A non-landing evidence pointer and compliance gated intelligent question and answer method, system, device and computer program product for a standard regulation query system
By constructing an evidence pointer pool and implementing non-local controls, the problem of standard texts being easily stored and output illegally in the standard regulation query system is solved. This achieves unified control of compliance boundaries and traceability of source information, reduces the risk of AI output, and provides audit support for the control of user-uploaded content and abnormal access.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 杨旺霖
- Filing Date
- 2026-05-11
- Publication Date
- 2026-07-31
AI Technical Summary
The standard and regulation query system is prone to problems such as the standard text being stored or output in violation of regulations, AI-generated answers generating details of restricted clauses, difficulty in uniformly controlling compliance boundaries for different content types, difficulty in tracing the status of official sources, and difficulty in timely managing abnormal access and user-uploaded content.
This paper presents a non-local evidence pointer and compliance-gated intelligent question-answering method for standard regulation query systems. By acquiring user input questions and request context, it identifies standard numbers and regulation types, constructs an evidence pointer pool, implements non-local control, performs authoritative source routing and output gating detection, records audit logs, and achieves control over user-uploaded content.
It avoids the risks of local storage and display of restricted standard text, ensures unified control of compliance boundaries, achieves traceability of standard status and source information, reduces the risks of specific terms in AI output, and provides verifiable evidence of abnormal access.
Smart Images
Figure CN122489714A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of standard and regulatory information retrieval, natural language processing, content compliance control, enterprise safety production compliance management, and artificial intelligence question answering, and particularly to a non-landable evidence pointer and compliance-gated intelligent question answering method, system, device, and computer program product for standard and regulatory query systems. Background Technology
[0002] Scenarios involving enterprise safety production, oil and gas stations, hazardous chemicals, special equipment, fire protection, lightning protection, explosion protection, occupational health, and environmental protection involve a large number of laws, regulations, national standards, industry standards, local standards, and safety technical specifications. The related content is abundant, changes rapidly, and comes from diverse sources, leading on-site personnel to frequently ask questions in conversational language.
[0003] Existing standards and regulations query systems typically use standard numbers, standard names, or full-text keywords as entry points and return document lists, text excerpts, or download results to users. In the context of mandatory standards, recommended standards, industry standards, and local standards, this approach is prone to copyright and proprietary publishing risks due to the local storage, caching, display, or secondary dissemination of the full text of the standard, specific clauses, parameters, tables, or appendices.
[0004] While general-purpose large language models can generate natural language responses, if they are directly accessed from the full text of a standard or a fragment of a clause, they may output specific clauses, parameters, tables, or appendices. If there is a lack of source constraints, it is easy to fabricate standard names, standard numbers, clause numbers, or applicable conclusions, making on-site judgments and compliance boundaries unreliable.
[0005] The compliance boundaries differ for different content types. Laws and regulations can provide source guidance or compliance display within legally publicly available channels. Mandatory national standards are suitable for retaining official guidance and a simplified scope of application, while recommended national standards, industry standards, and local standards are more suitable for providing only basic indexes and official external links. If the system cannot form a unified technical control chain between data collection, retrieval, AI input, AI output, and external link redirection, the risk of unauthorized storage or output may still be reintroduced in subsequent iterations.
[0006] Therefore, a technical solution is needed that can complete the identification of standard and regulatory issues, content type determination, official source routing, evidence pointer construction, AI-based compliant responses, user upload control, and audit traceability without storing the restricted standard text locally. Summary of the Invention
[0007] The technical problem this invention aims to solve is: addressing the issues in standard and regulatory query systems where standard texts are easily stored or output illegally, AI-generated responses easily generate restricted clause details, compliance boundaries for different content types are difficult to control uniformly, official source status is difficult to trace, and abnormal access and user-uploaded content are difficult to manage in a timely manner. This invention provides a non-landable evidence pointer and compliance-gated intelligent question-and-answer method, system, device, and computer program product for standard and regulatory query systems.
