Construction plan intelligent examination system and method

Abstract

The present disclosure provides a construction scheme intelligent review system and method. The system comprises: a document analysis module, which analyzes and processes the construction scheme file to generate a text string; a knowledge base retrieval module, which retrieves and processes the text string based on a technical knowledge base, a safety knowledge base and a quality knowledge base to generate three types of retrieval results, which are also managed by positive and negative review opinion knowledge bases; a review opinion generation module, which reasons and processes the text string and the three types of retrieval results based on the data in the positive and negative review opinion knowledge bases to generate review opinions; and a human-computer interaction dialogue module, which processes dialogue response of the review opinions to generate dialogue responses. The present disclosure can realize differentiated intelligent review of construction schemes in three dimensions of technology, safety and quality, and give consideration to continuous optimization of review opinion quality.

Description

Technical Field

[0001] This disclosure relates to the interdisciplinary field of artificial intelligence and engineering document review, and in particular to an intelligent review system and method for construction plans. Background Technology

[0002] Construction plans, as technical guidance documents for on-site construction, are core technical documents for ensuring project safety, quality, and schedule. However, the traditional manual review model has significant shortcomings.

[0003] Current construction plan review technologies each have their own technical shortcomings in terms of review efficiency and quality. For example, the traditional manual review model relies entirely on the personal experience and professional skills of reviewers for the preparation, review, and approval of construction plans. While this type of model has the advantage of flexible human judgment, it suffers from difficulties in standardizing review criteria, overlooking key risks, simplistic retrieval of historical plans, and inconvenience in comparing plans, resulting in overall low review efficiency. Especially against the backdrop of rapid development and strict regulation of nuclear power, the ability of participating units to prepare and review construction plans faces challenges, and the traditional manual review model can no longer meet the needs of large-scale, high-standard construction plan reviews.

[0004] As another approach, rule-based automated review methods proposed in recent years compare construction plans against a pre-compiled rule base one by one. However, while this type of approach has the advantages of fast review speed and uniform standards, the rule base has limited coverage and is difficult to adapt to the diversity and complexity of construction plans. It also lacks semantic understanding of unstructured technical descriptions, resulting in insufficient review depth and making it easy to overlook potential technical risks.

[0005] As another approach, document review methods based on natural language processing technology proposed in recent years divide construction plans into blocks and then match them with standard specification texts based on similarity. However, while this type of approach has a certain degree of semantic matching capability, it has a single review dimension and lacks the ability to conduct differentiated reviews from multiple professional dimensions such as technology, safety, and quality. At the same time, review experience cannot be accumulated and reused, and each review starts from scratch, resulting in the inability to continuously improve the review quality.

[0006] Furthermore, existing construction plan review systems lack a mechanism to distill the quality evaluation experience from review comments into reusable knowledge, and the review systems lack the ability for self-iterative optimization. At the same time, review results are difficult to directly translate into training resources, creating a disconnect between review and training, resulting in low efficiency in learning and mastering construction plans. Summary of the Invention

[0007] This disclosure is based on the above circumstances and aims to provide an intelligent review system and method for construction plans, which can realize differentiated intelligent review of construction plans in three professional dimensions: technology, safety, and quality, so as to improve review efficiency and the uniformity of review standards. Through a knowledge base of positive and negative review opinions, the review system has the ability to self-optimize based on historical review experience, taking into account both the continuous optimization of the quality of review opinions and the automatic generation of training resources.

[0008] According to a first aspect of this disclosure, a construction plan intelligent review system is provided, comprising: a document parsing module M1, configured to receive uploaded construction plan files, parse the construction plan files, and generate and output a text string; a knowledge base retrieval module M2, configured to receive the text string output by the document parsing module M1, and perform retrieval processing on the text string based on a technical knowledge base, a safety knowledge base, and a quality knowledge base, thereby generating and outputting technical retrieval results, safety retrieval results, and quality retrieval results, wherein the knowledge base retrieval module M2 further includes a positive review opinion knowledge base and a negative review opinion knowledge base, wherein the positive review opinion knowledge base stores review opinion data marked as positive, and the negative review opinion knowledge base stores review opinion data marked as negative; and a review opinion generation module M3, configured to receive the text string output by the document parsing module M1 and the technical retrieval results, the safety retrieval results, and the quality retrieval results output by the knowledge base retrieval module M2, wherein the review opinion generation module M3... The system is configured to perform reasoning processing on the text string, the technology search results, the security search results, and the quality search results based on data in the positive review opinion knowledge base and the negative review opinion knowledge base, thereby generating and outputting review opinions; and a human-computer interaction dialogue module M4 is configured to receive the review opinions output by the review opinion generation module M3, perform dialogue response processing on the review opinions, thereby generating and outputting dialogue responses.

[0009] The technical knowledge base, the safety knowledge base, and the quality knowledge base correspond to the technical, safety, and quality dimensions in the construction plan review, respectively, and each is independently constructed and stores reference data for its corresponding professional dimension.

[0010] In another preferred embodiment, the system further includes a training and assessment module M5, which is configured to receive the review opinions output by the review opinion generation module M3, process the review opinions to generate training and assessment content, and thereby generate and output training and assessment content.

[0011] In another preferred embodiment, the document parsing module M1 is further configured to perform database entry processing on the reference files in the technical knowledge base, the security knowledge base, and the quality knowledge base, that is, to parse the files in the knowledge base and save them as text strings.

[0012] In another preferred embodiment, the knowledge base retrieval module M2 includes a knowledge base comprising a technical knowledge base, a security knowledge base, a quality knowledge base, a positive review opinion knowledge base, and a negative review opinion knowledge base.

[0013] In another preferred embodiment, the knowledge base is divided into a local knowledge base and an external knowledge base according to the data source and maintenance method.

[0014] In another preferred embodiment, the local knowledge base includes the positive review opinion knowledge base and the negative review opinion knowledge base. The data in the local knowledge base is generated by the review opinion generation module M3, labeled and categorized by the feedback module M6, and stored therein. It is dynamically updated through an iterative closed loop between the review opinion generation module M3 and the feedback module M6. The local knowledge base is directly managed by the knowledge base retrieval module M2.

[0015] In another preferred embodiment, the external knowledge base includes the technical knowledge base, the security knowledge base, and the quality knowledge base. The data in the external knowledge base is imported from at least one external system, including an external collaboration platform, a quality management system, and a security management system. The knowledge base retrieval module M2 performs the retrieval process on the external knowledge base through the external knowledge base interface. The external knowledge base interface performs vector similarity search based on the text string and returns the matching document fragment as the retrieval result.

[0016] In another preferred embodiment, the data in the external knowledge base is maintained by periodically acquiring new and updated files. The new files are converted into document fragments and stored in the external knowledge base, and the updated files are converted into updated document fragments to replace the original data in the external knowledge base.

[0017] In another preferred embodiment, the data in the local knowledge base is continuously accumulated with each iteration of the iterative loop, and the review opinion generation module M3 performs the reasoning processing based on all the data accumulated in the local knowledge base in each iteration.

[0018] In another preferred embodiment, the technical knowledge base includes a construction plan knowledge base and a technical specification knowledge base; the safety knowledge base includes a health, safety and environment (HSE) risk knowledge base; and the quality knowledge base includes a quality information notification case knowledge base, a standards and specifications knowledge base, and an experience feedback case knowledge base.

[0019] In another preferred embodiment, when the knowledge base retrieval module M2 performs the retrieval process, it sends a request message to the knowledge base interface. The request message includes a query text field, a return quantity field, and a similarity threshold field. The knowledge base interface performs a vector similarity search based on the query text field and returns the matching document fragments and optionally the corresponding similarity scores as the retrieval results.

[0020] In another preferred embodiment, the knowledge base retrieval module M2 is configured to perform the retrieval process on the technical knowledge base, the security knowledge base, and the quality knowledge base in parallel.

[0021] In another preferred embodiment, the retrieval processing performed by the knowledge base retrieval module M2 on the text string further includes combining the text string with the retrieval results into a prompt word string, and outputting the prompt word string to the review opinion generation module M3.

[0022] In another preferred embodiment, the review comment generation module M3 sends the prompt word string to the application programming interface (API) of the large language model. The request message of the API includes a model name field, a prompt word field, and a maximum output length field.

[0023] In another preferred embodiment, the knowledge base retrieval module M2 is configured to periodically acquire newly added and updated reference files, convert the newly added reference files into document fragments and store them in the corresponding knowledge base, and convert the updated reference files into updated document fragments to replace the original data in the corresponding knowledge base.

[0024] In another preferred embodiment, the review opinion generation module M3 performs the inference processing by calling a large language model, which is integrated through local deployment or application programming interface (API) calls.

[0025] According to the intelligent construction plan review system described above in this disclosure, the document parsing module M1 parses and processes the construction plan documents to generate text strings. The knowledge base retrieval module M2 retrieves the text strings based on the technical knowledge base, safety knowledge base, and quality knowledge base to generate three types of retrieval results. The review opinion generation module M3 generates review opinions based on the data in the positive review opinion knowledge base and the negative review opinion knowledge base. This system can construct a complete intelligent review link from document input to review opinion output, enabling differentiated review of construction plans in the three professional dimensions of technology, safety, and quality, improving the review coverage and the uniformity of review standards. At the same time, the positive and negative review opinion knowledge bases enable the review system to have self-optimization capabilities based on historical review experience. Furthermore, the training and assessment module M5 transforms the review results into training resources, realizing the linkage between review and training. Through workflow connections between modules, local / external classification management of the knowledge base, vector similarity retrieval mechanism, and large language model integration, the system possesses comprehensive capabilities for structured data flow, dynamic maintenance of the knowledge base, and intelligent reasoning processing.

[0026] According to another aspect of this disclosure, a construction plan intelligent review system is provided. Based on the above system, the knowledge base retrieval module M2 is further configured to perform retrieval processing on the text string based on the positive review opinion knowledge base and the negative review opinion knowledge base, thereby further generating and outputting positive retrieval results and negative retrieval results. The positive retrieval results and the negative retrieval results, together with the technical retrieval results, the safety retrieval results, and the quality retrieval results, are output to the review opinion generation module M3, so that the review opinion generation module M3 can perform reasoning processing based on these retrieval results.

[0027] According to the intelligent review system for construction plans described in this disclosure, the knowledge base retrieval module M2 can simultaneously search five knowledge bases, using positive and negative historical review opinions as a source of enhanced retrieval. This allows the review opinion generation module M3 to refer to the quality evaluation experience of historical review opinions during reasoning and processing, thereby improving the accuracy and consistency of the review opinions.

