An intelligent operation and maintenance question and answer system and method, an electronic device, and a storage medium

Abstract

This invention discloses an intelligent operation and maintenance question-answering system, method, electronic device, and storage medium. The invention extracts slot information through structured question rewriting and denoising, terminology standardization, and tag-based clue classification to achieve accurate judgment. It relies on a phased state machine intelligent agent workflow to process intent recognition and invocation in stages using a global state JSON file. An external rule engine is introduced to parse natural language rules into structured rule files for hard-coded judgment. Combined with a layered knowledge base for iterative management and storage of the entire process data, this invention solves problems in existing systems such as inaccurate classification due to text fragmentation, attention dilution from long rules in large models, high accuracy in numerical judgment, poor output stability, random decay of rule adjustment effects, and inability to retain results and stage blockages in linear workflows. This improves the accuracy of question classification, the precision of rule judgment, output stability, and the traceability of responses, while reducing the cost of rule iteration maintenance and human resource operation.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and intelligent operation and maintenance technology, and in particular to an intelligent operation and maintenance question-and-answer system, method, electronic device and storage medium. Background Technology

[0002] Insurance business spans the entire business chain, from underwriting and approval to payment and claims settlement. Back-office staff generate numerous system error reports, business process questions, and data anomaly-related maintenance consultation requests during daily operations. Traditional maintenance models rely entirely on manual work order processing, requiring multiple steps from problem acceptance and assignment to final response, resulting in low overall response efficiency and high labor costs.

[0003] In recent years, large language models and intelligent agent technologies have been gradually applied to the insurance operations and maintenance question-answering field. Currently, the mainstream solutions in the industry can be divided into three categories. The first is the RAG knowledge base-based question-answering solution, which vectorizes business operation manuals and standardized FAQ documents and stores them in a vector database. Relevant knowledge is retrieved through semantic retrieval, and then a large model generates a response. The second is the MCP model context protocol tool call solution, which builds standardized tool API nodes on the system side. The large model directly reads the flat natural language business rules, autonomously decides to call the corresponding interface to query business data, and generates a response based on the rule text. The third is the ReAct linear intelligent agent workflow solution, which uses a single-round linear process combining reasoning and action to complete the problem processing in one go. It does not break down tasks into stages and cannot retain intermediate state data.

[0004] The above solutions still have significant shortcomings in actual business scenarios. In MCP solutions, a single interface typically comes with a large number of long, flat rules. The length of a single rule text is considerable, making it prone to attention dilution when the large model reads it in its entirety. Even with the correct tool selection, logical errors frequently occur in the final response, and business personnel cannot independently adjust the rule logic; continuous optimization is insufficient to address the root cause of the problem. Insurance operations involve a large amount of standardized field judgment logic, such as code length verification, field specified character matching, parameter value comparison, package number prefix matching, and business status flag identification. Large models have high judgment errors regarding precise numerical values ​​and fixed strings, making it difficult to meet the verification requirements of highly standardized financial businesses such as payment settlement and fee calculation.

[0005] Meanwhile, existing systems exhibit inconsistent tool selection and varying response wording when faced with the same question under different model versions, temperature parameters, or context lengths. This results in poor output stability and fails to meet the regulatory requirements of insurance finance businesses for traceable, auditable, and consistent results. Minor adjustments to business rules, such as changes to fee settlement conditions or joint guarantee judgment logic, can cause the large model's understanding of the rules to change randomly. Previously stable calls and responses may experience unpredictable accuracy declines, and existing solutions lack quantitative regression verification methods to identify the effect decay caused by iterations in advance. Traditional linear ReAct workflows are black-box, one-time executions without task phase breakdown. When user information is missing, parameters are incorrect, or business status conflicts occur, the entire process must be rerun. It cannot retain preliminary results such as intent recognition and order number extraction, nor does it support interrupting the task at any stage to clarify and supplement information to the user. Overall, it lacks traceability and auditability.

[0006] In addition, questions from business personnel often include background information on departmental transfers, descriptions of operational processes, subjective guesses, and redundant polite phrases, resulting in severe text fragmentation. The existing system relies solely on semantic matching of the original text for classification, which is insufficient in terms of classification accuracy in ambiguous scenarios. This makes it easy to make incorrect calls to business interfaces and makes it difficult to accurately locate the user's real problem.

[0007] Therefore, there is an urgent need to provide a technical solution to address the above problems. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides an intelligent operation and maintenance question-and-answer system, method, electronic device, and storage medium.

[0009] Firstly, the present invention provides an intelligent operation and maintenance question-and-answer system, the technical solution of which is as follows: The structured question rewriting module is used to perform noise reduction and terminology standardization on the original natural language questions input by the user, extract business slot information, and organize the business slot information into standardized structured questions; The tag clue classification module is used to perform tag matching processing on the standardized structured questions, generate tag clues, and input the tag clues into the classification model to obtain the business category classification results; The phased state machine intelligent agent workflow module is used to determine user intent based on the business category classification results using a global state JSON file as a carrier, extract business parameters from the standardized structured questions, verify the legality of the business parameters, generate a tool call sequence based on the user intent, the business parameters and the verification results, send the tool call sequence to the external rule engine MCP tool call module to receive the returned rule judgment results, and generate a response based on the rule judgment results. The external rule engine MCP tool call module is used to parse natural language rule text into structured rule files and store them in a versioned manner during the offline stage. During the online stage, it receives the tool call sequence, calls the business interface according to the tool call sequence and obtains business data through the business parameters, loads the structured rules that match the business data from the structured rule file, performs hard-coded logic judgment on the business data, and outputs the rule judgment result. The hierarchical knowledge base iteration management module is used to store knowledge data, provide reference knowledge for the structured question rewriting module, and store the answers and full-process data.

[0010] The beneficial effects of the intelligent operation and maintenance question-and-answer system of the present invention are as follows: The system of this invention extracts slot information by rewriting and denoising structured questions and standardizing terminology, and achieves accurate judgment by combining tag clue classification. It relies on a phased state machine intelligent agent workflow to process intent recognition and invocation in stages using a global state JSON file. An external rule engine is introduced to parse natural language rules into structured rule files for hard-coded judgment. Combined with a hierarchical knowledge base, it iteratively manages and stores the entire process data. This solves the problems of inaccurate classification due to text fragmentation, dilution of attention for large models and long rules, high error in accurate numerical judgment, poor output stability, random decay of rule adjustment effects, and inability to retain results and stage blockages in linear workflows in existing systems. It improves the accuracy of question classification, the precision of rule judgment, the stability of output and the traceability of responses, and reduces the cost of rule iteration and maintenance and human operation.

[0011] Based on the above solution, the intelligent operation and maintenance question and answer system of the present invention can be further improved as follows.

[0012] In one alternative approach, the structured problem rewriting module is specifically used for: Retrieve business process documents, internal enterprise terminology dictionaries, and operation manuals from the hierarchical knowledge base iteration management module as reference knowledge; The original natural language problem is denoised by using specialized rewritten prompts to remove emotional descriptions, irrelevant background descriptions, and subjective speculations about cause and effect. Replace the non-standard expressions in the original natural language problem with the company's internal standard terminology; Extract the problem category, core anomaly type, employee ID, business order number, business system to which it belongs, and user-predicted cause of the fault as the business slot information; The business slot information is organized into a standardized structured problem in a standardized JSON format.

[0013] The advantages of adopting the above optional methods are as follows: by further introducing business process documents and enterprise terminology dictionaries, the original questions are denoised and the terminology is standardized, fragmented text is converted into structured JSON format, key business slot information is extracted, ambiguity is reduced, and the accuracy of subsequent classification and processing is improved.

[0014] In one alternative approach, the tag clue classification module is specifically used for: Pre-store business category tags, proprietary business term tags, and operational behavior tags, with each tag configured with a matching weight score; The script performs three matching logics on the standardized structured problem: exact matching, synonym matching, and word order-independent matching, generating tag clues that include tag name, text matching similarity, business category, frequency of occurrence, and business importance level.

[0015] The beneficial effects of adopting the above optional methods are as follows: by further performing precise matching, synonym matching and word order-independent matching on structured questions through multi-dimensional tag matching and weight configuration, tag clues containing similarity and importance levels are generated, enriching the classification basis and improving the accuracy of business category determination.

[0016] In one alternative approach, the tag clue classification module is specifically used for: The label clues and historical classification samples are input into the classification model to obtain the business category classification result; wherein, the matching weight scores of each label on which the classification model outputs the business category classification result are iteratively adjusted through a configured automatic machine learning dynamic learning interface.

[0017] The advantages of adopting the above optional approach are: further inputting label clues and historical classification samples into the model, and dynamically adjusting the label weights through an automatic machine learning interface, so that the model can continuously adapt to business changes, reduce manual intervention, and maintain classification accuracy.

[0018] In one alternative approach, the phased state machine agent workflow module is specifically used for: The question-and-answer process is divided into four stages: intent recognition, slot filling, data verification, and processing planning. The global state JSON file stores the stage identifiers, output results of each stage, historical dialogue records, and a list of missing parameters for the four stages. The next action is automatically determined based on the current execution stage and the output results of each stage. The next action is one of the following: calling business tools, initiating clarification to the user, or the process is terminated normally.

[0019] The advantages of adopting the above optional approach are: further dividing the question and answer process into four stages and recording the results of each stage in a global state JSON file, realizing staged execution and state retention, supporting blocking and initiating clarification at any stage, avoiding repeated running of the entire process, and improving the traceability of processing.