[0008] To address the aforementioned technical issues, this invention provides a non-local evidence pointer and compliance-gated intelligent question-answering method for standard regulation query systems, comprising the following steps: obtaining a natural language question and request context input by the user; identifying the standard number, regulation type, and extracting on-site scenario keywords from the natural language question to obtain a question feature set; determining the target content type based on the question feature set, and reading the inclusion boundary, display boundary, artificial intelligence output boundary, and external link boundary corresponding to the target content type from a risk strategy table.
[0009] An evidence pointer pool is constructed within the defined inclusion boundaries. This pool includes only one or more of the following: standard number, standard name, publication date, implementation date, status information, source identifier, official link, scenario tag, and a minimized summary that is allowed to be displayed. For recommended national standards, industry standards, and local standards, the evidence pointer pool does not store the full text of the standard, specific clauses, parameters, tables, and appendices.
[0010] Implement non-local control over restricted standard content, prohibiting the storage of the full text of the standard, specific clauses, parameters, tables, and appendices in local databases, caches, logs, and AI contexts. This non-local control can be achieved through field whitelisting, pre-write detection, cache disabling, log anonymization, and existing data auditing.
[0011] Based on the target content type and the official link, authoritative source routing is performed to obtain or display source status information, and session-level destruction conditions are set for temporarily obtained public metadata. The public metadata may include standard status, publication date, implementation date, repeal or replacement relationship, source name, and official query entry.
[0012] The evidence pointers in the evidence pointer pool are scored according to metadata matching degree, on-site scene matching degree, standard state weight, and source credibility to obtain the target evidence pointer set. The target evidence pointer set is used to limit the source range of the artificial intelligence response, but does not include the restricted standard text.
[0013] The target evidence pointer set and compliance gating instructions are input into the large language model to generate response results that do not include specific clauses, parameters, tables, and appendices of the restricted standards. The compliance gating instructions are used to limit the large language model to output only general conclusions, source guidance, and on-site handling suggestions, and must not output the original text of the restricted standards or details that can replace the original text.
[0014] The output gating test is performed on the response results, and a conclusion, source guidance, and on-site handling suggestions are output after the test passes. The output gating test includes detection of the number of suspected standard original text characters, detection of the proportion of suspected original text citations, detection of intent to request clause details, detection of intent to output parameter tables, and detection of the number of consecutive follow-up questions on the same standard.
[0015] Record audit logs for access, redirection, output gating, and abnormal requests. These audit logs may include the time and network address of user access to the standard index page, official external link redirection records, AI requests and output gating results, abnormal access frequency, content fingerprints, watermark identifiers, and honeypot alert trigger records.
[0016] Compared with existing technologies, this invention has at least the following beneficial effects: First, by transforming the compliance boundaries of different content types into machine-executable rules through a risk strategy table, it avoids the reintroduction of restricted standard text storage and display in subsequent iterations; Second, by replacing the localized content library containing standard text with an evidence pointer pool, the system can retain retrieval, routing, and response capabilities while reducing the risks of local retention and secondary dissemination; Third, by using authoritative source routing and session-level destruction mechanisms, the standard status and source information are traceable without forming a long-term cache; Fourth, by using output gating detection, it reduces the probability of artificial intelligence outputting specific clauses, parameters, tables, or appendices; Fifth, by using user upload control and auditing traces, it provides verifiable evidence during abnormal access, complaint handling, and internal inspections. Attached Figure Description
[0017] Figure 1 This is a schematic diagram of the system structure provided in an embodiment of the present invention.
[0018] Figure 2 This is a schematic diagram of the compliance-gated intelligent question-answering method provided in an embodiment of the present invention.
[0019] Figure 3 This is a schematic diagram of the authoritative source routing and non-landing control process provided in the embodiments of the present invention.