[0028] According to another aspect of this disclosure, a construction plan intelligent review system is provided. Based on the above system, the review opinion generation module M3 is further configured to perform pre-reasoning processing on the text string, the technical search results, the safety search results, and the quality search results, thereby generating and outputting pre-review opinions. Furthermore, based on the data associated with the pre-review opinions in the positive review opinion knowledge base and the negative review opinion knowledge base, the pre-review opinions are further optimized and reasoned to generate and output optimized review opinions.

[0029] In another preferred embodiment, after generating and outputting the pre-review opinion, the review opinion generation module M3 outputs the text string, each search result, and the pre-review opinion to the knowledge base retrieval module M2; the knowledge base retrieval module M2 performs searches in the positive review opinion knowledge base and the negative review opinion knowledge base based on the received data, thereby generating and outputting the positive search results and the negative search results to the review opinion generation module M3; the review opinion generation module M3 performs optimization reasoning processing on the pre-review opinion based on the positive search results and the negative search results, thereby generating and outputting optimized review opinions.

[0030] According to the intelligent review system for construction plans described in this disclosure, by dividing the reasoning process into two stages—pre-reasoning and optimized reasoning—preliminary review opinions can be generated quickly in the first stage. In the second stage, targeted optimization is performed based on data in the positive and negative review opinion knowledge base that is precisely associated with the pre-review opinions. This ensures review efficiency while improving the quality and relevance of the review opinions. Furthermore, a loop-like interaction between the review opinion generation module M3 and the knowledge base retrieval module M2 enables secondary precise retrieval driven by the pre-review opinions.

[0031] According to another aspect of this disclosure, a construction plan intelligent review system is provided. Based on the above system, it further includes an opinion feedback module M6, configured to receive the review opinions output by the review opinion generation module M3 and the annotation information for the review opinions, classify and process the review opinions and the annotation information, thereby generating and outputting positive review opinion data and / or negative review opinion data, and storing the positive review opinion data in the positive review opinion knowledge base and the negative review opinion data in the negative review opinion knowledge base.

[0032] In another preferred embodiment, the review opinion generation module M3 and the feedback module M6 form an iterative closed loop, which is configured to repeatedly perform the following processes: the review opinion generation module M3 generates and outputs a current review opinion based on the current data in the positive review opinion knowledge base and the negative review opinion knowledge base; the feedback module M6 receives the current review opinion and the annotation information for the current review opinion, classifies the current review opinion and the annotation information, thereby generating and outputting positive review opinion data and / or negative review opinion data, and stores the positive review opinion data in the positive review opinion knowledge base and the negative review opinion data in the negative review opinion knowledge base, thereby updating the data in the positive review opinion knowledge base and the negative review opinion knowledge base; and in the next iteration, the review opinion generation module M3 generates and outputs the next review opinion based on the updated data in the positive review opinion knowledge base and the negative review opinion knowledge base.

[0033] In another preferred embodiment, the annotation information for the current review opinion is input by the user through the human-computer interaction dialogue module M4.

[0034] In another preferred embodiment, the review opinions and the annotation information, after being classified, can also be classified as low-value opinion data. The low-value opinion data is review opinion data that lacks substantive content. The low-value opinion data is neither stored in the positive review opinion knowledge base nor in the negative review opinion knowledge base.

[0035] In another preferred embodiment, the iterative closed loop iterates more than twice, and the data in the positive review opinion knowledge base and the negative review opinion knowledge base accumulates continuously with each iteration.

[0036] According to the intelligent construction plan review system described in this disclosure, by adding an opinion feedback module M6 to construct a labeling, classification, and dual-database storage mechanism for review opinions, the quality evaluation experience of review opinions can be precipitated into reusable knowledge, enabling the review system to have self-iterative optimization capabilities. Furthermore, through the iterative closed loop between the review opinion generation module M3 and the opinion feedback module M6, the data in the positive and negative review opinion knowledge bases continuously accumulates with each iteration. At the same time, by classifying review opinions into three categories—positive, negative, and low-value—refined experience precipitation is achieved. Low-value opinion data is not stored in any knowledge base to avoid noise interference with review quality.

[0037] According to another aspect of this disclosure, a construction plan intelligent review system is provided. Based on the above system, it further includes an additional conditions module M7, configured to receive input of additional conditions and output the additional conditions to the review opinion generation module M3. The review opinion generation module M3 constrains the reasoning process based on the additional conditions, wherein the additional conditions include at least one of the following: output format requirements, language style requirements, and scope of review focus.

[0038] In another preferred embodiment, the additional condition module M7 outputs the additional conditions to the review opinion generation module M3 before the review opinion generation module M3 generates and outputs the review opinion, that is, before the review opinion generation module M3 performs analysis and processing, and the review opinion generation module M3 imposes constraints based on the additional conditions during the reasoning process.

[0039] In another preferred embodiment, after the review opinion generation module M3 generates and outputs the pre-review opinion, the additional condition module M7 outputs the additional conditions to the review opinion generation module M3 before performing optimization reasoning processing (which may be before or after combining positive and negative review opinions). The review opinion generation module M3 adjusts the generated review opinion based on the additional conditions, thereby generating and outputting the adjusted review opinion, such as the optimized review opinion.

[0040] In another preferred embodiment, the additional conditions module M7 may also output the additional conditions to the review opinion generation module M3 after the review opinion generation module M3 generates and outputs the optimized review opinion. The review opinion generation module M3 adjusts the optimized review opinion based on the additional conditions, thereby generating and outputting the final review opinion.

[0041] According to the intelligent review system for construction plans described in this disclosure, by adding an additional condition module M7, the review opinion generation module M3 is constrained by the additional conditions for reasoning processing. This enables flexible response to personalized review needs, making the generated review opinions more aligned with user intent in terms of output format, language style, and scope of review focus. Furthermore, through a multi-temporal mechanism of additional conditions constraining before reasoning, adjusting after reasoning, or further adjusting optimized review opinions, precise control over the direction of review opinion generation is achieved.

[0042] According to another aspect of this disclosure, a construction plan intelligent review system is provided. Based on the above system, the document parsing module M1 performs the parsing process of the construction plan document as follows: preprocessing the text of the construction plan document to extract an effective range; performing structural parsing on the text within the effective range to obtain the heading level; grouping chapter content based on the heading level to generate a chapter tree relationship; merging and splitting chapters in the chapter tree relationship under length constraints to generate a general segment list; and extracting targeted content from the general segment list based on keywords corresponding to the target review dimension, thereby generating the text string.

[0043] In another preferred embodiment, the construction plan file received by the document parsing module M1 is input in the form of a file object, which includes a file name, a file type identifier, a file access path, and a file size field.

[0044] In another preferred embodiment, in the step of the document parsing module M1 performing the structural parsing of the text of the construction plan document to obtain the heading level, when the heading level is not obtained, the document parsing module M1 splits the entire text of the construction plan document using newline characters as delimiters, and accumulates the split text segments in a loop based on a length threshold. When the accumulated length exceeds the length threshold, the current segment is output, thereby generating the general segment list as backup segments.

[0045] In another preferred embodiment, in the step of merging and splitting chapters in the chapter tree relationship under the length constraint, when the length of a single chapter exceeds the length constraint, the document parsing module M1 retains the title of the chapter at the beginning of each segment after the chapter is split, so that each segment carries the title information of its respective chapter.

[0046] In another preferred embodiment, the targeted content extraction of the general segmented list based on keywords corresponding to the target review dimension includes: receiving a structured content list containing text blocks and table blocks; identifying text blocks containing the keywords and having a heading level; extracting the numbers of the text blocks as starting numbers; collecting text blocks and table blocks after the starting numbers until the numbers exceed the prefix range of the starting numbers; and reorganizing the collected text blocks and table blocks into coherent text according to the format. Targeted content extraction involves extracting content related to the keywords corresponding to the target review dimension from the general segmented list.

[0047] In another preferred embodiment, the document parsing module M1, the knowledge base retrieval module M2, the review opinion generation module M3, the human-computer interaction dialogue module M4, and the training and assessment module M5 are connected through a workflow engine. After the upstream module has completed its execution and output the results, it passes structured data to the downstream module. The structured data includes at least one of file objects, text strings, and JSON arrays.

[0048] In another preferred embodiment, the human-computer interaction dialogue module M4 is configured to annotate the review comments from the review comment generation module M3.

[0049] In another preferred embodiment, the human-computer interaction dialogue module M4 is further configured to receive follow-up questions input by the user, and perform allocation processing based on the follow-up questions, such as retrieving the original text basis corresponding to the review opinion, thereby generating and outputting the original text source information.

[0050] In another preferred embodiment, the training assessment content includes learning points and / or assessment questions.

[0051] In another preferred embodiment, the training and assessment module M5 generates learning points based on the review comments, generates assessment questions based on the review comments, pushes the learning points and assessment questions to the education and training platform according to the training matrix, and dynamically updates the learning points and assessment questions based on the exam results and experience feedback information.

[0052] In another preferred embodiment, the construction plan document is a construction plan document for a nuclear energy EPC project.

[0053] According to the intelligent construction plan review system described in this disclosure, the construction plan document can be parsed into semantically complete and logically independent text segments through the six-step semantic segmentation process of the document parsing module M1. This ensures that the review unit maintains contextual coherence while matching the review focus of each intelligent review module. Furthermore, the robustness of document parsing is enhanced through a backup segmentation strategy, the title carrying mechanism ensures that each segment retains contextual information after segmentation, the workflow engine enables structured data transfer between modules, the source inquiry function improves the credibility of the review results, and the dynamic update mechanism of the training and assessment module enables continuous linkage between review and training.

[0054] Furthermore, this disclosure also provides an intelligent review method for construction plans, comprising: a first step of receiving an uploaded construction plan file, parsing the construction plan file to generate and output a text string; a second step of receiving the text string, performing retrieval processing on the text string based on a technical knowledge base, a safety knowledge base, and a quality knowledge base to generate and output technical retrieval results, safety retrieval results, and quality retrieval results; a third step of receiving the text string and the technical retrieval results, the safety retrieval results, and the quality retrieval results, performing reasoning processing on the text string and the technical retrieval results, the safety retrieval results, and the quality retrieval results based on review opinion data marked as positive stored in a positive review opinion knowledge base and review opinion data marked as negative stored in a negative review opinion knowledge base to generate and output review opinions; and a fourth step of receiving the review opinions, performing dialogue response processing on the review opinions to generate and output a dialogue response.