[0020] In one alternative approach, the phased state machine agent workflow module is specifically used for: During the intent recognition stage, the business category to which the user's question belongs is determined based on the business category classification result. When the confidence difference of multiple categories is less than or equal to a preset threshold, a candidate category is generated and a clarification question is sent to the user. During the slot filling stage, the policy number, employee number, and organization code are extracted from the standardized structured question as the business parameters. When the core business parameters are missing, the process is blocked and a supplementary information query is initiated to the user. During the data verification phase, the validity of the document number, the operator's business permissions, and the document business status are verified. If the verification fails, a standardized exception message is returned. The standardized tool call sequence, data query SQL statements, and business troubleshooting solutions are output during the processing planning phase.

[0021] The advantages of adopting the above-mentioned optional methods are: to further set up refined processing logic at each stage, including low confidence clarification, missing parameter blocking, permission and status verification and standardized output, to ensure that the data in each link is complete and compliant, reduce erroneous calls, and improve the overall processing quality.

[0022] In one alternative approach, the external rules engine MCP tool invocation module is specifically used for: In the offline phase, the natural language rule text is structured and parsed using a high-specification large language model. The parsing results are split into an execution strategy part and a business rule part. The business rule part is structured and stored in the rule base. Conflict detection and regression testing are performed on the natural language rule text. During the online phase, grouping rules matching the current business scenario are loaded, so that the large language model loads only the rule text within the current business group at a time. The script consumes the business data returned by the business interface, executes the hard-coded logic judgment of the structured rules, and outputs the matching rule description as the rule judgment result.

[0023] The advantages of adopting the above optional approach are as follows: in the offline stage, natural language rules are further structured and conflict detection and regression testing are performed. In the online stage, rules are loaded according to business groups and hard-coded judgments are executed by scripts, avoiding attention dilution and numerical judgment errors in large models, and improving the stability of rule execution and judgment accuracy.

[0024] Secondly, this invention provides an intelligent operation and maintenance question-and-answer method, employing the intelligent operation and maintenance question-and-answer system provided by this invention. The technical solution of this method is as follows: The user-input natural language question is subjected to noise reduction and terminology standardization, business slot information is extracted, and the business slot information is organized into a standardized structured question. The standardized structured problem is processed by label matching to generate label clues, and the label clues are input into the classification model to obtain the business category classification results; Using a global state JSON file as a carrier, the user intent is determined based on the business category classification results. Business parameters are extracted from the standardized structured questions, and the legality of the business parameters is verified. A tool call sequence is generated based on the user intent, the business parameters, and the verification results. The tool call sequence is sent to the external rule engine MCP tool call module to receive the returned rule judgment results, and a response is generated based on the rule judgment results. In the offline phase, the natural language rule text is parsed into a structured rule file and stored in a versioned manner. In the online phase, the tool call sequence is received, the business interface is called according to the tool call sequence, and business data is obtained through the business parameters. The structured rules that match the business data are loaded from the structured rule file, hard-coded logic is performed on the business data, and the rule judgment result is output. It provides reference knowledge for the structured question rewriting module and stores the responses and full-process data.

[0025] The beneficial effects of the intelligent operation and maintenance question-and-answer method of the present invention are as follows: The method of this invention extracts slot information by rewriting and denoising structured questions and standardizing terminology, and achieves accurate judgment by combining tag clue classification. It relies on a phased state machine intelligent agent workflow to process intent recognition and invocation in stages using a global state JSON file. An external rule engine is introduced to parse natural language rules into structured rule files for hard-coded judgment. Combined with a hierarchical knowledge base, iterative management and storage of full-process data solves the problems of inaccurate classification due to text fragmentation, dilution of attention for large models and long rules, high error in accurate numerical judgment, poor output stability, random decay of rule adjustment effects, and inability to retain results and stage blockages in linear workflows in existing systems. It improves the accuracy of question classification, the precision of rule judgment, the stability of output and the traceability of responses, and reduces the cost of rule iteration and maintenance and human operation.

[0026] Thirdly, the technical solution of an electronic device according to the present invention is as follows: It includes a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps of the intelligent operation and maintenance question-and-answer method of the present invention.

[0027] Fourthly, the technical solution of a computer-readable storage medium provided by the present invention is as follows: The computer-readable storage medium stores instructions that, when read, cause the computer-readable storage medium to perform the steps of the intelligent operation and maintenance question-and-answer method of the present invention.

[0028] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other objects, features and advantages of the present invention more obvious and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0029] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a schematic diagram of an embodiment of the intelligent operation and maintenance question-and-answer system of the present invention; Figure 2 A schematic diagram of the two-layer architecture of the external rule engine MCP tool calling module for offline parsing and online execution; Figure 3 This is a flowchart illustrating an embodiment of an intelligent operation and maintenance question-and-answer method according to the present invention; Figure 4 This is a schematic diagram of an embodiment of an electronic device according to the present invention. Detailed Implementation

[0030] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0031] Figure 1 A schematic diagram of an embodiment of an intelligent operation and maintenance question-and-answer system 100 provided by the present invention is shown. Figure 1 As shown, the intelligent operation and maintenance Q&A system 100 includes: The structured question rewriting module 101 is used to perform noise reduction and terminology standardization on the original natural language question input by the user, extract business slot information, and organize the business slot information into a standardized structured question.

[0032] Among them, raw natural language questions refer to: original question texts entered by users in everyday natural language without any structured processing, containing redundant descriptions, emotional expressions and non-standard terms; for example, a user enters "Hello, I am employee A from the finance department. This morning I was processing the handling fee for institution 021003, package number BX20260512003, but the system prompted that the handling fee settlement failed. I suspect that the joint guarantee ratio is incorrect. Please help me check the reason. I adjusted the handling fee rules yesterday and I don't know if there is a problem. I need to settle the account urgently, thank you."

[0033] Among them, business slot information refers to the key business elements extracted from the user's original natural language question, including the question category, core anomaly type, employee number, business order number, business system to which it belongs, and the user's inferred cause of the failure; for example, from the above user input, the organization code is 021003, the package number is BX20260512003, the business type is payment and fee accounting, the anomaly type is fee settlement failure, and the user's inferred cause is that the joint guarantee ratio configuration is abnormal and the fee rule has become invalid after modification.

[0034] The standardized structured question refers to a structured question text organized according to a preset format after the original natural language question has undergone noise reduction and terminology standardization. It includes the cleaned core question description, business type, anomaly type, slot information, and noise information annotation. For example, after processing the above user input, the output standardized structured question is "Institution 021003, Packaging No. BX20260512003, Fee Settlement Failed, Query the Reason for Failure, Check the Joint Guarantee Ratio and Fee Rule Configuration", the business type is "Payment and Fee Calculation", and the anomaly type is "Fee Settlement Failed".

[0035] The tag clue classification module 102 is used to perform tag matching processing on the standardized structured problem, generate tag clues, and input the tag clues into the classification model to obtain the business category classification result.

[0036] Among them, the tag clues refer to the tag set generated after multi-mode tag matching processing of standardized structured questions. Each tag carries a matching similarity and weight score, which is used to assist the classification model in determining the business category. For example, after tag matching of the above standardized structured questions, tag clues are generated containing the tag "handling fee" (matching similarity 0.98, weight 10, frequency of occurrence 2, importance level is core), the tag "joint guarantee" (matching similarity 0.95, weight 8, frequency of occurrence 1, importance level is important), the tag "package number" (matching similarity 1.0, weight 7, frequency of occurrence 1, importance level is important), and the tag "settlement failure" (matching similarity 0.96, weight 9, frequency of occurrence 1, importance level is core), with a total score of 34.

[0037] The classification model refers to a machine learning model used to automatically determine the business category to which a user's problem belongs. Its inputs are tag clues and historical classification samples, and its output is the business category classification result. For example, after inputting the aforementioned tag clues and historical abnormal fee samples into the classification model, the model outputs the classification result "Abnormal Fee Calculation" with a confidence level of 0.978. The business category classification result refers to the determination result of the business category to which the user's problem belongs, along with its confidence level. For example, the classification model determines that the aforementioned user problem belongs to the "Abnormal Fee Calculation" category with a confidence level of 0.978.

[0038] The phased state machine intelligent agent workflow module 103 is used to determine the user intent based on the business category classification result using a global state JSON file as a carrier, extract business parameters from the standardized structured questions, verify the legality of the business parameters, generate a tool call sequence based on the user intent, the business parameters and the verification result, send the tool call sequence to the external rule engine MCP tool call module 104 to receive the returned rule judgment result, and generate a response based on the rule judgment result.

[0039] The global state JSON file refers to a unified data carrier stored in JSON format that runs through the entire workflow of the phased state machine intelligent agent. It includes the session identifier, the current execution stage, the stage history, the output results of each stage, the dialogue record, and the list of missing parameters. For example, a global state JSON file is generated for this question-and-answer session. The file records the session identifier as S202606180923001, the current stage as init, and the data of each stage is updated sequentially as the process progresses.

[0040] Here, user intent refers to the business purpose corresponding to the user question determined in the intent identification phase; for example, the intent of the above user question is determined as "root cause query for handling fee settlement failure". Business parameters refer to the key parameters extracted from the standardized structured question for subsequent business interface calls, including policy number, employee number, organization code, etc.; for example, organization code 021003, package number BX20260512003, employee number and affiliated finance department are extracted from the above standardized structured question as business parameters.

[0041] The tool call sequence refers to the standardized instruction sequence generated during the processing planning phase for calling external business interfaces, including the interface name and its input parameters. For example, generating a tool call sequence includes a payment package document query interface, with the input parameters being the package number BX20260512003 and the organization code 021003. The rule judgment result refers to the judgment conclusion output by the external rule engine after performing structured rule judgment on the business data, including the triggered rule identifier, rule description, judgment pass / fail status, judgment details, and processing suggestions. For example, the rule engine outputs a judgment result with the triggered rule identifier R2-08, the rule description "The joint guarantee ratio must not be lower than 0.7; below the threshold triggers fee settlement interception," the judgment result being "failed," and the judgment details being "The joint guarantee ratio of the document is 0.65, which is less than the standard threshold of 0.7, and the settlement conditions are not met." The processing suggestions include adjusting the joint guarantee ratio to 0.7 or higher and re-initiating the fee settlement process.

