Electronic government enterprise service intelligent response method based on AI cue word engineering

Abstract

The invention relates to the crossing field of artificial intelligence technology and e-government, and discloses an e-government enterprise service intelligent response method based on AI cue word engineering, comprising: constructing and maintaining a government knowledge graph and a scenarized cue word template library; receiving and analyzing a service request input by an enterprise, and extracting an enterprise feature tag and a business intention; selecting a target cue word template based on the matching degree of the enterprise feature tag and the scene tag of the template in the scene cue word template library, and generating an adapted cue word template; retrieving associated knowledge data from the government affair knowledge graph according to the service request, and filling the adapted cue word template with the retrieved associated knowledge data to generate a structured cue word; inputting the structured cue word into an AI model to drive the AI model to generate a government affair service response; and collecting multi-dimensional evaluation feedback of government affair service response, and carrying out iterative optimization. The demand understanding precision is improved, and the risk of core information misjudgment is reduced.

Description

Technical Field

[0001] This invention relates to the intersection of artificial intelligence technology and e-government, specifically to an intelligent response method for e-government enterprise services based on AI prompt word engineering. Background Technology

[0002] With the deepening of the "streamlining administration, delegating power, and improving services" reform and the advancement of digital government construction, utilizing artificial intelligence (AI) technology, especially large language modeling (LLM), to improve the efficiency and quality of government services has become an important direction. Currently, some government service platforms have introduced AI customer service or Q&A systems to handle the massive inquiries from enterprises.

[0003] However, when existing technical solutions are applied to government service scenarios that are highly specialized, complex in process, and subject to dynamic policies, a series of core problems that urgently need to be addressed are exposed:

[0004] High deviation rate in understanding needs and weak adaptability to scenarios: General AI models lack guidance from government affairs knowledge, resulting in inaccurate understanding of industry-specific terminology, implicit conditions, and complex intentions input by enterprises. The existing system's prompts are simplistic and lack structured design for different scenarios such as "policy consultation," "cross-regional business processing," and "material pre-review," leading to a mismatch between responses and actual needs, especially insufficient support for startups or complex application scenarios.

[0005] The professionalism and compliance of responses are difficult to guarantee: Government services require answers to strictly adhere to the latest policy provisions and standard procedures. Existing AI systems rely on static data from model training, and their prompts lack effective "knowledge constraints" and "timeliness verification" mechanisms. This leads to frequent issues in responses, such as incorrect policy document references, outdated versions, omissions of process nodes, or lack of cross-departmental coordination, posing compliance risks.

[0006] Insufficient ability to analyze complex cross-domain requirements: Enterprise needs often involve collaboration across levels (national, provincial, municipal) and business departments (such as taxation, customs, and human resources). Existing solutions lack a mechanism to decompose such complex requirements and reorganize them according to business logic. The prompts do not embed cross-domain reasoning logic chains, resulting in one-sided AI-generated responses, broken processes, and an inability to form a complete closed-loop guidance.

[0007] The system is rigid and lacks continuous evolution capabilities: After traditional systems go live, the prompt word strategy and knowledge base updates rely on manual intervention, making it impossible to quantify and evaluate response quality or perform closed-loop optimization. When policies and regulations change or new business models emerge, the system has a long adaptation cycle and cannot achieve feedback-based self-iteration, resulting in long-term service performance degradation.

[0008] The service model is passive and lacks foresight: the existing system only provides a "question and answer" response and fails to build a profile of enterprises based on multi-dimensional data (such as development stage and business status) and predict their potential government needs. It is unable to realize the transformation from a proactive service model of "people looking for policies" to "policies looking for people".

[0009] Therefore, there is an urgent need for an intelligent response method and system that can deeply integrate knowledge from the government sector, dynamically adapt to enterprise characteristics, accurately guide AI models, and possess self-optimization and proactive service capabilities, in order to fundamentally solve the above problems. Summary of the Invention

[0010] This invention provides an intelligent response method for e-government enterprise services based on AI prompt word engineering, aiming to solve the technical problems existing in the current e-government service AI system, such as inaccurate understanding of needs, unprofessional response content, weak ability to handle complex cross-domain needs, poor system adaptability, and passive service mode.

[0011] This invention is achieved through the following technical solution:

[0012] A smart response method for e-government enterprise services based on AI prompt word engineering is proposed, running on a government intelligent response system. The system includes a government knowledge and prompt word template management module, an enterprise needs analysis and strategy matching module, a dynamic prompt word generation and AI interaction module, and a feedback evaluation and iterative optimization module. The method includes:

[0013] Step S1: The government knowledge and prompt word template management module constructs and maintains a government knowledge graph and a scenario-based prompt word template library. The government knowledge graph dynamically integrates entity nodes including policy nodes, business process nodes, enterprise tag nodes, and material nodes, as well as the relationships between entity nodes. The scenario-based prompt word template library stores multiple prompt word templates including role definitions, knowledge constraints, reasoning logic chains, and output format constraints.

[0014] Step S2: The enterprise demand parsing and strategy matching module receives and parses the service request input by the enterprise, extracts the enterprise feature tags and business intent; and selects a target prompt word template based on the matching degree between the enterprise feature tags and the scene tags of the templates in the scene-based prompt word template library. The content of the target prompt word template is dynamically adapted and adjusted according to the enterprise feature tags to generate an adapted prompt word template.

[0015] Step S3: The dynamic prompt word generation and AI interaction module retrieves relevant knowledge data from the government knowledge graph according to the service request, and fills the retrieved relevant knowledge data into the adapted prompt word template to generate structured prompt words; the structured prompt words are then input into the AI ​​model to drive the AI ​​model to generate a government service response.

[0016] Step S4: The feedback evaluation and iterative optimization module collects multi-dimensional evaluation feedback on the government service response, analyzes the root causes of the government service response quality based on the multi-dimensional evaluation feedback, and iteratively optimizes the corresponding prompt word templates in the scenario-based prompt word template library according to the root causes of the quality.

[0017] As an optimization, the specific process of constructing a government knowledge graph is as follows:

[0018] Through predefined API interface specifications, policy data, business process data, and enterprise tag data are collected in real time or near real time from government business systems, including at least tax, market supervision, and customs.

[0019] The system employs a matching algorithm based on policy text keywords and an association rule mining algorithm based on enterprise historical business data to automatically construct and maintain the association relationships between the entity nodes.

[0020] The relationships include: the applicability relationship between policy nodes and enterprise tag nodes, the prerequisite relationship between business process nodes, and the relationship between business nodes across different government departments.

[0021] As an optimization, the enterprise feature tags include the enterprise's industry, size level, qualification status, years of establishment, and development stage; dynamic adaptation and adjustment are performed based on the enterprise feature tags, including one of the following strategies:

[0022] When a company's characteristic tag includes "established for ≤1 year", a terminology simplification adjustment strategy is triggered.

[0023] When the business intent is identified to involve two or more different government business areas, a cross-domain reasoning logic enhancement adjustment strategy is triggered.

[0024] When the development stage in the enterprise's characteristic tag is identified as a preset specific stage, the policy package associated with the specific stage is triggered to adjust the strategy.