[0055] In another preferred embodiment, the technical knowledge base, the safety knowledge base, and the quality knowledge base correspond to the technical dimension, safety dimension, and quality dimension in the construction plan review, respectively, and each is independently constructed and stores reference data for the corresponding professional dimension.

[0056] In another preferred embodiment, the method further includes a fifth step: receiving the review comments, processing the review comments to generate training and assessment content, thereby generating and outputting training and assessment content.

[0057] In another preferred embodiment, in the second step, the text string is further searched based on the positive review opinion knowledge base and the negative review opinion knowledge base, thereby generating and outputting positive and negative search results. The positive and negative search results, together with the technical search results, the security search results, and the quality search results, are used for the reasoning process in the third step.

[0058] In another preferred embodiment, in the third step, the text string, the technology search results, the security search results, and the quality search results are first subjected to pre-reasoning processing to generate and output a pre-review opinion. Then, based on the data associated with the pre-review opinion in the positive review opinion knowledge base and the negative review opinion knowledge base, the pre-review opinion is subjected to optimization reasoning processing again to generate and output an optimized review opinion.

[0059] In another preferred embodiment, after generating and outputting the pre-review opinion, the text string, the search results, and the pre-review opinion are used to search in the positive review opinion knowledge base and the negative review opinion knowledge base, thereby generating and outputting the positive search results and the negative search results. Based on the positive search results and the negative search results, the pre-review opinion is optimized and reasoned to generate and output optimized review opinion.

[0060] In another preferred embodiment, after the third step, the method further includes: receiving the review comments and annotation information for the review comments; classifying the review comments and the annotation information to generate and output positive review comment data and / or negative review comment data; storing the positive review comment data in the positive review comment knowledge base and storing the negative review comment data in the negative review comment knowledge base; and in the third step of the next iteration, generating and outputting the next review comment based on the updated data in the positive review comment knowledge base and the negative review comment knowledge base.

[0061] In another preferred embodiment, the review opinions and the annotation information, after being classified, can also be classified as low-value opinion data. The low-value opinion data is review opinion data that lacks substantive content. The low-value opinion data is neither stored in the positive review opinion knowledge base nor in the negative review opinion knowledge base.

[0062] In another preferred embodiment, the annotation information for the current review opinion is input by the user through a dialog response.

[0063] In another preferred embodiment, in the third step, the reasoning process is constrained based on additional conditions, wherein the additional conditions include at least one of the following: output format requirements, language style requirements, and scope of review focus.

[0064] In another preferred embodiment, the additional conditions are input before the examination opinion is generated, i.e., before the examination opinion generation module M3 performs analysis and processing, to constrain the reasoning process; or they are input after the pre-examination opinion is generated, but before the optimization reasoning process (which may be before or after combining positive and negative examination opinions), to adjust the generated pre-examination opinion, thereby generating and outputting an adjusted examination opinion, such as the optimized examination opinion. The additional conditions may also be input after the optimized examination opinion is generated and output, to adjust the optimized examination opinion, thereby generating and outputting the final examination opinion.

[0065] In another preferred embodiment, in the first step, the parsing process includes: preprocessing the text of the construction plan document to extract a valid range; performing structural parsing on the text within the valid range to obtain the heading level; grouping chapter content based on the heading level to generate a chapter tree relationship; merging and splitting chapters in the chapter tree relationship under length constraints to generate a general segment list; and extracting targeted content from the general segment list based on keywords corresponding to the target review dimension, thereby generating the text string.

[0066] In another preferred embodiment, when the title level cannot be obtained, the entire text of the construction plan document is split using a newline character as a delimiter, and the split text fragments are accumulated cyclically based on a length threshold. When the accumulated length exceeds the length threshold, the current segment is output, thereby generating the general segment list as a backup segment.

[0067] In another preferred embodiment, when the length of a single chapter exceeds the length constraint, the title of the chapter is retained at the beginning of each segment after the chapter is divided, so that each segment carries the title information of the chapter to which it belongs.

[0068] In another preferred embodiment, the retrieval process in the second step includes: sending a request message to the knowledge base interface as the query text, the request message including a query text field, a return quantity field, and a similarity threshold field; the knowledge base interface performing a vector similarity search based on the query text field and returning the matching document fragments and optionally the corresponding similarity scores as the retrieval results.

[0069] In another preferred embodiment, in the second step, the text string is combined with the search result to form a prompt word string, and in the third step, the prompt word string is sent to the application programming interface of the large language model for the inference processing, wherein the large language model is integrated by means of local deployment or application programming interface call.

[0070] In another preferred embodiment, the review comments are marked in the fourth step.

[0071] In another preferred embodiment, in the fourth step, a follow-up question input by the user is received, and allocation processing is performed based on the follow-up question, such as retrieving the original text basis corresponding to the review opinion, thereby generating and outputting the original text source information.

[0072] In another preferred embodiment, in the fifth step, the training assessment content includes learning points and / or assessment questions, and the learning points and assessment questions are generated based on the review comments, the learning points and assessment questions are pushed to the education and training platform according to the training matrix, and the learning points and assessment questions are dynamically updated based on the examination results and experience feedback information.

[0073] In another preferred embodiment, the steps are connected via a workflow engine, wherein after the upstream step has been fully executed and outputs the result, it passes structured data to the downstream step, the structured data including at least one of file objects, text strings, and JSON arrays.

[0074] According to the intelligent review method for construction plans described in this disclosure, a complete intelligent review process is achieved through five steps corresponding to the system, from document input to review opinion output and then to the generation of training and assessment content. Through various preferred examples, the method implements technical means such as five-database synchronous retrieval, two-stage reasoning, iterative closed-loop feedback, additional condition constraints, six-step semantic segmentation, vector similarity retrieval, large language model integration, original source inquiry, and dynamic updates of training, so that the method has the same technical effect as the system.

[0075] In addition, this disclosure also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, which is executed by the processor to implement the above-described intelligent review method for construction plans.

[0076] This disclosure also provides a computer program product, including a computer program that is executed by a processor to implement the above-described intelligent review method for construction plans.

[0077] This disclosure also provides a computer-readable storage medium storing a computer program, which is executed by a processor to implement the above-described intelligent review method for construction plans. Attached Figure Description

[0078] For a more complete understanding of the invention, reference can be made to the following description of exemplary embodiments taken in conjunction with the accompanying drawings, in which: Figure 1 This is a schematic diagram of the main architecture of the intelligent review system for construction plans according to one embodiment of the present disclosure.

[0079] Figure 2 This is a flowchart illustrating the operation of an intelligent review system for construction plans according to another embodiment of this disclosure.

[0080] Figure 3 This is a flowchart illustrating the semantic segmentation process of the document parsing module according to one embodiment of this disclosure. Detailed Implementation

[0081] The present invention will be further described below with reference to specific embodiments and accompanying drawings. More details are set forth in the following description in order to provide a full understanding of the present invention. However, the present invention can obviously be implemented in many other ways different from those described herein. Those skilled in the art can make similar extensions and derivations based on actual application situations without departing from the spirit of the present invention. Therefore, the scope of protection of the present invention should not be limited by the content of this specific embodiment.

[0082] First, refer to Figure 1 and Figure 2 The main architecture of the intelligent review system for construction plans according to one embodiment of this disclosure will be described.

[0083] Figure 1 A schematic diagram of the main architecture of a construction plan intelligent review system according to an embodiment of this disclosure is shown. Figure 1 As shown, the intelligent review system for construction plans mainly includes a document parsing module M1, a knowledge base retrieval module M2, a review opinion generation module M3, a human-computer interaction dialogue module M4, and a training and assessment module M5.

[0084] The document parsing module M1, knowledge base retrieval module M2, review comment generation module M3, human-computer interaction module M4, and training and assessment module M5 are connected through a workflow engine. After the upstream module has fully executed and output its results, it passes structured data to the downstream module. This structured data includes at least one of the following: file objects, text strings, and JSON arrays.

[0085] The workflow engine described here refers to a software engine that organizes data processing flows in the form of a node graph. Each node represents a processing module, and nodes are connected via data lines. The output data of the upstream node serves as the input data for the downstream node. The workflow engine is responsible for executing the processing tasks of each module sequentially or in parallel according to the dependencies between nodes.

[0086] As an example, the aforementioned workflow engine may employ the Dify intelligent agent workflow platform, but this disclosure is not limited to a specific workflow engine implementation.

[0087] 1.1 Document Parsing Module M1

[0088] The document parsing module M1 is configured to receive uploaded construction plan files, parse and process the construction plan files, and generate and output text strings.

[0089] Specifically, the document parsing module M1 receives the uploaded construction plan file, and according to the file format (e.g., PDF, Word, scanned document, etc.), calls the corresponding document parsing engine to extract the text content from the file. The extracted text content is then combined into a plain text string, the number of characters of which is equal to the number of characters in the original document. The text string generated by the document parsing module M1 is then transmitted to the entry point of the knowledge base retrieval module M2.

[0090] The construction plan document mentioned here is input in the form of a file object. The file object is a JSON data structure that always contains four fields: name (string filename), mime_type (string file MIME type, such as text / plain, application / pdf), url (the access path of the string file in temporary storage), and size (integer file size in bytes). This file object is transmitted completely to the input port of the document parsing module M1 by the workflow engine at the same time as triggering its execution.

[0091] The document parsing module M1 processes construction plan documents through semantic segmentation. This semantic segmentation integrates multi-level directory structures, chapter functional attributes, and professional boundaries to formulate a semantically complete and logically independent segmentation strategy. This ensures that the review unit maintains contextual coherence while matching the review focus of each intelligent review module.

[0092] Reference Figure 3 The semantic segmentation process described above specifically includes the following steps S1 to S6: S1. Text preprocessing and effective range extraction: The original text string arg1 is split into a list of paragraph-level strings temp_list using the split method with double newline characters "\n\n" as the delimiter. This list is then traversed, and the string find method is used to check if each paragraph contains the substring "# directory" or "# directory", marking the first matching position as start_index. Subsequently, temp_list is sliced ​​using temp_list[start_index:], discarding any non-text paragraphs such as cover pages or approval information that may be present at the beginning. The remaining paragraph list is then reassembled into a single string temp_str using "\n\n".join, which is then passed as the payload to the subsequent structure parsing function.