[0042] The response refers to the final reply to the user generated based on the rule judgment results, which includes an explanation of the root cause of the failure and handling suggestions. For example, the generated response is: "Regarding the failure of fee settlement for institution 021003 and package number BX20260512003, the root cause of the failure is that the current document co-insurance ratio is 0.65, which is lower than the minimum threshold of 0.7 for the R2 group fee settlement rule, and the settlement is automatically blocked. The handling suggestion is to adjust the document co-insurance ratio to 0.7 or higher, and re-initiate the fee settlement after modification. If the problem persists after adjustment, a maintenance work order can be submitted to review the rule configuration."

[0043] The external rule engine MCP tool call module 104 is used to parse natural language rule text into structured rule files and store them in a versioned manner during the offline stage, receive the tool call sequence during the online stage, call the business interface according to the tool call sequence and obtain business data through the business parameters, load the structured rules that match the business data from the structured rule file, perform hard-coded logic judgment on the business data, and output the rule judgment result.

[0044] The offline phase refers to the stage within the non-real-time question-and-answer processing window where natural language rule text is pre-parsed, structurally converted, and versioned for storage. This phase does not participate in the online real-time question-and-answer process. For example, in the offline phase, business personnel submit newly added joint guarantee ratio determination rules. After structural parsing of the rule, a structured rule file is generated and stored in the rule base, with the version number marked as V2.1. A structured rule file is a rule file generated after the natural language rule text has been structurally parsed, which can be directly read and executed by the rule executor. It contains structured fields such as rule groups, rule identifiers, judgment conditions, and processing suggestions. For example, parsing the "handling fee settlement rule for printed but unpaid documents" into a structured rule file includes rule group R2, rule identifier R2-08, judgment condition "joint guarantee ratio greater than or equal to 0.7", and processing suggestion "trigger settlement interception when the ratio is lower than the threshold".

[0045] The online phase refers to the stage of receiving user questions in real time and completing question-and-answer processing. During this phase, pre-parsed structured rule files are invoked for rule determination, but no structured rule parsing is performed. For example, after a user submits a question, the tool's call sequence is received in real time during the online phase, and the business interface is invoked to obtain data and load the pre-parsed structured rule file for determination. The business interface refers to a standardized interface for data interaction with external business systems, used to obtain the real-time status and data of business documents; for example, calling the payment / receipt package document query interface to obtain the business data corresponding to package number BX20260512003. Business data refers to the real-time status and field information of business documents obtained from external business systems by calling business interfaces. For example, the payment and receipt package document query interface returns business data, which includes package number BX20260512003, organization code 021003, business status "printed but not paid", joint guarantee ratio of 0.65, standard ratio of 0.7, fee standard of 1200, calculated fee of 780, system error code FEE001, and error message "joint guarantee ratio is lower than the minimum settlement threshold of 0.7".

[0046] Structured rules refer to rule files that have been parsed offline and stored in a rule base, containing clear decision logic and triggering conditions. These files include rule groups, version numbers, trigger rule identifiers, rule descriptions, decision conditions, and processing suggestions. For example, a structured rule might include rule group R2, version number V2.1, rule identifier R2-08, rule description "The joint guarantee ratio must not be lower than 0.7; below this threshold, fee settlement interception will be triggered," decision condition "The joint guarantee ratio is greater than or equal to 0.7," and processing suggestion "Adjust the joint guarantee ratio to 0.7 or higher." Hard-coded logic decision-making refers to a method where script code directly executes logical operations such as numerical comparisons and string matching, rather than relying on a large language model for understanding or reasoning. For example, a script might read the joint guarantee ratio of 0.65 and the standard ratio of 0.7 from business data, perform numerical comparison logic, determine whether 0.65 is greater than or equal to 0.7, and output a decision result of "fail."

[0047] The hierarchical knowledge base iteration management module 105 is used to store knowledge data, provide reference knowledge for the structured question rewriting module 101, and store the answers and full-process data.

[0048] Among them, knowledge data refers to the collective term for various types of structured knowledge stored in the hierarchical knowledge base, including standard question and answer knowledge, process table knowledge, and tagged fallback documents; for example, the hierarchical knowledge base stores high-frequency FAQ question and answer pairs, operation process tables of the core business system of auto insurance, operation manuals with additional business tags, and policy documents.

[0049] The reference knowledge refers to the business process documents, internal terminology dictionaries, and operation manuals retrieved from the hierarchical knowledge base when rewriting structured questions. For example, the operation manual of the core business system for auto insurance, the internal terminology dictionary, and payment business process documents can be retrieved from the hierarchical knowledge base as a reference for rewriting user questions.

[0050] The full-process data refers to all intermediate state data and output results generated at each stage of the phased state machine intelligent agent workflow, from the user inputting the original natural language question to the generation of the final answer. For example, the session identifier of this question-and-answer session, the JSON data of each stage, the output results of the rule engine, and the final answer are uniformly stored in the hierarchical knowledge base.

[0051] The technical solution of this embodiment achieves accurate judgment by rewriting and denoising structured questions and extracting slot information through terminology standardization, combined with tag clue classification. It relies on a phased state machine intelligent agent workflow to process intent recognition and invocation in stages using a global state JSON file. An external rule engine is introduced to parse natural language rules into structured rule files for hard-coded judgment. Combined with a hierarchical knowledge base, iterative management and storage of full-process data solves the problems of inaccurate classification due to text fragmentation, dilution of attention for large models and long rules, high error in accurate numerical judgment, poor output stability, random decay of rule adjustment effects, and inability to retain results and stage blockages in linear workflows in existing systems. It improves the accuracy of question classification, the precision of rule judgment, the stability of output and the traceability of responses, and reduces the cost of rule iteration and maintenance and human operation.

[0052] In an alternative embodiment, the structured problem rewriting module 101 is specifically used for: The business process documents, internal terminology dictionaries, and operation manuals are retrieved from the hierarchical knowledge base iteration management module 105 as reference knowledge.

[0053] Among them, business process documents refer to standardized documents describing the specific operational procedures, system functional modules, and business rules of insurance business, used as background reference knowledge when rewriting structured questions; for example, retrieving the payment and collection business process document records the complete operational steps, preconditions, and common error messages for commission settlement. An internal terminology dictionary refers to a dictionary recording the company's standard internal terms and their correspondences, used to replace non-standard expressions in the user's original natural language questions with unified internal standard terms; for example, the internal terminology dictionary records the mapping relationship between "issuance system" and "auto insurance core business system". Operation manuals refer to guiding documents describing the specific operational steps and functional descriptions of the business system, used to assist in identifying the system operation scenarios involved in the user's questions when rewriting structured questions; for example, retrieving the auto insurance core business system operation manual includes specific step instructions for operations such as package order query and commission settlement.

[0054] The original natural language problem is denoised by using specialized rewritten prompts to remove emotional descriptions, irrelevant background descriptions, and subjective speculations about causality.

[0055] Among them, the special rewriting prompt words refer to prompt word templates specifically designed for structured question rewriting tasks, which are used to guide the large language model to complete text denoising and terminology standardization processing according to preset rules; for example, calling the special rewriting prompt words to guide the large language model to perform operations such as deleting emotional descriptions, standardizing terminology, and extracting business slots for the user's original question.

[0056] Emotional descriptions refer to redundant content in the user's original natural language questions that express subjective emotions, exclamations, or polite tones, without carrying substantive business information. For example, "hello," "thank you," and "urgently settling the bill" are identified as emotional descriptions in user questions and will be deleted during noise reduction. Irrelevant background descriptions refer to redundant content in the user's original natural language questions that is unrelated to core business anomalies, such as personal introductions, departmental transfer backgrounds, and operational process reviews. For example, "I am employee A from the finance department," "processed this morning," and "adjusted the handling fee rules yesterday, not sure if there was a problem" are identified as irrelevant background descriptions and will be deleted during noise reduction. Subjective causal speculation refers to non-deterministic inferences and subjective guesses based on personal experience in the user's original natural language questions, which are unverified. For example, "suspect the joint guarantee ratio is incorrect" and "not sure if there is a problem" are identified as subjective causal speculation and will be deleted during noise reduction.

[0057] Replace the non-standard expressions in the original natural language problem with the company's internal standard terminology.

[0058] Non-standard expressions refer to informal, non-standard business terminology abbreviations or colloquialisms used in users' original natural language questions, which are inconsistent with the company's internal standard terminology. For example, "issuance system" in a user's question is a non-standard expression; the company's internal standard terminology should be "auto insurance core business system." Internal standard terminology refers to standardized business terminology used uniformly in the company's internal business documents, system interfaces, and process specifications. For example, the company's internal standard terminology uniformly standardizes "issuance system" as "auto insurance core business system."

[0059] Extract the problem category, core anomaly type, employee number, business order number, business system to which it belongs, and user-predicted cause of failure as the business slot information.

[0060] The problem category refers to the major business area to which the user's original natural language question belongs, used to initially define the business scope of the problem; for example, the problem category extracted from the user's question is "Payment and Fee Calculation". The core anomaly type refers to the specific error or anomaly type encountered by the user during business operations, used to pinpoint the specific manifestation of the problem; for example, the core anomaly type extracted from the user's question is "Fee Settlement Failure". The employee ID refers to the unique identification number assigned to each employee by the company, used to identify the specific operator in system operation logs and business documents; for example, the employee ID and their affiliated finance department are extracted from the user's question. The business order number refers to the number used in the business system to uniquely identify a business transaction or document, including policy number, package number, payment order number, etc.; for example, the business order number extracted from the user's question is package number BX20260512003. The affiliated business system refers to the specific system name involved in the business operation processing the user's question; for example, the affiliated business system extracted from the user's question is "Auto Insurance Core Business System". User-speculated causes of failure refer to possible reasons for anomalies proposed by users based on subjective judgment in the original natural language questions, which are speculative contents to be verified; for example, user-speculated causes of failure extracted from user questions are "abnormal configuration of joint guarantee ratio" and "invalid after modification of handling fee rules".