[0025] As an optimization, when the cross-domain inference logic enhancement adjustment strategy is triggered, in step S2, the enterprise requirement parsing and strategy matching module also performs the following:

[0026] Based on the preset business domain priority rules and process timing constraints, service requests involving multiple business domains are decomposed into multiple atomic tasks;

[0027] Based on the prerequisite and association relationships between business process nodes defined in the government knowledge graph, the multiple atomic-level tasks are logically reorganized to form a cross-domain task execution chain;

[0028] The logical structure of the cross-domain task execution chain is integrated into the adapted prompt word template as an enhanced reasoning logic chain.

[0029] As an optimization, in step S3, before generating the structured prompt words, the dynamic prompt word generation and AI interaction module also performs the following:

[0030] Based on the enterprise identifier, query the business process nodes and material submission nodes that the enterprise has handled within a preset historical time period in the government knowledge graph;

[0031] If the business or material associated with the current service request matches the historical records, the reusable historical business logic or material information will be used as additional knowledge data and populated into the adapted prompt word template.

[0032] As an optimization, in step S3, after the AI ​​model generates the government service response, the dynamic prompt word generation and AI interaction module also performs the following:

[0033] The policy basis and processing conditions in the response content will be compared and verified in real time with the latest data of the corresponding nodes in the government knowledge graph.

[0034] If the verification finds that the policy reference is incorrect, the version is outdated, or the conditions are not met, a correction instruction is generated and the AI ​​model is driven to regenerate the response content based on the correction instruction.

[0035] As an optimization, in step S4, the multidimensional evaluation feedback includes automatic evaluation scores and manual feedback scores. The automatic evaluation scores are calculated based on a preset quantitative indicator system, which includes the accuracy of policy citation, the completeness of process nodes, and the compliance of output format.

[0036] Iterative optimization based on quality root causes includes: weighting and fusing the automatic evaluation score with the human feedback score to obtain a comprehensive evaluation result; when the comprehensive evaluation result is lower than a preset threshold, initiating an optimization process for the corresponding template in the contextualized prompt word template library, the optimization process including analyzing error types, modifying template elements, and conducting effect verification tests.

[0037] As an optimization, the response method also includes:

[0038] The enterprise demand analysis and strategy matching module, based on the enterprise profile derived from multi-source data, uses a machine learning model to predict the types of government affairs demands that the enterprise may trigger within a preset time period in the future.

[0039] The predicted types of government service needs are matched with scenario tags in the scenario-based prompt word template library, and corresponding service prompt information or pre-filled consultation service responses are generated proactively.

[0040] As an optimization, the process of building an enterprise profile by integrating multi-source data includes:

[0041] Collect and integrate data from various sources, including government business systems, legitimate commercial data platforms, and publicly available information on the Internet; perform deduplication, anonymization, field alignment, and missing value imputation on the collected data to form a structured set of enterprise characteristic data;

[0042] Based on the aforementioned enterprise feature data set, a machine learning model for demand prediction is trained.

[0043] As an optimization, the government knowledge and prompt word template management module includes a policy dynamic perception engine; the policy dynamic perception engine continuously monitors government policy release channels, and when it detects a change in policy status such as effectiveness, revision, or repeal, it automatically triggers the following linked operations:

[0044] Update the status attributes of the corresponding policy nodes in the government knowledge graph;

[0045] The policy status change event is pushed to the contextualized prompt word template library, triggering a synchronous update of the knowledge constraint content in the associated prompt word template.

[0046] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0047] This invention improves the accuracy of demand understanding. By using entity recognition guidance and role definition with prompts specific to government scenarios, it can reduce misunderstandings of complex enterprise needs and minimize the risk of misjudging core information. For example, it can more accurately pinpoint the core demands in applications for R&D subsidies from newly registered specialized and innovative enterprises in 2024, which contain multiple conditions.

[0048] This invention ensures the professionalism and compliance of the response. The knowledge-enhanced prompts ensure that the AI ​​strictly references the latest policy document number, version ID, and provides a complete description of cross-departmental processes, reducing the push of outdated policy information and errors in the material list, making the response content more in line with government service standards.

[0049] This invention enhances the ability to handle complex requests. The logical chain guidance in the prompts improves the completeness of responses to compound requests (such as overlapping policy matching) and cross-domain requests, avoiding one-sided answers. For example, for the request of micro and small enterprises to apply for both R&D subsidies and income tax incentives, it can clearly break down the applicable conditions and application paths of the two policies; for the cross-domain request of cross-border e-commerce tax refunds, it can fully analyze the process connections between customs, tax, and foreign exchange departments.

[0050] This invention optimizes scenario adaptability and user experience. Scenario-based response templates can improve the efficiency of enterprise information acquisition: information search is more convenient in policy consultation scenarios, the direction of correction is clearer in material pre-review scenarios, the description of the process is clearer in progress query scenarios, and the process connection is easier to understand in cross-domain business scenarios.

[0051] This invention enhances the system's adaptability. The dynamic iterative closed loop can shorten the system's adaptation cycle to new policies. The prompt word template is continuously optimized through user feedback, improving the long-term service quality of the system. The policy dynamic perception engine enables real-time response to policy updates, avoiding the push of outdated information.

[0052] This invention covers the needs of enterprises throughout their entire life cycle. Through an enterprise life cycle needs model and corresponding prompts, it can proactively predict and meet the potential needs of enterprises at different stages of development. For example, it can push high-tech enterprise cultivation policies to growing enterprises and listing guidance policies to enterprises that have reached revenue targets, thereby improving the foresight and comprehensiveness of the service. Attached Figure Description

[0053] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0054] Figure 1 A system architecture diagram for implementing the method of the present invention;

[0055] Figure 2 A flowchart illustrating the prompt word template structure and dynamic generation process;

[0056] Figure 3 This is a schematic diagram of the node relationships in a government knowledge graph.

[0057] Figure 4 Optimize the closed-loop timing diagram for AI response. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0059] Before introducing this invention, the following keywords are defined:

[0060] Prompt Engineering refers to a technical approach that guides AI to generate high-quality responses that meet expectations by designing structured input text (prompts) to clarify the AI's role, task objectives, knowledge constraints, output format, and reasoning logic.

[0061] Enhanced prompts for government knowledge: These are prompts that embed professional knowledge in the government field (policy document number, business process, enterprise tag, policy version, cross-domain association, life cycle stage, etc.) and are adapted to the rules of government scenarios. They are characterized by clear roles, specific constraints, and clear logic.

[0062] Dynamic adjustment strategy: refers to a set of rules that optimize the basic prompt word template in real time based on enterprise attributes (such as size, industry, development stage) and demand characteristics (such as urgency, complexity, and whether it is cross-domain), so as to achieve personalized prompt word generation.

[0063] Policy Dynamics Perception Engine: This refers to a mechanism that connects to the government policy release system in real time, captures changes in the policy's effective / repealed / revised status, and automatically triggers updates to prompt word templates and reviews historical responses.

[0064] Cross-domain requirements: These refer to enterprise requirements that involve multiple levels (national-provincial-municipal) and multiple business areas (taxation + finance + human resources and social security, etc.), requiring the interconnection and processing of business processes across multiple departments.

[0065] Enterprise lifecycle demand model: This model predicts and matches corresponding government service needs based on different development stages of an enterprise from registration to growth to transformation, and guides demand in stages through prompt word engineering.

[0066] To enable those skilled in the art to implement the present invention, the technical solutions in the embodiments of the present invention will be clearly, completely, and in detail below with reference to the accompanying drawings and claims. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention.