[0093] S2. Document Structure Parsing and Heading Level Retrieval: The `mistune.create_markdown()` function instantiates a Markdown renderer and converts the `temp_str` obtained in step S1 into an HTML document `html`. Using the BeautifulSoup constructor and specifying the `html.parser` parser, a Document Object Model (DOM) tree is constructed. All heading tags in the tree are extracted at once using `soup.find_all(["h1","h2","h3","h4","h5","h6"])`, returning a list of heading nodes named `headers`. The `name` attribute value of each element (e.g., "h2") provides its hierarchy depth, and the `get_text()` method provides the heading text.

[0094] S3. Chapter Content Aggregation and Title Tree Generation: First, if the headers list is not empty, the previous_sibling attribute is used to backtrack all sibling nodes before the first title, collect their text content and store it in the list pre_content, and add it to the sections list as a record item with level 0 and empty title. Then, for each title node in headers, the next_sibling chain is used to traverse the subsequent sibling nodes until another title node in headers is encountered; during this period, the text of the node with the name attribute is extracted and stored in the content list content. Thus, a dictionary sequence sections containing the structure {"level": int(header.name[1]), "title": title, "content": content} is constructed, forming a chapter tree relationship.

[0095] S4. Semantic unit merging and splitting under length constraints: Initialize the merged text string `temp_text` to empty, and the output list `merge_list` to empty. Iterate through the `sections` list, and for each chapter object, assign its title (marked with "#"). The `level + " " + title` format and the content paragraphs (joined using "\n\n".join(content)) are combined to form the complete Markdown block `all_content`. The processing branches are as follows: If the length of the string after concatenating temp_text and all_content does not exceed the preset chunking threshold chunk_size, then perform cumulative concatenation and continue to the next chapter.

[0096] If the threshold is exceeded, first add the accumulated temp_text to merge_list and clear temp_text; then check if the length of all_content itself is less than chunk_size: if so, set it as the starting point of the new temp_text; otherwise, fix the chapter title as the paragraph beginning, and concatenate each paragraph in content with the title. Whenever the accumulated length reaches the threshold, output the current segment and retain the title to start again, ensuring that the context title is still carried after the large chapter is split. This process is completed through continuous if-else and continue control flow.

[0097] S5. General segmented list generation and backup strategy: The elements in the merge_list obtained in step S4 are all stripped using the strip() function to remove whitespace and filter out empty strings, resulting in the final general segment list merge_list. If this list is empty (e.g., the document is missing a title, causing the segmentation to fail), the backup function split_by_return(text, chunk_size) is called: this function first splits the entire text using "\n\n", then accumulates the segments through a loop. When the accumulated length exceeds chunk_size, the current segment is output and reset, ensuring a minimum number of segments are output. Finally, a non-empty segment list is returned.

[0098] S6. Targeted content extraction based on specific review dimensions: The function receives a JSON string `arg2`, which is parsed by `json.loads` into a list of objects named `content_list`. Each object contains fields such as `type` (text / table), `text` (text content), and `text_level` (heading level). It maintains a boolean flag `start_rel` and a prefix number `start_para`. The function iterates through `content_list`. When start_rel is false, check if the current block is a text block containing the keyword "quality" and whose text_level is not empty; if it is, add the block to the result set rel_text_list and use the regular expression r"\d+(\.\d+). Extract the number from the beginning of the text as start_para, and set start_rel to true.

[0099] When start_rel is true, extract the number of the subsequent text block (if any). If the number exists and is not prefixed with start_para (i.e., it crosses the current chapter sequence), exit the collection state; otherwise, continue collecting text blocks and table blocks.

[0100] After the traversal is complete, the collected `rel_text_list` is reorganized by type: text blocks are padded with `#` symbols according to `text_level` to convert them into Markdown heading format; table blocks are concatenated with the `table_caption`, `table_body`, and `table_footnote` fields. Finally, a coherent plain text `all_rel_content` is returned as the focused content for a specific review domain. The output of step S6 is a context-coherent text specifically serving a particular review domain, which can be directly used by the large language model for review in that domain.

[0101] The target review dimensions mentioned here refer to technical, safety, or quality dimensions. Different review dimensions correspond to different keywords. For example, when the target review dimension is quality, the keyword is "quality"; when the target review dimension is safety, the keyword is "safety". By extracting targeted content from the general segmented list based on the keywords corresponding to the target review dimension, text content related to specific review dimensions can be extracted from the construction plan, enabling the subsequent review opinion generation module M3 to generate specialized review opinions for specific dimensions.

[0102] Furthermore, the document parsing module M1 is also configured to process reference documents in the technical, safety, and quality knowledge bases for import. This involves parsing the reference documents in the knowledge bases using the same semantic segmentation process described above, saving them as text strings before storing them in the corresponding knowledge base. This means that semantic segmentation is used not only for uploading and parsing construction plans but also for processing various reference documents related to technology, safety, and quality when importing them into the knowledge base.

[0103] As an example, the document parsing module M1 has completed the intelligent parsing and optimization of over 3,000 historical construction plans in practical engineering applications, and incorporated them into the knowledge base as a reference for review. The document parsing module M1 deploys multi-channel document parsing services, supporting the parsing of multiple images, texts, and tables in scanned signed construction plans, thus improving the accuracy of text recognition and processing efficiency.

[0104] As an example, the document parsing module M1 also provides editing functions, allowing users to edit the parsed text segments to correct potential errors or adjust their formatting.

[0105] As a document parsing module M1, it can, for example, employ a learned model formed by learning from a pre-defined dataset using machine learning or deep learning techniques. Specifically, a dataset is constructed by using construction plan documents as samples and the segmentation results corresponding to each sample as labels. Based on this, a pre-defined model (e.g., neural network, decision tree, support vector machine, etc.) is trained using these datasets to form the document parsing module M1, which serves as the learned model.

[0106] Preferably, the document parsing module M1 imposes stronger length constraints on the parsing of technology-related content than on security and quality-related content; for example, the parsing length of each segment of technology-related content is shorter than that of security and quality-related content. Similarly, the document parsing module M1 imposes stronger length constraints on the parsing of technical knowledge base documents than on the parsing of security and quality knowledge base documents; for example, the parsing length of each segment of the technical knowledge base is shorter than that of the security and quality knowledge base documents.

[0107] In another preferred approach, the technical, quality, and safety-related content is first segmented and parsed using a uniform length constraint, and then the technical-related segments are further segmented and parsed.

[0108] 1.2 Knowledge Base Retrieval Module M2

[0109] The knowledge base retrieval module M2 is configured to receive a text string output by the document parsing module M1 and send this text string as a query term to the knowledge base interface via an HTTP POST request. The knowledge base interface performs a vector similarity search in the vector database based on the query term and returns matching document fragments as the search results. The request message includes a query text field (query), a return quantity field (top_k), and a similarity threshold field (score_threshold). For example, the return quantity field top_k has a fixed value of 5, and the similarity threshold field score_threshold has a fixed value of 0.7. However, these values ​​are not restrictive and can be adjusted according to actual needs.

[0110] The response message or search results returned by the knowledge base interface contain an array of records. Each element of the array is a JSON object that always contains a content field (the retrieved text block) and a similarity score field (score, the similarity score).

[0111] The knowledge base retrieval module M2 also performs retrieval processing on text strings by combining the text string with the retrieval results into a prompt word string, and then outputs the prompt word string to the review opinion generation module M3.

[0112] The knowledge base of the knowledge base retrieval module M2 includes a technology (T) knowledge base, a safety (S) knowledge base, a quality (Q) knowledge base, a positive review opinion knowledge base, and a negative review opinion knowledge base. The technology, safety, and quality knowledge bases mentioned here correspond to the technical, safety, and quality dimensions in construction plan review, respectively, and each is independently constructed and stores reference data for its corresponding professional dimension. The positive review opinion knowledge base stores review opinion data marked as positive, and the negative review opinion knowledge base stores review opinion data marked as negative.

[0113] The knowledge base retrieval module M2 includes knowledge bases categorized into local knowledge bases and external knowledge bases based on data source and maintenance method.

[0114] In the first implementation, the technical knowledge base, security knowledge base, quality knowledge base, positive review opinion knowledge base, and negative review opinion knowledge base are all included in the external knowledge base. The external knowledge base is a vector database set up in an external system, such as deployed on an external webpage, which stores all the content of reference data for the three professional dimensions (T / S / Q) and historical positive and negative review opinion data. As an example, the knowledge base retrieval module M2 accesses the external knowledge base interface via an HTTP POST request. The external knowledge base interface performs a vector similarity search in the vector database based on the query terms, returning document fragments semantically similar to the query terms and their corresponding similarity scores; this vector database stores the three types of reference data (T / S / Q) and historical positive and negative review opinion data, with a data volume of several thousand entries. The external knowledge base connects to the local knowledge base through an external knowledge base interface (i.e., the external knowledge base API). The local knowledge base sends a text string as a query term to the external knowledge base through the external knowledge base interface, and the external knowledge base performs a vector similarity search and returns the matching document fragments as the search results.

[0115] The local knowledge base serves as a temporary storage for search results and text strings retrieved from external knowledge bases during the operation of the knowledge base retrieval module M2. Specifically, after receiving a text string output by the document parsing module M1, the knowledge base retrieval module M2 sends this text string as a query term to the external knowledge base interface and optionally to the local knowledge base for retrieval. The search results returned by the external knowledge base (i.e., the records array), along with the original text string output by the document parsing module M1, are stored in the internal storage area of ​​the knowledge base retrieval module M2. For example, the external knowledge base API returns a JSON response body containing a records array; each element of this array is a JSON object, which always contains two fields: content (a string, the retrieved text block) and score (a floating-point number, the similarity score); the knowledge base retrieval module M2 stores the entire records array, along with the original text string output by the document parsing subsystem, in the internal shared storage area of ​​the workflow. This internal storage area is the local knowledge base. Essentially, it is a temporary data storage area within the knowledge base retrieval module M2. It is used to temporarily store intermediate data during the data transfer process between nodes of the workflow engine. When the review opinion generation module M3 starts, it reads the aforementioned data from this internal storage area.