[0061] The business slot information is organized into a standardized structured problem in a standardized JSON format.

[0062] Among the above optional methods, further noise reduction and terminology standardization of the original questions can be achieved by introducing business process documents and enterprise terminology dictionaries, transforming fragmented text into structured JSON format, extracting key business slot information, reducing ambiguity, and improving the accuracy of subsequent classification and processing.

[0063] In an alternative embodiment, the tag clue classification module 102 is specifically used for: Business category tags, proprietary business term tags, and operational behavior tags are pre-stored, and each tag is configured with a matching weight score.

[0064] The tags are categorized into three main types: **Business Category Tags:** These tags mark the primary business category to which a user's question belongs, covering areas such as premiums, commissions, claims, and management fees. For example, the tag library pre-stores business category tags like "commissions," "premiums," and "claims." **Professional Business Terminology Tags:** These tags mark professional terms specific to the insurance business, such as "package number," "pre-issued invoice," and "joint insurance." For example, the tag library pre-stores proprietary business term tags like "joint insurance," "package number," and "pre-issued invoice." **Operational Behavior Tags:** These tags mark the specific operational actions performed by the user in the business system, such as "print," "cancel," and "actual receipt." For example, the tag library pre-stores operational behavior tags like "settlement," "print," "cancel," and "actual receipt."

[0065] Among them, the matching weight score refers to the numerical value pre-configured for each business tag to represent the importance of the tag in the classification judgment; for example, the tag library configures a weight of 10 for the "handling fee" tag, a weight of 8 for the "joint guarantee" tag, and a weight of 7 for the "package number" tag.

[0066] The script performs three matching logics on the standardized structured problem: exact matching, synonym matching, and word order-independent matching, generating tag clues that include tag name, text matching similarity, business category, frequency of occurrence, and business importance level.

[0067] Exact matching refers to a matching method where, when performing label matching on standardized structured questions, the original string of the label name is compared word by word to determine if a match exists. For example, performing exact matching on "handling fee" in a standardized structured question will find the string "handling fee" in the text, resulting in a successful match. Synonym matching refers to a matching method where words in the text are compared with synonyms of the label names to determine if a match exists. For example, performing synonym matching on a standardized structured question will identify "commission" as a synonym of "handling fee" and result in a successful match. Word order-independent matching refers to a matching method where the order of words in the text is disregarded when performing label matching on standardized structured questions; only the presence of all required words in the text is considered. For example, performing word order-independent matching on a standardized structured question will determine a successful match if both "handling fee" and "settlement failure" appear in the text, regardless of their order.

[0068] Here, "tag name" refers to the name identifier of a predefined business tag in the tag library; for example, the tag names in the tag lead are "handling fee", "joint guarantee", "packaging number", and "settlement failure". "Text matching similarity" refers to the quantified value of the similarity between the matched text and the tag name when performing tag matching on a standardized structured question; for example, matching "handling fee" in "handling fee settlement failure" with the tag "handling fee" using exact matching yields a similarity of 0.98. "Business category" refers to the business category level to which each tag in the tag lead belongs; for example, the "handling fee" tag in the tag lead belongs to the business category "collection and payment - handling fee accounting", and the "joint guarantee" tag belongs to the business category "reinsurance - joint guarantee". "Frequency of occurrence" refers to the number of times the same tag name is matched in the same standardized structured question; for example, the "handling fee" tag appears twice in the tag lead, and the "joint guarantee" tag appears once. Business importance level refers to the level of a tag based on the coreness of the business attribute it represents, including three levels: core, important, and general. For example, in the tag leads, the importance level of the tags "handling fee" and "settlement failure" is core, while the importance level of the tags "joint guarantee" and "package number" is important.

[0069] Among the above optional methods, further multi-dimensional tag matching and weight configuration are used to perform precise matching, synonym matching and word order-independent matching on structured questions, generating tag clues that include similarity and importance levels, enriching the classification basis and improving the accuracy of business category determination.

[0070] In an alternative embodiment, the tag clue classification module 102 is specifically used for: The label clues and historical classification samples are input into the classification model to obtain the business category classification result; wherein, the matching weight scores of each label on which the classification model outputs the business category classification result are iteratively adjusted through a configured automatic machine learning dynamic learning interface.

[0071] Among them, the automatic machine learning dynamic learning interface refers to: an interface reserved for automatically adjusting the matching weight scores of each label online, and automatically optimizing the weight configuration by continuously accumulating classification feedback data; for example, through the automatic machine learning dynamic learning interface, based on the classification effect feedback of the classification model on historical samples, the matching weight score of the "Lian Gong Bao" label is automatically adjusted from 8 to 9.

[0072] In the above-mentioned optional methods, the label clues and historical classification samples are further input into the model, and the label weights are dynamically and iteratively adjusted through an automatic machine learning interface, so that the model can continuously adapt to business changes, reduce manual intervention, and maintain classification accuracy.

[0073] In an alternative embodiment, the phased state machine agent workflow module 103 is specifically used for: The question-and-answer process is divided into four stages: intent recognition, slot filling, data verification, and processing planning. The global state JSON file stores the stage identifiers, output results of each stage, historical dialogue records, and a list of missing parameters for the four stages.

[0074] The intent recognition stage refers to the first stage of the phased state machine intelligent agent workflow. This stage determines the user intent based on the business category classification results. For example, upon entering the intent recognition stage, the user intent is determined to be "root cause query for fee settlement failure" based on the business category classification results, with a confidence level of 0.978 and no candidate classifications requiring clarification. The slot filling stage refers to the second stage of the phased state machine intelligent agent workflow. This stage extracts key business parameters from standardized structured questions and stores the extraction results in a global state JSON file. For example, upon entering the slot filling stage, the organization code 021003, packaging number BX20260512003, finance department, and employee number are extracted from the standardized structured questions and stored in the global state JSON file, with no missing parameters. The data verification phase refers to the third stage of the phased state machine intelligent agent workflow. This stage verifies the legality of the extracted business parameters, including the legality of the document number, the operator's business permissions, and the document's business status. For example, upon entering the data verification phase, it verifies that the package number BX20260512003 is valid, the operator has the authority to query fee settlement, and the document's business status is "printed but not paid," all of which pass the verification. The processing planning phase refers to the fourth stage of the phased state machine intelligent agent workflow. This stage generates a tool call sequence, a data query SQL statement, and a business investigation and processing plan based on user intent, business parameters, and verification results. For example, upon entering the processing planning phase, it generates a tool call sequence containing the call to the payment package document query interface, an SQL statement with the query statement "SELECT * FROM t_pay_fee_package WHERE package_no='BX20260512003'", and a business investigation and processing plan for joint guarantee ratio verification.

[0075] The stage identifier refers to the unique marker in the global state JSON file used to identify the current stage of the question-and-answer process. For example, the stage identifiers recorded in the global state JSON file are init, intent_recognize, slot_extract, data_verify, and plan_tool. The output results of each stage refer to the intermediate output data generated after the execution of each stage of the phased state machine agent workflow, including intent recognition results, slot filling results, data validation results, and processing planning results. For example, the global state JSON file records the output results of each stage, including the main intent "Fee Settlement Failure Root Cause Query" and confidence level of 0.978 output by the intent recognition unit, the organization code 021003 and packaging number BX20260512003 output by the slot filling unit, the validation pass flag and document status "Printed but Not Paid" output by the data validation unit, and the tool call sequence and SQL statements output by the processing planning unit. Historical dialogue records refer to the complete interaction history between the user and the system, including the original question entered by the user each time and the response content returned each time; for example, the dialogue records in the global state JSON file contain the user's original question text and the generated final response content. Missing parameter list refers to the list of core business parameters identified during the slot filling phase that were not provided in the user's input but are required for subsequent processes; for example, if the packaging number is not extracted during the slot filling phase, the missing parameter list records "packaging number" as a missing parameter, pauses the process, and sends a supplementary query to the user.

[0076] The next action is automatically determined based on the current execution stage and the output results of each stage. The next action is one of the following: calling business tools, initiating clarification to the user, or the process is terminated normally.

[0077] The next action refers to the type of subsequent operation automatically determined by the phased state machine intelligent agent workflow based on the current execution stage and the output results of each stage. This includes three types: calling business tools, initiating clarification to the user, and normal process termination. For example, after the processing planning stage is completed, the next action is determined to be calling business tools, followed by calling the payment and receipt package document query interface.

[0078] In the above optional methods, the question-and-answer process is further divided into four stages and the results of each stage are recorded in a global state JSON file to achieve staged execution and state retention. It supports blocking and initiating clarification at any stage, avoids running the entire process repeatedly, and improves the traceability of processing.

[0079] In an alternative embodiment, the phased state machine agent workflow module 103 is specifically used for: During the intent recognition stage, the business category to which the user's question belongs is determined based on the business category classification results. When the confidence difference between multiple categories is less than or equal to a preset threshold, a candidate category is generated and a clarification question is sent to the user.

[0080] Among them, the business category refers to the primary business category to which the user question belongs, which is used to determine the user's intent during the intent identification stage; for example, the business category to which the above user question belongs is determined to be "payment and fee accounting".