[0067] This invention provides an intelligent response method for e-government enterprise services based on AI prompt word engineering. This method runs on a specially designed intelligent e-government response system, such as... Figure 1 As shown. Logically, the system comprises four main modules: a government knowledge management module, a prompt word template management module, and a [missing information - likely related to government knowledge management and prompt word template management]. Figure 1 The system comprises three modules: a government affairs knowledge and prompt word template management module, an intelligent response engine module (including an enterprise needs analysis and strategy matching module, a dynamic prompt word generation and AI interaction module), and a feedback optimization hub module (i.e., a feedback evaluation and iterative optimization module). These modules work together to achieve the entire process from knowledge construction to intelligent response and then to closed-loop optimization.

[0068] (I) Overall Method Process

[0069] The implementation process of this method strictly follows the steps below, and is executed separately or collaboratively by the four main modules:

[0070] Step S1: The government knowledge and prompt word template management module constructs and maintains a government knowledge graph and a scenario-based prompt word template library. The government knowledge graph dynamically integrates entity nodes including policy nodes, business process nodes, enterprise tag nodes, and material nodes, as well as the relationships between entity nodes. The scenario-based prompt word template library stores multiple prompt word templates including role definitions, knowledge constraints, reasoning logic chains, and output format constraints.

[0071] Step S2: The enterprise demand parsing and strategy matching module receives and parses the service request input by the enterprise, extracts the enterprise feature tags and business intent; and selects a target prompt word template based on the matching degree between the enterprise feature tags and the scene tags of the templates in the scene-based prompt word template library. The content of the target prompt word template is dynamically adapted and adjusted according to the enterprise feature tags to generate an adapted prompt word template.

[0072] Step S3: The dynamic prompt word generation and AI interaction module retrieves relevant knowledge data from the government knowledge graph according to the service request, and fills the retrieved relevant knowledge data into the adapted prompt word template to generate structured prompt words; the structured prompt words are then input into the AI ​​model to drive the AI ​​model to generate a government service response.

[0073] Step S4: The feedback evaluation and iterative optimization module collects multi-dimensional evaluation feedback on the government service response, analyzes the root causes of the government service response quality based on the multi-dimensional evaluation feedback, and iteratively optimizes the corresponding prompt word templates in the scenario-based prompt word template library according to the root causes of the quality.

[0074] I. Government Affairs Knowledge and Prompt Text Template Management Module

[0075] This module forms the system's knowledge foundation, responsible for building, maintaining, and dynamically updating the domain knowledge base and intelligent response guidance templates. It specifically includes the following sub-modules:

[0076] 1.1 Government Knowledge Graph Submodule

[0077] The core task of this submodule is to build and maintain a structured, dynamically updated knowledge graph of the government domain (such as...). Figure 3 (As shown).

[0078] (1) Construction of basic data sources: Before constructing the knowledge graph, the three core knowledge bases are sorted out and structured.

[0079] Policy and Regulation Database: Stored in a structured format by policy type (tax / subsidy / qualification) - applicable entities - effective date - clause content - document number - version ID - status (effective / repealed / revised). For example, storing the "Announcement of the Ministry of Finance and the State Taxation Administration on Further Improving the Policy of Additional Deduction for R&D Expenses (No. 7 of 2024)," whose policy type is tax, applicable entities include micro and small enterprises, and version ID is V1.0.

[0080] Business Process Library: A structured process chain is built using the following format: Item - Prerequisites - Steps / Nodes - Required Documents - Time Limits - Related Businesses - Involved Levels. For example, the prerequisite for opening a corporate social security account is completion of business registration, and the required documents include a copy of the business license, etc.

[0081] Enterprise Tag Library: Define and maintain attribute tags such as industry (software / manufacturing), size (small / medium-sized), qualifications (high-tech / specialized and innovative), establishment time, revenue status, and development stage (startup / growth / mature).

[0082] (2) Dynamic data collection: Data is collected from various government business systems in real time or near real time through predefined API interfaces to update the above knowledge base.

[0083] Data source and interface:

[0084] Collect policy data from the tax bureau system (API such as xxx.xxx.xxx / tax-policy, GET request), and return it in JSON format, containing fields such as policy ID, document number, effective date, and clause content.

[0085] Collect enterprise tag data from the market supervision bureau system (API such as xxx.xxx.xxx / market-enterprise, GET request), and return JSON format, which includes fields such as enterprise ID, name, industry, size, qualifications, and establishment time.

[0086] Collect business process data from customs and other business systems (API such as xxx.xxx.xxx / customs-business, POST request), and return it in XML format, containing fields such as business ID, name, type, processing requirements, and associated materials.

[0087] Update frequency and quality parameters:

[0088] Policy data: A real-time synchronization mechanism is adopted (data changes in the departmental system are synchronized to the knowledge graph within 1 hour).

[0089] Business process data: It adopts a mechanism of updating once a week (executed at 3:00 AM every Monday, updating changes such as business steps, material lists, and time limits).

[0090] Enterprise profile data: It adopts a mechanism of updating once a day (executed at 2:00 AM every day, updating changes such as enterprise qualifications, revenue status, and development stage).

[0091] Each update requires recording the amount of data and setting a target data validation pass rate of ≥98%.

[0092] (3) Knowledge graph construction and relationship mining:

[0093] Node and Relationship Definitions: In a graph database (such as Neo4j), define seven core entity node types: Policy, Business Process, Enterprise, Material, Region, Hierarchy, and Implementing Department. Define five core relationship types to achieve cross-domain association: Applicable relationship (e.g., Policy A applies to micro and small enterprises), Precedence relationship (e.g., Business C precedes Business D), Association relationship (e.g., cross-border e-commerce tax refund business is associated with customs, tax, and foreign exchange departments), Hierarchical inclusion relationship, and Business cross-relationship. Each relationship must be labeled with its association basis (e.g., policy clauses, business rules).

[0094] Automatic relation construction algorithm:

[0095] The keyword matching algorithm based on policy text first performs word segmentation and part-of-speech tagging on the policy text to extract keywords related to business type, enterprise attributes, and implementing department. Then, through preset keyword-node mapping rules (e.g., customs declaration corresponding to customs declaration business node), it automatically establishes the association between policy nodes and business nodes, and enterprise nodes. The matching results must indicate the specific location of the keywords in the policy text. The target keyword matching accuracy of this algorithm is ≥95%.

[0096] An association rule mining algorithm based on enterprise historical business data (Apriori algorithm): This algorithm collects the enterprise's historical business data for the past 3 years and uses the Apriori algorithm to perform association analysis on the data, calculating the support and confidence between different business processes. When the confidence reaches a preset threshold (default 80%), an association relationship between the two business nodes is automatically established. The target coverage of this algorithm's association rule mining is ≥90%.

[0097] Visual Editing and Version Management: The system provides visual editing tools, including modules for adding, deleting, and modifying nodes, adjusting relationship weights, managing graph versions, and querying and displaying relationships. It supports users manually adding nodes, deleting invalid nodes, and modifying node attributes, and records operation logs. All modifications automatically generate version numbers (in YYYYMMDD-V format), supporting version rollback and difference comparison. Versions are retained for 5 years.