[0116] The local knowledge base described here as a temporary knowledge base does not mean that it does not store any persistent data. In one implementation, the local knowledge base may also store a portion of reference data (e.g., a small amount of data), such as data with low replacement probability (i.e., data that is not easily updated or iterated), or more important data (i.e., data with wide applicability). This portion of reference data can be a subset of all reference data stored in the external knowledge base, or it can be data different from the external knowledge base. When the external knowledge base stores all the data content of the local knowledge base and the data volume is larger, for example, the local knowledge base stores dozens of data entries while the external knowledge base stores all the content (e.g., thousands of data entries), in actual operation, the knowledge base retrieval module M2 prioritizes performing retrieval processing on the external knowledge base through the external knowledge base interface to obtain more comprehensive and richer retrieval results. When the local knowledge base stores data different from the external knowledge base, the knowledge base retrieval module M2 can choose to perform retrieval processing on either the local or external knowledge base first, or it can perform retrieval processing on both simultaneously and merge the retrieval results.

[0117] The data in the external knowledge base is maintained by periodically acquiring new and updated files. New files are converted into document fragments and stored in the external knowledge base, while updated files are converted into updated document fragments to replace the original data in the external knowledge base. The data in the positive and negative review opinion knowledge bases in the external knowledge base accumulates with each iteration of the iterative loop, and the review opinion generation module M3 performs reasoning processing based on all the accumulated data in the positive and negative review opinion knowledge bases in each iteration.

[0118] In the second implementation, the local knowledge base includes a positive review opinion knowledge base and a negative review opinion knowledge base. Data in the local knowledge base is generated by the review opinion generation module M3, labeled and categorized by the feedback module M6, and then stored. It is dynamically updated through an iterative closed loop between the review opinion generation module M3 and the feedback module M6. The local knowledge base is directly managed by the knowledge base retrieval module M2.

[0119] The external knowledge base includes a technical knowledge base, a security knowledge base, and a quality knowledge base. Data in the external knowledge base is imported from at least one external system, including an external collaboration platform, a quality management system, and a security management system. The knowledge base retrieval module M2 performs retrieval processing on the external knowledge base through an external knowledge base interface. This interface performs vector similarity searches based on text strings and returns matching document fragments as the retrieval results.

[0120] The data in the external knowledge base is maintained by periodically retrieving new and updated files. New files are converted into document fragments and stored in the external knowledge base, while updated files are converted into updated document fragments to replace the original data in the external knowledge base.

[0121] The data in the local knowledge base continues to accumulate with each iteration of the iterative closed loop, and the review opinion generation module M3 performs reasoning processing based on all the data accumulated in the local knowledge base in each iteration.

[0122] The following provides a detailed explanation of the specific content of each knowledge base.

[0123] The technical knowledge base includes a construction plan knowledge base and a technical specification knowledge base. The construction plan knowledge base contains over 3,000 construction plans published across various projects on the enterprise's business collaboration and management platform, all processed using the aforementioned semantic segmentation algorithm. The technical specification knowledge base extracts construction-related specifications from the technical specifications of the projects where the plans are located.

[0124] The safety knowledge base includes the HSE risk knowledge base. Built upon data from the HSE management module of the enterprise's business collaboration and management platform, the HSE risk knowledge base integrates high-risk operation judgment standards, the HSE risk database (including 82 general operation categories in the "General Operation Risk Database," 225 high-risk operations in the "Full-Cycle Safety Production High-Risk Operation Standardization Database," and a TOP risk list), providing standardized basis for contractors to prepare job hazard analyses (JHAs) online.

[0125] The quality knowledge base includes a quality information notification case knowledge base, a standards and specifications knowledge base, and an experience feedback case knowledge base. The quality information notification case knowledge base covers 380 quality information notification cases since 2023. The standards and specifications knowledge base covers the standards and specifications applicable to the projects in which the solutions are implemented. The experience feedback case knowledge base compiles internal and external experience feedback cases, professional experience feedback cases, top-tier quality case studies from regulatory bodies, and a series of construction process guidelines.

[0126] The positive review opinion knowledge base includes 131 sets of expert review opinions on major first-level construction plans for various projects, as well as high-quality AI review opinions that have undergone manual iterative review. The negative review opinion knowledge base includes unfavorable AI review opinions that have undergone manual iterative review.

[0127] 1.3 Review Comments Generation Module M3

[0128] The review opinion generation module M3 is configured to receive the text string output by the document parsing module M1 and the technical search results, security search results, and quality search results output by the knowledge base retrieval module M2. The review opinion generation module M3 is configured to perform reasoning processing on the text string, technical search results, security search results, and quality search results based on the data in the positive review opinion knowledge base and the negative review opinion knowledge base, thereby generating and outputting review opinions.

[0129] Specifically, the review opinion generation module M3 reads two pieces of data from the workflow's internal shared storage: one is a UTF-8 encoded document parsing string, and the other is a list of knowledge base retrieval results in JSON array format, such as {"content":"Extremely Dangerous Construction Plan", "score": 0.89}. These two parts are combined into a complete prompt string. For example, the prompt string might be: "You are a nuclear power plant construction expert. The user-uploaded document content is as follows: {text}, and the relevant information retrieved from the construction plan database is as follows: {knowledge}. Please generate a review and analysis summary of the construction plan based on the above information."

[0130] The review comment generation module M3 sends the aforementioned prompt string to the application programming interface (API) of the large language model. The request message from this API includes a model name field (model), a prompt field (prompt), and a maximum output length field (max_tokens). For example, the maximum output length field has a fixed value of 2000.

[0131] The review comment generation module M3 performs inference processing by calling a large language model. The large language model is integrated either through local deployment or via API calls.

[0132] As one example, the large language model in this disclosure can be the Qwen3-235B-A22B-Thinking large language model released in July 2025, with a total of 235 billion parameters and a context processing capacity of 262,144 tokens. As another example, the DeepseekR1 large language model released in January 2025 can be used, with a total of 671 billion parameters and a maximum context processing capacity of 256,000 tokens. However, this is not limiting; other systems and methods integrated into the Agent workflow that use large language models for construction plan review are all within the scope of this disclosure.

[0133] The review comments generation module M3 has four main functions: understanding context, logical reasoning, calling the knowledge base, and finally generating natural language answers and structured output.

[0134] The examination opinion generation module M3 performs reasoning processing based on data from the positive and negative examination opinion knowledge bases. This means that during the reasoning process, the module M3 can access historical examination opinion data from both knowledge bases and use this data to guide the generation direction of the current examination opinion. The data in the positive examination opinion knowledge base allows the module M3 to reference the output direction and logical patterns of historical high-quality examination opinions when generating its own opinions; the data in the negative examination opinion knowledge base helps the module M3 avoid generating examination opinions containing factual errors or logical biases.

[0135] 1.4 Two Path Variations of Review Comment Generation Module M3

[0136] In the above embodiments, the examination opinion generation module M3 performs reasoning processing based on data from the positive examination opinion knowledge base and the negative examination opinion knowledge base. Regarding the way data from the positive and negative examination opinion knowledge bases enters the examination opinion generation module M3, this disclosure provides the following two variations.

[0137] In the first variation, for example, corresponding to the first implementation of the knowledge base retrieval module M2 described above, such as... Figure 1 As shown, the knowledge base retrieval module M2 performs retrieval processing on text strings based on a technical knowledge base, a security knowledge base, and a quality knowledge base, thereby generating and outputting technical retrieval results, security retrieval results, and quality retrieval results. The knowledge base retrieval module M2 is configured to simultaneously perform retrieval processing on text strings based on both a positive review opinion knowledge base and a negative review opinion knowledge base, thereby further generating and outputting positive and negative retrieval results. These positive and negative retrieval results, along with the technical, security, and quality retrieval results, are then output to the review opinion generation module M3, enabling the review opinion generation module M3 to perform reasoning processing based on these retrieval results. Of course, this first modification also applies to the second implementation of the knowledge base retrieval module M2 described above.

[0138] Specifically, in this first variation, the knowledge base retrieval module M2 uses text strings as query terms and performs vector similarity searches in five knowledge bases: the technical knowledge base, the security knowledge base, the quality knowledge base, the positive review opinion knowledge base, and the negative review opinion knowledge base. The retrieval results from the five knowledge bases are respectively used as technical retrieval results, security retrieval results, quality retrieval results, positive retrieval results, and negative retrieval results, and are output together to the review opinion generation module M3. The knowledge base retrieval module M2 can perform retrieval processing in parallel on the technical (T) knowledge base, the security (S) knowledge base, the quality (Q) knowledge base, the positive review opinion knowledge base, and the negative review opinion knowledge base.

[0139] The review opinion generation module M3 receives all five categories of search results, combines them with text strings to form complete prompt word strings, and then sends them to the large language model for inference processing. Positive search results provide the review opinion generation module M3 with a reference direction for historical high-quality review opinions, while negative search results indicate review errors and logical deviations that need to be avoided.

[0140] In this first variation, the data in the positive and negative review opinion knowledge base is processed by the knowledge base retrieval module M2 and then enters the review opinion generation module M3 as retrieval results. In other words, the positive and negative review opinion knowledge base functions as a pre-retrieval enhancement source. The advantage of this approach is that all retrieval results are simultaneously input into the large language model in a single reasoning process, enabling the large language model to comprehensively reference T / S / Q professional reference data and historical positive and negative review opinions within the same context, generating comprehensive and accurately oriented review opinions.

[0141] In the second deformation, such as Figure 2 As shown, for example, corresponding to the second embodiment of the knowledge base retrieval module M2 described above, the review opinion generation module M3 is configured to perform pre-reasoning processing on text strings and technical retrieval results, security retrieval results, and quality retrieval results to generate and output pre-review opinions. Furthermore, based on data associated with the pre-review opinions in the positive and negative review opinion knowledge bases, the pre-review opinions are further optimized through reasoning processing to generate and output optimized review opinions. Of course, this second variation also applies to the first embodiment of the knowledge base retrieval module M2 described above.

[0142] The knowledge base retrieval module M2 can perform retrieval processing on the technology (T) knowledge base, security (S) knowledge base, and quality (Q) knowledge base in parallel. This means that a text string can be sent to three knowledge bases for retrieval at the same time, and the retrieval results from all three knowledge bases can be returned simultaneously, thereby improving retrieval efficiency.

[0143] Specifically, in this second variation, the reasoning process of the examination opinion generation module M3 is divided into two stages. In the first stage, the examination opinion generation module M3 performs pre-reasoning processing based on text strings and T / S / Q three-category search results to generate a pre-examination opinion. In the second stage, the examination opinion generation module M3 performs optimized reasoning processing on the pre-examination opinion based on data associated with the pre-examination opinion in the positive and negative examination opinion knowledge bases to generate an optimized examination opinion.

