Construction scheme intelligent auditing method based on rule engine and large model

Abstract

The application provides a building construction scheme intelligent auditing method based on a rule engine and a large model, and belongs to the technical field of construction scheme intelligent auditing. The method performs Open XML structured analysis on a construction scheme document, constructs a chapter tree and divides the chapter tree into content blocks with paths; content auditing integrates structured retrieval and vector semantic retrieval, automatically extracts an engineering geographic area for regional elimination items, dynamically filters a suitable elimination item database for semantic comparison; style auditing adopts a three-level backtracking mechanism of character level, paragraph level and document default style to obtain actual effective styles, and combines a tolerance threshold and semantic font matching for accurate verification; and the auditing result is automatically inserted based on an XML offset position <w:comment>The method can effectively break through the identification bottleneck of style inheritance deviation, realize dynamic adaptation of regional specifications, and improve the depth, accuracy and automation level of the review.< / w:comment>

Description

Technical Field

[0001] This invention relates to the field of intelligent computer document processing technology, and in particular to an intelligent review method for building construction plans based on rule engines and large models. Background Technology

[0002] In the field of construction engineering, technical documents such as construction plans and specialized plans are usually submitted in the form of Word documents. Their content not only covers substantive technical elements such as construction techniques, equipment and material selection, but also includes a large amount of content with mandatory or standardized requirements regarding formatting and layout. These documents often need to simultaneously meet national and local standards, industry specifications, and template requirements of the construction unit or competent authority; therefore, the review of document content compliance and format standardization is of a high standard.

[0003] Regarding document format verification, existing technologies lack the ability to deeply penetrate complex style nesting, focusing primarily on text content deduplication or simple style name matching, making it difficult to perform refined verification of deep document formatting. In Word documents, the final presentation of text is determined by multiple layers of styles. Conventional methods can only identify explicit style definitions and cannot handle deviations caused by style inheritance. For example, when a document does not explicitly set a font but inherits from paragraph or document default settings, existing technologies often fail to accurately extract the actual rendered style, making it difficult to identify deep formatting errors such as "pseudo-SimSun" or non-standard mixing of Chinese and Western fonts. Regarding content compliance review, existing technologies cannot effectively address the regional differences and dynamic updates in building codes. National and local housing and construction departments regularly publish "lists of phased-out processes, equipment, and materials," and different geographical regions (such as provinces) have clear regional requirements for prohibited or restricted processes. Traditional review technologies based on keywords or simple search enhancements lack a deep understanding of engineering semantics, especially the inability to automatically identify the geographical location of a construction plan and compare it with a dynamically updated regional phased-out database. This results in the review process relying heavily on human experience, making it easy to misuse outdated standards or prohibited processes within restricted areas, thus posing potential risks to project quality and safety.

[0004] Therefore, there is an urgent need for an intelligent auditing method that can both penetrate the underlying XML of Word to perform full-element style validation and combine geographic location semantics to perform accurate compliance screening, so as to meet the needs of accurate review of elimination items in complex engineering scenarios. Summary of the Invention

[0005] The main objective of this invention is to provide an intelligent review method for building construction plans based on a rule engine and a large model, which solves the problem of insufficient review depth and lag caused by frequent updates of regulations and elimination items related to geographical regions, and improves the retrieval and matching accuracy of construction plans.

[0006] To address the aforementioned technical problems, the present invention provides an intelligent review method for building construction plans based on a rule engine and a large model, comprising the following steps:

[0007] S1. Parse the underlying XML structure of the target construction plan document, construct a chapter tree, and divide it into content blocks with chapter path tags; S2. Construct a dual-path retrieval system using both structured indexes and semantic vector indexes for each content block to obtain a set of candidate content. S3. Identify the geographical region associated with the target construction plan document, dynamically filter the elimination item database based on the geographical region, and perform semantic comparison between the filtered elimination item information and the candidate content set to output the content review results. S4. Extract the actual effective styles based on a three-level backtracking mechanism of character level, paragraph level and document default level, compare them with the template style library, and output the style review results; S5. Convert the content review results and style review results into XML offset positions, and inject annotation tags at the corresponding positions in the target construction plan document.

[0008] Preferably, in step S1, parsing the XML underlying structure of the target construction plan document specifically includes: Traverse paragraph labels and table elements in a Word document, identify heading styles and their levels, and construct a chapter tree representing the parent-child relationship between chapters; Each content block generated by dividing the content by chapter nodes contains a chapter path label, a chapter number, and the main text.

