessment.

[0092] 2) Image tampering detection

[0093] The large-scale model FakeShield is used for identification. FakeShield is a multimodal framework for interpretable image forgery detection and localization. By combining the visual and language understanding capabilities of a multimodal large language model (MLLM), it can not only determine the authenticity of images and locate tampered areas, but also explain the basis for the judgment in natural language.

[0094] 3) Knowledge Base (RAG)

[0095] Built on LangChain and ChromaDB. The core idea is to first load, segment, and quantize external documents into ChromaDB, and then, when a user asks a question, retrieve relevant content and provide it as context to the large language model, thereby generating more accurate and timely answers.

[0096] 4) Natural Language to Structured Query Language (NL2SQL)

[0097] The NL2SQL session system, built on LangChain and LangGraph, is an intelligent data query solution that converts natural language into SQL queries and supports multi-turn dialogue interaction.

[0098] 5) OCR

[0099] Supports Qwen-VL and PaddleOCR recognition.

[0100] (III) Beneficial Effects

[0101] Compared with traditional solutions, the present invention has the following advantages:

[0102] 1. Achieve intelligent closed-loop throughout the GRC process: Construct a complete closed loop of "perception-identification-early warning-response-evolution", break the fragmentation dilemma of the traditional GRC system, and realize the reconstruction of the technical architecture from passive compliance to proactive immunity;

[0103] 2. Improve GRC efficiency and accuracy: Intelligent analysis of unstructured data is achieved through multimodal large models, increasing the risk identification rate to over 80%, the audit coverage rate from less than 30% to 100%, and the response speed from several days / weeks to real-time / minutes, completely eliminating blind spots in control.

[0104] 3. Reduce reliance on human resources and costs: Transform expert experience into replicable and iterative system capabilities, avoid experience gaps caused by staff turnover, and free up human resources to invest in high-value risk analysis and decision support work;

[0105] 4. Possesses self-evolution capability: Through the feedback evolution layer, the knowledge base, rule base, and model are continuously iterated, enabling rapid response to business changes and industry policy adjustments, and rule updates can be completed without a large investment in technology research and development;

[0106] 5. High adaptability and scalability: Adopting a modular architecture, it can flexibly expand specialized intelligent agents according to the enterprise's size and business needs, adapt to the GRC management and control needs of multiple fields such as manufacturing, and at the same time support seamless integration with the enterprise's existing business systems, reducing implementation costs. Attached Figure Description

[0107] Figure 1 This is a diagram illustrating the overall architecture of the GRC-AI intelligent agent of this invention.

[0108] Figure 2 This is a flowchart of the GRC-AI intelligent agent implementation method of the present invention;

[0109] Figure 3 This is a flowchart illustrating the operation of the financial intelligent agent in Example 1 of this invention.

[0110] Figure 4 This is a flowchart illustrating the operation of the contract smart agent in Example 2 of this invention. Detailed Implementation

[0111] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This demonstrates the overall architecture of the invention. Figure 2 The method flow of the present invention is shown. Figure 3 and Figure 4 The operational processes of the financial intelligent agent and the contract intelligent agent are shown respectively.

[0112] Implementation Example 1: Application of GRC-AI Intelligent Agent in Financial Reimbursement Risk Control in Manufacturing (see...) Figure 3 )

[0113] like Figure 3 As shown, the operation flow of the financial intelligence agent in this embodiment is as follows:

[0114] 1. Knowledge Base Construction: Collect financial reimbursement management systems, expense standards, historical cases of fraudulent reimbursement, tax compliance rules, etc. in the manufacturing industry, and build a dedicated RAG knowledge base for GRC by combining internal enterprise systems;

[0115] 2. Prompt word optimization: Design prompt word templates based on the characteristics of industry expense control. First, define the roles of the large model and give the task requirements of the roles. Then, give the behavioral requirements for the large model to perform. Next, give the precautions in the analysis and how to define the analysis level. Finally, give strict requirements for the data format of the output report to facilitate subsequent processing.

[0116] 3. Deployment of Intelligent Agent Matrix: Deploy financial intelligent agents and chart / data intelligent agents. The financial intelligent agent can connect to the enterprise OA system to obtain expense reimbursement data; the chart / data intelligent agent is used to generate expense reimbursement risk control statistical reports.