[0081] The multi-class confidence difference refers to the numerical difference in confidence scores between the two classification results with the highest confidence scores output by the classification model. For example, if the classification model outputs "abnormal handling fee calculation" with a confidence score of 0.978 and "normal settlement query" with a confidence score of 0.021, the difference between the two is 0.957, which is greater than 0.1, and no clarification is required from the user. The preset threshold refers to the critical value used to determine whether to ask the user for clarification regarding the multi-class confidence difference. For example, if the preset threshold for the confidence difference is 0.1, a clarification question is triggered when the difference between the two categories with the highest confidence scores is less than or equal to 0.1.

[0082] Candidate categories refer to the two categories with the highest confidence levels selected when the difference in confidence levels among multiple categories is less than or equal to a preset threshold. For example, if the classification model outputs "Fee Settlement Failure Query" with a confidence level of 0.48 and "Fee Settlement Success Query" with a confidence level of 0.47, and the difference is 0.01, which is less than the threshold of 0.1, these two categories are presented to the user as candidate categories for selection. Clarification questions refer to questions posed to the user to confirm the business category to which the problem belongs when the difference in confidence levels among multiple categories is less than or equal to a preset threshold. For example, sending the user a message such as, "Is your problem a fee settlement failure query or a fee settlement success query? Please confirm."

[0083] During the slot filling stage, the policy number, employee number, and organization code are extracted from the standardized structured question as the business parameters. When the core business parameters are missing, the process is blocked and a supplementary information query is initiated to the user.

[0084] Among them, information supplement inquiry refers to: when core business parameters are missing, a question is sent to the user to obtain the specific value of the missing parameter; for example, when the user does not provide the package number, the user is sent "Please provide the package number for which you need to inquire about the reason for the failure of the handling fee settlement".

[0085] During the data verification phase, the validity of the document number, the operator's business permissions, and the document's business status are verified. If the verification fails, a standardized exception message is returned.

[0086] The document number validity refers to the result of validating the document number in the business parameters by performing format and existence checks. For example, verifying that package number BX20260512003 conforms to the package number coding standard and exists in the system results in a valid verification. Operator business permissions refer to the result of verifying whether the operator initiating the query has the authority to query the corresponding business data. For example, verifying that the operator corresponding to employee number 021003 has the authority to query the handling fee settlement data results in an authorized verification. Document business status refers to the current state of the business document in the business process, such as "printed but not paid," "paid," or "voided." For example, during the data verification phase, if the document business status of package number BX20260512003 is confirmed as "printed but not paid," rule group R2 is locked accordingly.

[0087] Standardized exception prompts refer to the preset format of exception prompts returned when data verification fails, clearly indicating the specific reason for the verification failure; for example, when the package number BX20260512003 does not exist, the standardized exception prompt "Package number BX20260512003 does not exist in the system, please confirm whether the package number is correct" is returned.

[0088] The standardized tool call sequence, data query SQL statements, and business troubleshooting solutions are output during the processing planning phase.

[0089] The standardized tool call sequence refers to the formatted API call instructions output during the planning phase, including the interface name and input parameter list. For example, the output standardized tool call sequence might have an interface name of "Payment and Receipt Package Document Query Interface" and input parameters of package number BX20260512003 and organization code 021003. The data query SQL statement refers to the structured query statement generated during the planning phase for retrieving relevant business data from the business database. The business investigation and handling plan refers to the complete investigation and handling plan for user issues output during the planning phase, including the tools to be called, the rule groups to be executed, and the problem handling steps. For example, the output business investigation and handling plan might include calling the payment and receipt package document query interface to obtain data, loading R2 grouping rules to determine the joint guarantee ratio, and outputting handling suggestions based on the determination results.

[0090] Among the above-mentioned optional methods, further refined processing logic can be set at each stage, including low confidence clarification, missing parameter blocking, permission and status verification, and standardized output, to ensure that the data in each link is complete and compliant, reduce erroneous calls, and improve the overall processing quality.

[0091] In an alternative embodiment, the external rule engine MCP tool invocation module 104 is specifically used for: In the offline phase, the natural language rule text is structured and parsed using a high-specification large language model. The parsing results are split into an execution strategy part and a business rule part. The business rule part is structured and stored in the rule base. Conflict detection and regression testing are then performed on the natural language rule text.

[0092] Among them, the high-specification large language model refers to a language model with a large number of parameters that can perform structured parsing and logical reasoning of natural language rules. In the offline stage, it is used to convert business rules described in natural language into structured rule files. For example, in the offline stage, the natural language rule text "Settlement rule for handling fees of printed but unpaid documents: Trigger settlement interception when the joint guarantee ratio is lower than 0.7" is input into the high-specification large language model. After the model performs structured parsing of the rule text, it outputs a structured rule file, which includes the rule grouping as R2, the rule identifier as R2-08, the judgment condition as "the joint guarantee ratio is greater than or equal to 0.7" and the processing suggestion as "trigger settlement interception when the ratio is lower than the threshold". After the parsing is completed, the structured rule file is stored in the rule base and the version number is marked as V2.1.

[0093] The execution strategy section refers to the strategic content describing the timing and conditions for rule invocation, obtained by the external rule engine after parsing the natural language rule text. This part is judged by the large language model during the online phase. For example, after parsing the natural language rule, the execution strategy section records "trigger the R2 grouping rule when the document's business status is 'printed but not paid'". The business rule section refers to the business rule content describing the specific judgment conditions and judgment logic, obtained by the external rule engine after parsing the natural language rule text. This part is structured and stored in the rule base and hard-coded for execution by the rule executor. For example, after parsing the natural language rule, the business rule section records "the joint guarantee ratio must be greater than or equal to 0.7" and "trigger fee settlement interception when the joint guarantee ratio is lower than 0.7".

[0094] The rule base refers to a storage system used to centrally store and manage structured rule files, supporting version management, conflict detection, and regression testing of rule files; for example, the rule files of the R2 group are stored in the rule base and marked with the version number V2.1.

[0095] Conflict detection and regression testing refer to the quality assurance operations performed by the rule interpreter when performing structured parsing of natural language rule text. Conflict detection is used to identify logical contradictions between new rules and existing rules. Conflict detection includes condition exclusion detection and condition overlap detection. When the judgment conditions of a new rule and the judgment conditions of an existing rule have a logical mutual exclusion relationship or inclusion relationship, a conflict alarm is triggered. Regression testing is used to verify that the overall effect has not degraded after the rule change. Regression testing compares the judgment results before and after the rule change by loading a historical question and answer sample set. When the difference rate of the judgment results exceeds a preset threshold, a regression alarm is triggered. For example, when parsing a new joint protection ratio rule, conflict detection is performed to confirm that the new rule has no logical conflict with the existing R2 grouping rule. Regression testing is performed to confirm that the rule update does not affect the judgment results of other business scenarios.

[0096] During the online phase, grouping rules matching the current business scenario are loaded, so that the large language model loads only the rule text within the current business group at a time. The script consumes the business data returned by the business interface, executes the hard-coded logic judgment of the structured rules, and outputs the matching rule description as the rule judgment result.

[0097] The rule description refers to the text information in the rule judgment result output by the rule executor that describes the content of the triggered rule and the basis for the judgment; for example, the rule executor outputs the rule description "The joint guarantee ratio shall not be lower than 0.7, and if it is lower than the threshold, the handling fee settlement will be intercepted".

[0098] In the above-mentioned optional methods, natural language rules are further structured and conflict detection and regression testing are performed in the offline stage. In the online stage, rules are loaded by business group and hard-coded judgment is performed by script, which avoids attention dilution and numerical judgment error of large models and improves the stability of rule execution and judgment accuracy.

[0099] In another embodiment of the intelligent operation and maintenance question-and-answer system of the present invention, the system includes: a structured question rewriting module 101, a tag clue classification module 102, a phased state machine intelligent agent workflow module 103, an external rule engine MCP tool invocation module 104, and a hierarchical knowledge base iteration management module 105. The complete business execution chain is as follows: The user's original question is input into the structured question rewriting module 101. The structured question rewriting module 101 outputs a standardized structured question to the tag clue classification module 102. The tag clue classification module 102 generates tag clues and inputs them into the classification model. It outputs the business category classification results to the phased state machine intelligent agent workflow module 103. The phased state machine intelligent agent workflow module 103 uses a global state JSON file as a carrier to sequentially execute the intent recognition stage, slot filling stage, data verification stage, and processing planning stage. In the processing planning stage, it generates a tool call sequence and sends it to the external rule engine MCP tool call module 104. The external rule engine MCP tool call module 104 calls the business interface to obtain business data and performs structured rule judgment. It outputs the rule judgment results and returns them to the phased state machine intelligent agent workflow module 103. The phased state machine intelligent agent workflow module 103 generates a response based on the rule judgment results and outputs it to the user. At the same time, it stores the entire question and answer result in the database and drives the hierarchical knowledge base iteration management module 105 to automatically iterate.

[0100] After obtaining the original natural language question input by the user, the structured question rewriting module 101 retrieves business process documents, internal terminology dictionaries, and operation manuals from the hierarchical knowledge base iteration management module 105 using RAG retrieval as reference knowledge. It then uses dedicated rewriting prompts to perform text denoising on the original natural language question, removing emotional descriptions, irrelevant background descriptions, and subjective causal speculations. Non-standard expressions in the original natural language question are replaced with standard internal terminology. From the denoised and standardized question, the question category, core anomaly type, employee number, business order number, affiliated business system, and user-specified cause of the fault are extracted as business slot information. This business slot information is organized into a standardized JSON format structured question and output. By introducing business process documents and an enterprise terminology dictionary, the structured question rewriting module 101 denoises and standardizes the terminology of the original question, converting fragmented text into structured JSON format, extracting key business slot information, reducing ambiguity, and improving the accuracy of subsequent classification and processing.