[0098] (4) Policy Timeliness Management: The system sets a timeliness threshold, only referencing policies that are effective within 365 days of the current time, and marking policies that are about to expire (within 30 days of expiration) (e.g., marking them as about to expire next to the policy node). Expired policies must be marked as abolished in the knowledge graph the day after they expire. The threshold for long-term effective policies can be set to ≤730 days.

[0099] 1.2 Scenario-based Template Design Submodule

[0100] This submodule designs and manages the core instructions used by the system response, namely the contextualized prompt word templates.

[0101] (1) Core elements of the template: Each prompt word template contains 7 core structured elements:

[0102] Role definition: For example, you are a business service consultant at the XX Municipal Government Service Center, specializing in the interpretation of tax incentive policies.

[0103] Knowledge constraint: If the answer is based solely on a policy that took effect after January 1, 2024, the policy document number and version ID must be indicated.

[0104] Input parsing guidance: The model focuses on extracting key information such as the company's industry tags, establishment time, business type, and development stage.

[0105] Reasoning logic chain: guide AI thinking step by step, such as 1. determine whether the enterprise meets the applicable policy conditions; 2. if it does, list the application materials; 3. prompt the deadline; 4. for cross-domain requirements, break down atomic tasks and associate implicit dependencies.

[0106] Output format constraints: Specify the response structure, such as presenting it in three levels of headings: 1. Policy basis; 2. Applicable conditions; 3. Processing steps (with time limits for each step); Cross-domain requirements must specify the departments and hierarchical relationships involved.

[0107] Error handling: Specifies how to handle uncertain information, such as clearly indicating 'It is recommended to consult the XX department (phone: XXX)'.

[0108] Lifecycle guidance: Guide AI to proactively associate with applicable policies at each stage of a company's development.

[0109] (2) Template Classification and Management: Templates are classified and stored according to scenario tags such as policy consultation - tax incentives, material pre-review - qualification certification, cross-border business - cross-border e-commerce tax refund, and life cycle - enterprise growth stage. Each tag is associated with the corresponding business scenario description and the number of applicable templates, and the tags can be dynamically expanded. The system monitors template satisfaction (target value ≥ 85%) and template error rate (target value ≤ 3%) as the basis for iteration, with the iteration cycle being monthly (7 working days) and quarterly (15 working days).

[0110] 1.3 Policy Dynamics Perception Engine Submodule

[0111] This submodule is responsible for ensuring the timeliness of the knowledge base and templates. It continuously monitors the government policy release system and automatically triggers the following actions when it detects a change in policy status, such as implementation, revision, or repeal:

[0112] Update the status attributes of the corresponding policy nodes in the government knowledge graph.

[0113] The change event will be pushed to the prompt word template management module, triggering a synchronous update of the knowledge constraint content in the associated prompt word templates. When policies change dynamically, the relevant templates must be updated within 24 hours.

[0114] 1.4 Enterprise Profile Data Governance Submodule

[0115] This submodule builds high-quality enterprise profiles and trains demand prediction models to support proactive services.

[0116] Multi-source data fusion and governance:

[0117] Data sources: A multi-dimensional data collection integrating government data (approximately 1.72 million records from the market supervision bureau, tax bureau, customs, etc.), legitimate business data (approximately 14.3 million records of enterprise annual reports from 2015 to 2024, acquired in cooperation with data exchanges), and internet data (approximately 210 million records of publicly available recruitment and bidding information from 2018 to 2024, obtained through compliant web crawlers).

[0118] Data governance: Strict deduplication, anonymization (in accordance with GB / T35273-2020), outlier handling (using σ=3 truncation for continuous features), field alignment (uniform mapping to the industry code GB / T4754-2017), and missing value imputation (using the MICE method for fields with a missing value rate higher than 15%) are performed on the original data.

[0119] Data output: The final result is a structured profile table covering 10.24 million enterprises, with a data volume of approximately 3.6GB, containing 237 feature fields (such as revenue, tax payment, intellectual property, number of job openings, investment and financing rounds, etc.).

[0120] Demand prediction model training and deployment:

[0121] Training data preparation: The input feature (X) is a 237-dimensional profile feature after processing. After feature engineering (standardization, one-hot encoding, derived features, industry trend index extraction, and high-dimensional feature screening), 186 effective features are retained. The output label (Y) is 42 categories of potential needs labeled by government experts (a total of 684,000 positive samples), which are used to form a training set of 2,736,000 samples through 1:3 downsampling.

[0122] Model Structure and Training: A hybrid network of Wide & Deep + Self-Attention is adopted. The Wide side retains interpretable cross features, while the Deep side is a 3-layer fully connected network (512 / 256 / 128 hidden layers, Dropout=0.3). The AdamW optimizer (lr=1e-3) is used with FocalLoss (γ=2) as the loss function, iterating for 80 epochs, and the IntegratedGradients method is used to provide feature interpretability.

[0123] Performance Validation and Deployment: The model achieved a Micro-F1 score of 0.851 and an AUC of 0.923 on the test set. In production environment backtesting, it achieved an accuracy of 85.4% in predicting requirements three months in advance. The model was deployed in ONNX format on a government intranet GPU server (T4×2), with an average prediction time of 38 milliseconds. II. Enterprise Requirements Analysis and Strategy Matching Module

[0124] This module serves as the system's intelligent sensing and decision-making hub, responsible for accurately understanding enterprise inputs (including proactively anticipated needs) and matching the optimal response strategy.

[0125] 2.1 Multimodal Input Processing Submodule

[0126] This module supports businesses to input service requests through various methods, including text, voice (ASR to text), and image (OCR recognition). It standardizes unstructured input into a format of request text plus additional information (such as the emotion tag 'urgent'), filters redundant information, and extracts the core requirements.

[0127] 2.2 NLP Feature Extraction Submodule

[0128] A pre-trained model in the government sector (such as ERNIE-Gov2.0) is used to perform in-depth analysis of the standardized requirements and output structured feature labels.

[0129] Entity recognition and intent classification: Accurately identifies key entities such as company name, business type (e.g., R&D expense deduction), level (national-provincial-municipal), and relevant business department (e.g., taxation + technology). The confidence threshold for entity recognition is set to ≥0.85.

[0130] Feature tag generation: Output a structured tag set containing fields such as industry, size, qualification status, years of establishment, revenue status, development stage, demand type, urgency, and whether it is cross-domain.

[0131] 2.3 Intelligent Matching and Strategy Engine Submodule

[0132] Template matching: Based on extracted enterprise feature tags (e.g., industry: software size: micro-enterprise; demand type: material query; urgency: high; development stage: growth stage; cross-domain: yes) and template scene tags, the matching degree is calculated using a cosine similarity algorithm. An example of weight allocation is as follows: business type 30%, industry 15%, urgency 10%, enterprise size 10%, historical interactions 5%, cross-domain: 15%, development stage 15%. The template matching similarity threshold is set to ≥0.75; if this threshold is met, the template is directly applied.

[0133] Dynamically adjust strategy library: Built-in strategy library, which triggers adjustments in real time based on feature tags.

[0134] Terminology popularization strategy: When a company with an establishment period of ≤1 year (newly established enterprise) is identified, professional terms such as tax settlement are converted to annual tax return settlement.

[0135] Enhanced adjustment strategy for cross-domain reasoning logic: Triggered when the business involves two or more different government business areas.

[0136] Policy package-related adjustment strategy: triggered when the enterprise's development stage is identified as growth stage or revenue exceeds 50 million.

