Construction plan generation method and device, electronic equipment and storage medium

Abstract

The invention relates to the technical field of artificial intelligence, and provides a construction plan generation method and device, electronic equipment and a storage medium. According to the method, firstly, a construction design document of a target project is preprocessed to obtain a coding vector, then an intention analysis model responds to a received construction plan construction instruction to perform intention analysis on the coding vector, and an initial construction subtask set is obtained; the intention analysis module also retrieves each sub-task in the initial construction sub-task set, and adjusts a sub-task division mode based on a retrieval result until a target sub-task set meeting a preset condition is obtained; and finally, inputting the target sub-task set into an inference model, integrating the construction period and the construction path of each construction sub-task by the inference model based on the constraint conditions in the target sub-task set, and obtaining a construction plan of the target project, thereby improving the automatic generation efficiency of the construction plan, and enabling the generated construction plan to better meet the requirements of users.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a construction plan generation method, device, electronic device and storage medium. Background Art

[0002] In the construction industry, the preparation of construction plans usually requires a lot of manual experience and time. Project managers need to extract key information from hundreds of pages of technical documents such as construction organization design, and then combine their own experience to develop a construction schedule. This process is cumbersome and susceptible to human factors. There are often problems such as insufficient information extraction and incomplete knowledge utilization, resulting in insufficient planning or omission of important details. In addition, traditional construction plan preparation relies on project management software or manual methods, and complex constraints such as task dependency and resource constraints still need to be set manually, which increases the workload. Although some digital tools have been introduced in the construction industry at this stage, the degree of automation is still limited, and the quality of the construction plan depends on the experience level of the compiler.

[0003] With the development of artificial intelligence and large model technology, people have begun to explore the application of LLM (Large Language Model) in construction planning. RAG (Retrieval-augmented Generation), as a technology that combines knowledge retrieval and text generation, has shown potential in areas such as knowledge question answering. RAG retrieves relevant information from external knowledge bases and provides it to language models for reference, thereby reducing the "hallucination" phenomenon of the model and improving the accuracy of the answer. This idea makes it possible to integrate domain documents into LLM. For example, the latest research applies RAG to project documents, making the model's answers more credible and well-documented.

[0004] However, the current AI (Artificial Intelligence) retrieval enhancement system for the construction field still has shortcomings. On the one hand, the PDF (Portable Document Format) document format, such as the construction organization plan, is complex, containing charts, multi-column layouts, etc., and it is not easy to directly extract text. Irregular PDF formats, scanned OCR (Optical Character Recognition) errors, etc. will cause the text obtained in the retrieval stage to be incomplete or noisy. These problems reduce the accuracy of information retrieval. On the other hand, even if the general large language model is connected to the retrieval data, it may still be insufficient in logical reasoning and planning, and it is difficult to ensure the rationality of the output construction plan in terms of timing and resource allocation. For example, the model may miss certain sequential constraints or have errors in the numerical calculation of the document. These problems are difficult to overcome under a single LLM architecture, because a single model may not be able to take into account both dialogue understanding and complex reasoning at the same time. Summary of the invention

[0005] In view of this, the embodiments of the present application provide a construction plan generation method, device, electronic device and storage medium to solve the problems in the prior art of low construction plan generation efficiency and low availability of the generated plan.

[0006] A first aspect of an embodiment of the present application provides a construction plan generation method, comprising: Obtain the construction design document of the target project, and pre-process the construction design document to obtain the encoding vector of the design document; In response to receiving an instruction to construct a construction plan for a target project, inputting a coding vector of a design document into an intent parsing model, performing intent parsing on the coding vector based on a knowledge graph of the project domain and preset parsing rules, and obtaining an initial construction subtask set; the initial construction subtask set includes at least one initial construction subtask and constraints of each initial construction subtask, wherein the constraints include at least one of the following: a priority of each initial construction subtask, resource constraint information of each initial construction subtask, and dependencies between different initial construction subtasks; Retrieve each initial construction subtask in the intention parsing model to obtain retrieval results for each initial construction subtask; the retrieval results at least include construction requirements and historical construction tasks corresponding to the initial construction subtask; In response to determining that the similarity between the task attributes and constraints of the target initial construction subtask and the search result is less than or equal to a first preset similarity threshold, re-performing intent parsing on the encoding vector of the design document to obtain an updated subtask set; the target initial construction subtask is at least one construction subtask in the initial construction subtasks; In response to determining that the similarity between the task type and constraint condition of each subtask in the current construction subtask and the search result is greater than a first preset similarity threshold, determining the current subtask set as the target subtask set; wherein the current subtask set is the subtask set to be compared with the current search result of the intent parsing model, and the current construction subtask is the construction subtask in the current subtask set; The target subtask set is input into the reasoning model, and the construction period and construction path of each construction subtask are integrated based on the constraints to obtain the construction plan of the target project.

[0007] A second aspect of the embodiments of the present application provides a construction plan generating device, including: An acquisition module is configured to acquire a construction design document of a target project and pre-process the construction design document to obtain a coding vector of the design document; The parsing module is configured to, in response to receiving an instruction to construct a construction plan for a target project, input a coding vector of a design document into an intent parsing model, perform intent parsing on the coding vector based on a knowledge graph of the project domain and preset parsing rules, and obtain an initial construction subtask set; the initial construction subtask set includes at least one initial construction subtask and constraints of each initial construction subtask, and the constraints include at least one of the following: a priority of each initial construction subtask, resource constraint information of each initial construction subtask, and dependencies between different initial construction subtasks; A retrieval module is configured to search each initial construction subtask in the intention parsing model to obtain a retrieval result of each initial construction subtask; the retrieval result at least includes the construction requirements and historical construction tasks corresponding to the initial construction subtask; The iteration module is configured to, in response to determining that the similarity between the task attributes and constraints of the target initial construction subtask and the search result is less than or equal to a first preset similarity threshold, re-parse the encoding vector of the design document to obtain an updated subtask set; the target initial construction subtask is at least one construction subtask in the initial construction subtask; The iteration module is further configured to determine that the task type and constraint conditions of each subtask in the current construction subtask are similar to the search result and are greater than a first preset similarity threshold, and determine that the current subtask set is a target subtask set; wherein the current subtask set is a subtask set to be compared with the current search result of the intent parsing model, and the current construction subtask is a construction subtask in the current subtask set; The reasoning module is configured to input the target subtask set into the reasoning model, integrate the construction period and construction path of each construction subtask based on the constraint conditions, and obtain the construction plan of the target project.

[0008] According to a third aspect of an embodiment of the present application, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program.

[0009] According to a fourth aspect of an embodiment of the present application, a computer-readable storage medium is provided, which stores a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0010] Compared with the prior art, the embodiments of the present application have the following beneficial effects: the embodiments of the present application first pre-process the construction design document of the target project to obtain a coding vector, and then the intent parsing model responds to the received instruction to build a construction plan to perform intent parsing on the coding vector to obtain an initial set of construction subtasks; the intent parsing module also retrieves each subtask in the initial set of construction subtasks, and adjusts the subtask division method based on the retrieval results until a target subtask set that meets the preset conditions is obtained; finally, the target subtask set is input into the inference model, and the inference model integrates the duration and construction path of each construction subtask based on the constraints in the target subtask set to obtain a construction plan for the target project. The method uses the intent parsing model to perform intent parsing and task retrieval on the construction design document to obtain a reasonably divided subtask set, and then uses an independent inference model to schedule and plan the path of each task in the subtask set, thereby improving the efficiency of automatic generation of the construction plan, and the generated construction plan is more in line with user needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0012] Figure 1 It is a flowchart of a construction plan generation method provided in an embodiment of the present application.

[0013] Figure 2 It is a flowchart of a method for searching basic knowledge subtasks in an intention parsing model to obtain search results provided in an embodiment of the present application.