[0020] Figure 4 This is a schematic diagram of the user upload control and auditing process provided in an embodiment of the present invention. Detailed Implementation
[0021] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. The described embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.
[0022] Example 1: Compliance-Gated Intelligent Question Answering Method. The system obtains the natural language question and request context input by the user. The request context may include the user's network address, session identifier, historical follow-up questions, target standard number, page entry point, and user role. The natural language question can be a standard number, regulation name, equipment name, on-site inspection question, hazard assessment question, or inspection preparation question.
[0023] The system performs standard number recognition, regulation type recognition, and on-site scenario keyword extraction for natural language processing. Standard number recognition can standardize numbers such as GB, GB / T, GB / Z, AQ, AQ / T, TSG, SY, SY / T, HJ, HJ / T, and GBZ, and can also perform fuzzy matching between the user-input numeric core number and the numeric portion of the standard number. On-site scenario keywords can include equipment name, work scenario, hazard type, inspection action, and management topic.
[0024] The system determines the target content type based on a set of problem characteristics. Target content types can include laws and regulations, mandatory national standards, recommended national standards, industry standards, local standards, and user-uploaded content. The risk policy table configures allowed fields, prohibited fields, word count thresholds, redirect prompts, caching policies, and manual review policies for different target content types.
[0025] When the target content type is laws and regulations, the risk strategy table may allow the retention of legally publicly available information with verifiable sources, or provide links to official sources; when the target content type is mandatory national standards, the risk strategy table may allow the retention of the standard number, name, publication and implementation date, official link, and a general scope of application not exceeding a preset word limit; when the target content type is recommended national standards, industry standards, or local standards, the risk strategy table may only allow the retention of the standard number, name, publication and implementation date, status, and official link, and may not retain content summaries, specific clauses, parameters, tables, and appendices.
[0026] The system constructs an evidence pointer pool based on the risk strategy table. Evidence pointers differ from the evidence text; they are used to describe verifiable sources and minimize metadata, and are not intended to replace the original standard text. Evidence pointers may include the standard number, standard name, publication date, implementation date, current or repealed status, official source name, official link, scenario tag, and credibility identifier.
[0027] The system performs non-persistent control before data is written. Non-persistent control includes field whitelisting and pre-write content detection. Field whitelisting prevents the addition of standard full text, clauses, parameters, tables, and appendix fields to the data table. Pre-write content detection is used to identify suspected standard text, clause numbers, parameter tables, appendix identifiers, and continuous long text; when restricted rules are matched, the system refuses to write, truncates the content, or converts the content into an official link.
[0028] The system can also be configured with existing data audit tasks. These tasks periodically scan the database, object storage, cache, and logs to check for restricted standard text. If any unauthorized content is found, the system generates an audit event and prompts the user to address it within a preset timeframe.
[0029] The system routes authoritative sources based on the type of target content. The authoritative source routing table can map national standards to official sources of national standards, safety production industry standards to competent authorities or official release channels, special equipment safety technical specifications to official sources of market supervision, ecological environment standards to official sources of ecological environment, and laws and regulations to official sources of laws and regulations.
[0030] In one implementation, the system only temporarily requests state or public metadata from official sources within a user session. This temporary metadata is filtered by fields before entering the AI context and destroyed when the response generation ends, the session expires, or the request fails. The system does not preload, pre-render, snapshot, transcode, or cache the full text of restricted standards for extended periods.
[0031] The system scores the pool of evidence pointers. The score may include scores for matching standard or core number, standard name, on-site scenario, current validity status, mandatory or recommended type, official source credibility, and risk of continued user questioning. The system selects the target set of evidence pointers based on the overall score.
[0032] The system inputs the target evidence pointer set and compliance gating instructions into the large language model. The compliance gating instructions can require the model to first output a general judgment, then output source guidance and on-site handling suggestions; and explicitly prohibit the output of specific clauses, parameters, tables, appendices, or long excerpts that can replace the original text of the restricted standards.