[0009] Preferably, in step S2, the dual-path retrieval specifically includes: The first candidate set is generated by matching metadata between the name of the review item and the title or summary of the content block; Calculate the cosine similarity between the review item vector and the content block vector, and filter out content blocks that exceed a preset threshold to generate a second candidate set; The first and second candidate sets are deduplicated and semantically reordered using a large language model to determine the final relevant content blocks.

[0010] Preferably, in step S3, the dynamic filtering of the elimination item database by geographical region specifically includes: Geographic entities are extracted from the project overview section of the target construction plan document using a large language model; Dynamic SQL queries are generated based on geographic location entities to filter out lists of prohibited or restricted processes, equipment, and materials applicable to the entire country or a specific geographic region from the elimination item database.

[0011] Preferably, step S3 also includes a standard validity verification, specifically: extracting the standard name and standard number from the candidate content set, comparing and verifying them with the standard database, and generating a failure review conclusion when the standard name does not exist or the standard number is not in the current version status corresponding to the database.

[0012] Preferably, in step S4, the three-level backtracking mechanism specifically includes: when extracting style attributes of the target document, if the current character-level style definition is empty, backtracking to the corresponding paragraph-level style; if the paragraph-level style is still empty, backtracking to the document's global default style; when a clear attribute definition is obtained at any level, backtracking stops, and the attribute is used as the actual effective style. The output of the style review result also includes intelligent font matching judgment: when comparing the actual effective font with the template font, if the two are semantically equivalent mappings of Chinese font names, it is judged as a match; if the template requires a Chinese font but the actual effective style is a Western font and the content contains Chinese, it is judged as a mismatch. The output of the style review result also includes tolerance comparison of numerical attributes: for font size, line spacing, and indentation attributes, the absolute value of the difference between the actual effective style value and the template baseline value is calculated, and only when the absolute value of the difference is greater than the preset tolerance threshold δ, it is judged as a style inconsistency. The actual effective style also includes representative styles for table elements. By traversing specific rows in the table that are not header rows, the mode of the appearance of style attributes in each cell is counted, and the mode is used as the baseline value of the actual effective style for the table element.

[0013] Preferably, in step S5, injecting annotation tags at the corresponding positions in the target construction plan document specifically includes: Based on the content block ID and XML offset position recorded in the audit conclusion, the audit comments are encapsulated in a format conforming to the OpenXML standard. <w:comment>The node is then mapped back to the underlying .xml code of the target construction plan document.

[0014] Beneficial effects: 1. Through a three-level style backtracking mechanism of "character level - paragraph level - document default level", this invention can penetrate the complex XML nesting structure of Word documents to obtain the actual rendering style of the text, rather than just matching the explicit style name. The system can accurately identify "pseudo-SimSun", non-standard mixed Chinese and Western text, and implicit formatting errors caused by style inheritance, and can discover minor typesetting deviations that are difficult to detect by manual review.

[0015] 2. This invention reduces the "illusion" risk of large language models. It automatically identifies the geographical location of projects through large language models and dynamically drives structured SQL to pre-filter the database of eliminated items. This "semantic recognition + dynamic retrieval" model reduces the massive national standards to a regional subset highly relevant to the project. This not only solves the problem of frequent updates to building codes and large regional differences, but also effectively suppresses the phenomenon of large language models generating false compliance conclusions through precise contextual constraints.

[0016] 3. Balancing the "breadth" and "depth" of retrieval, the system employs a dual-path concurrent mode: structured retrieval based on "chapter path + abstract" and semantic retrieval based on "vector space." Structured retrieval ensures that chapter logic is not lost, while vector retrieval ensures semantic connections across chapters. Through deduplication and reordering of the two sets of results, the system can provide reviewers with logically complete contextual information, effectively avoiding the semantic fragmentation problem caused by traditional RAG technology when slicing long documents.