[0144] The "data associated with the pre-examination opinion" mentioned here refers to historical examination opinion data that is semantically similar to the technical topics, examination conclusions, or examination basis involved in the pre-examination opinion, in the positive examination opinion knowledge base and the negative examination opinion knowledge base.

[0145] In one embodiment of this second variation, after generating and outputting the preliminary review opinion, the review opinion generation module M3 outputs the text string, each search result, and the preliminary review opinion to the knowledge base retrieval module M2, such as a local knowledge base, more specifically a positive review opinion knowledge base and a negative review opinion knowledge base; the knowledge base retrieval module M2 performs searches in the positive review opinion knowledge base and the negative review opinion knowledge base based on the received data, thereby generating and outputting positive search results and negative search results to the review opinion generation module M3; the review opinion generation module M3 performs optimization reasoning processing on the preliminary review opinion based on the positive search results and the negative search results, thereby generating and outputting optimized review opinions.

[0146] In this second variation, data from the positive and negative review opinion knowledge base is introduced only after the preliminary review opinion is generated, meaning the knowledge base functions as a post-optimization and enhancement source. The advantage of this approach is that the first stage rapidly generates preliminary review opinions, and the second stage uses the preliminary review opinion to accurately retrieve historical positive and negative opinions for targeted optimization, thereby improving the quality and relevance of the review opinions while ensuring review efficiency.

[0147] The first and second variations mentioned above can be used independently or in combination. When used in combination, the knowledge base retrieval module M2 simultaneously searches and outputs five types of retrieval results in the five knowledge bases during the initial retrieval phase (first variation). The review opinion generation module M3 performs pre-reasoning processing based on the five types of retrieval results to generate pre-review opinions, and then performs optimization reasoning processing based on data in the positive and negative review opinion knowledge bases that are more accurately related to the pre-review opinions (second variation), thereby generating the final optimized review opinions.

[0148] 1.5 Human-Computer Interaction Dialogue Module M4

[0149] The human-computer interaction dialogue module M4 is configured to receive the review comments output by the review comment generation module M3, process the review comments through dialogue response, and thus generate and output the dialogue response.

[0150] Specifically, the human-computer interaction dialogue module M4 receives structured or unstructured review comments output by the review comment generation module M3, converts them into natural language dialogue responses, and engages in multi-turn, context-coherent interactive dialogues with the user. A user feedback collector is embedded in the dialogue interface, and the large language model receives the feedback results and performs corresponding feedback learning.

[0151] The M4 human-computer interaction dialogue module provides users with, for example, analytical fragments of the construction plan under review, and corresponding modification suggestions, including, for example, positive, negative, and invalid annotation selections, for users, such as experts, to review. The M4 human-computer interaction dialogue module also provides dialog boxes for users to input and engage in interactive dialogue.

[0152] The human-computer interaction dialogue module M4 is further configured to receive follow-up questions from the user, for example, via a dialog box. The human-computer interaction dialogue module M4 extracts prompt words from the user input. The human-computer interaction dialogue module M4 is further configured to determine the relevance between the prompt words and the review comments, and select different processing paths based on the relevance determination result.

[0153] When the prompt words are associated with the review comments, the human-computer interaction dialogue module M4 outputs the prompt words to the review comment generation module M3. The review comment generation module M3 uses the prompt words as prompt word constraints, and re-reasons the construction plan based on the prompt word constraints, thereby generating and outputting updated review comments.

[0154] When the prompt words are not related to the review comments, the human-computer interaction dialogue module M4 outputs the prompt words to the knowledge base retrieval module M2 or the external processing module. The knowledge base retrieval module M2 performs a search based on the prompt words at least one of the following: technical knowledge base, security knowledge base, quality knowledge base, positive review comment knowledge base, and negative review comment knowledge base, thus outputting the corresponding search results. Alternatively, the external processing module performs an external query based on the prompt words, outputting the corresponding query results, and then outputs the search or query results to the review comment generation module M3 to generate and output the corresponding review comments. As an example, the human-computer interaction dialogue module M4 retrieves the original text supporting the review comments based on follow-up requests, thereby generating and outputting the original source information. The original source follow-up function mentioned here means that users can ask about the generation process and original text supporting the review comments through an interactive window. The system's intelligent agent remembers the supporting documents and review logic of the review process, thereby assisting users in judging the reference value of the comments. As an example, the aforementioned external processing module is an external artificial intelligence service interface, which performs natural language question-and-answer processing based on the prompt words to output query results.

[0155] The determination of relevance mentioned here is based on whether the prompt words contain keywords that match the technical subject matter, examination conclusions, or examination basis of the examination opinion. If the prompt words contain technical subject matter related to the current examination opinion, they are deemed relevant; if the prompt words do not contain any content related to the current examination opinion, they are deemed irrelevant.

[0156] The human-computer interaction dialogue module M4, together with the review opinion generation module M3 and / or the knowledge base retrieval module M2, forms a constrained iterative closed loop. This constrained iterative closed loop is configured to repeatedly execute the following processes: the human-computer interaction dialogue module M4 receives the current prompt word and outputs it to the review opinion generation module M3 and / or the knowledge base retrieval module M2; the review opinion generation module M3 constrains the reasoning process based on the current prompt word and / or the knowledge base retrieval module M2 performs a retrieval based on the current prompt word, thereby generating and outputting the current review opinion; and the human-computer interaction dialogue module M4 receives the adjusted prompt word and outputs it to the review opinion generation module M3 and / or the knowledge base retrieval module M2, so that the review opinion generation module M3 and / or the knowledge base retrieval module M2 re-performs the reasoning process and / or retrieval based on the adjusted prompt word, thereby generating and outputting the adjusted review opinion.

[0157] The above-mentioned constraint iteration closed loop iterates more than twice, and the review opinion generation module M3 generates and outputs the corresponding review opinion based on the latest prompt words in each iteration until the output review opinion meets the predetermined conditions.

[0158] The aforementioned constraint iterative closed loop and the feedback module M6 iterative closed loop can be executed in parallel or sequentially.

[0159] As an example, the system of this disclosed embodiment is embedded with the intranet Zhihuixing AI assistant, adopts a front-end and back-end separation architecture, and the front-end provides a visual operation interface to support manual intervention, result correction and feedback loop.

[0160] 1.6 Training and Assessment Module M5

[0161] The training and assessment module M5 is configured to receive review comments from the review comment generation module M3, process these comments to generate training and assessment content, and then generate and output the training and assessment content. This training and assessment content includes learning points and / or assessment questions.

[0162] Specifically, the training and assessment module M5 generates learning points based on review comments, and then generates assessment questions based on those comments. The learning points and assessment questions are pushed to the education and training platform according to the training matrix, and are dynamically updated based on exam results and experience feedback.

[0163] The training and assessment module M5 includes the following seven steps during runtime: Step 1: Establish a construction plan training and assessment matrix for the construction sector: Based on the level and specialty of the construction plan, clarify the training targets and the learning and assessment requirements for each category of personnel, which will guide the training of construction plans.

[0164] Step 2: Upload the construction plan to the AI ​​AI: Upload the final version of the construction plan to the AI ​​AI. The AI ​​will automatically generate key learning points and training materials for the construction plan. The AI-generated key learning points will be reviewed by experts who will provide feedback. The AI ​​will then iterate and optimize based on the feedback to ensure the rationality and accuracy of the key learning points.

[0165] Step 3: AI automatically generates construction plan assessment questions: The AI ​​automatically generates different types of exam questions based on prompts (including the key points of the assessment, question types, and number of questions); the AI-generated questions are reviewed by experts and iteratively optimized to form an assessment question bank.

[0166] Step 4: Create a construction plan training system: Create a construction plan training system category on the company's education and training platform, including: training year, plan name, category and specialty, applicable project department, etc.; upload training courseware (including: construction plan documents and AI-generated learning points); create a construction plan question bank and upload AI-generated exam questions.

[0167] Step 5: Create a construction plan training course: Following the classification of the construction plan training system, create a construction plan training course, and link the training materials and exams; upon completion of the course and passing the exam, a certificate of qualification will be automatically obtained.

[0168] Step Six: Regularly Push Construction Plan Learning: Based on the training plan list and training matrix, regularly create construction plan learning tasks and push them to designated personnel, requiring them to complete the course learning and examination within a limited time; those who fail the examination must retake it, and only those who pass the examination can participate in relevant work activities.

[0169] Step 7: Dynamic Optimization and Iteration: Regularly feed back frequently missed questions and construction experience feedback to the AI ​​and the upgraded construction plan course content to achieve dynamic iteration and optimization, and improve the accuracy of learning points and test question generation.

[0170] As an example, since its implementation, the prompt words have been upgraded and iterated three times, resulting in 12 sets of question banks with more than 890 questions.

[0171] 1.7 Feedback Module M6 and Additional Conditions Module M7

[0172] Furthermore, based on the above embodiments, the intelligent review system for construction plans disclosed herein may also include a feedback module M6 and an additional conditions module M7, such as... Figure 2 As shown.

[0173] (1) Feedback module M6

[0174] The feedback module M6 is configured to receive review opinions from the review opinion generation module M3 and the annotation information for those review opinions, classify and process the review opinions and annotation information, thereby generating and outputting positive review opinion data and / or negative review opinion data, and storing the positive review opinion data in the positive review opinion knowledge base and the negative review opinion data in the negative review opinion knowledge base.

[0175] The annotation information mentioned here is input by the user through the human-computer interaction dialogue module M4. Review comments can also be sent to the feedback module M6 via the human-computer interaction dialogue module M4, or directly fed back to the feedback module M6 from the review comment generation module M3. Specifically, after reviewing the AI-generated review comments through the human-computer interaction dialogue module M4, experts evaluate the comments, determine whether they are positive or negative, and input annotation information through the human-computer interaction dialogue module M4.

[0176] After the feedback module M6 categorizes and processes the review comments and annotation information, these comments and annotation information can be further classified as low-value opinion data. Low-value opinion data refers to review comments that lack substantive content, and it is neither stored in the positive review comment knowledge base nor the negative review comment knowledge base.

[0177] Specifically, the review opinions are divided into three categories: positive opinions, which refer to AI review opinions that are logically rigorous and of significant value, and it is recommended to strengthen this type of output in the future. The corresponding review opinion data is stored in the positive review opinion knowledge base; negative opinions, which refer to AI review opinions that contain factual errors or logical deviations and need to be optimized to avoid recurrence. The corresponding review opinion data is stored in the negative review opinion knowledge base; and low-value opinions, which refer to AI review opinions that lack substantive content and have limited reference value. These are neither stored in the positive nor negative review opinion knowledge base.