[0033] The system performs output gating checks on the model's responses. Output gating checks may include detecting the number of suspected standard original text characters, the proportion of suspected original text citations, the intent to request details of clauses, the intent to output parameter tables, and the number of consecutive follow-up questions on the same standard. If the detection results exceed a preset threshold, the system blocks the response and returns the system to the official query guidance.
[0034] In one implementation, the threshold for the number of suspected original text characters can be set to no more than 300 characters, and the threshold for the proportion of suspected original text citations can be set to no more than a preset proportion. When the same user continuously asks a preset number of follow-up questions about specific clauses, parameters, or tables of the same standard in the same session, the system will stop answering further detailed questions.
[0035] Example 2: User-Uploaded Content Control. After receiving files, pasted text, or form content uploaded by users, the system identifies standard numbers, standard names, clause numbers, parameter tables, appendix identifiers, and suspected restricted text features. For upload requests that match recommended national standards, industry standards, or local standards, the system either blocks the upload or transfers it to manual review.
[0036] For user-uploaded content that is allowed to be retained, the system only saves metadata, source links, upload history, review status, and content fingerprints. Content fingerprints can be used for subsequent source tracing and duplicate detection, but not for reconstructing the restricted standard text. For content that fails manual review, the system records the reason for rejection and clears temporary files.
[0037] Example 3: Audit Tracking and Abnormal Access Control. The system records standard index page access, official external link redirects, AI requests, output gating results, user-uploaded review results, and abnormal access frequency. For users who access the same standard more than a preset number of times in a single day, make batch requests to multiple standards, frequently trigger clause detail blocking, or are suspected of scraping, the system will implement rate limiting, CAPTCHA pre-processing, redirect to the official platform, or suspend the service.
[0038] The system can add content fingerprints, watermarks, or honeypot hints to allowed output prompts or general information. These content fingerprints, watermarks, or honeypot hints do not alter official standard content, nor are they used to forge the original standard text; rather, they are used to identify copying and propagation paths and abnormal access behavior.
[0039] Example 4: Intelligent Question Answering System. This system includes a question parsing module, a risk strategy module, an evidence pointer construction module, a non-local control module, an authoritative source routing module, an evidence scoring module, an answer generation module, an output gating module, a user upload management module, and an audit trail module.
[0040] The issue parsing module is used to acquire user input and generate a set of issue features. The risk strategy module is used to read the inclusion boundaries, display boundaries, AI output boundaries, and external link boundaries based on the target content type. The evidence pointer construction module is used to generate a pool of evidence pointers that do not contain restricted standard text. The non-local control module is used to restrict field structure, detect written content, and disable caching.
[0041] The authoritative source routing module identifies official data sources and outputs status and source information. The evidence scoring module selects the target set of evidence pointers from the evidence pointer pool. The answer generation module generates a summary answer based on the target set of evidence pointers. The output gating module blocks answers containing restricted details. The user upload control module intercepts or reviews user-uploaded content. The auditing and logging module records access, redirects, gating, uploads, and abnormal behavior.
[0042] Example 5: Electronic Device and Computer Program Product. This example provides an electronic device including a processor and a memory. The memory stores a computer program, which, when executed by the processor, implements the methods described in any of the above examples. This example also provides a computer program product, which includes a computer program that, when executed by a processor, implements the methods described in any of the above examples.
[0043] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the present invention. Those skilled in the art can modify, substitute, or combine the above embodiments without departing from the technical concept of the present invention, and such modifications, substitutions, or combinations should all fall within the protection scope of the present invention.