[0017] 4. Reduced manual review costs: By establishing a coordinate mapping relationship between the review conclusion and the XML offset position of the original document, this invention supports directly injecting AI review comments into the original document in the form of annotations. Users do not need to frequently switch between the review system and the original file to search, achieving accurate problem location and one-click source tracing. Attached Figure Description

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a structural diagram of the intelligent review method for building construction plans based on rule engines and large models of the present invention; Figure 2 This is a flowchart of the intelligent review method for building construction plans based on a rule engine and a large model, which is the basis of this invention. Detailed Implementation

[0019] Example 1 like Figure 1 As shown, the intelligent review method for building construction plans based on rule engines and large models of the present invention consists of operations S110~S130, including: Operating S110, the management and control layer, provides a framework for resource governance, security, and collaboration to ensure the standardization and controllability of the audit process; Operating S120, the core capability layer, encapsulates the technical methods of intelligent auditing, and provides capability support for upper-layer applications in the form of microservices; The S130 is an interactive application layer, the interface between the user and the system, providing an intuitive and convenient operating experience.

[0020] According to an embodiment of the present invention, in operation S110, the management and control layer is mainly responsible for the allocation of underlying system resources, the isolation of user permissions, and the status governance of the review process. After receiving the construction plan document uploaded by the front end, the system first generates a globally unique file_id through the file management module. The original document is stored in a highly available file server, and a related record is established in the structured database to record the document status. The system implements differentiated permission policies for reviewers at different levels. The management layer is responsible for defining the access permissions of the standard template library, ensuring that only authorized users can perform CRUD operations on the "elimination item database" and the "standard specification database," thus ensuring the authority of the review benchmark. The system uses an asynchronous task queue to manage computationally intensive tasks such as vectorization of long documents and semantic recognition of Large Language Models (LLMs), supporting task queuing, progress monitoring, and anomaly cleanup to ensure the stability of the system under high-concurrency scenarios.

[0021] According to an embodiment of the present invention, in operation S120, the core capability layer achieves accurate review through heterogeneous data processing and multi-dimensional semantic comparison. The core engine identifies heading styles and constructs a chapter tree by traversing the underlying XML elements of the Word document. Each content block carries chapter path information and a summary. The system simultaneously constructs a dual-path retrieval system, including structured database retrieval based on SQL and semantic vector retrieval based on embedding technology, and calculates a comprehensive relevance score through a mathematical mapping model. The two search results are then weighted, merged, and sorted as shown in equation (1): (1) in, To assess the relevance of the feature space, cosine similarity is used to calculate the audit item vector. With content block vector The semantic distance, with a value range of [0,1]. For metadata attribute mapping, based on the BM25 or keyword hit weight of the review item name and the content block title / summary, the value range is [0,1]. and For the weighting coefficients, satisfying The classic value is , It can be adjusted according to the review scenario.

[0022] When reviewing construction processes or material plans, the core layer triggers a scenario-based review process: The LLM is driven to perform Named Entity Recognition (NER) on the content of the "Project Overview" section, thereby extracting the geographic entities to which the project belongs; Based on the identified geographic entities, the core layer dynamically generates SQL statements to retrieve a list of prohibited and restricted items that meet the criteria of "applicable nationwide" or "applicable to specific provinces" from the elimination item database; Regional elimination items refer to lists of prohibited or restricted technologies, equipment, and materials regularly published by national and local housing and urban-rural development departments. These lists include three categories: eliminated technologies, prohibited equipment, and restricted materials. The system automatically retrieves the latest regional elimination item data from the official website of the Ministry of Housing and Urban-Rural Development and the websites of provincial and municipal housing and urban-rural development departments every 24 hours. After data retrieval, the system uses MD5 verification to ensure data integrity and compares the new data with the old data, updating only records that have changed. The elimination item database includes the following fields: elimination item ID, type (technology / equipment / material), name, applicable region, reason for elimination, alternative solution, and update time.

[0023] The filtered elimination item information and the candidate content set are input into the LLM (Local Management Module). A prompt guides the LLM to determine if the construction plan carries a risk of violating regional restrictions. An SQL query generated by the LLM based on geographic entities is used to filter applicable elimination items from the regional elimination item database. The query automatically includes two conditions: "nationwide" and the identified province, ensuring that both nationally applicable elimination items and locally specific elimination items are included. An audit conclusion is output only when a match exists in the elimination item database. The audit conclusion undergoes dual verification: first, verification through the elimination item database, and second, semantic similarity verification. If both verifications pass, the conclusion is confirmed. Confidence levels are then calculated. As shown in the following formula: (2) in, The score is matched against the database for the elimination item. A successful match is assigned a value of 1, and a failed match is assigned a value of 0. The semantic similarity is calculated using cosine similarity, ranging from 0 to 1. The conclusion is output when the confidence level of the review decision is higher than 0.8. Through this mechanism, the system effectively avoids the "illusion" of LLM generating non-existent specifications and also solves the problem of the complexity of manually maintaining regional rules.