[0101] After obtaining the standardized structured questions output by the structured question rewriting module 101, the tag clue classification module 102 reads three pre-stored tag systems: business category tags, proprietary business term tags, and operational behavior tags. Each tag is configured with a matching weight score. Through a script, it performs three matching logics on the standardized structured questions: exact matching, synonym matching, and word order-independent matching. This generates tag clues containing tag names, text matching similarity, business category, frequency of occurrence, and business importance level. These tag clues, along with historical classification samples, are input into the classification model to comprehensively determine the business category to which the user's question belongs, and outputs the business category classification result. The matching weight scores of each tag used by the classification model when outputting the business category classification result are iteratively adjusted through a reserved automatic machine learning dynamic learning interface to continuously optimize classification accuracy. The tag clue classification module 102, through multi-dimensional tag matching and weight configuration, performs exact matching, synonym matching, and word order-independent matching on the structured questions, generating tag clues containing similarity and importance levels, enriching the classification basis and improving the accuracy of business category determination.

[0102] The phased state machine intelligent agent workflow module 103, after acquiring the business category classification results output by the tag clue classification module 102 and the standardized structured questions output by the structured question rewriting module 101, uses a global state JSON file as the data carrier for the entire process, permanently storing the intermediate processing results of each stage. The global state JSON file stores the session identifier, the current execution stage, stage history, output results of each stage, historical dialogue records, and a list of missing parameters. The output results of each stage are stored in different fields of the global state JSON file: the output results of the intent recognition stage are stored in the intent_result field, the output results of the slot filling stage are stored in the slot_cache field, the output results of the data validation stage are stored in the check_result field, and the output results of the processing planning stage are stored in the plan_result field. The phased state machine intelligent agent workflow module 103 divides the question-and-answer process into four independent sub-workflow units: the intent recognition stage, the slot filling stage, the data validation stage, and the processing planning stage. The phased state machine intelligent agent workflow module 103 breaks down the question-and-answer process into four phases and records the results of each phase in a global state JSON file, enabling phased execution and state retention. It supports blocking and initiating clarification at any phase, avoiding repeated running of the entire process and improving processing traceability.

[0103] The phased state machine intelligent agent workflow module 103's overall architecture consists of an MCP node paired with independent sub-workflow units. The MCP node acts as the phase management decision center, receiving user input and a global state JSON file. Based on the current execution phase and the output results of each phase, it automatically determines the next action. After each sub-workflow unit finishes processing, it invokes the state machine to perform state transitions and returns to the MCP node. The global state JSON file records the current phase identifier and the output results of each phase. The MCP node determines the next execution path by reading the global state JSON file. The state machine is responsible for phase advancement, parsing response information, organizing dialogue JSON, and determining whether to advance the phase, revert to the slot filling phase, or initiate interaction with the user based on missing parameter markers, validation results, and user confirmation results. Interaction methods include providing the user with two options, guiding questions for user input, and initiating supplementary information queries. The state machine supports process pause, phase rollback, and breakpoint resumption.

[0104] In the intent recognition phase, the phased state machine intelligent agent workflow module 103 determines the business category to which the user's question belongs based on the business category classification results. When the difference in confidence scores among multiple categories is greater than 0.1, the classification result with the highest confidence score is taken as the user's intent. When the difference in confidence scores among multiple categories is less than or equal to 0.1, the two classification results with the highest confidence scores are output as candidate classifications, and a clarification question is raised to the user to confirm the user's intent. After the user confirms, the user's intent is determined and recorded in the global state JSON file.

[0105] During the slot filling phase, the phased state machine intelligent agent workflow module 103 extracts the policy number, employee number, and organization code as business parameters from the standardized structured problem. When the core business parameters are missing, the missing parameter list is recorded in the global state JSON file, the process is blocked, and an information supplementation query is initiated to the user. After the user supplements the information, the process continues based on the extraction results.

[0106] During the data verification phase, the phased state machine intelligent agent workflow module 103 performs verification on the extracted business parameters, including document number validity verification, operator business permission verification, and document business status verification. If the verification fails, a standardized exception message is returned. If the verification passes, the verification result is recorded in the global status JSON file.

[0107] During the planning phase, the phased state machine intelligent agent workflow module 103 outputs standardized tool call sequences, data query SQL statements, and business troubleshooting solutions based on the user intent, business parameters, and verification results recorded in the global state JSON file.

[0108] The external rule engine MCP tool invocation module 104 adopts a two-layer architecture of offline rule parsing and online hard-coded rule engine, completely decoupling the business rule judgment logic from the large language model. The external rule engine MCP tool invocation module 104 decomposes the processing of the MCP node based on three principles: decomposing the processing stages of the large language model to focus on one task at a time; reducing the scale of text consumed by the large language model to concentrate its attention; and using hard-coded logic for judgments that can be made without relying on the large language model's understanding. Specifically, the decomposition method is as follows: the original MCP node's task of decision-making plus response is split into two parts; the MCP node only handles the tool invocation and rule adoption decisions, while response processing is handled by another large language model node; the interface return data is described in natural language state, and rules are hierarchically organized by state grouping, priority arrangement, and mutual exclusion descriptions; the rule matching logic is externalized, using code judgments where possible, and the judgment results are output to the large language model for processing.

[0109] like Figure 2As shown, the overall architecture of the external rule engine MCP tool invocation module 104 splits the original MCP and tool library into four stages: rule configuration, rule interpreter, rule structuring, and rule execution. The offline stage is the rule plug-in link, responsible for rule preparation and structuring. In this stage, business personnel write strategies, constraints, and tool invocation logic in natural language and centrally condense them into rule text. The rule interpreter uses a high-specification large language model to perform structured parsing of the natural language rule text, conduct conflict detection and coding planning, identify new and modified rules, form a structured intermediate representation, and solidify the structured intermediate representation into a rule.json format structured rule file and write it into the rule library for version management. The rule.json format structured rule file contains rule group identifier, rule version number, trigger rule identifier, rule description, judgment condition expression, and processing suggestion list. The online phase is the main execution chain, responsible for real-time question and answer processing. In this phase, the MCP entry point receives user input and performs intent recognition, determining whether rule-driven tool calls are needed and extracting key entities and parameters. The MCP tool call chain executes business interface calls according to the APIs pointed to in the MCP tool call and obtains the interface return data. The rule executor reads the rule.json file in the rule base, inputs the interface return data into the rule executor, consumes the interface return data using a script, executes the hard-coded logic of structured rules, and generates a rule description upon rule matching. This rule description is then input into the large language model for result fusion. After fusing the tool results, the large language model continues to call tools or generate readable explanations and final responses. In the online phase, rules are fixed in rule.json and executed only by the rule executor; the large language model is responsible for scheduling and interpretation. The rule-attached chain runs in the offline phase. Business personnel only maintain natural language rules; the rule interpreter is responsible for coding, and rule.json serves as the sole execution source. Rule writing and rule execution are decoupled, and rule.json becomes the stable rule foundation for the MCP node, preventing rule adjustments from causing invisible degradation to online behavior.

[0110] In the offline phase, when the external rule engine MCP tool invocation module 104 performs structured parsing of natural language rule text using a high-specification large language model, it splits the parsing result into an execution strategy part and a business rule part. The execution strategy part is determined by the large language model during the online phase, while the business rule part is structured and stored in the rule base. Conflict detection and regression testing are then performed on the natural language rule text. Conflict detection identifies logical contradictions between newly added rules and existing rules, while regression testing verifies that the overall effect has not degraded after rule changes. In the online phase, the external rule engine MCP tool invocation module 104 receives the tool invocation sequence output from the processing planning phase, invokes the business interface according to the sequence, obtains business data through business parameters, and loads grouped rules matching the current business scenario from the rule base. This ensures that the large language model loads only the rule text within the current business group at a time, effectively avoiding the attention dilution problem caused by reading all long rule texts at once. A script consumes the business data returned by the business interface, executes the hard-coded logic judgment of the business rule part, and outputs the matching rule description as the rule judgment result. Hard-coded logic checks directly execute numerical comparisons and string matching through script code, ensuring that all numerical calculations, string matching, and field logic checks are performed by the underlying code, completely avoiding errors in large language model logic checks. Business rules are fixed in static JSON files, unaffected by model versions or generated temperature parameters, eliminating rule execution drift at its source. Modifications to business rules only require offline re-parsing and generating of a new version of rule.json; the online Q&A process remains unaffected, preventing the issue of random accuracy drops after rule iterations.

[0111] The hierarchical knowledge base iteration management module 105 establishes a three-tiered hierarchical knowledge asset system, covering full-link question-and-answer support and automated iterative optimization. The first layer is a standard question-and-answer knowledge base, storing frequently asked standardized FAQs. When a user asks a question, this layer is searched first, and a successful match results in a quick and direct response, reducing the overhead of large language model calls. The second layer is a process table knowledge base, storing complete business operation process tables, including system menu paths, business prerequisites, common error messages, and standardized processing steps. All content is bound to standardized business tags. The third layer is a tagged fallback document library, where all operation manuals and business policy documents are tagged with process tags, system tags, and operation type tags. Documents are stored in slices according to business scenarios, and the sliced ​​text and metadata tags are used to generate vectors. A contextual embedding model is used for vectorization processing to optimize retrieval accuracy in scenarios with mixed long and short texts. The hierarchical knowledge base iteration management module 105 provides the structured question rewriting module 101 with business process documents, internal terminology dictionaries, and operation manuals required for RAG retrieval as reference knowledge. It also receives responses and full-process data generated by the phased state machine intelligent agent workflow module 103, storing these data in the hierarchical knowledge base. The hierarchical knowledge base iteration management module 105 automatically captures maintenance work orders through an online RPA program, incorporating problems that the system cannot automatically resolve into the knowledge base to be supplemented. After obtaining the annotation results of incorrect responses from maintenance personnel, it writes the content to the knowledge base with one click based on the annotation results, completing the content update. It also periodically supplements the intelligent agent workflow, rule base, and hierarchical knowledge base with failed samples through offline review of typical failure cases, continuously iterating and optimizing the overall system performance.