[0014] Figure 3 It is a flowchart of a method for searching independent construction subtasks in an intention parsing model to obtain search results provided in an embodiment of the present application.

[0015] Figure 4It is a flowchart of a method for searching collaborative construction subtasks in an intention parsing model to obtain search results provided in an embodiment of the present application.

[0016] Figure 5 It is a flowchart of a method for integrating the construction period and construction path of each construction subtask based on constraint conditions in a reasoning model provided in an embodiment of the present application.

[0017] Figure 6 It is a schematic diagram of a construction plan generating device provided in an embodiment of the present application.

[0018] Figure 7 It is a schematic diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0019] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0020] A construction plan generation method and device according to an embodiment of the present application will be described in detail below with reference to the accompanying drawings.

[0021] As mentioned above, although the construction industry has introduced some digital tools, the degree of automation is still limited, and the quality of construction plans depends on the experience of the compiler. At the same time, the current AI (Artificial Intelligence) retrieval enhancement system for the construction field still has shortcomings.

[0022] In view of this, the embodiment of the present application provides a construction plan generation method, which first pre-processes the construction design document of the target project to obtain a coding vector, and then the intent parsing model responds to the received construction plan instruction to perform intent parsing on the coding vector to obtain an initial construction subtask set; the intent parsing module also retrieves each subtask in the initial construction subtask set, and adjusts the subtask division method based on the retrieval results until a target subtask set that meets the preset conditions is obtained; finally, the target subtask set is input into the reasoning model, and the reasoning model integrates the duration and construction path of each construction subtask based on the constraints in the target subtask set to obtain a construction plan for the target project. The method uses the intent parsing model to perform intent parsing and task retrieval on the construction design document to obtain a reasonably divided subtask set, and then uses an independent reasoning model to schedule and plan the path of each task in the subtask set, thereby improving the efficiency of automatic generation of the construction plan, and the generated construction plan is more in line with user needs.

[0023] Figure 1 Schematic diagram of a construction plan generation method provided in an embodiment of the present application. Figure 1 As shown, the method comprises the following steps: In step S101, the construction design document of the target project is obtained, and the construction design document is preprocessed to obtain a coding vector of the design document.

[0024] In step S102, in response to receiving an instruction to build a construction plan for the target project, the coding vector of the design document is input into the intent parsing model, and the coding vector is parsed based on the knowledge graph of the project field and preset parsing rules to obtain an initial set of construction subtasks.

[0025] Among them, the initial construction subtask set includes at least one initial construction subtask and constraints of each initial construction subtask, and the constraints include at least one of the following: the priority of each initial construction subtask, the resource constraint information of each initial construction subtask and the dependency relationship between different initial construction subtasks.

[0026] In step S103, each initial construction subtask is searched in the intention parsing model to obtain the search results of each initial construction subtask.

[0027] The search results at least include the construction requirements and historical construction tasks corresponding to the initial construction subtask.

[0028] In step S104, in response to determining that the similarity between the task attributes and constraints of the target initial construction subtask and the retrieval result is less than or equal to a first preset similarity threshold, the encoding vector of the design document is re-intent parsed to obtain an updated subtask set.

[0029] The target initial construction subtask is at least one construction subtask in the initial construction subtasks.

[0030] In step S105 , in response to determining that the similarity between the task type and constraint condition of each subtask in the current construction subtask and the search result is greater than a first preset similarity threshold, the current subtask set is determined to be a target subtask set.

[0031] Among them, the current subtask set is the subtask set that is compared with the current retrieval result of the intent parsing model, and the current construction subtask is the construction subtask in the current subtask set.

[0032] In step S106, the target subtask set is input into the reasoning model, and the construction period and construction path of each construction subtask are integrated based on the constraint conditions to obtain the construction plan of the target project.

[0033] In some embodiments of the present application, the method may be executed by a terminal device or by a server.

[0034] In some embodiments of the present application, the construction design document of the target project may be first obtained, and the construction design document may be preprocessed to obtain the encoding vector of the design document. The target project is a project for which a construction plan needs to be generated.

[0035] During preprocessing, considering that design documents are usually in PDF format, they may contain interfering elements such as multi-column layout, headers and footers. Mature PDF parsing tools can be used to ensure high-fidelity extraction. After extraction, the content is formatted and cleaned, such as removing irrelevant information such as page numbers and redundant headers. Then, a semantic block strategy is used to divide long documents into several content blocks. Blocks are divided into chapters or paragraphs to maintain the semantic integrity of each block of content and avoid breaking up related content. At the same time, the block size is dynamically adjusted according to the upper limit of the model input length to ensure that each block of text is not too long to affect the retrieval effect. Meta information such as the chapter title and page number can also be stored for each block to facilitate result tracing.

[0036] After the block division is completed, each document block can be encoded into a vector through an embedding model, and the obtained encoded vector can be stored in a vector database. In one example, the vector database can be, for example, a FAISS database. The embedding model can be an Embedding model to ensure that the vector representation has good expression ability for construction field terms.

[0037] When an instruction to build a construction plan for a target project is received, the encoding vector of the design document can be input into the intent parsing model, which will perform intent parsing based on the knowledge graph of the project domain and preset parsing rules to obtain an initial set of construction subtasks.

[0038] That is, the intention parsing model can receive user instructions, such as "generate a construction plan for the target project", and then obtain the encoding vector of the design document based on the instruction, and perform task decomposition on the encoding vector to obtain an initial construction subtask set including multiple initial construction subtasks. At the same time, the initial construction subtask set also includes the constraints of each initial construction subtask.

[0039] In some implementations, the constraint condition may include at least one of the following: a priority of each initial construction subtask, resource constraint information of each initial construction subtask, and dependencies between different initial construction subtasks.

[0040] The intention parsing module can also search each initial construction subtask in the initial construction subtask set to obtain a search result for each initial construction subtask, and the search result at least includes the construction requirements and historical construction tasks corresponding to the initial construction subtask.

[0041] That is to say, the vector database can also store the encoding vectors of historical construction plans, as well as knowledge such as laws, regulations, and construction standards in the field where the target project is located. For each initial construction subtask, its related laws, regulations, construction standards, and historical construction tasks can be retrieved. Then, each initial construction subtask is compared with its corresponding retrieval results.

[0042] In one example, the similarity between the task attributes and constraints of each initial construction subtask and the corresponding search results can be calculated. If the similarity is less than or equal to the first preset similarity threshold, it can be considered that the initial construction subtask has deviations or insufficient information. At this time, the intent parsing model can re-parse the encoding vector of the document involved to obtain an updated subtask set. The steps of parsing planning-retrieval comparison can be iteratively performed until the similarity between the task attributes and constraints of all initial construction subtasks and the corresponding search results is greater than the first preset similarity threshold. Among them, the value of the first preset similarity threshold can be set according to actual needs and is not limited here.

[0043] If it is determined that the similarity between the task type and constraint condition of each subtask in the current construction subtask and the search result is greater than the first preset similarity threshold, the current subtask set is determined to be the target subtask set. The current subtask set is the subtask set compared with the search result of the intent parsing model, and the current construction subtask is the construction subtask in the current subtask set.

[0044] In one example, the intent parsing model can be implemented by the large model Qwen2.5. When a user raises a requirement (e.g., "generate a construction plan for the target project"), Qwen2.5 parses the user's intent to decompose the complex requirement into subtasks. In the retrieval enhancement process, Qwen2.5 constructs a retrieval query based on the information currently needed and obtains the relevant document block content from the vector retrieval module. For each aspect of the construction plan, including construction technology, schedule, and resource allocation, Qwen2.5 can raise subqueries one by one. For example, it can first query "the process flow of basic construction" and then query "the duration data of similar projects" to obtain the required knowledge in stages. The retrieved document content will be provided to Qwen2.5 as contextual knowledge. With the powerful long text generation and instruction following capabilities of Qwen2.5, it will also record the retrieved key information in the subtask set for subsequent weaving into the construction plan. At the same time, Qwen2.5 can interact with the user to confirm project parameters or preferences based on the conversation context to ensure that the generated plan meets actual needs.