[0024] The system employs a comprehensive style review and a three-level backtracking mechanism, delving into the underlying XML attributes. It recursively backtracks from character-level styles (Run), paragraph-level styles (Paragraph), and document default styles (Doc Defaults) to obtain the actually effective style attributes (such as font and line spacing). The three-level backtracking mechanism means that when extracting style attributes from the target document, if the current character-level style is empty, it backtracks to the corresponding paragraph-level style; if the paragraph-level style is also empty, it backtracks to the document's global default style; and it stops backtracking when a clear attribute definition is obtained at any level, treating that attribute as the actual effective style. This mechanism effectively solves the style nesting problem caused by style inheritance in Word documents. For complex elements such as tables, the system calculates the representative style of the element by counting the mode of styles in non-empty cells. A tolerance threshold δ is introduced, comparing the extracted actual style value with the standard template style value. If the absolute value of the difference exceeds δ, it is determined that the styles are inconsistent, and its offset position in the XML is recorded.

[0025] According to an embodiment of the present invention, in operation S130, the interactive application layer directly faces the end user and is responsible for transforming the logical operations of the core layer into operable and visualized results. Through the WebUI, users can flexibly select review dimensions according to the characteristics of the current project, while the interactive layer encapsulates the user's selection as configuration parameters and passes them to the core capability layer. In the core capability layer, based on the content block ID associated with the review result, the starting offset of the content block in document.xml is retrieved. The string is then searched for in the text of the content block, and its relative offset within the content block is calculated. By combining the starting offset and the relative offset of the content block, the XML offset position of the fragment is obtained. After receiving the XML offset position returned by the core layer, the system calls the Office component interface to automatically insert annotation tags at the corresponding positions in the original document. <w:comment>Users can view highlighted style deviations and semantic suggestions in real time through the interactive interface. The interactive layer automatically aggregates all review conclusions and generates a solution review report. Finally, the annotated review draft and the solution review report in PDF format are output through the download interface.

[0026] Example 2 like Figure 2 As shown, the system uploads construction plan documents through the front-end interface. The preferred document format is Word (.docx). After receiving the document, the system generates a unique file identifier file_id and stores the original document on the file server for subsequent parsing, annotation injection, and result backtracking.

[0027] The system performs structured parsing on the Open XML structure of the original construction plan document, by traversing the paragraphs in the Word document ( <w:p>The system uses table elements to identify "Heading" or "title" styles and extracts the first three levels of heading structure. It constructs a chapter tree structure to represent the parent-child relationship between chapters, splits the chapter content, and generates multiple content blocks. Each content block contains at least chapter path information and body text.

[0028] After parsing the chapter structure, the system segments the document content based on chapter nodes and the first three levels of headings, dividing the document content into multiple content chunks. Each content chunk is a structure containing at least: chapter path tags (e.g., ["Chapter 1", "Section 1", "1.1 Project Overview"]), chapter number, and main text content. Table cell content in the document is converted into tab-delimited text. For images, their caption text and position index in the document are recorded for subsequent review result location and annotation injection. For each content chunk, an LLM (Language Modeling) abstract is generated. In this embodiment, the large language model can be the ChatGLM3-6B model, which has Chinese semantic understanding, named entity recognition, and text reordering capabilities. The file_id and chunk keys, number, content, and abstract are stored in the file content table of a MySQL database. The BGE (BAAI General Embedding) model is used to convert the content of each content chunk into a high-dimensional vector. The high-dimensional vector, along with its corresponding chunk_id and file_id, is stored in the ChromaDB vector database to construct a semantic index and a vector index, which will be used to support subsequent semantic similarity-based retrieval.

[0029] After document parsing, the system enters the intelligent content review stage. For preset review items, the system performs a dual search: structured retrieval based on chapter titles and abstracts, and semantic retrieval based on vector similarity. The structured information-based search retrieves the titles and abstracts of all content blocks from the file content table using the file_id, forming a candidate metadata list. The semantic vector-based search calculates the similarity between the semantic vector of the review item and the semantic vector of the content block, filtering content blocks with a similarity score no lower than a preset threshold. The results of both search methods are input into the LLM (Local Management Module), which uses constructed prompts to deduplicate and re-sort the content, outputting the final set of relevant content block IDs and their corresponding content.