[0112] The technical solution of this embodiment will be described in detail below through a complete example.

[0113] The example business scenario is an anomaly query in the accounting of payment and collection fees. The user's original question is: "Hello, this is Xiao Li from the finance department. This morning, I was processing the fees for institution 021003, package number BX20260512003, but the system indicated that the fee settlement failed. I suspect it's due to an incorrect joint guarantee ratio. Could you please help me check the cause? I adjusted the fee rules yesterday, but I don't know if that caused the problem. I need to settle the accounts urgently, thank you." This original question contains a lot of redundant wording and vague expressions. The system needs to clean, classify, perform multi-stage verification, and use a rule engine to determine the accuracy of the response.

[0114] The complete end-to-end workflow is as follows: the user's original question is input into the structured question rewriting module 101, which outputs a standardized JSON-formatted structured question; the tag clue classification module 102 obtains the standardized structured question and outputs weighted tag JSON; the phased state machine intelligent agent workflow module 103 starts, sequentially executing four units: intent recognition stage, slot filling stage, data verification stage, and processing planning stage, while synchronously updating the global state JSON; the external rule engine MCP tool call module 104 schedules the payment and receipt document query interface to obtain business data; the external rule engine loads R2 grouping rules for hard-coded judgment; the large language model integrates the judgment results to generate a response; and the entire process of data is entered into the database to update the knowledge base.

[0115] The input to the structured question rewriting module 101 is the user's original natural language text, and the output is a standardized structured question JSON. The standardized structured question JSON contains the original question text, the cleaned question description, the business type, the exception type, the slot information, and noise information annotations. After cleaning, the issue is identified as "Institution 021003, Package Number BX20260512003, Fee Settlement Failed, Inquire about the Reason for Failure, Check Joint Insurance Ratio and Fee Rule Configuration"; the business type is "Payment and Collection - Fee Calculation"; the anomaly type is "Fee Settlement Failed"; the slot information includes the employee name as Xiao Li, the department as Finance Department, the institution code as 021003, the package number as BX20260512003, the business system as Auto Insurance Core Business System, the user's speculation that the cause of the failure is an abnormal joint insurance ratio configuration and the failure of the modified fee rules, and the request is to inquire about the root cause of the settlement failure; the noise information label includes polite greetings, urgent settlement requests, and a description of yesterday's operation background; the standard terminology mapping includes "Issuance System" corresponding to "Auto Insurance Core Business System".

[0116] The input to the tag clue classification module 102 is the cleaned text of the question and slot information in the standardized structured question JSON output by the structured question rewriting module 101. The output is a weighted tag clue JSON. The weighted tag clue JSON contains the source text, tag list, total weight score, final business category, and model confidence. The tag list includes the tag "handling fee", which is a business category tag with a similarity of 0.98, weight of 10, frequency of occurrence of 2, and importance level of core; the tag "joint guarantee", which is a proprietary business term tag with a similarity of 0.95, weight of 8, frequency of occurrence of 1, and importance level of important; the tag "packaging number", which is a proprietary business term tag with a similarity of 1.0, weight of 7, frequency of occurrence of 1, and importance level of important; and the tag "settlement failure", which is an operation exception tag with a similarity of 0.96, weight of 9, frequency of occurrence of 1, and importance level of core. The total weight score is 34, the final business category is "abnormal accounting of payment and collection fees", and the model confidence level is 0.978.

[0117] The global state JSON of the phased state machine serves as the state carrier throughout the process, continuously updated with each phase. The initial state is identified as `init`. After the intent recognition phase, the state is updated; the session identifier is S202606180923001, the current phase is `intent_recognize`, and the phase history includes `init` and `intent_recognize`. The intent result contains the intent graph "Fee settlement failure root cause query", with a confidence level of 0.978. Alternative intents are empty, no confirmation is required, the slot cache is empty, the validation result is empty, the planning result is empty, the missing parameter list is empty, and the status is "Running". After the slot filling phase, the state is updated; the current phase is `slot_extract`, and the phase history includes `init`, `intent_recognize`, and `slot_extract`. The slot cache contains the organization code 021003, packaging number BX20260512003, department (Finance Department), and employee name (Xiao Li). The missing parameter list is empty. After the data verification phase, the status is updated. The current phase is data_verify, and the phase history is init, intent_recognize, slot_extract, and data_verify. The verification result includes a valid package number, authorized personnel, and a document business status of "printed but not paid." After the processing planning phase, the status is updated. The current phase is plan_tool, and the phase history is init, intent_recognize, slot_extract, data_verify, and plan_tool. The planning result includes a tool call list, which contains a payment package document query interface. The input parameters are package number BX20260512003 and organization code 021003. The rule group is R2, and the query SQL statement is "SELECT * FROM t_pay_fee_package WHERE package_no='BX20260512003'". The status changes to "ready to call tool".

[0118] After the external rule engine MCP tool calls module 104 to call the payment and receipt package document query interface, the business API returns document data, which includes package number BX20260512003, organization code 021003, business status "printed but not paid", joint guarantee ratio 0.65, standard ratio 0.7, fee standard 1200, calculated fee 780, system error code FEE001 and error message "joint guarantee ratio is lower than the minimum settlement threshold of 0.7". After the external rule engine loads the hard-coded R2 group rule for judgment, it outputs the rule judgment result, which includes the rule group R2-Printed Unpaid Document Handling Fee Rule, rule version V2.1, trigger rule identifier R2-08, rule description "The joint guarantee ratio shall not be lower than 0.7, and if it is lower than the threshold, the handling fee settlement will be intercepted". The judgment result is that it fails, and the judgment details are "The joint guarantee ratio of the document is 0.65, which is less than the standard threshold of 0.7, and the settlement conditions are not met". The handling suggestions include adjusting the joint guarantee ratio to 0.7 or above and re-initiating the handling fee settlement process. The original data contains a joint guarantee ratio of 0.65 and a standard ratio of 0.7.

[0119] After integrating the rule engine's judgment results, the large language model outputs a standardized JSON response and a user-readable text response. The standardized JSON response includes the session identifier S202606180923001, the user's question "Institution 021003 Package No. BX20260512003 Fee Settlement Failure Query Reason", the root cause of the fault "The joint guarantee ratio of this packaged document is 0.65, lower than the minimum settlement threshold of 0.7 in the business rules, the system intercepts the fee settlement", and three handling suggestions: adjust the joint guarantee ratio of this document to 0.7 or higher, re-initiate the fee settlement operation after the ratio is modified, and if the error persists after modification, submit an operation and maintenance work order to review the rule configuration. The audit information includes the rule version V2.1, rule group R2, and query time June 18, 2026, 09:23:15. The user-readable response text is as follows: "The investigation results regarding the failed fee settlement for institution 021003 and package number BX20260512003 are as follows: The root cause of the failure is that the current document co-insurance ratio is 0.65, which is lower than the minimum threshold of 0.7 for the R2 group fee settlement rule, and the system automatically intercepts the settlement; the recommended solution is to adjust the document co-insurance ratio to 0.7 or higher, and then re-initiate the fee settlement. If the problem persists after adjustment, a maintenance work order can be submitted to review the rule configuration; this judgment is based on the fee rules for printed but unpaid documents in version V2.1, and the results are auditable and traceable." In this example, the state machine uniquely identifies the session using session identifier S202606180923001. The complete state transition chain is as follows: The initial state `init` receives the user's original query and starts the structured rewriting module; it then transitions to the `intent_recognize` stage, reading tag clues to complete the business category determination, with sufficient confidence and no need for clarification; next, it transitions to the `slot_extract` stage, extracting core parameters such as the organization and packaging number, with no missing parameters and no process interruption; finally, it transitions to the `data_verify` stage, verifying the document's legality and the operator's permissions. All verifications pass, the document's business status is set to "printed but not paid," and rule group R is locked. 2. The process transitions to the `plan_tool` stage, generating corresponding API call instructions for querying payment and receipt documents, and the status changes to `ready_call_tool`. It then transitions to the `call_mcp` stage, calling the business interface to retrieve the original document data. Next, it transitions to the `rule_execute` stage, where the R2 grouping rules are hard-coded to complete numerical comparison and judgment, outputting standardized root causes of failures. Finally, it transitions to the `generate_reply` stage, where the large language model integrates the rule judgment results and outputs a readable reply. The process ends in the `finish` state, with all JSON, logs, and Q&A results from all stages being stored in the database and automatically pushed to the hierarchical knowledge base for iteration. If the package number is missing, the slot extraction stage may remain in the `slot_extract` state, marking the missing parameter list and proactively initiating a clarification question to the user. The process will not be directly interrupted or have intermediate data destroyed; it can resume from where it left off after the parameters are supplemented.

[0120] Figure 3 The diagram illustrates a flowchart of an embodiment of an intelligent operation and maintenance question-and-answer method provided by the present invention. This method employs the intelligent operation and maintenance question-and-answer system 100 provided by the present invention. Figure 3 As shown, the method includes the following steps: S1. Perform noise reduction and terminology standardization on the original natural language question input by the user, extract business slot information, and organize the business slot information into a standardized structured question; S2. Perform label matching processing on the standardized structured problem to generate label clues, and input the label clues into the classification model to obtain the business category classification results; S3. Using the global state JSON file as a carrier, determine the user intent based on the business category classification results, extract business parameters from the standardized structured questions, verify the legality of the business parameters, generate a tool call sequence based on the user intent, the business parameters and the verification results, send the tool call sequence to the external rule engine MCP tool call module 104 to receive the returned rule judgment results, and generate a response based on the rule judgment results. S4. In the offline stage, the natural language rule text is parsed into a structured rule file and stored in a versioned manner. In the online stage, the tool call sequence is received, the business interface is called according to the tool call sequence and the business data is obtained through the business parameters, the structured rule matching the business data is loaded from the structured rule file, the hard-coded logic judgment is performed on the business data, and the rule judgment result is output. S5. Provide reference knowledge for the structured question rewriting module 101 and store the response and full-process data.