[0117] 4. Intelligent operation throughout the entire process:

[0118] 1) Retrieve relevant reimbursement standards from the RAG knowledge base based on the reimbursement form information, and extract the reimbursement standards (such as meal allowance: RMB 100 / day in first-tier cities and RMB 60 / day in second-tier and below cities) into structured rules;

[0119] 2) To identify whether the voucher is forged (if so), first call the dedicated image recognition model to determine whether the electronic voucher has been modified, then call the OCR model to recognize all the data of the electronic voucher, including key data such as the date of payment, the amount of payment, the consumption scenario, and the payee. Then, analyze the logical relationship between the data obtained by the large model to determine whether it has been modified again.

[0120] 3) The risk assessment module calls the reimbursement standards and rules library in the RAG knowledge base, compares the reimbursement data, and identifies the risk of violations (such as a single business entertainment meal cost of 320 yuan, which exceeds the control standard of 200 yuan per person and is judged as high risk).

[0121] 4) Output risk warning and assessment reports, clarifying the reasons for violations, risk levels, etc., and provide handling suggestions (supplementary materials explaining the substance of the business are required).

[0122] 5) Following the company's risk registration response process, the intelligent agent automatically initiates the OA process to complete the closed loop of risk events.

[0123] 5. Autonomous Evolution and Iteration: The feedback evolution layer collects feedback on risk handling (such as confirming that the expense is for special business entertainment, and the rules need to be adjusted). The feedback analysis module extracts this information, and the iteration update module adds the rule "special business entertainment expenses need to be declared and approved in advance" to the rule base and optimizes the risk warning word template to improve the recognition accuracy of special scenarios.

[0124] Implementation Example 2: Application of GRC-AI Intelligent Agent in Contract Review in Manufacturing (see...) Figure 4 )

[0125] like Figure 4 As shown, the operation flow of the contract intelligence agent in this embodiment is as follows:

[0126] 1. Knowledge Base Construction: Collect model contract texts for the manufacturing industry, laws and regulations, industry compliance requirements, historical contract dispute cases, etc., to build a dedicated RAG knowledge base for GRC, and enter core rules for contract review (such as payment nodes, liquidated damages mechanisms, intellectual property clauses, etc.).

[0127] 2. Prompt word optimization: Design prompt word templates based on the characteristics of the manufacturing industry. First, define the role of the large model, then give the behavioral requirements for the large model, then give the precautions in the analysis and how to define the analysis level, and finally give strict requirements for the data format of the output report to facilitate subsequent processing;

[0128] 3. Deployment of Intelligent Agent Matrix: Deploy contract intelligent agents and chart intelligent agents to complete the core work of contract review and visualize the analysis results. Contract intelligent agents can directly connect to the enterprise contract management system to realize automatic capture and intelligent parsing of contract documents;

[0129] 4. Intelligent operation throughout the entire process:

[0130] 1) Users upload unformatted contract files or retrieve them directly from the enterprise contract management system. The contract intelligence agent calls the unstructured data parsing module to extract key information such as basic information, financial terms, business terms, and legal terms from the contract.

[0131] 2) The risk assessment module combines compliance rules in the RAG knowledge base to identify risk points in contracts (such as ambiguous payment terms and unreasonable penalty clauses), while quantifying potential benefits (such as calculating the amount that can be claimed based on the penalty clauses).

[0132] 3) Output a contract review report, including risk classification, benefit quantification results, and specific clause modification suggestions, to assist legal personnel and business personnel in contract negotiation or execution;

[0133] 4) Following the enterprise's risk registration response process, the intelligent agent automatically initiates the OA process to complete the closed loop of risk events.

[0134] 5. Autonomous Evolution and Iteration: Collect suggestions from legal professionals for revising review reports, update contract review rules and prompt word templates in the knowledge base, optimize the model's ability to analyze complex contract clauses, and improve the efficiency and accuracy of contract review.

[0135] Specific implementation of model fine-tuning:

[0136] In this embodiment, the fine-tuning strategy for the multimodal large model is as follows:

[0137] The Qwen2.5-14B-Instruct model was used as the base model, and fine-tuned based on 100,000 GRC domain-specific labeled data. The fine-tuning employed LoRA (Low-Rank Adaptation) technology to reduce memory usage and improve training efficiency.