[0045] In some embodiments of the present application, the determined target subtask set may be input into the reasoning model, and the duration and construction path of each construction subtask may be integrated based on the constraint conditions to obtain a construction plan for the target project.

[0046] In one example, the reasoning model can be implemented by the large model DeepSeek-R1. After Qwen2.5 generates a set of subtasks, it schedules each construction subtask and optimizes the path through reasoning calculations, which can be completed by DeepSeek-R1, which is good at reasoning. For example, for construction schedule scheduling, it may be necessary to deduce the task sequence based on the sequence of construction processes, or to calculate the duration of a sub-project based on previous data. These can all be solved by DeepSeek-R1.

[0047] In some embodiments of the present application, the reasoning model can generate a construction plan by itself after completing reasoning tasks such as scheduling and path optimization, or it can transmit scheduling information, path information, etc. to the intent parsing model, and the intent parsing model will generate a complete construction plan.

[0048] According to the technical solution provided by the embodiment of the present application, the construction design document of the target project is first preprocessed to obtain a coding vector, and then the intent parsing model responds to the received instruction to build a construction plan to perform intent parsing on the coding vector to obtain an initial set of construction subtasks; the intent parsing module also retrieves each subtask in the initial set of construction subtasks, and adjusts the subtask division method based on the retrieval results until a target subtask set that meets the preset conditions is obtained; finally, the target subtask set is input into the inference model, and the inference model integrates the duration and construction path of each construction subtask based on the constraints in the target subtask set to obtain a construction plan for the target project. The method uses the intent parsing model to perform intent parsing and task retrieval on the construction design document to obtain a reasonably divided subtask set, and then uses an independent inference model to schedule and plan the path of each task in the subtask set, thereby improving the efficiency of automatic generation of the construction plan, and the generated construction plan is more in line with user needs.

[0049] In some embodiments of the present application, the task attributes of each construction subtask may include basic knowledge subtasks, independent construction subtasks, and collaborative construction subtasks. Different construction subtasks may be configured with different priorities. In one example, when generating an initial set of construction subtasks, the highest priority may be configured for the basic knowledge subtask to ensure that background support can be provided for other construction subtasks; then, the priority of each collaborative construction subtask is determined based on the dependency relationship with other construction subtasks, and the priority may be dynamically adjusted during subsequent retrieval and reasoning; finally, the priority of the independent construction subtask is determined.

[0050] Figure 21 is a flow chart of a method for retrieving basic knowledge subtasks in an intention parsing model to obtain search results provided by an embodiment of the present application. Figure 2 As shown, the method comprises the following steps: In step S201, a target encoding vector corresponding to a target construction subtask is obtained.

[0051] In step S202, the target coding vector is used to search the knowledge graph of the project field in a keyword matching manner to obtain a basic knowledge search result set.

[0052] In step S203, the retrieval results in the basic knowledge retrieval result set whose vector similarity with the target coding vector is greater than a second preset similarity threshold are determined as the retrieval results of the target construction subtask.

[0053] The target construction subtask is any construction subtask in the initial construction subtask set or the updated subtask set.

[0054] In some embodiments of the present application, task retrieval for basic knowledge subtasks may be performed by first obtaining a target coding vector corresponding to a target construction subtask, and then using the target coding vector to search in a knowledge graph of the field in which the project is located in a keyword matching manner to obtain a basic knowledge retrieval result set. Finally, the retrieval result in the basic knowledge retrieval result set whose vector similarity with the target coding vector is greater than a second preset similarity threshold is determined as the retrieval result of the target construction subtask.

[0055] Among them, a fixed second preset similarity threshold can be set according to the actual situation, and all the search results in the basic knowledge search result set whose vector similarity with the target coding vector is greater than the second preset similarity threshold are used as the search results of the basic knowledge subtask. The second preset similarity threshold can also be set dynamically according to the search situation, and the N search results in the basic knowledge search result set whose vector similarity with the target coding vector is the highest are used as the search results of the basic knowledge subtask, where N is a positive integer.

[0056] That is to say, for basic knowledge tasks, the vector retrieval module in the intent parsing model can be called to directly perform keyword matching and high-dimensional vector similarity calculation based on the pre-built construction domain knowledge base. The retrieval module can use a hierarchical grouping strategy (for example, by chapter, theme, or construction stage) to sort the results and output them in the form of weights. Furthermore, the retrieval results can be simplified and formatted before output, and labeled with "background knowledge" according to preset rules.

[0057] Figure 31 is a flow chart of a method for retrieving independent construction subtasks in an intention parsing model to obtain search results provided by an embodiment of the present application. Figure 3 As shown, the method comprises the following steps: In step S301, the task identifier of the target construction subtask is determined, and a search template is determined based on the task identifier.

[0058] The task identifier is used to characterize the task type of an independent construction subtask.

[0059] In step S302, a search template is called to retrieve a set of historical similar tasks.

[0060] The historical similar task set includes historical independent construction subtasks that have the same task identifier as the target construction subtask.

[0061] In step S303, the similarity between the task parameters of each task in the historical similar task set and the constraint conditions of the target construction subtask is calculated.

[0062] In step S304, it is determined that the tasks in the historical similar task set whose calculated similarity is greater than a third preset similarity threshold are the search results of the target construction subtask.

[0063] In some embodiments of the present application, task retrieval for independent construction subtasks may be performed by first determining the task identifier of the target construction subtask, and then determining a retrieval template based on the task identifier. In one example, different retrieval templates may be configured for different types of independent construction subtasks, and the retrieval template corresponding to the task type of the target construction subtask may be directly called during retrieval.

[0064] Next, the search template is called to retrieve a historical similar task set, which includes historical independent construction subtasks with the same task identifier as the target construction subtask. The similarity between the task parameters of each task in the historical similar task set and the constraint conditions of the target construction subtask can also be calculated, and the tasks in the historical similar task set whose calculated similarity is greater than a third preset similarity threshold are determined as the search results of the target construction subtask.

[0065] In other words, since the structure of the independent construction subtask is relatively complete, its corresponding search template can be called to perform multi-dimensional comparison on the key parameters involved in the task to obtain the search results.

[0066] Figure 4 1 is a flow chart of a method for retrieving collaborative construction subtasks in an intention parsing model to obtain search results provided by an embodiment of the present application. Figure 4 As shown, the method comprises the following steps: In step S401, the task identifiers of the target construction subtask and its coordinated construction subtask are determined, and a search template is determined based on the task identifiers.

[0067] In step S402, the corresponding search template is called for each construction subtask to retrieve the respective historical similar task sets.

[0068] The historical similar task set includes historical construction subtasks having the same task identifier as each construction subtask, and the historical construction subtasks are independent construction subtasks or collaborative construction subtasks.

[0069] In step S403, the constraints of the target construction subtask and its coordinated construction subtasks are obtained, and component matching is performed on the historical similar task sets of each construction subtask based on the constraints to obtain an initial matching result.

[0070] In step S404, in response to determining that the similarity between the initial matching result and the design document is less than a second preset similarity threshold, at least one of the constraint condition and the target search template is adjusted, and search matching is performed again after the adjustment is completed.

[0071] The target retrieval template is at least one retrieval template among the retrieval templates corresponding to the target construction subtask and its coordinated construction subtask.

[0072] In step S405, in response to determining that the similarity between the matching result and the design document is greater than or equal to a second preset similarity threshold, a task in a historical similar task set corresponding to the target construction subtask is determined as a search result of the target construction subtask.

[0073] In some embodiments of the present application, task retrieval for collaborative construction subtasks may be performed by first determining the task identifiers of the target construction subtask and its collaborative construction subtasks, determining a retrieval template based on the task identifiers, and then calling the corresponding retrieval template for each construction subtask to retrieve the respective historical similar task sets. In other words, the historical similar task sets corresponding to each construction subtask with a dependency relationship may be retrieved for the collaborative construction subtasks.