[0030] After obtaining the relevant content, different scenario-based review processes are initiated based on the type of review item, including review of geographically excluded items, verification of the validity of standards and specifications, and review of general semantic compliance.

[0031] During the review of geographically restricted items, when the review item involves construction technology, equipment plans, or material plans, the system first identifies the geographical region of the project from the relevant content of the project overview. If it cannot be identified, semantic extraction is performed from the project name or project description. If it cannot be clearly identified, a nationally applicable rule is used as a fallback strategy. Subsequently, the restricted item database is filtered according to the geographical region, retaining only those restricted items applicable nationwide or to that specific geographical region. For example: SELECT The system uses the elimination_items function (FROM elimination_items WHERE type = '[process / equipment / material]' AND (province ='national' OR province = '[identified province]'). Finally, the system inputs the filtered elimination item information, candidate content set, and review points into a large model for semantic comparison and analysis. It then outputs a structured review result indicating whether the elimination requirement has been triggered. The review result includes the title, highlighted content, and review outcome.

[0032] During the verification of the validity of standards and specifications, when the audit item involves the basis for compilation or standards and specifications, the system extracts the standard name and standard number from the relevant content and compares and verifies them with the standard database. When the standard name does not exist or the standard number does not match, the system generates a corresponding audit failure result and records the triggering content location.

[0033] When conducting general semantic compliance audits, for audit items that do not fall under the above special scenarios, the system inputs the candidate content and corresponding audit points into the large model for semantic analysis, and requires the large model to output the audit conclusion, trigger content, and reason explanation in a preset format.

[0034] The system performs a global style depth audit based on a three-level style backtracking mechanism, including: First, read the current text. <w:r>under the node <w:rpr>If a style attribute within a tag has a clearly defined attribute, then that attribute will be used as the actual style and backtracking will stop. If character level <w:rpr>If defined as empty, then its parent node paragraph level is read. <w:ppr>If a style attribute exists with an explicit definition, then that attribute will be used as the actual effective style, and backtracking will stop. If paragraph level <w:ppr>If defined as empty, it will ultimately trace back to the document properties file styles.xml. <w:docdefaults>The global default style of the document within the tag, and this attribute is used as the actually effective style.

[0035] Compare the actually effective style information extracted from the target document with the template styles in the template style library, and conduct the following different operation comparisons respectively: Intelligent font matching: During the comparison, not only string matching but also semantic judgment are carried out. For example, when the reference font is "Songti" and the actual font is "SimSun" (both are Chinese fonts), it can be regarded as a match; when the reference is a Chinese font, the actual is "Arial" and the text content contains Chinese, it is determined as a mismatch ("false Songti" error).

[0036] Tolerance comparison: For numerical attributes such as font size and spacing, set a tolerance threshold δ, and the classic value is set to 0.5pt. Only when the absolute value of the difference is greater than δ is it recorded as inconsistent.

[0037] Complex element processing: For tables, compare their border styles, alignment methods, and font consistency between the table headers and content rows; for pictures, compare their caption styles, alignment methods, and wrapping methods.

[0038] Style review is the comparison and verification of style attributes such as fonts, line spacing, and spacing in construction documents. The system unifies and summarizes the intelligent content review results and style review results to form structured review result data, including the problem location, triggering content, and review conclusion. When the actual style is within the tolerance threshold δ of the template style, it is considered to meet the requirements; if it exceeds the tolerance threshold, it is considered not to meet the requirements. According to the integrated review results, the system automatically inserts annotation information at the corresponding location in the original construction plan document to achieve the precise association between the review opinions and the original content. After completing the annotation injection, the system automatically generates a review report, which includes the inspection results and problem descriptions of each review item, and finally completes the review and outputs the results.