[0137] For example:

[0138] High urgency level → Add priority processing time limits and expedited processing instructions to the template.

[0139] Newly established businesses → trigger a strategy of simplifying terminology, such as converting the annual tax settlement into annual tax filing.

[0140] Cross-domain requirements trigger cross-domain reasoning guidance strategies, requiring the response to clearly define the order of related business processes and implicit dependencies.

[0141] Enterprises registered for 3 years or more → trigger the matching template for the high-tech enterprise cultivation policy.

[0142] Enterprise revenue exceeds 50 million → triggers related listing guidance policy package template.

[0143] 2.4 Cross-Domain Requirement Decomposition - Reorganization of Submodules

[0144] This submodule starts when the cross-domain policy is triggered. Its specific logic and parameters are as follows:

[0145] Decomposition Algorithm and Parameters: A rule-based hierarchical decomposition algorithm is adopted. The rule base includes business domain priority rules, process sequence constraint rules, and departmental authority and responsibility boundary rules. Core parameters include: business domain matching threshold (≥0.8), sequence constraint confidence (≥0.95 for strong constraints), and departmental collaboration coefficient (≥0.7 indicates the need for multi-departmental collaboration).

[0146] Business area priority rules: used to determine the order of processing different business operations, such as tax business taking precedence over financial business.

[0147] Process sequence constraint rules: Define the mandatory order of business processes, such as qualification certification taking precedence over subsidy application.

[0148] Departmental authority and responsibility boundary rules: Clearly define the department to which a task belongs, such as the market supervision bureau being responsible for verifying enterprise registration information, and the science and technology bureau being responsible for the identification of research and development projects.

[0149] Business priority quantification: Business priority weight is calculated using the formula: Business Priority Weight = Policy Level Coefficient × 0.6 + Processing Frequency Percentage × 0.4. The policy level coefficient is set as follows: National level 1.0, Provincial level 0.9, Municipal level 0.8, County level 0.7. Processing Frequency Percentage = (Number of times this business has been processed in the past year ÷ Total number of times all businesses have been processed in the past year) × 100%.

[0150] For example: Tax-related matters (national-level policies, coefficient 0.9, frequency percentage 45%) have a weight of 0.72; financial matters (provincial-level policies, coefficient 0.8, frequency percentage 30%) have a weight of 0.6. Therefore, tax-related matters have a higher priority.

[0151] Time Conflict Resolution: When strong time constraints (such as prerequisite qualification certification) conflict with an enterprise's urgent needs (such as an upcoming subsidy deadline), a manual review process is automatically triggered, as follows:

[0152] Identify conflict types (time-constraint type and emergency requirement type) and generate a "Conflict Description Report" (including estimated duration, remaining duration, and impact analysis).

[0153] The report is pushed to relevant departments (such as qualification certification departments and subsidy disbursement departments), and the review response time limit is set to 2 hours.

[0154] Government officials determine the handling method (allowing parallel processing, processing first and supplementing later, rejection, etc.). Based on the review results, the system adds a manual review suggestion module to the prompt template, clearly indicating the conclusion, path, and contact information of the responsible department. A "Conflict Explanation Report" is generated, with a review response time limit set at 2 hours.

[0155] Atomic task division: The atomic-level tasks generated by decomposition must meet the following requirements: single business area, single processing objective, and the ability to independently output a clear conclusion (e.g., compliant / non-compliant). For example, the identification of specialized and innovative enterprises + R&D subsidy application can be decomposed into atomic tasks such as qualification verification (Science and Technology Bureau), subsidy condition matching (Industry and Information Technology Bureau / Science and Technology Bureau), and cross-verification of materials.

[0156] Reorganization and Logical Integration: Based on the prerequisites and relationships in the knowledge graph, and combined with the enterprise's urgency level (trigger time limit compressed by 30% when extremely high), atomic tasks are logically reorganized to form a cross-domain task execution chain. The logical structure of this chain is then used as an enhanced reasoning logic chain and integrated into the adapted prompt word template.

[0157] Basic order: Based on the business prerequisite relationships in the knowledge graph. For example, subsidy condition matching requires qualification verification results as input, so qualification verification is a prerequisite.

[0158] Dynamic Adjustment: Adjustments are made dynamically based on the enterprise's needs. If an enterprise requests priority access to subsidy amount calculations, the subsidy conditions will be matched and executed ahead of schedule, provided that strong time constraints are met. If the urgency level is extremely high, the task chain will be marked as expedited, triggering time reduction rules for each step (e.g., shortening by 30%).

[0159] Cross-level processing: Tasks involving multiple levels such as national, provincial, and municipal are reorganized according to the hierarchical inclusion relationship in the knowledge graph to ensure that national policies are implemented before provincial supporting measures are implemented.

[0160] The logical structure of the cross-domain task execution chain generated during the reorganization and logical integration is transformed into a structured, clearly defined description of reasoning steps. This description is then integrated into the adapted prompt word template as an enhanced reasoning logic chain, thereby guiding the AI ​​model to perform cross-domain parsing and response according to this complex logic.

[0161] 2.5 Demand Prediction Triggering and Proactive Response Submodule

[0162] This submodule calls the prediction model deployed in submodule 1.4 to implement proactive services.

[0163] Prediction and Tiered Triggering: Input the latest profile features of the enterprise into the model to obtain the probability of demand prediction for the next 0-12 months.

[0164] When the predicted probability is ≥0.75, the system automatically matches the requirement with the contextualized prompt word template library.

[0165] When the probability is between 0.6 and 0.75, the generated to-do items are pushed to government staff for manual review.

[0166] If the probability is less than 0.6, it will not be triggered.

[0167] Proactive response generation: For automatically triggered requests, the system calls the matching template, combines it with the company's current profile data, and automatically generates pre-filled consultation service responses or prompts, which are proactively pushed to the company through channels such as the message center.

[0168] III. Dynamic Prompt Generation and AI Interaction Module

[0169] This module is the execution center of the system, responsible for synthesizing precise instructions and driving AI to generate high-quality, compliant final responses.

[0170] 3.1 Prompt Word Assembly Engine Submodule

[0171] This submodule performs a four-step precision assembly process:

[0172] Step 1: Variable Extraction: Extract four core variables from the parsing results: entity variables (including company name, establishment time, size, revenue, development stage, etc.), business variables (including business type, departments involved, cross-domain identifiers, etc.), policy variables (including policy effective time range, document number constraints, etc.), and output variables (including special format requirements, terminology conversion requirements, etc.).

[0173] For example: For consultation on policies regarding additional deduction of R&D expenses and cultivation of high-tech enterprises, the variables extracted are: Company Name = XX Company, Establishment Time = 2023, Size = Micro and Small Enterprise, Revenue = 60 million, Development Stage = Growth Stage, Business Type = Additional Deduction of R&D Expenses + Cultivation of High-tech Enterprises, Involved Departments = Tax Bureau + Science and Technology Bureau, Policy Effective Date = After January 1, 2024, Terminology Conversion Requirement = Simplification of Additional Deduction Ratio, Cross-Domain Identifier = Yes.

[0174] Step 2: Variable Mapping: Accurately map variables to preset placeholders (such as [role / department], [policy / time constraint]) in the adapted template.

[0175] For example: the [Role Department] in the template maps to the Tax Bureau and the Science and Technology Bureau; the [Company Basic Information] maps to XX Company (established in 2023, a micro-enterprise, with revenue of 60 million, in the growth stage); and the [Policy Time Constraint] maps to policies that will take effect after January 1, 2024.