[0074] Next, the constraints of the target construction subtask and its collaborative construction subtasks can be obtained, and the historical similar task sets of each construction subtask can be matched based on the constraints to obtain an initial matching result. The initial matching result can be compared with the design document. If the similarity between the initial matching result and the design document is less than a second preset similarity threshold, at least one of the constraints and the target search template is adjusted, and search matching is performed again after the adjustment is completed.

[0075] The steps of searching, matching and comparing with the design document may be iteratively performed until the similarity between the obtained matching result and the design document is greater than or equal to a second preset similarity threshold. At this point, the task in the historical similar task set corresponding to the target construction subtask may be determined as the search result of the target construction subtask.

[0076] Comparing the matching result with the design document may be to extract the construction requirements from the design document using the intent parsing model, and when extracting the construction requirements, the laws and regulations in the field of the project, the knowledge of construction standards, and the resource availability information input by the user may also be combined. If the matching result obtained by combining the construction sub-projects with dependencies has a similarity with the construction requirement that is less than a second preset similarity threshold, it may be considered that the matching combination method is unreasonable and needs to be readjusted.

[0077] In other words, since collaborative construction subtasks are related to other construction subtasks, they need to be searched as a whole. At this time, the search strategy can be divided into "intra-group search" and "cross-group collaborative search". "Intra-group search" means first conducting an independent search within each construction subtask and confirming the search results separately. "Component collaborative construction" means matching the search results of each construction subtask between groups.

[0078] During retrieval, the intent parsing model can use the results of high-priority tasks as context based on the dependency graph between collaborative construction subtasks and other construction subtasks with dependencies, promote retrieval requests for subsequent tasks, and dynamically adjust retrieval templates, such as adding dependency constraints, to obtain better retrieval results. For the dependency graph of some complex tasks, the intent parsing model can also call the reasoning model when necessary to complete partial matching combination operations. At the same time, the intent parsing model can also adopt a phased integration strategy, entering the intermediate merging stage after the initial independent retrieval, aggregating the boundary data of each task to form a comprehensive context, and performing a secondary retrieval again to optimize the consistency of the results.

[0079] In some embodiments of the present application, adjusting at least one of the constraints and the target retrieval template may be to adjust at least one of the following: modifying the priority of the target construction subtask and its collaborative construction subtasks; adding the dependency relationship of the target construction subtask and its collaborative construction subtasks in the target retrieval target.

[0080] The technical solution provided in the embodiment of the present application introduces an iterative mechanism of planning-retrieval-replanning in the intention recognition stage, and continuously corrects the initially generated intention decomposition solution, so that the recognition result has higher accuracy and adaptability.

[0081] That is to say, after the user inputs the instruction, the intention parsing model can first plan the design document and generate a preliminary intention disassembly plan consisting of an initial set of construction subtasks. This process includes: Deep semantic feature extraction: Perform word segmentation, part-of-speech tagging and grammatical analysis on the input text in advance, and then combine it with the standardization of professional terminology in the construction field to obtain a multi-dimensional semantic vector.

[0082] Candidate intent generation: Utilize domain knowledge graphs and preset rules, such as identifying implicit requirements such as "construction progress", "resource allocation", and "safety measures" based on the context, split the input into different subtasks, and temporarily summarize them into three types of candidate intents: basic knowledge, independent construction, and collaborative construction.

[0083] Preliminary task dependency construction: For each subtask, build a dependency graph between tasks, such as whether a task needs to obtain background standards and regulatory information as support first, or whether a fine-grained query depends on previous query results, so as to clarify each level of intent.

[0084] After generating a preliminary intent solution, the retrieval phase begins. The intent parsing model can call the retrieval module or knowledge base interface to verify and supplement the key information of each subtask. The key steps include: Multi-source data verification: For each candidate intent, construct a corresponding search query, such as retrieving the latest construction standards, relevant regulations, or historical engineering cases, and determine whether the original intent split is accurate based on the returned relevant data.

[0085] Feedback correction: In the retrieved results, the system compares the keywords and context information extracted from the input text based on preset rules. If a deviation or insufficient information is found in a certain intention, the re-planning module is automatically called to make adjustments, such as adding subdivision steps or re-determining task priorities.

[0086] When the preliminary search results show that there are inconsistencies or omissions in intent recognition, the intent parsing model enters the re-planning stage and iterates the "planning-retrieval-re-planning" process. In each iteration, the intent parsing model can modify the original intent decomposition plan based on the information obtained in the previous round of retrieval, and can add, adjust or delete certain subtasks. For example, if it is found that the resource constraint information involved in a certain requirement is not fully reflected, the system will add a "resource query" subtask to the original plan.

[0087] The intent parsing model can also perform adaptive task sorting. In each round of iteration, the intent parsing model will update the task dependency graph and priority allocation in real time based on the feedback of the search results, ensuring that key intents, such as construction condition verification and regulatory proofreading, always maintain the highest priority, laying the foundation for more accurate intent confirmation in the future.

[0088] After several iterations, when the retrieval results reach a high degree of consistency with the intent planning and no further optimization is required, the intent parsing model will output the final identified intent category and its corresponding subtask structure, and identify the dependencies between tasks, providing accurate contextual input for subsequent construction plan generation.

[0089] Using this approach, deep semantic feature extraction and knowledge graph rule screening no longer rely solely on keyword matching, but instead combine context, grammatical structure and domain expertise to generate candidate intents. At the same time, the adaptive task dependency graph and feedback correction mechanism introduced in the process can dynamically adjust the priority of different subtasks, effectively distinguish between basic knowledge queries and fine-grained demand confirmation, and ensure that the multi-level intent of complex inputs is fully identified, thereby improving the accuracy of task decomposition for construction projects.

[0090] Figure 5 1 is a flow chart of a method for integrating the duration and construction path of each construction subtask based on constraint conditions in a reasoning model provided by an embodiment of the present application. Figure 5 As shown, the method comprises the following steps: In step S501, a target subtask set is input into the reasoning model.

[0091] In step S502, the inference model obtains available resource information.

[0092] The available resource information includes the total available resources within the estimated construction period of the target project and the available resources within a preset time interval.

[0093] In step S503, the reasoning model calls a constraint solving algorithm to determine key subtasks and bottleneck subtasks in the target subtask set based at least on the available resource information.

[0094] In step S504, the inference model calls the path planning algorithm to schedule each construction subtask in the target subtask set based on the constraints and the key subtasks and bottleneck subtasks, and obtain the duration and construction path of each construction subtask.

[0095] In some embodiments of the present application, when the reasoning model integrates the construction period and construction path of each construction subtask based on constraint conditions, it can first obtain the target subtask set and available resource information, and the available resource information includes the total available resources within the estimated construction period of the target project and the available resources within a preset time interval.

[0096] Next, the reasoning model can call the constraint solving algorithm to determine the key subtasks and bottleneck subtasks in the target subtask set based on at least the available resource information. The reasoning model can also call the path planning algorithm to schedule each construction subtask in the target subtask set based on the constraints and the key subtasks and bottleneck subtasks to obtain the duration and construction path of each construction subtask.

[0097] As mentioned above, the reasoning model can be implemented by the large model DeepSeek-R1. Before calling DeepSeek-R1 for reasoning, the existing construction plan data can be structurally annotated. Specifically, it includes: Information extraction: Use pre-trained information extraction models to extract key information such as task name, construction period, process logic, resource usage (including manpower, equipment and materials) from historical construction plans and construct them into structured data.

[0098] Keyword extraction: Use word segmentation, TF-IDF and BERT semantic matching algorithms to generate a keyword list for each task segment, such as "foundation construction", "main beam casting", "equipment scheduling", etc.