Claims
1. A non-landed evidence pointer and compliance gated intelligent question answering method for a standard regulation query system, characterized in that, include: Obtain the user's input in natural language and the request context; The natural language problem is subjected to standard number identification, regulation type identification, and on-site scenario keyword extraction to obtain a problem feature set; The target content type is determined based on the set of problem features, and the inclusion boundary, display boundary, artificial intelligence output boundary and external link boundary corresponding to the target content type are read from the risk strategy table. Within the inclusion boundary, an evidence pointer pool is constructed, which includes only one or more of the following: standard number, standard name, publication date, implementation date, status information, source identifier, official link, scenario tag, and the minimum summary that can be displayed. Implement non-local controls on restricted standard content, prohibiting the storage of the full text of the standard, specific clauses, parameters, tables, and appendices in local databases, caches, logs, and artificial intelligence contexts; Based on the target content type and the official link, authoritative source routing is performed to obtain or prompt source status information, and session-level destruction conditions are set for temporarily obtained public metadata. The evidence pointers in the evidence pointer pool are scored according to metadata matching degree, on-site scene matching degree, standard state weight and source credibility to obtain the target evidence pointer set; Input the target evidence pointer set and compliance gating instructions into the large language model to generate response results that do not contain specific clauses, parameters, tables and appendices of the restricted standards; Perform output gating detection on the response results, and output conclusions, source guidance, and on-site handling suggestions after the detection passes; Record audit logs for access, redirection, output gating, and abnormal requests.
2. The method of claim 1, wherein, The target content types include one or more of the following: laws and regulations, mandatory national standards, recommended national standards, industry standards, local standards, and user-uploaded content; the risk strategy table configures allowed fields, prohibited fields, word count thresholds, redirection prompts, caching strategies, and manual review strategies for different target content types.
3. The method according to claim 1, characterized in that, The non-local control measures include: restricting data table structure through field whitelists; detecting keywords in the full text of the standard, clauses, parameters, tables, and appendices before writing; rejecting writing, de-identifying and truncating, or converting text that hits the restricted rules into official links; and periodically scanning existing data to discover illegally stored content.
4. The method according to claim 1, characterized in that, The authoritative source routing includes: mapping national standards, industry standards, local standards, special equipment safety technical specifications, ecological environment standards, and laws and regulations to their corresponding official data sources; standardizing the status, publication date, implementation date, repeal or replacement relationship returned by the official data sources; and outputting the status to be verified and the official query portal when the official data sources are unavailable.
5. The method according to claim 1, characterized in that, The target evidence pointer set does not contain restricted standard text, and the scoring includes one or more of the following: standard number or core number matching score, standard name matching score, on-site scenario matching score, current valid status score, mandatory or recommended type weight, official source credibility weight, and user continuous follow-up question risk weight.
6. The method according to claim 1, characterized in that, The output gating detection includes: detecting the number of suspected standard original text characters in the response result, the proportion of suspected original text citations, the intent of requesting clause details, the intent of outputting parameter tables, and the number of consecutive follow-up questions on the same standard; when any detection result exceeds a preset threshold, the response result is blocked and official query guidance is output.
7. The method according to claim 1, characterized in that, The method also includes user-uploaded content control steps: identifying standard numbers, standard names, and restricted text features in user-uploaded content; intercepting or manually reviewing upload requests that match recommended national standards, industry standards, or local standards; and saving only metadata, source links, upload records, review status, and content fingerprints for content that is allowed to be retained.
8. The method according to claim 1, characterized in that, The audit logs include the time and network address of user access to the standard index page, official external link redirection records, artificial intelligence requests and output gating results, abnormal access frequency, content fingerprints, watermark identifiers, and honeypot prompt trigger records; the audit logs are used to prove that the system only provides compliance guidance and official source redirection.
9. A non-local evidence pointer and compliance-gated intelligent question-and-answer system for standard regulatory query systems, characterized in that: include: The system comprises a problem analysis module, a risk strategy module, an evidence pointer construction module, a non-local control module, an authoritative source routing module, an evidence scoring module, an answer generation module, an output gating module, a user upload control module, and an audit trail module; each module is configured to perform the method described in any one of claims 1 to 8.
10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by the processor, implements the method according to any one of claims 1 to 8.