[0176] Step 3: Rule Validation: Check variable completeness (whether required variables are missing), format compliance (e.g., time format is YYYY year), and logical consistency (e.g., whether cross-domain business has been associated with the corresponding department variables). The system requires a variable extraction completeness rate ≥ 98%, a variable mapping accuracy rate ≥ 99%, and a rule validation response time ≤ 1 second. If the validation fails, return supplementary information from the requirements parsing step.

[0177] Step 4: Dynamic Filling and Historical Association: For example Figure 2As shown, the validated variables are substituted into the template to automatically generate structured prompts. Simultaneously, additional rules are triggered based on variable attributes (e.g., if the cross-domain identifier is "yes," a cross-domain process connection explanation module is automatically added). Furthermore, based on historical business association results, the following parsing logic is dynamically integrated (i.e., based on the enterprise identifier, the cross-domain business processing records of the enterprise for the past 3 years are automatically retrieved, business type association patterns, departmental collaboration nodes, and material reuse information from historical business are extracted, and this is integrated into the current prompt generation process to achieve linked parsing of historical data and current needs):

[0178] Departmental Collaboration Node Warning: If historical business processes have encountered departmental connection breaks (such as tax audits requiring verification by the technology department), the warning message will pre-annotate the collaborative node description: The technology department's R&D project verification must be completed first. This extracts historical departmental connection breaks from cross-domain business processes (e.g., tax department document review requires the technology department to issue an R&D project verification certificate first). When the current requirement involves the same departmental combination, the preconditions for the collaborative node will be pre-annotated in the warning message. For example, if a company previously experienced delays in tax audits due to failure to complete customs declaration verification when processing cross-border e-commerce tax refunds, and now, when inquiring about export tax refunds, the parsing logic triggers a warning that customs declaration status verification must be completed first, adding a customs-tax collaborative node description to the output format.

[0179] Material Reuse Information Prompt: If materials submitted in previous business transactions and still valid (such as a list of R&D personnel, copies of the company's business license, and high-tech enterprise certificates) match the current requirements, the prompt word "Material List" will be automatically marked: "The list of R&D personnel can reuse the materials from business number 202405XX, and does not need to be submitted again." For example, if a company submitted a list of R&D personnel when applying for a science and technology subsidy three months ago, and is now inquiring about R&D expense deduction, the parsing logic will trigger a material reuse prompt, and the prompt word "Required Materials" will indicate that the list of R&D personnel can reuse the materials from business number 202405XX.

[0180] Additional rule trigger: Automatically add modules based on variable attributes. For example, if the cross-domain identifier is "Yes", a cross-domain process connection description module will be automatically added to the template.

[0181] After completing all the above operations, the final structured prompts are generated. The system performance metrics are set as follows: variable extraction completeness ≥ 98%, variable mapping accuracy ≥ 99%, and rule validation response time ≤ 1 second.

[0182] In other words, after the structured prompt words are generated but before they are sent to the AI ​​model, the prompt word assembly engine submodule performs a series of post-processing operations on them to ensure that their format is absolutely compliant and secure, and to prevent the AI ​​model from failing to call or producing erroneous output due to problems with the prompt words themselves.

[0183] Special character filtering: Scan the prompt text and remove or escape illegal characters, control characters, or excessively long spaces that may be misinterpreted by the AI ​​model as control instructions or cause API call errors.

[0184] Length verification: Calculate the number of tokens or character length of the prompt and compare it with the maximum input limit of the connected AI model. If the prompt length exceeds the limit, the engine will automatically trigger a length compression strategy, such as: simplifying descriptive text without affecting core instructions and key variables; or reconstructing some content with more concise sentence structure to ensure that the length of the final output prompt is within the model's safety threshold.

[0185] Final format confirmation: After the above processing is completed, the engine will perform a final check according to the preset template format specifications to ensure that the structure is complete, the variables are filled correctly, and there are no format errors, thereby generating a final prompt word with a compliant format for subsequent modules to call.

[0186] 3.2 Government Affairs Big Data Model Interface Submodule

[0187] Multi-model scheduling: Connects to large-scale government-specific models, supporting the scheduling of different models based on the complexity of the requirements (regular / cross-domain), with a single response time controlled to ≤3 seconds. For regular requirements, the system quickly invokes the appropriate model; for identified complex cross-domain requirements, it automatically schedules models with stronger logical reasoning capabilities for processing. The system controls the single response time to ≤3 seconds; if a timeout occurs, it automatically switches to a backup model.

[0188] Multi-turn dialogue management: If the AI ​​response contains missing information (missing ≥1 core variable) or ambiguity (≥1 point of ambiguity), follow-up questions will be automatically triggered. If the enterprise does not respond within 15 minutes, a timeout reminder will be sent and the dialogue context will be saved.

[0189] 3.3 Response Verification and Correction Submodule

[0190] Real-time compliance verification: After generating the initial response, the knowledge graph interface is immediately invoked to compare the policy document number, version ID, application conditions, and materials list in the response with the latest data in the knowledge graph. Verification criteria include: the policy status is in effect, the application conditions are completely consistent with the terms and conditions, and the materials list is complete (missing information must be ≤0).

[0191] Error Correction and Regeneration: Verification criteria include: the policy document number and version ID are authentic and valid (matching the effective status node); the application conditions are completely consistent with the policy clauses (e.g., the criteria for determining micro and small enterprises strictly correspond to the "Regulations on the Classification Standards for Small and Medium-sized Enterprises"); and the material list is complete (compare with the required material fields for this business in the business process library; the missing rate must be ≤0). If an incorrect policy reference, outdated version, or inconsistent conditions / materials are found, the verification is deemed unsuccessful, and a specific correction instruction is generated (e.g., if the referenced policy has been repealed, please answer according to Article Z of the latest policy "XXX"). The AI ​​model is then driven to regenerate the response content based on the correction instruction (this process can be repeated up to 3 times). The system requires a compliance verification pass rate of ≥99.5%.

[0192] Output readability optimization: For startups, a step-by-step approach with case analogies is used (e.g., just like individuals need to fill in income details when filing taxes, companies need to submit a list of R&D project expenses); for established companies, a data-driven approach with professional terminology is used (e.g., the R&D expense deduction ratio is 175%, meaning that 1 million RMB in R&D expenses can be deducted from 1.75 million RMB in taxable income).

[0193] Automatically split long sentences to ensure that the average length of a single sentence is ≤30 characters.

[0194] Key information such as deadlines and contact numbers are presented in bold and highlighted format to ensure 100% coverage of key information.

[0195] IV. Feedback Evaluation and Iterative Optimization Module

[0196] This module is the evolutionary hub of the system, driving continuous optimization through quantitative evaluation and root cause analysis. Its optimization closed-loop timeline is as follows: Figure 4 As shown.

[0197] 4.1 Feedback Acquisition Submodule

[0198] We collect feedback from enterprise users (such as useful / useless tags) through web pages and mini-programs, provide government personnel with a backend for reviewing and marking issues, and automatically collect error cases into a unified issue database.