[0099] Rule verification: Perform preliminary checks on the extracted data using built-in rules, such as ensuring that the sequence of key tasks complies with engineering logic and eliminating abnormal data.

[0100] Build prompt word templates: After integrating annotations and keyword information, preset prompt templates with different formats according to different task types. For example, for a duration calculation task, the prompt word template may include: "Consider that Task A (duration X days) must be completed after Task B; Task C can be done in parallel with Task D, and the resource limit is... Please calculate the overall duration and critical path." For resource allocation tasks, the template embeds “Construction resource bottleneck: XX equipment, YY material, total resource quantity Z, please refer to the allocation logic…”.

[0101] The prompt template not only contains static instructions, but also allows dynamic filling: through the aforementioned search and annotation results, the extracted parameters, numerical information and keywords are filled into the template to ensure that DeepSeek-R1 receives complete, structured and accurate information.

[0102] Prompt word optimization feedback mechanism: The system is designed with a loop feedback mechanism. When the output generated by DeepSeek-R1 does not meet expectations, Qwen2.5 will adjust the details in the prompt template based on the comparison results, such as adding more constraints or restating the prompt, and call the inference model again.

[0103] During this process, the system retains the intermediate results of the reasoning process to analyze the cause of the error and modify the prompt words and data structure format in a targeted manner.

[0104] Taking the construction field as an example, in this field, for example, the system can be considered to combine the common structure of construction projects, and divide the entire project into construction sections, such as foundation, main body, decoration and other factors, and generate corresponding reasoning templates for each stage. The template can list, for example: the main tasks and subtasks of each stage; the logical dependencies between tasks, such as "the main body template can be installed only after the foundation pouring is completed"; the resource constraints set for each task, such as the number of manpower, materials, equipment and the use cycle; the timing requirements for task execution and the allowed parallel combinations.

[0105] In specific reasoning, DeepSeek-R1 first performs constraint solving based on the template, such as calculating the parallelism of tasks within each construction stage and identifying key tasks and bottlenecks.

[0106] Next, DeepSeek-R1 gradually expands the task calculation in the order of chain thinking: The first step is to build a preliminary task network based on the task sequence and dependency relationship; In the second step, the feasible task start time and duration are calculated by combining the resource requirements of each task and the daily available resources through the built-in linear programming module; The third step is to adjust the task combination and arrangement through iterative feedback until all constraints are met; The fourth step is to output the critical path diagram and task execution schedule.

[0107] The model also generates explanatory text to explain how the constraints (such as resource limitations, time dependencies, and construction segment divisions) used in the reasoning process affect task sequencing and overall construction duration.

[0108] In some embodiments of the present application, both the intent parsing model and the reasoning model can be interactive models. When the intent parsing model performs intent parsing on the coding vector, in response to receiving the first interactive information, the coding vector is updated based on the first interactive information, and the updated coding vector is re-performed intent parsing. When the reasoning model integrates the duration and construction path of each construction subtask, in response to receiving the second interactive information, the constraint condition is updated based on the second interactive information, and the duration and construction path of each construction subtask are integrated based on the updated constraint condition to obtain the construction plan of the target project.

[0109] In order to improve the focus of the two models in the construction field, fine-tuning training can also be carried out based on a large amount of domain corpus to make the model more sensitive to professional terminology, construction processes and resource scheduling; at the same time, the latest construction cases and data are regularly introduced to update the training parameters.

[0110] At the prompt word level, an automatic evaluation module can be set up to dynamically adjust the prompt template based on the matching degree between the inference result and the historical correct case. For example, a rule engine can be set up, and the rules include but are not limited to: Task sequence rules: such as "foundation construction must be completed before the main project can be started"; Resource allocation rules: such as "a single device cannot be scheduled for multiple tasks in the same time period"; Construction period verification rules: Compared with historical data, the construction period of a certain stage must not be lower than the empirical lower limit.

[0111] After the reasoning is generated, the generated construction plan is automatically checked through the rule engine. If any illogicality or resource conflicts are found, it is automatically fed back to Qwen2.5 for correction or rescheduling.

[0112] The rule engine also supports a manual intervention mechanism, allowing users to make modifications to the generated results and provide feedback to the system for subsequent model retraining and prompt word optimization. After viewing the initially generated construction plan, users can ask questions about certain tasks, such as "Is the resource arrangement for Task B reasonable?" The system automatically records the question and calls a two-way feedback mechanism to provide an explanatory answer based on historical data and the constraints of the rule engine. If necessary, the construction plan part of a specific task can be regenerated, and the details of different plans can be displayed by comparison to allow users to participate in the final decision. At the same time, the feedback data is used as a sample for a new round of fine-tuning training to continuously improve the optimization capabilities and engineering adaptability of the generated results.

[0113] The following takes the intention parsing model as Qwen2.5 and the reasoning model as DeepSeek-R1 as an example to explain in detail the implementation process of the construction plan generation method provided in the embodiment of the present application.

[0114] Step 1. Document parsing and knowledge base construction: Import PDFs such as construction organization plans into the system. PyMuPDF and other libraries can be used to parse PDFs, and multi-column layouts and OCR recognition errors can be handled when extracting text. Regular cleaning is performed on the extracted plain text to remove irrelevant content such as headers, footers, and directories. Semantic blocks are divided according to the natural structure of the document (chapter / paragraph), and each block is attached with a source identifier. Each block is then encoded into a vector storage using the Embedding model. It is preferred to use a pre-trained Chinese domain model to obtain embeddings to improve the representation of architectural terms. Finally, a knowledge base index is established in the vector space to support subsequent similarity retrieval.

[0115] Step 2: User interaction and task analysis: Users make task requests through the interface. For example, input: "Please generate a detailed construction schedule for this project based on the uploaded plan." Qwen2.5 receives the request, analyzes the sentence structure and intent, and recognizes that the task type is "plan generation" rather than a simple question and answer. Based on this, it prepares an internal work plan, including the information categories to be found (such as process flow, schedule, resources, etc.), as well as the format structure of the final output (such as the text chapter framework).

[0116] Step 3: Multiple rounds of searching and information acquisition: Qwen2.5 searches item by item according to the predetermined information requirement list: - First search: Generate query Q1: "Overall construction arrangement" and submit it to the vector search module. The module returns several most relevant text blocks, such as paragraphs describing the overall layout of the construction organization and project overview statements in the document. Qwen2.5 parses these contents and extracts useful information into the temporary knowledge pool, such as construction zoning method, overall construction period requirements, etc.

[0117] - Second search: Based on the information obtained, Qwen2.5 decides to query Q2: "Construction phase division and sequence". The search returns the description paragraphs about construction phases or milestones in the document. After reading, the model obtains knowledge such as phase division principles and sequence relationships. If a paragraph in the document mentions a specific phase sequence (such as "foundation → main body → decoration"), the model records it for subsequent sorting logic.

[0118] - The third search: Continue to ask questions about duration and resources. For example, Q3: "Planned duration of each sub-project", retrieves content containing duration data of each part of similar projects, or the text of the schedule example table in the document. The model extracts the duration estimate of each task from it. Since there are multiple numbers involved, Qwen2.5 submits the task-duration pair list to DeepSeek-R1 for further reasoning and calculation after understanding it (see step 4 for details here).

[0119] - Subsequent retrieval: Qwen2.5 can perform multiple rounds of queries according to task requirements, such as safety management measures, quality assurance measures, resource allocation standards, etc., and extract the key points of the results in each round. During the whole process, if the results of a certain round of retrieval are not satisfactory (such as low relevance), the model can optimize the query terms and retry the retrieval. When the knowledge pool covers the main information required for planning, the retrieval phase ends. It is worth emphasizing that this multi-round step-by-step retrieval is equivalent to providing the model with contextually enriched information, making the subsequent generation more comprehensive and accurate.