[0039] The above embodiments are only the preferred technical solutions of the present invention and should not be regarded as limitations on the present invention. The protection scope of the present invention should be the technical solutions recorded in the claims, including the equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, the equivalent replacement improvements within this scope are also within the protection scope of the present invention.< / w:docdefaults> < / w:ppr> < / w:ppr> < / w:rpr> < / w:rpr> < / w:r> < / w:p> < / w:comment> < / w:comment>

Claims

1. A method for intelligent review of building construction plans based on a rule engine and a large model, characterized in that, Includes the following steps: S1. Parse the underlying XML structure of the target construction plan document, construct a chapter tree, and divide it into content blocks with chapter path tags; S2. Construct a dual-path retrieval system using both structured indexes and semantic vector indexes for each content block to obtain a set of candidate content. S3. Identify the geographical region associated with the target construction plan document, dynamically filter the elimination item database based on the geographical region, and perform semantic comparison between the filtered elimination item information and the candidate content set to output the content review results. S4. Extract the actual effective styles based on a three-level backtracking mechanism of character level, paragraph level and document default level, compare them with the template style library, and output the style review results; S5. Convert the content review results and style review results into XML offset positions, and inject annotation tags at the corresponding positions in the target construction plan document.

2. The intelligent review method for building construction plans based on rule engine and large model as described in claim 1, characterized in that the steps are as follows: In S1, the XML underlying structure parsing of the target construction plan document specifically includes: Traverse paragraph labels and table elements in a Word document, identify heading styles and their levels, and construct a chapter tree representing the parent-child relationship between chapters; Each content block generated by dividing the content by chapter nodes contains a chapter path label, a chapter number, and the main text.

3. The intelligent review method for building construction plans based on rule engines and large models according to claim 1, characterized in that, In step S2, the dual-path retrieval specifically includes: The first candidate set is generated by matching metadata between the name of the review item and the title or summary of the content block; Calculate the cosine similarity between the review item vector and the content block vector, and filter out content blocks that exceed a preset threshold to generate a second candidate set; The first and second candidate sets are deduplicated and semantically reordered using a large language model to determine the final relevant content blocks.

4. The intelligent review method for building construction plans based on rule engines and large models according to claim 1, characterized in that, In step S3, the dynamic filtering of the elimination database by geographical region specifically includes: Geographic entities are extracted from the project overview section of the target construction plan document using a large language model; Dynamic SQL queries are generated based on geographic location entities to filter out lists of prohibited or restricted processes, equipment, and materials applicable to the entire country or a specific geographic region from the elimination item database.

5. The intelligent review method for building construction plans based on rule engines and large models according to claim 1, characterized in that, In step S4, the three-level backtracking mechanism specifically includes: When extracting style attributes from the target document, if the current character-level style is defined as empty, the style will be traced back to the paragraph-level style. If the paragraph-level style is still empty, it will revert to the document's global default style; When a clear attribute definition is obtained at any level, backtracking stops and that attribute is used as the actual effective style.

6. The intelligent review method for building construction plans based on rule engines and large models according to claim 5, characterized in that, The output of the style review results also includes intelligent font matching judgment: when comparing the actual effective font with the template font, if the two are semantically equivalent mappings of Chinese font names, they are judged as a match; if the template requires Chinese fonts but the actual effective style is a Western font and the content contains Chinese characters, it is judged as a mismatch.

7. The intelligent review method for building construction plans based on rule engines and large models according to claim 5, characterized in that, The output of the style review results also includes a tolerance comparison of numerical attributes: For font size, line spacing, and indentation attributes, calculate the absolute value of the difference between the actual effective style value and the template baseline value. Only when the absolute value of the difference is greater than the preset tolerance threshold δ, it is determined that the style is inconsistent. The tolerance threshold δ can be dynamically configured according to different audit dimensions.

8. The intelligent review method for building construction plans based on rule engines and large models as described in claim 5, characterized in that, in practice... The effective styles also include representative styles for table elements. By traversing specific rows in the table that are not header rows, the mode of each cell's style attribute is counted, and the mode is used as the baseline value for the actual effective styles of the table element.

9. The intelligent review method for building construction plans based on rule engines and large models according to claim 1, characterized in that, In step S5, injecting annotation tags at the corresponding locations in the target construction plan document specifically includes: Based on the content block ID and XML offset position recorded in the audit conclusion, the audit comments are encapsulated in a format conforming to the Open XML standard. <w:comment> The node is then mapped back to the underlying .xml code of the target construction plan document.< / w:comment> 10. The intelligent review method for building construction plans based on rule engines and large models according to claim 1, characterized in that, Step S3 also includes standard validity verification, specifically: extracting the standard name and standard number from the candidate content set and comparing them with the standard database. If the standard name does not exist or the standard number is not in the current version status in the database, a failure review conclusion is generated.

Images