[0138] The fine-tuning parameters are set as follows:

[0139] • Learning rate: 2e-4

[0140] • Number of training epochs: 5

[0141] LoRA rank: 8

[0142] • LoRA scaling parameter (alpha): 16

[0143] After the above fine-tuning, the model's accuracy on semantic understanding tasks in the GRC domain improved by approximately 15%.

[0144] Specific implementation of the RAG system:

[0145] In this embodiment, the knowledge base is constructed and retrieved using the following configuration:

[0146] Text segmentation: RecursiveCharacterTextSplitter is used, with a block size of 512 characters and a block overlap of 64 characters to ensure the semantic integrity of the text blocks.

[0147] Vector embedding: The bge-base-zh-v1.5 model is used to convert text blocks into 768-dimensional vectors, which are stored in the ChromaDB vector database.

[0148] Vector retrieval: Converts user questions into vectors of the same dimension, uses a cosine similarity algorithm for matching, and returns the top 10 results with the highest similarity.

[0149] Specific implementation of OCR dual verification:

[0150] In this embodiment, the OCR recognition of the voucher image adopts a dual verification mechanism: Qwen-VL is used as the main recognition engine, and Paddle OCR is used as the backup recognition engine. When the confidence level of the main engine is lower than a set threshold, the backup engine is automatically called for secondary recognition, and the final recognition data is output by combining the results of the two engines, thereby improving the recognition accuracy of complex vouchers.

[0151] Effect verification:

[0152] To verify the technical effectiveness of this invention, the applicant conducted a simulated operation test in a test environment based on anonymized historical business data from a manufacturing company. The test dataset included: 12,030 financial reimbursement documents, 510 contract documents, and 1,240 scanned vouchers.

[0153] The test results are as follows:

[0154] Test Project Invention Solution Traditional approach (comparison) Improvement effect Expense approval accuracy 92.5% With manual sampling covering 30% of the sample, the recognition rate is less than 20%. Accuracy increased by more than 4 times Contract review time 6 minutes / serving 50 minutes / serving Efficiency increased by more than 8 times Accuracy of credential tampering detection 90% The accuracy rate of manual visual inspection is less than 60%. Accuracy improved by more than 50%

[0155] The test results above show that the present invention can significantly improve the processing efficiency and accuracy of GRC services, and achieve the expected technical effect.

Claims

1. A GRC-AI intelligent agent system based on a multimodal large model, characterized in that, include: The multimodal data parsing module is used for semantic understanding and structured extraction of unstructured data; The risk assessment module is used to perform rule comparison and anomaly identification on the extracted data based on the knowledge base. The processing and execution module is used to automatically trigger the process processing of the business system based on the analysis results; The feedback evolution module is used to transform the disposal results into knowledge base update data.

2. The system according to claim 1, characterized in that, The multimodal data parsing module includes: The image tampering detection unit is used to perform pixel-level forgery recognition of voucher images based on a multimodal large model; The OCR recognition unit is used to extract structured data from images; The data logic verification unit is used to perform cross-field logical consistency verification on the OCR recognition results.

3. The system according to claim 2, characterized in that, The image tampering detection unit uses the FakeShield multimodal framework and outputs tampering region location and tampering type explanation text.

4. The system according to claim 2, characterized in that, The data logic verification unit is configured as follows: Extract multiple fields from the OCR recognition results and verify the consistency of time sequence, amount, and business logic among the fields.

5. A GRC-AI intelligent implementation method based on a multimodal large model, characterized in that, include: Analyze unstructured data using a multimodal large model to extract structured features; Risk events are identified by rule matching based on the extracted features from the GRC knowledge base. The business system will automatically trigger the handling process based on the risk level. Collect and process the results, and iteratively update the knowledge base and rule base.

6. The method according to claim 5, characterized in that, The parsing of unstructured business data includes: Perform pixel-level tamper detection on the voucher image; Perform OCR recognition on the voucher image to extract structured data; Perform cross-field logical consistency verification on OCR recognition results.

7. The method according to claim 5, characterized in that, The iterative updates to the knowledge base and rule base include: Obtain expert confirmation of the handling results; Extract scene features from confirmation information; Combine scene features into structured rule expressions; Add the regular expressions to the rule base and optimize the associated prompt word templates.

8. The method according to claim 5, characterized in that, Also includes: The NL2SQL engine converts natural language queries into SQL statements. Execute SQL statements to generate business charts and statistical data.

9. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method of any one of claims 5 to 8.

Images