[0120] Step 4: Inference model assisted calculation: During the information acquisition process, for the content that needs to be inferred, the system immediately calls DeepSeek-R1 for processing: - Critical path calculation: Based on the task duration data retrieved by Q3, DeepSeek-R1 receives the task list and its sequence (the order of stages extracted by the previous step Qwen2.5, etc.) as input, and calculates the critical path and total duration of the entire project through internal chain reasoning. Its method is similar to CPM (Critical Path Method): accumulating duration in a tree relationship to find the longest path. DeepSeek-R1 can verify each step of the calculation and avoid missing dependencies due to its self-reflection and inspection capabilities. The calculation results are returned to Qwen2.5, including a list of critical tasks and corresponding durations.

[0121] - Resource matching verification: If the document provides personnel / equipment configuration standards, such as how many workers or machines are required for each process, Qwen2.5 can request DeepSeek-R1 to check the rationality of task and resource matching in the specific plan after obtaining these rules. For example, given a list of tasks and their respective resource arrangements, the inference model determines whether there is a resource conflict or shortage based on the rules. If a problem is found (for example, two parallel tasks exceed the upper limit of equipment configuration), DeepSeek-R1 will output prompts and suggest adjustment plans. Qwen2.5 will modify the draft plan accordingly to balance resources.

[0122] - Safety interval reasoning: Some construction processes require a safety interval (for example, concrete must be cured for 7 days before the next step can be taken). When document knowledge indicates such constraints, Qwen2.5 calls DeepSeek-R1 when scheduling to check whether the time difference between adjacent tasks meets the requirements, ensuring that the plan meets safety specifications.

[0123] In this way, DeepSeek-R1 plays the role of an intelligent calculator and auditor, supporting rigorous calculations that Qwen2.5 is unable to handle. Compared with letting the dialogue model barely complete complex reasoning, the introduction of this module significantly improves the accuracy and professionalism of the results.

[0124] Step 5: Generate a draft plan: After the information and reasoning results are complete, Qwen2.5 enters the content generation stage. It writes the text paragraph by paragraph based on the established plan document structure: - General description: Summarize the project overview and overall construction deployment, quote the retrieved project overview data, and describe the overall idea of ​​construction organization.

[0125] - Construction schedule: List the planned start and completion time (or duration) of the main construction phases and each sub-project. Here, Qwen2.5 incorporates the critical path results calculated by DeepSeek-R1 into the text, for example, stating "The total construction period is expected to be X months, and the critical path is pile foundation → basement structure → main structure...etc."

[0126] - Construction method: For each stage, briefly describe the construction technology and key points of the process. This part is directly taken from the construction plan description paragraph in the document, and Qwen2.5 makes necessary restatements to make the language coherent and smooth. When quoting professional terms, keep them consistent with the original text.

[0127] - Resource allocation plan: describes the personnel organization of the project department, the plan for the investment of mechanical equipment, etc. Qwen2.5 generates statements according to the configuration standards provided in the document, such as "This project adopts two-shift construction and invests several large equipment (2 tower cranes, 2 construction elevators...)". The data used are all from the search content or reasoning verification to ensure accuracy.

[0128] - Safety and quality measures: List the safety construction measures and quality control points required in the document, which are extracted by the model and written into the plan one by one.

[0129] - Progress guarantee measures: Combined with the documents and reasoning conclusions, propose measures to ensure progress, such as increasing parallel working faces, focusing on key line monitoring, etc. DeepSeek-R1's previous analysis can provide which tasks are critical tasks and need special attention.

[0130] - Schedules or diagrams: If necessary, Qwen2.5 can also generate structured content such as descriptions of construction progress Gantt charts or resource planning tables as required. Since Qwen2.5 is good at outputting JSON or table formats, if the user requires data format output, the format can be specified in the dialogue, and the model can organize the content according to the structure. During the generation process, Qwen2.5 always follows instructions and user preferences: for example, if the user requires a formal tone and clear organization, the model will adjust the writing style accordingly (this is also an area where the Qwen series models excel). The generated draft language conforms to the specifications of engineering technical documents, with clear paragraphs and prominent points.

[0131] Step 6, manual review and feedback correction: After the generation is completed, the user can browse the entire draft construction plan. Regarding the content, the user can ask "Why is a certain task scheduled at this time?" The system will retrieve relevant explanations to enhance persuasiveness. If the user thinks that a certain part needs to be modified (for example, the duration is too long and you want to shorten it), you can raise it in the dialogue. The system will try to adjust according to the user's requirements: on the one hand, modify the text description, and on the other hand, compress the duration of the corresponding task or adjust the order through the reasoning model without violating the hard constraints, and then generate a modified plan fragment for the user to review. This feedback-correction cycle can be carried out many times until the user is satisfied.

[0132] Step 7. Final output: After the user confirms the plan, click Export to obtain the final construction plan file. This file can be used as part of the formal construction organization design. While saving, the system also stores the structured plan data of the project for future query or version tracing. In addition, these confirmed plan data can also feed back to the system's knowledge base, continuously enriching the corpus of the dialogue model, making it more comfortable with the planning of similar projects.

[0133] The above implementation process includes the following key steps: Implementation of semantic vector retrieval: A retrieval method based on semantic vectors is used instead of simple keyword matching. By vectorizing PDF text blocks offline to build indexes, the speed and accuracy of retrieval are greatly improved. Vector retrieval can identify synonymous expressions and semantically related content fragments. For example, when a user queries "tower crane installation steps", even if the document uses the wording "crane installation plan", semantic retrieval can still find the corresponding paragraph, which is difficult to do with keyword retrieval. In order to improve the accuracy, a cross-coding Rerank step is also introduced: for the first N blocks obtained from the preliminary retrieval, Qwen2.5 calculates the relevance score after splicing the query and the block content to exclude individual semantically related but context-independent results. In this way, the reference content finally provided to the generation module is more in line with the problem.

[0134] Interface design of the dialogue and reasoning model: During implementation, a communication protocol can be designed for Qwen2.5 and DeepSeek-R1, similar to the "tool usage" mode. Qwen2.5 can construct a prompt with a special mark internally to request DeepSeek-R1 to perform calculations. For example:<call_reasoner> Calculate critical path: Enter... <end>After receiving the prompt, DeepSeek-R1 will give the calculation process and results in a rigorous way. Then we parse its output and feed the results back to Qwen2.5 for further processing. This interface ensures that the information exchange between the two models has clear boundaries to avoid mutual interference. In practice, by tuning the prompts, the output format of DeepSeek-R1 is easy to parse. For example, the key path results are fixed in JSON format, including a list of key tasks and the total duration value. In this way, after Qwen2.5 gets the JSON format data, it can directly reference the data to generate a description. In addition, a timeout and verification mechanism is set. If the inference model does not respond for a long time or the output does not conform to the expected format, Qwen2.5 will retry the request or adopt a downgrade solution, such as trying to simplify the reasoning by itself. The robust design of the interface ensures smooth collaboration of the entire system.

[0135] Long context processing and window optimization: Construction documents may be very large, and the input length of the model is limited, so a combination of sliding windows and summary compression can be used for optimization. For each retrieved document block, if the length is close to the upper limit of the model input and the content is complicated, you can first use Qwen2.5 to generate a compressed summary for it, or only extract sentences that are highly relevant to the query to reduce the input size. When generating dialogues, the context provided to the model in a single round of interaction is also controlled not to exceed its 8K-token capability range. When a large amount of text needs to be quoted, it can be generated in segments and then the results can be spliced. In addition, the Qwen2.5 model with smaller parameters can be used to perform retrieval and partial generation tasks, and the large model focuses on complex paragraph generation to improve response efficiency. These engineering strategies ensure that the system can handle more than 100 pages of documents and still maintain a fast response speed.