[0199] 4.2 Root Cause Analysis Submodule

[0200] Multidimensional quantitative automatic evaluation: The system automatically scores each AI response, with the following indicator system and weights:

[0201] Policy citation accuracy (30%): (1 - number of incorrect citations ÷ total number of citations) × 100%. Error judgment criteria: If any of the following occurs in the policy cited in the response: document number error, version ID error, citation of an expired policy, or mismatch between the content of the clause and the original text, it will be counted as an incorrect citation.

[0202] Process node completeness (30%): (Actual number of nodes ÷ Standard number of nodes) × 100%. Error judgment criteria: If the response is missing a non-minor node of the business process defined by the knowledge graph (such as missing the foreign exchange verification node in the cross-border e-commerce tax refund process), missing one or more nodes will be judged as incomplete. Missing minor nodes (such as the number of copies of materials) will only be recorded and will not affect the completeness judgment.

[0203] Cross-domain resolution completeness (10%): (Actual number of departments covered ÷ Number of necessary departments) × 100%. Error judgment criteria: For cross-domain business requirements, if the response does not mention or correctly handles the necessary related departments marked in the knowledge graph (e.g., cross-border e-commerce tax refunds do not involve the State Administration of Foreign Exchange), missing one or more of them is judged as incomplete. This indicator accounts for 10% of the total score.

[0204] Lifecycle matching degree (10%): (Number of matched policies ÷ Number of policies that should be matched) × 100%. Error judgment criteria: For enterprises in a specific development stage (such as the growth stage), if the response does not actively match or omits the core policies of that stage (such as the high-tech enterprise cultivation policy), it is judged as not meeting the matching degree standard. This indicator accounts for 10% of the total score.

[0205] Output format compliance (20%): (Number of compliant modules ÷ Total number of modules) × 100%. Error judgment criteria: If the response lacks the core output modules specified in the template (such as policy basis, processing steps), or the module order is incorrect, it is judged as non-compliant. Automatic evaluation of total score calculation: Based on the different contributions of the above five indicators to the response quality, the feedback optimization central module calculates the final score according to the following weighting formula:

[0206] The automatic assessment score is calculated as follows: policy citation accuracy × 30% + process node completeness × 30% + cross-domain resolution completeness × 10% + lifecycle matching degree × 10% + output format compliance × 20%.

[0207] The total score is 100 points, and the system sets ≥80 points as the automatic passing score.

[0208] Quantification of human feedback:

[0209] Enterprise feedback score: Collected through tags such as useful / useless. Useful feedback scores 100 points, useless feedback scores 0 points. If users further point out specific issues such as policy errors, missing processes, incomplete cross-domain resolution, or unmet stage requirements, 20 points will be deducted for each item; difficult-to-understand descriptions will deduct 10 points. The final enterprise feedback score is calculated as follows: Enterprise Feedback Score = (Total Score for Each Item ÷ Total Number of Tags) × 100%.

[0210] Government staff feedback score: Government staff can mark areas for improvement in the review backend. Core errors (such as incorrect policy basis) deduct 30 points per item, and general defects (such as vague material descriptions) deduct 10 points per item. If there are no areas for improvement, the score is 100. Calculate the government staff feedback score as follows: Government staff feedback score = 100 - (Total deductions ÷ Total number of feedback items × 100). Total manual feedback score = Enterprise feedback score × 60% + Government staff feedback score × 40%, with a passing score of 80.

[0211] Comprehensive assessment and attribution:

[0212] The overall evaluation score is calculated as follows: (Automatic evaluation score × 60%) + (Manual feedback score × 40%). A score of ≥85 is excellent, 70-84 requires improvement, and <70 is unsatisfactory.

[0213] Root cause analysis is performed based on the scores of each sub-indicator. When the overall evaluation score of a certain type of response is poor, the system attributes the cause based on the scores of each sub-indicator. For example, if the accuracy of policy citation is less than 80%, it is attributed to insufficient knowledge constraints in the template. The optimization direction is to strengthen the policy timeliness verification instructions in the template (such as real-time verification instructions for policy document number and version ID) and add rules to compare the cited policy with the effective status of the knowledge graph.

[0214] If the cross-domain parsing integrity is less than 80%, it is determined to be a vulnerability in the cross-domain decomposition-reorganization model. It is necessary to supplement the related department nodes in the knowledge graph or adjust the algorithm threshold, and adjust the department synergy coefficient threshold of the decomposition algorithm (from 0.7 to 0.65).

[0215] If the lifecycle matching degree is less than 80%, it is determined that the lifecycle guidance rules are missing. A corporate stage-policy mapping table needs to be added to the template. For example, a corporate stage-policy tag mapping table can be added to the template (e.g., revenue of more than 50 million + established for more than 3 years → matching listing guidance policies).

[0216] If more than 20% of the feedback is that the terminology is difficult to understand, it is determined that the terminology conversion table is incomplete, and the industry terminology will be supplemented with colloquial explanations (such as input tax credit → the tax paid when purchasing goods can be deducted from the tax payable).

[0217] 4.3 Template Iteration Engine Submodule

[0218] Optimization trigger: Regular iterations occur once a week. For templates with a comprehensive evaluation score <70, emergency optimization will be initiated (to be completed within 24 hours); templates with a score between 70 and 84 will be included in the regular optimization list.

[0219] Effectiveness Verification and Deployment: Modify template elements based on root cause analysis results. After modification, conduct A / B testing to verify effectiveness: randomly select 50 identical business scenarios, generate responses using both the old and new templates, and compare key business metrics such as overall evaluation rate, cross-domain parsing completeness rate, and enterprise useful feedback rate. The new template must improve these metrics by ≥10% compared to the old template before it can be officially replaced and updated to the scenario-based prompt word template library.

[0220] Version management: Save complete template history versions and iteration logs. When the policy dynamic perception engine detects policy changes, the knowledge constraint content of the relevant templates must be synchronized and updated within 24 hours.

[0221] The following example, using a small software company's consultation on the 2024 R&D expense deduction and high-tech enterprise cultivation policies, illustrates the implementation process of this invention.

[0222] Scenario: Company XX Software Technology Co., Ltd., established in June 2023, has 15 employees and revenue of 60 million yuan. Through the government service mini-program, they input: "We are a small company that develops software. How much more tax can we deduct from our R&D expenses this year? What materials do we need to submit? Also, our company is doing quite well now, and we would like to know how to cultivate high-tech enterprises?"

[0223] Execution process:

[0224] S1: The government knowledge management module has built a knowledge graph, which includes the effective "Announcement No. 7 of 2024" (additional deduction of 175%) and "XX Province High-tech Enterprise Cultivation Measures (2024 Edition)", and is associated with tags such as micro and small enterprises, software industry, and growth stage.

[0225] S2: The intelligent response engine module parses the request and extracts feature tags: {Industry: Software, Size: Micro-enterprise, Years Established: 1, Development Stage: Growth Stage, Revenue: 60 million}. The prompt word template management module matches the policy consultation - tax incentives + enterprise life cycle - growth stage template, and triggers the new enterprise popularization and growth stage policy guidance strategies to generate an adapted template.

[0226] S3: The intelligent response engine module retrieves information from the knowledge graph, including the additional deduction ratio, the list of required materials, and the policy conditions for high-tech enterprise cultivation, and then populates this information into the adapted template. The generated structured prompts explicitly require the AI, acting as a joint consultant to the tax bureau and the science and technology bureau, to explain the ratio using the common-sense analogy of deducting 175 yuan from a 100 yuan tax payment, and to list the required materials and high-tech enterprise cultivation instructions in bullet points. Based on this, the AI ​​model generates a response, which is then verified for compliance.