[0178] The review opinion generation module M3, the feedback module M6, and the knowledge base retrieval module M2 constitute an iterative closed loop. This iterative closed loop is configured to repeatedly execute the following processes: the review opinion generation module M3 generates and outputs the current review opinion based on the current data in the positive review opinion knowledge base and the negative review opinion knowledge base; the feedback module M6 receives the current review opinion and the annotation information for the current review opinion, classifies the current review opinion and the annotation information, thereby generating and outputting positive review opinion data and / or negative review opinion data, and stores the positive review opinion data in the positive review opinion knowledge base and the negative review opinion data in the negative review opinion knowledge base, thereby updating the data in the positive review opinion knowledge base and the negative review opinion knowledge base; and in the next iteration, the review opinion generation module M3 generates and outputs the next review opinion based on the updated data in the positive review opinion knowledge base and the negative review opinion knowledge base.

[0179] The iterative closed loop iterates more than twice, and the data in the positive review opinion knowledge base and the negative review opinion knowledge base accumulates continuously with each iteration.

[0180] The feedback module M6 can be part of the review feedback generation module M3; the two can also be independent modules. The feedback module M6 can call the large language model for processing.

[0181] The iterative closed loop mentioned here refers to an iterative closed-loop mechanism of "AI initial review - expert review - knowledge accumulation". The review results generated by the intelligent agent are labeled and fed back to the AI ​​for iteration, which effectively improves the quality of AI review output.

[0182] To verify the effectiveness of the iterative closed-loop mechanism, 11 standardized construction plans for civil engineering were selected as pilot samples, and the three professional brains (T / S / Q) were used for review. Based on the review matrix of construction, safety, and quality professionals, the generated opinions were tagged and classified into three categories: positive opinions, negative opinions, and low-value opinions. Among them, positive and negative opinions were synchronized to the corresponding knowledge base, realizing the dynamic updating and continuous optimization of the knowledge base.

[0183] (2) Additional Conditions Module M7

[0184] The additional conditions module M7 is configured to receive additional conditions as input and output them to the review opinion generation module M3. The review opinion generation module M3 adds additional conditions to the original content to constrain the reasoning process. These additional conditions include output format requirements, language style requirements, review focus (e.g., limited to technical items, quality items, and / or safety items), time frame, reviewers, etc.

[0185] As an example, additional conditions might include phrases like "You must be professional when generating content about security" or "You must be concise when you speak." These additional conditions are used to regulate the output direction of the review comments generation module M3, making the generated review comments more in line with the user's needs.

[0186] If the additional conditions module M7 outputs additional conditions to the examination opinion generation module M3 before M3 generates and outputs the examination opinion, i.e. before M3 performs analysis and processing, then M3 applies constraints based on these additional conditions during its reasoning process. If the additional conditions module M7 outputs additional conditions to the examination opinion generation module M3 after M3 generates and outputs the pre-examination opinion, but before performing optimization reasoning (which can be done before or after combining positive and negative examination opinions), then M3 adjusts the generated examination opinion based on these additional conditions, thereby generating and outputting an adjusted examination opinion, such as an optimized examination opinion.

[0187] The additional conditions module M7 can also output additional conditions to the examination opinion generation module M3 after the examination opinion generation module M3 generates and outputs the optimized examination opinion. In this case, the examination opinion generation module M3 adjusts the optimized examination opinion based on the additional conditions, thereby generating and outputting the final examination opinion.

[0188] According to the intelligent construction plan review system described in the above embodiments of this disclosure, the document parsing module M1 parses and processes the construction plan document to generate a text string. The knowledge base retrieval module M2 retrieves the text string based on the technical knowledge base, safety knowledge base, and quality knowledge base to generate three types of retrieval results. The review opinion generation module M3 uses data from the positive review opinion knowledge base and the negative review opinion knowledge base to reason about the text string and the three types of retrieval results to generate review opinions. This system can construct a complete intelligent review link from document input to review opinion output, enabling differentiated review of construction plans in the three professional dimensions of technology, safety, and quality, improving the review coverage and the uniformity of review standards. At the same time, the positive and negative review opinion knowledge bases enable the review system to have self-optimization capabilities based on historical review experience. Furthermore, the training and assessment module M5 transforms the review results into training resources, realizing the linkage between review and training.

[0189] According to the intelligent review system for construction plans described in the first modification of this disclosure, the knowledge base retrieval module M2 can simultaneously search five knowledge bases, using positive and negative historical review opinions as a source of enhanced retrieval. This allows the review opinion generation module M3 to refer to the quality evaluation experience of historical review opinions during reasoning and processing, thereby improving the accuracy and consistency of review opinions and achieving continuous optimization of the quality of review opinions.

[0190] According to the intelligent review system for construction plans described in the second modification of this disclosure, by dividing the reasoning process into two stages, pre-reasoning and optimization reasoning, preliminary review opinions can be generated quickly in the first stage, and precise optimization can be performed in the second stage based on a knowledge base of positive and negative review opinions, thereby improving the quality and relevance of review opinions while ensuring review efficiency.

[0191] According to the additional condition module M7 described in the above embodiments of this disclosure, by implementing branch routing processing of additional conditions, the system can intelligently select the processing path according to the type of additional conditions, and flexibly respond to personalized review requirements.

[0192] In the above embodiments and variations, in order to improve the accuracy of review, the review opinion generation module M3 adopts a differentiated processing strategy for the content corresponding to different review dimensions in the text string, such as focusing on reviewing the more important technical content in the construction plan.

[0193] Specifically, for the content corresponding to the security and quality dimensions, the review opinion generation module M3 takes the entire text string (when its corresponding content is relatively small) or a longer segment as input, performs reasoning processing, and generates review opinions for the security and quality dimensions. This is because the review focus of the security and quality dimensions is relatively concentrated, and the input of the whole or a longer segment is sufficient to meet the requirements of granularity and accuracy of the review.

[0194] For the content related to the technical dimensions, the review comment generation module M3 adopts a segmented iterative process, such as... Figure 2 As shown. The segmented iterative processing described here refers to the following: the text string output by the document parsing module M1 has been divided into N segments after semantic segmentation. The review opinion generation module M3 performs iterative processing on these N segments in sequence. Each segment undergoes a complete "retrieval → reasoning" process, thereby generating the review result corresponding to that segment, and summarizing the N review results corresponding to the N segments into a technical dimension review opinion, where N is an integer greater than 1.

[0195] Specifically, in the i-th iteration (1 ≤ i ≤ N), the review opinion generation module M3 outputs the i-th segment to the knowledge base retrieval module M2. The knowledge base retrieval module M2 performs retrieval processing on the i-th segment based on the technical knowledge base, thereby generating and outputting the technical retrieval result corresponding to the i-th segment. The review opinion generation module M3 receives the i-th segment and its corresponding technical retrieval result, combines the segment and the technical retrieval result into a prompt word string, and sends it to the large language model for inference processing, thereby generating and outputting the review result corresponding to the i-th segment.

[0196] After completing all N iterations, the review opinion generation module M3 summarizes and merges the N review results corresponding to the N segments to generate a complete review opinion from a technical perspective. The summarization and merging methods can include splicing the review results according to the original order of the segments, or integrating review results that are duplicated or contradictory after deduplication and consistency verification.

[0197] Furthermore, in a preferred embodiment, the user can input technical specification review-related prompts and / or technical review prompts, such as professional dimensions, through the additional conditions module M7. The professional dimension information mentioned here includes, but is not limited to, civil engineering, installation, and crane engineering. The review opinion generation module M3 and / or the knowledge base retrieval module M2 can process the data by overlaying prompts on existing content. These prompts can be input before the review opinion generation module M3 reviews the data, or after pre-reasoning processing. For example, when the knowledge base retrieval module M2 performs a search in the technical knowledge base, it limits the search scope to reference data corresponding to that professional dimension based on the professional dimension information, thereby avoiding the retrieval of reference data from other professional fields and improving search efficiency and the relevance of search results.

[0198] As an example, when a user uploads a construction plan document for the installation profession, they can simultaneously input "This is a construction plan for the installation profession" as the professional dimension information through the additional conditions module M7. When the knowledge base retrieval module M2 performs retrieval processing in the technical knowledge base, it only retrieves reference data for construction plans and technical specifications related to the installation profession, and does not retrieve reference data for civil engineering and hoisting professions, thereby reducing the interference of irrelevant search results on the reasoning process.

[0199] The above-described segmented iterative processing can be used independently or in combination with the first variation (simultaneous retrieval of five databases) and the second variation (two-stage reasoning). When combined with the first variation, in each iteration, the knowledge base retrieval module M2 simultaneously searches the technical knowledge base, the positive review opinion knowledge base, and the negative review opinion knowledge base. The retrieval results, along with the current segment, are input into the review opinion generation module M3 for reasoning processing. When combined with the second variation, in each iteration, the review opinion generation module M3 first performs pre-reasoning processing based on the current segment and the corresponding technical retrieval results, and then performs optimized reasoning processing based on the data related to the pre-review opinions in the positive and negative review opinion knowledge bases, thereby generating the optimized review result corresponding to that segment.

[0200] According to the intelligent construction plan review system described in this embodiment, by employing segmented iterative processing in the technical dimension, the system leverages the semantic understanding advantages of large language models for short texts to ensure thorough and in-depth review of each technical section. This enhances the precision and accuracy of technical review opinions, while maintaining overall review of safety and quality dimensions to improve processing efficiency. This achieves optimized differentiated review strategies across different dimensions. Furthermore, by introducing prompts such as professional dimension information as additional conditions to constrain the knowledge base retrieval scope, the retrieval results can be more focused on a narrower direction, further improving the targeting and efficiency of the review.

[0201] According to the feedback module M6 described in the above embodiments of this disclosure, by constructing an iterative closed loop of "AI preliminary review - expert review - knowledge accumulation", the quality evaluation experience of review opinions is accumulated into reusable knowledge and the review system has the ability to self-iterate and optimize.

[0202] Based on the intelligent construction plan review system described in the above embodiments, a method for intelligent construction plan review according to an embodiment of this disclosure will be described. This method includes: The first step is to receive the uploaded construction plan file, parse and process the construction plan file, and then generate and output a text string.

[0203] Specifically, in the first step, the document parsing module M1 receives the construction plan file uploaded in the form of a file object, and parses the text of the construction plan file through semantic segmentation processing (including six sub-steps from S1 to S6) to generate a text string.