[0136] Professional terminology and format processing: There are a large number of special terms and fixed expressions in the construction field, such as "construction joints", "rebar binding", "concrete strength C30", etc. By expanding the domain vocabulary of Qwen2.5, professional vocabulary lists can be added to the prompt template to stabilize its translation and generation. In addition, in terms of output format, special attention should be paid to the generation of tables and lists: for example, construction plans often list tasks and time in tabular form. We can let the model directly output Markdown tables or specific delimiter formats, and then render them into the final table to ensure alignment. For unit conversion (such as days to months), DeepSeek-R1 can verify whether the units are consistent. This kind of detailed processing makes the generated plan document format standardized and professional, which is convenient for direct use.

[0137] Integration with existing systems: The system that implements the construction plan generation method provided in the embodiment of the present application can also provide an API interface, which can be seamlessly connected to the company's existing project management platform. For example, documents and parameters can be uploaded through the API and the plan results can be returned. This enables companies to embed AI-assisted compilation into their own workflows. In addition, for highly sensitive project data, the system can also be deployed and run locally in a private manner to ensure information security.

[0138] The construction plan generation method provided in the embodiment of the present application can be implemented by a construction plan generation system, which at least includes an intention parsing model and a reasoning model. Typical usage scenarios of the system may include: Preparation of construction plan for new projects: Before the project starts, the project chief engineer or planning engineer needs to prepare a detailed construction organization plan. When using this system, they can first upload the PDF materials such as the construction organization design plan, technical specifications, etc. related to the project, and enter the basic profile of the project (such as the scale of the project, the construction period requirements, etc.). Then, request the system to generate a construction plan through the dialogue interface. The system will interact with the user to ask for necessary parameters (such as the start date of construction, the number of parallel working surfaces, etc.), and then automatically generate a plan containing the division of construction phases, the main tasks and duration of each phase, milestone nodes, resource allocation tables, etc. After the user obtains the first draft, he can further ask the basis for certain arrangements, such as "Why does the main structure construction plan take X days?" The system will retrieve the document basis to answer. After several rounds of adjustments and confirmations, the user can export the final plan for approval or on-site disclosure. The whole process is greatly accelerated compared to the traditional preparation method, and the construction plan provided by this system is well-founded and reduces repeated modifications.

[0139] Dynamic adjustment of the construction process: If there are changes in the plan during the construction process, such as shortening the construction period and adjusting the construction sequence, users can also use this system to reorganize the plan. For example, if a delay in a certain process causes the subsequent plan to be rescheduled, the user enters the new situation into the system and asks how to adjust the schedule. The system will retrieve the original plan and industry knowledge, recalculate the critical path based on the reasoning model, generate an adjusted schedule and highlight the changed parts for the project manager to make a decision. This "real-time interactive" plan adjustment assistance is different from the past adjustment that completely relied on manual experience, and can converge to an optimized plan more quickly.

[0140] Knowledge Q&A and training: Construction companies can also use this system for engineering knowledge Q&A or new employee training. When faced with a heavy construction organization design, new engineers can ask questions through the system in the form of dialogue, such as "What construction method is used for the foundation construction of this project? What are the advantages and disadvantages?" The system will extract the corresponding paragraphs from the document and give an explanation. For example, in a training scenario, the teaching plan allows students to use the system to query construction plan cases for typical projects. When answering, the system can explain the source of each part of the plan step by step to help students deepen their understanding. In these scenarios, Qwen2.5's dialogue capabilities ensure that the answers are easy to understand, while DeepSeek-R1 ensures that questions involving engineering calculations are answered correctly (such as how to calculate the amount of steel bars, etc.).

[0141] Project bidding and scheme comparison: During the bidding stage, it is necessary to quickly prepare a construction organization plan to respond to the requirements of the bidding documents. Using this system, bidders can input the technical conditions provided by the bidding and the company's previous similar engineering schemes, and let the system automatically generate a draft construction organization plan for this project. Since the system can fully call on the successful experience in historical schemes (such as efficient construction technology or optimized staffing), the generated scheme is competitive. The bidder can then make minor adjustments based on the project's particularities to finalize the draft, greatly improving the efficiency of preparing bidding documents. Similarly, when comparing multiple schemes, users can let the system quickly generate multiple sets of draft plans based on different focuses, and then compare the pros and cons and make a choice based on them.

[0142] The above scenarios show the various ways in which construction companies can apply this system. Whether it is preliminary planning, process adjustment or knowledge support, the system can provide users with efficient and reliable assistance through a friendly dialogue interface and powerful retrieval and reasoning capabilities. This will significantly lower the threshold for construction plan preparation, allowing project teams to devote more energy to decision-making and innovation, while leaving the tedious information search and calculation tasks to the intelligent system.

[0143] All the above optional technical solutions can be arbitrarily combined to form optional embodiments of the present application, which will not be described one by one here.

[0144] The following is an embodiment of the device of the present application, which can be used to execute the embodiment of the method of the present application. For details not disclosed in the embodiment of the device of the present application, please refer to the embodiment of the method of the present application.

[0145] Figure 6 Schematic diagram of a construction plan generation device provided in an embodiment of the present application. Figure 6 As shown, the device comprises: The acquisition module 601 is configured to acquire the construction design document of the target project and pre-process the construction design document to obtain the encoding vector of the design document.

[0146] The parsing module 602 is configured to, in response to receiving an instruction to build a construction plan for a target project, input the coding vector of the design document into the intent parsing model, perform intent parsing on the coding vector based on the knowledge graph of the project domain and preset parsing rules, and obtain an initial construction subtask set; the initial construction subtask set includes at least one initial construction subtask and constraints of each initial construction subtask, and the constraints include at least one of the following: the priority of each initial construction subtask, the resource constraint information of each initial construction subtask, and the dependency relationship between different initial construction subtasks.

[0147] The retrieval module 603 is configured to search each initial construction subtask in the intention parsing model to obtain the retrieval results of each initial construction subtask; the retrieval results at least include the construction requirements and historical construction tasks corresponding to the initial construction subtask.

[0148] Iteration module 604 is configured to, in response to determining that the similarity between the task attributes and constraints of the target initial construction subtask and the retrieval results is less than or equal to a first preset similarity threshold, re-intent-parse the encoding vector of the design document to obtain an updated subtask set; the target initial construction subtask is at least one construction subtask in the initial construction subtasks.

[0149] The iteration module 604 is also configured to determine that the current subtask set is a target subtask set in response to determining that the similarity between the task type and constraint conditions of each subtask in the current construction subtask and the retrieval results is greater than a first preset similarity threshold, wherein the current subtask set is a subtask set that is compared with the current retrieval results of the intent parsing model, and the current construction subtask is a construction subtask in the current subtask set.

[0150] The reasoning module 605 is configured to input the target subtask set into the reasoning model, integrate the construction period and construction path of each construction subtask based on the constraint conditions, and obtain the construction plan of the target project.

[0151] According to the technical solution provided by the embodiment of the present application, the construction design document of the target project is first preprocessed to obtain a coding vector, and then the intent parsing model responds to the received instruction to build a construction plan to perform intent parsing on the coding vector to obtain an initial set of construction subtasks; the intent parsing module also retrieves each subtask in the initial set of construction subtasks, and adjusts the subtask division method based on the retrieval results until a target subtask set that meets the preset conditions is obtained; finally, the target subtask set is input into the inference model, and the inference model integrates the duration and construction path of each construction subtask based on the constraints in the target subtask set to obtain a construction plan for the target project. The method uses the intent parsing model to perform intent parsing and task retrieval on the construction design document to obtain a reasonably divided subtask set, and then uses an independent inference model to schedule and plan the path of each task in the subtask set, thereby improving the efficiency of automatic generation of the construction plan, and the generated construction plan is more in line with user needs.