[0227] S4: After receiving a clear answer, the company provides helpful feedback. The feedback optimization hub module records this interaction.

[0228] (Extended Prediction): Meanwhile, the company profile shows that the company's revenue has exceeded 50 million and it is in the growth stage. The demand prediction model predicts with high confidence that it has a need for high-tech enterprise certification. The system can proactively push relevant in-depth interpretation articles or application guidance.

[0229] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart response method for e-government enterprise services based on AI prompt word engineering, running in a government intelligent response system, characterized in that, The system includes a government knowledge and prompt word template management module, an enterprise needs analysis and strategy matching module, a dynamic prompt word generation and AI interaction module, and a feedback evaluation and iterative optimization module. The method includes: Step S1: The government knowledge and prompt word template management module constructs and maintains a government knowledge graph and a scenario-based prompt word template library. The government knowledge graph dynamically integrates entity nodes including policy nodes, business process nodes, enterprise tag nodes, and material nodes, as well as the relationships between entity nodes. The scenario-based prompt word template library stores multiple prompt word templates including role definitions, knowledge constraints, reasoning logic chains, and output format constraints. Step S2: The enterprise demand parsing and strategy matching module receives and parses the service request input by the enterprise, extracts the enterprise feature tags and business intent; and selects a target prompt word template based on the matching degree between the enterprise feature tags and the scene tags of the templates in the scene-based prompt word template library. The content of the target prompt word template is dynamically adapted and adjusted according to the enterprise feature tags to generate an adapted prompt word template. Step S3: The dynamic prompt word generation and AI interaction module retrieves relevant knowledge data from the government knowledge graph according to the service request, and fills the retrieved relevant knowledge data into the adapted prompt word template to generate structured prompt words; the structured prompt words are then input into the AI ​​model to drive the AI ​​model to generate a government service response. Step S4: The feedback evaluation and iterative optimization module collects multi-dimensional evaluation feedback on the government service response, analyzes the root causes of the government service response quality based on the multi-dimensional evaluation feedback, and iteratively optimizes the corresponding prompt word templates in the scenario-based prompt word template library according to the root causes of the quality.

2. The intelligent response method for e-government enterprise services based on AI prompt word engineering as described in claim 1, characterized in that, The specific process of constructing a government knowledge graph is as follows: Through predefined API interface specifications, policy data, business process data, and enterprise tag data are collected in real time or near real time from government business systems, including at least tax, market supervision, and customs. The system employs a matching algorithm based on policy text keywords and an association rule mining algorithm based on enterprise historical business data to automatically construct and maintain the association relationships between the entity nodes. The relationships include: the applicability relationship between policy nodes and enterprise tag nodes, the prerequisite relationship between business process nodes, and the relationship between business nodes across different government departments.

3. A method for intelligent response of e-government enterprise services based on AI prompt word engineering as described in claim 1 or 2, characterized in that, The enterprise characteristic tags include the enterprise's industry, size level, qualification status, years of establishment, and development stage; dynamic adaptation and adjustment are performed based on the enterprise characteristic tags, including one of the following strategies: When a company's characteristic tag includes "established for ≤1 year", a terminology simplification adjustment strategy is triggered. When the business intent is identified to involve two or more different government business areas, a cross-domain reasoning logic enhancement adjustment strategy is triggered. When the development stage in the enterprise's characteristic tag is identified as a preset specific stage, the policy package associated with the specific stage is triggered to adjust the strategy.

4. The intelligent response method for e-government enterprise services based on AI prompt word engineering as described in claim 3, characterized in that, When the cross-domain inference logic enhancement adjustment strategy is triggered, in step S2, the enterprise requirement parsing and strategy matching module also performs the following: Based on the preset business domain priority rules and process timing constraints, service requests involving multiple business domains are decomposed into multiple atomic tasks; Based on the prerequisite and association relationships between business process nodes defined in the government knowledge graph, the multiple atomic-level tasks are logically reorganized to form a cross-domain task execution chain; The logical structure of the cross-domain task execution chain is integrated into the adapted prompt word template as an enhanced reasoning logic chain.

5. The intelligent response method for e-government enterprise services based on AI prompt word engineering according to claim 1, characterized in that, In step S3, before generating structured prompts, the dynamic prompt generation and AI interaction module also performs the following: Based on the enterprise identifier, query the business process nodes and material submission nodes that the enterprise has handled within a preset historical time period in the government knowledge graph; If the business or material associated with the current service request matches the historical records, the reusable historical business logic or material information will be used as additional knowledge data and populated into the adapted prompt word template.

6. The intelligent response method for e-government enterprise services based on AI prompt word engineering according to claim 1, characterized in that, In step S3, after the AI ​​model generates the government service response, the dynamic prompt word generation and AI interaction module also performs the following: The policy basis and processing conditions in the response content will be compared and verified in real time with the latest data of the corresponding nodes in the government knowledge graph. If the verification finds that the policy reference is incorrect, the version is outdated, or the conditions are not met, a correction instruction is generated and the AI ​​model is driven to regenerate the response content based on the correction instruction.

7. The intelligent response method for e-government enterprise services based on AI prompt word engineering according to claim 1, characterized in that, In step S4, the multidimensional evaluation feedback includes automatic evaluation scores and manual feedback scores. The automatic evaluation scores are calculated based on a preset quantitative indicator system, which includes the accuracy of policy citation, the completeness of process nodes, and the compliance of output format. Iterative optimization based on quality root causes includes: weighting and fusing the automatic evaluation score with the human feedback score to obtain a comprehensive evaluation result; when the comprehensive evaluation result is lower than a preset threshold, initiating an optimization process for the corresponding template in the contextualized prompt word template library, the optimization process including analyzing error types, modifying template elements, and conducting effect verification tests.

8. The intelligent response method for e-government enterprise services based on AI prompt word engineering according to claim 1, characterized in that, Response methods also include: The enterprise demand analysis and strategy matching module, based on the enterprise profile derived from multi-source data, uses a machine learning model to predict the types of government affairs demands that the enterprise may trigger within a preset time period in the future. The predicted types of government service needs are matched with scenario tags in the scenario-based prompt word template library, and corresponding service prompt information or pre-filled consultation service responses are generated proactively.

9. The intelligent response method for e-government enterprise services based on AI prompt word engineering as described in claim 8, characterized in that, The process of building an enterprise profile by integrating multi-source data includes: Collect and integrate data from various sources, including government business systems, legitimate commercial data platforms, and publicly available information on the Internet; perform deduplication, anonymization, field alignment, and missing value imputation on the collected data to form a structured set of enterprise characteristic data; Based on the aforementioned enterprise feature data set, a machine learning model for demand prediction is trained.

10. The intelligent response method for e-government enterprise services based on AI prompt word engineering according to claim 1, characterized in that, The government knowledge and prompt word template management module includes a policy dynamic perception engine; the policy dynamic perception engine continuously monitors government policy release channels, and when it detects a change in policy status such as effectiveness, revision, or repeal, it automatically triggers the following linked operations: Update the status attributes of the corresponding policy nodes in the government knowledge graph; The policy status change event is pushed to the contextualized prompt word template library, triggering a synchronous update of the knowledge constraint content in the associated prompt word template.

Images