[0121] It should be noted that the beneficial effects of the intelligent operation and maintenance question-and-answer method provided in the above embodiments are the same as those of the intelligent operation and maintenance question-and-answer system 100 described above, and will not be repeated here. Furthermore, the method and system embodiments provided in the above embodiments belong to the same concept, and their specific implementation process can be found in the system embodiments, which will not be repeated here.

[0122] The intelligent operation and maintenance question and answer system 100 of the present invention can be a computer program (including program code) running on a computer device. For example, the intelligent operation and maintenance question and answer system 100 of the present invention is an application software that can be used to execute the corresponding steps in the intelligent operation and maintenance question and answer method of the present invention.

[0123] In some embodiments, the intelligent operation and maintenance question-and-answer system 100 of the present invention can be implemented in a combination of hardware and software. As an example, the intelligent operation and maintenance question-and-answer system 100 of the present invention can be a processor in the form of a hardware decoding processor, which is programmed to execute the intelligent operation and maintenance question-and-answer method of the present invention. For example, the processor in the form of a hardware decoding processor can be one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0124] The modules described in the embodiments of this invention can be implemented in software or hardware. The names of the modules are not, in some cases, limiting the scope of the module itself.

[0125] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-mentioned intelligent operation and maintenance question-and-answer methods. That is, an electronic device according to an embodiment of the present invention may include, but is not limited to: a processor and a memory; the memory is used to store the computer program; the processor is used to execute the intelligent operation and maintenance question-and-answer method shown in any embodiment of the present invention by calling the computer program.

[0126] In one alternative embodiment, an electronic device is provided, such as Figure 4 As shown, Figure 4 The illustrated electronic device 4000 includes a processor 4001 and a memory 4003. The processor 4001 and the memory 4003 are connected, for example, via a bus 4002. Optionally, the electronic device 4000 may further include a transceiver 4004, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 4004 is not limited to one type, and the structure of the electronic device 4000 does not constitute a limitation on the embodiments of the present invention.

[0127] Processor 4001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. Processor 4001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0128] Bus 4002 may include a path for transmitting information between the aforementioned components. Bus 4002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 4002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 4The bus 4002 is represented by only one thick line, but this does not mean that there is only one bus or one type of bus.

[0129] The memory 4003 may be ROM (Read Only Memory) or other types of static storage devices capable of storing static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices capable of storing information and instructions, or EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0130] The memory 4003 stores application code (computer program) for executing the present invention, and its execution is controlled by the processor 4001. The processor 4001 executes the application code stored in the memory 4003 to implement the content shown in the foregoing method embodiments.

[0131] Among them, electronic devices can also be terminal devices. A terminal device can be any terminal device that can install applications and access web pages through applications, including at least one of smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, smart TVs, and smart in-vehicle devices.

[0132] It should be noted that, Figure 4 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0133] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned intelligent operation and maintenance question-and-answer methods.

[0134] Alternatively, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, a floppy disk, and an optical data storage device, etc.

[0135] In an exemplary embodiment, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. The processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform the aforementioned intelligent operation and maintenance question-and-answer method.

[0136] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0137] It should be understood that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0138] The computer-readable storage medium provided in this invention can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0139] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the method shown in the above embodiments.

[0140] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0141] It should be noted that the terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and represent a limitation on a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of this application described herein can be implemented in an order other than that shown or described.

[0142] Those skilled in the art will recognize that this invention can be implemented as a system, method, or computer program product. Therefore, this invention can be specifically implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, this invention can also be implemented as a computer program product contained in one or more computer-readable media, which includes computer-readable program code.

[0143] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. An intelligent operation and maintenance question-and-answer system, characterized in that, include: The structured question rewriting module is used to perform noise reduction and terminology standardization on the original natural language questions input by the user, extract business slot information, and organize the business slot information into standardized structured questions; The tag clue classification module is used to perform tag matching processing on the standardized structured questions, generate tag clues, and input the tag clues into the classification model to obtain the business category classification results; The phased state machine intelligent agent workflow module is used to determine user intent based on the business category classification results using a global state JSON file as a carrier, extract business parameters from the standardized structured questions, verify the legality of the business parameters, generate a tool call sequence based on the user intent, the business parameters and the verification results, send the tool call sequence to the external rule engine MCP tool call module to receive the returned rule judgment results, and generate a response based on the rule judgment results. The external rule engine MCP tool call module is used to parse natural language rule text into structured rule files and store them in a versioned manner during the offline stage. During the online stage, it receives the tool call sequence, calls the business interface according to the tool call sequence and obtains business data through the business parameters, loads the structured rules that match the business data from the structured rule file, performs hard-coded logic judgment on the business data, and outputs the rule judgment result. The hierarchical knowledge base iteration management module is used to store knowledge data, provide reference knowledge for the structured question rewriting module, and store the answers and full-process data.

2. The intelligent operation and maintenance question-and-answer system according to claim 1, characterized in that, The structured problem rewriting module is specifically used for: Retrieve business process documents, internal enterprise terminology dictionaries, and operation manuals from the hierarchical knowledge base iteration management module as reference knowledge; The original natural language problem is denoised by using specialized rewritten prompts to remove emotional descriptions, irrelevant background descriptions, and subjective speculations about cause and effect. Replace the non-standard expressions in the original natural language problem with the company's internal standard terminology; Extract the problem category, core anomaly type, employee ID, business order number, business system to which it belongs, and user-predicted cause of the fault as the business slot information; The business slot information is organized into a standardized structured problem in a standardized JSON format.

3. The intelligent operation and maintenance question-and-answer system according to claim 1, characterized in that, The tag clue classification module is specifically used for: Pre-store business category tags, proprietary business term tags, and operational behavior tags, with each tag configured with a matching weight score; The script performs three matching logics on the standardized structured problem: exact matching, synonym matching, and word order-independent matching, generating tag clues that include tag name, text matching similarity, business category, frequency of occurrence, and business importance level.

4. The intelligent operation and maintenance question-and-answer system according to claim 3, characterized in that, The tag clue classification module is specifically used for: The label clues and historical classification samples are input into the classification model to obtain the business category classification result; wherein, the matching weight scores of each label on which the classification model outputs the business category classification result are iteratively adjusted through a configured automatic machine learning dynamic learning interface.

5. The intelligent operation and maintenance question-and-answer system according to claim 1, characterized in that, The phased state machine intelligent agent workflow module is specifically used for: The question-and-answer process is divided into four stages: intent recognition, slot filling, data verification, and processing planning. The global state JSON file stores the stage identifiers, output results of each stage, historical dialogue records, and a list of missing parameters for the four stages. The next action is automatically determined based on the current execution stage and the output results of each stage. The next action is one of the following: calling business tools, initiating clarification to the user, or the process is terminated normally.

6. The intelligent operation and maintenance question-and-answer system according to claim 5, characterized in that, The phased state machine intelligent agent workflow module is specifically used for: During the intent recognition stage, the business category to which the user's question belongs is determined based on the business category classification result. When the confidence difference of multiple categories is less than or equal to a preset threshold, a candidate category is generated and a clarification question is sent to the user. During the slot filling stage, the policy number, employee number, and organization code are extracted from the standardized structured question as the business parameters. When the core business parameters are missing, the process is blocked and a supplementary information query is initiated to the user. During the data verification phase, the validity of the document number, the operator's business permissions, and the document business status are verified. If the verification fails, a standardized exception message is returned. The standardized tool call sequence, data query SQL statements, and business troubleshooting solutions are output during the processing planning phase.

7. The intelligent operation and maintenance question-and-answer system according to claim 1, characterized in that, The external rule engine MCP tool calling module is specifically used for: In the offline phase, the natural language rule text is structured and parsed using a high-specification large language model. The parsing results are split into an execution strategy part and a business rule part. The business rule part is structured and stored in the rule base. Conflict detection and regression testing are performed on the natural language rule text. During the online phase, grouping rules matching the current business scenario are loaded, so that the large language model loads only the rule text within the current business group at a time. The script consumes the business data returned by the business interface, executes the hard-coded logic judgment of the structured rules, and outputs the matching rule description as the rule judgment result.

8. An intelligent operation and maintenance question-and-answer method, characterized in that, include: The user-input natural language question is subjected to noise reduction and terminology standardization, business slot information is extracted, and the business slot information is organized into a standardized structured question. The standardized structured problem is processed by label matching to generate label clues, and the label clues are input into the classification model to obtain the business category classification results; Using a global state JSON file as a carrier, the user intent is determined based on the business category classification results. Business parameters are extracted from the standardized structured questions, and the legality of the business parameters is verified. A tool call sequence is generated based on the user intent, the business parameters, and the verification results. The tool call sequence is sent to the external rule engine MCP tool call module to receive the returned rule judgment results, and a response is generated based on the rule judgment results. In the offline phase, the natural language rule text is parsed into a structured rule file and stored in a versioned manner. In the online phase, the tool call sequence is received, the business interface is called according to the tool call sequence, and business data is obtained through the business parameters. The structured rules that match the business data are loaded from the structured rule file, hard-coded logic is performed on the business data, and the rule judgment result is output. It provides reference knowledge for the structured question rewriting module and stores the responses and full-process data.

9. An electronic device, characterized in that, The electronic device includes a processor coupled to a memory, the memory storing at least one computer program, which is loaded and executed by the processor to enable the electronic device to implement the intelligent operation and maintenance question-and-answer method as described in claim 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which, when executed by a processor, implements the intelligent operation and maintenance question-and-answer method as described in claim 8.

Images