[0204] The second step involves receiving a text string and performing retrieval processing on the text string based on the technical knowledge base, security knowledge base, and quality knowledge base, thereby generating and outputting technical retrieval results, security retrieval results, and quality retrieval results.

[0205] Specifically, in the second step, the knowledge base retrieval module M2 sends a text string as a query term to the knowledge base interface via an HTTP POST request. The knowledge base interface performs a vector similarity search based on the query term and returns matching document fragments as the retrieval results. The knowledge base retrieval module M2 can perform retrieval processing on three knowledge bases in parallel. Furthermore, the knowledge base retrieval module M2 also manages positive review opinion knowledge bases and negative review opinion knowledge bases.

[0206] The third step involves receiving a text string along with technical search results, security search results, and quality search results. Based on the positively labeled review opinion data stored in the positive review opinion knowledge base and the negatively labeled review opinion data stored in the negative review opinion knowledge base, the text string, technical search results, security search results, and quality search results are processed to generate and output a review opinion.

[0207] Specifically, in the third step, the review opinion generation module M3 combines the text string with the search results into a prompt word string, which is then sent to the application programming interface of the large language model. The large language model performs reasoning processing based on the data in the positive review opinion knowledge base and the negative review opinion knowledge base to generate and output the review opinion.

[0208] The fourth step is to receive review comments, process the review comments through dialogue, and generate and output the dialogue response.

[0209] Specifically, in the fourth step, the human-computer interaction dialogue module M4 converts the review comments into natural language dialogue responses, engaging in multi-round interactive dialogue with the user. Additionally, optionally, in the fourth step, it can receive follow-up questions input by the user, retrieve the original text supporting the review comments based on the follow-up questions, and thus generate and output the original source information.

[0210] The fifth step is to receive review comments, process the review comments to generate training and assessment content, and then generate and output the training and assessment content.

[0211] Specifically, in the fifth step, the training and assessment module M5 generates learning points and assessment questions based on the review comments, pushes the learning points and assessment questions to the education and training platform according to the training matrix, and dynamically updates the learning points and assessment questions based on the exam results and experience feedback.

[0212] In the first variation of the method, in the second step, the text string is further retrieved based on the positive review opinion knowledge base and the negative review opinion knowledge base, thereby generating and outputting positive and negative search results, and using the positive and negative search results for reasoning processing in the third step.

[0213] In the second variation of the method, in the third step, the text string and the technical search results, security search results and quality search results are first pre-reasoned to generate and output a pre-examination opinion; then, based on the data related to the pre-examination opinion in the positive review opinion knowledge base and the negative review opinion knowledge base, the pre-examination opinion is further optimized and reasoned to generate and output an optimized review opinion.

[0214] Furthermore, following the third step, additional steps may be included: receiving examination opinions and annotation information for those opinions; classifying and processing the examination opinions and annotation information to generate and output positive and / or negative examination opinion data; and storing the positive examination opinion data in a positive examination opinion knowledge base and the negative examination opinion data in a negative examination opinion knowledge base. Subsequently, in the next third step, the next examination opinion is generated and output based on the updated data in the positive and negative examination opinion knowledge bases, thus forming an iterative closed loop.

[0215] The intelligent construction plan review system described in this disclosure has been in practical application since June 2025, embedded with the Zhihuixing AI assistant on the company's intranet. After application, the review time for a single construction plan can be shortened from 6 hours to 4 hours, improving efficiency by more than 30%. Version 1.0 ranks among the top three in application rate in the company's Zhihuixing intelligent agent application market. After the launch of version 2.0, it has been fully applied in multiple nuclear power projects, assisting engineers in shortening the efficiency of plan review, improving plan quality, and ensuring the safety and quality of construction operations. AI-assisted review of construction plans has reached the industry-leading level. The basic function of quality plan signing has been developed, and the company is actively promoting its application in multiple projects, with more than 8,000 online witness point notifications and witnesses initiated.

[0216] Specifically, this disclosure achieves the following technical effects: 1. High-quality analysis of construction plans: Solve the problem of analyzing multiple images, texts, and tables in scanned and signed construction plans, deploy multi-path document analysis services, improve the accuracy of text recognition and processing efficiency, and analyze more than 3,000 construction plans.

[0217] 2. Build three brains for safety, quality, and technology: build more than 10 knowledge bases of various types, complete complex tasks such as extraction, screening, and weighting, coordinate and sort out existing data, and build knowledge bases for standards and specifications, technical specifications, operational safety risks, and typical internal and external quality issues.

[0218] 3. Iterative training of intelligent agents: Establish a positive knowledge base (compliance standards, high-quality cases) and a reverse knowledge base (common problems, error cases) to improve the dimensions of AI judgment.

[0219] 4. Original source inquiry function: This allows users to easily find the original source evidence corresponding to the review comments, thereby improving the credibility of the results.

[0220] 5. Supports users to add questions using prompts, flexibly responding to personalized review and learning content generation needs.

[0221] Furthermore, it should be understood that this disclosure also provides an electronic device including a processor and a memory (or a computer-readable storage medium), the memory storing a computer program or instructions capable of implementing the intelligent review method for construction plans described in the above embodiments and their variations and any other suitable variations.

[0222] In addition, this disclosure also provides a computer program product, which includes a computer program or instructions for implementing the intelligent review method for construction plans described in the above embodiments and their variations and any other suitable variations.

[0223] In addition, this disclosure also provides a computer-readable storage medium storing a computer program or instructions that are executed by a processor to implement the intelligent review method for construction plans described in the above embodiments and their variations and any other suitable variations.

[0224] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Any variations and modifications can be made by those skilled in the art without departing from the spirit and scope of the invention. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention, without departing from the scope of the invention, fall within the protection scope defined by the claims of the present invention.

Claims

1. A construction plan intelligent review system, characterized in that, include: The document parsing module M1 is configured to receive uploaded construction plan files, parse and process the construction plan files, and generate and output text strings. The knowledge base retrieval module M2 is configured to receive the text string output by the document parsing module M1, and perform retrieval processing on the text string based on the technical knowledge base, security knowledge base, and quality knowledge base, thereby generating and outputting technical retrieval results, security retrieval results, and quality retrieval results. The knowledge base retrieval module M2 also includes a positive review opinion knowledge base and a negative review opinion knowledge base. The positive review opinion knowledge base stores review opinion data marked as positive, and the negative review opinion knowledge base stores review opinion data marked as negative. The review opinion generation module M3 is configured to receive the text string output by the document parsing module M1 and the technical retrieval result, the security retrieval result, and the quality retrieval result output by the knowledge base retrieval module M2. The review opinion generation module M3 is configured to perform reasoning processing on the text string, the technical retrieval result, the security retrieval result, and the quality retrieval result based on the data in the positive review opinion knowledge base and the negative review opinion knowledge base, thereby generating and outputting review opinions. as well as The human-computer interaction dialogue module M4 is configured to receive the review opinions output by the review opinion generation module M3, process the review opinions through dialogue response, and thereby generate and output a dialogue response.

2. The intelligent construction plan review system according to claim 1, characterized in that, The knowledge base retrieval module M2 is further configured to perform retrieval processing on the text string based on the positive review opinion knowledge base and the negative review opinion knowledge base, thereby generating and outputting positive and negative retrieval results. The positive and negative retrieval results, together with the technical retrieval results, the security retrieval results, and the quality retrieval results, are then output to the review opinion generation module M3, so that the review opinion generation module M3 can perform reasoning processing based on these retrieval results.

3. The intelligent construction plan review system according to claim 1, characterized in that, The review opinion generation module M3 is further configured to perform pre-reasoning processing on the text string, the technology search results, the security search results, and the quality search results, thereby generating and outputting pre-review opinions. Based on the data associated with the pre-review opinions in the positive review opinion knowledge base and the negative review opinion knowledge base, the pre-review opinions are further optimized and reasoned to generate and output optimized review opinions.

4. The intelligent construction plan review system according to claim 1, characterized in that, It also includes a feedback module M6, configured to receive the review opinions output by the review opinion generation module M3 and the annotation information for the review opinions, classify and process the review opinions and the annotation information, thereby generating and outputting positive review opinion data and / or negative review opinion data, and storing the positive review opinion data in the positive review opinion knowledge base and the negative review opinion data in the negative review opinion knowledge base.

5. The intelligent construction plan review system according to claim 1, characterized in that, It also includes an additional conditions module M7, configured to receive input of additional conditions and output the additional conditions to the review opinion generation module M3. The review opinion generation module M3 constrains the reasoning process based on the additional conditions, wherein the additional conditions include at least one of the following: output format requirements, language style requirements, and scope of review focus.

6. The intelligent review system for construction plans according to claim 1, characterized in that, The document parsing module M1 performs the following parsing process on the construction plan document: preprocessing the text of the construction plan document to extract a valid range; performing structural parsing on the text within the valid range to obtain the heading level; grouping chapter content based on the heading level to generate a chapter tree relationship; merging and splitting chapters in the chapter tree relationship under length constraints to generate a general segment list; and extracting targeted content from the general segment list based on keywords corresponding to the target review dimension to generate the text string.

7. A method for intelligent review of construction plans, comprising: The first step is to receive the uploaded construction plan file, parse the construction plan file, and generate and output a text string. The second step is to receive the text string, perform retrieval processing on the text string based on the technical knowledge base, security knowledge base, and quality knowledge base, thereby generating and outputting technical retrieval results, security retrieval results, and quality retrieval results. The third step involves receiving the text string, the technical retrieval results, the security retrieval results, and the quality retrieval results. Based on the positively labeled review opinion data stored in the positive review opinion knowledge base and the negatively labeled review opinion data stored in the negative review opinion knowledge base, the text string, the technical retrieval results, the security retrieval results, and the quality retrieval results are subjected to reasoning processing to generate and output review opinions. as well as The fourth step is to receive the review comments, process the review comments through a dialogue response, and generate and output a dialogue response.

8. An electronic device comprising a processor and a memory, the memory storing a computer program, characterized in that, The computer program is executed by the processor to implement the intelligent review method for construction plans as described in claim 7.

9. A computer program product, comprising a computer program, characterized in that, The computer program is executed by a processor to implement the intelligent review method for construction plans as described in claim 7.

10. A computer-readable storage medium storing a computer program, characterized in that, The computer program is executed by a processor to implement the intelligent review method for construction plans as described in claim 7.

Images