[0152] In some embodiments, the task attributes of each construction subtask include basic knowledge subtasks, independent construction subtasks, and collaborative construction subtasks; in response to determining that the target construction subtask is a basic knowledge subtask, the target construction subtask is searched in the intent parsing model to obtain a search result for the target construction subtask, including: obtaining a target coding vector corresponding to the target construction subtask; using the target coding vector to search in a knowledge graph of the project field in a keyword matching manner to obtain a basic knowledge search result set; determining that the search result in the basic knowledge search result set whose vector similarity with the target coding vector is greater than a second preset similarity threshold is the search result for the target construction subtask; wherein the target construction subtask is any construction subtask in the initial construction subtask set or the updated subtask set.

[0153] In some embodiments, in response to determining that a target construction subtask is an independent construction subtask, the target construction subtask is retrieved in the intent parsing model to obtain a retrieval result for the target construction subtask, including: determining a task identifier of the target construction subtask, and determining a retrieval template based on the task identifier; the task identifier is used to characterize the task type of the independent construction subtask; calling the retrieval template to retrieve a set of historical similar tasks, the set of historical similar tasks including historical independent construction subtasks with the same task identifier as the target construction subtask; calculating the similarity between task parameters of each task in the set of historical similar tasks and the constraints of the target construction subtask; determining that the tasks in the set of historical similar tasks whose calculated similarity is greater than a third preset similarity threshold are retrieval results for the target construction subtask.

[0154] In some implementations, in response to determining that a target construction subtask is a collaborative construction subtask, the target construction subtask is retrieved in the intent parsing model to obtain a retrieval result for the target construction subtask, including: determining a task identifier of the target construction subtask and its collaborative construction subtask, and determining a retrieval template based on the task identifier; calling the corresponding retrieval template for each construction subtask to retrieve each respective historical similar task set, the historical similar task set including historical construction subtasks having the same task identifier as each construction subtask, the historical construction subtasks being independent construction subtasks or collaborative construction subtasks; obtaining constraints for the target construction subtask and its collaborative construction subtask Conditions, component matching is performed on the historical similar task set of each construction subtask based on the constraint conditions to obtain an initial matching result; in response to determining that the similarity between the initial matching result and the design document is less than a second preset similarity threshold, at least one of the constraint conditions and the target retrieval template is adjusted, and retrieval matching is performed again after the adjustment is completed; the target retrieval template is at least one retrieval template in the retrieval templates corresponding to the target construction subtask and its collaborative construction subtasks; in response to determining that the similarity between the matching result and the design document is greater than or equal to the second preset similarity threshold, determining that the task in the historical similar task set corresponding to the target construction subtask is the retrieval result of the target construction subtask.

[0155] In some embodiments, adjusting at least one of the constraint conditions and the target retrieval template includes adjusting at least one of the following: modifying the priority of the target construction subtask and its collaborative construction subtasks; adding the dependency of the target construction subtask and its collaborative construction subtasks in the target retrieval target.

[0156] In some embodiments, the target subtask set is input into the reasoning model, and the duration and construction path of each construction subtask are integrated based on the constraint conditions, including: the target subtask set is input into the reasoning model; the reasoning model obtains available resource information, and the available resource information includes the total available resources within the estimated construction period of the target project and the available resources within a preset time interval; the reasoning model calls the constraint solving algorithm to determine the key subtasks and bottleneck subtasks in the target subtask set based on at least the available resource information; the reasoning model calls the path planning algorithm to schedule each construction subtask in the target subtask set based on the constraint conditions and the key subtasks and bottleneck subtasks to obtain the duration and construction path of each construction subtask.

[0157] In some embodiments, both the intent parsing model and the reasoning model are interactive models; when the intent parsing model performs intent parsing on the coding vector, in response to receiving first interactive information, the coding vector is updated based on the first interactive information, and the updated coding vector is re-performed intent parsing; when the reasoning model integrates the construction period and construction path of each construction subtask, in response to receiving second interactive information, the constraint conditions are updated based on the second interactive information, and the construction period and construction path of each construction subtask are integrated based on the updated constraint conditions to obtain the construction plan of the target project.

[0158] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0159] Figure 7 Schematic diagram of an electronic device provided in an embodiment of the present application. Figure 7 As shown, the electronic device 7 of this embodiment includes: a processor 701, a memory 702, and a computer program 703 stored in the memory 702 and executable on the processor 701. When the processor 701 executes the computer program 703, the steps in the above-mentioned method embodiments are implemented. Alternatively, when the processor 701 executes the computer program 703, the functions of the modules / units in the above-mentioned device embodiments are implemented.

[0160] The electronic device 7 may be a desktop computer, a notebook, a PDA, a cloud server, or other electronic device. The electronic device 7 may include, but is not limited to, a processor 701 and a memory 702. Those skilled in the art will appreciate that Figure 7 The electronic device 7 is merely an example and does not limit the electronic device 7 , and may include more or less components than those shown in the figure, or different components.

[0161] The processor 701 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0162] The memory 702 may be an internal storage unit of the electronic device 7, for example, a hard disk or memory of the electronic device 7. The memory 702 may also be an external storage device of the electronic device 7, for example, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the electronic device 7. The memory 702 may also include both an internal storage unit and an external storage device of the electronic device 7. The memory 702 is used to store computer programs and other programs and data required by the electronic device.

[0163] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units.

[0164] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.< / end>

Claims

1. A construction plan generation method, characterized in that: include: Obtaining a target project construction design document, and preprocessing the construction design document to obtain a coding vector of the design document; In response to receiving an instruction to construct a construction plan for a target project, inputting the encoding vector of the design document into an intent parsing model, performing intent parsing on the encoding vector based on a knowledge graph of the field in which the project is located and preset parsing rules, and obtaining an initial construction subtask set; the initial construction subtask set includes at least one initial construction subtask and constraints of each initial construction subtask; Searching each initial construction subtask in the intention parsing model to obtain search results for each initial construction subtask; In response to determining that the similarity between the task attributes and constraints of the target initial construction subtask and the search result is less than or equal to a first preset similarity threshold, re-performing intent parsing on the encoding vector of the design document to obtain an updated subtask set; the target initial construction subtask is at least one construction subtask in the initial construction subtasks; In response to determining that the similarity between the task type and the constraint condition of each subtask in the current construction subtask and the search result is greater than a first preset similarity threshold, determining the current subtask set as the target subtask set; The target subtask set is input into the reasoning model, and the construction period and construction path of each construction subtask are integrated based on the constraint conditions to obtain the construction plan of the target project.

2. The method according to claim 1, characterized in that The task attributes of each construction subtask include basic knowledge subtasks, independent construction subtasks, and collaborative construction subtasks; In response to determining that the target construction subtask is a basic knowledge subtask, the target construction subtask is searched in the intention parsing model to obtain a search result for the target construction subtask, including: Obtain the target encoding vector corresponding to the target construction subtask; Using the target encoding vector to search in the knowledge graph of the field where the project is located in a keyword matching manner, to obtain a basic knowledge search result set; Determine the search result in the basic knowledge search result set whose vector similarity with the target coding vector is greater than a second preset similarity threshold as the search result of the target construction subtask; The target construction subtask is any construction subtask in the initial construction subtask set or the updated subtask set.

3. The method according to claim 2, characterized in that In response to determining that the target construction subtask is an independent construction subtask, the target construction subtask is searched in the intention parsing model to obtain a search result of the target construction subtask, including: Determine the task identifier of the target construction subtask, and determine the search template based on the task identifier; the task identifier is used to characterize the task type of the independent construction subtask; Calling the search template to retrieve a historical similar task set, wherein the historical similar task set includes historical independent construction subtasks having the same task identifier as the target construction subtask; Calculating the similarity between the task parameters of each task in the historical similar task set and the constraint conditions of the target construction subtask; It is determined that the tasks in the historical similar task set whose calculated similarity is greater than a third preset similarity threshold are the search results of the target construction subtask.

4. The method according to claim 2, characterized in that: In response to determining that the target construction subtask is a collaborative construction subtask, the target construction subtask is searched in the intention parsing model to obtain a search result of the target construction subtask, including: Determine the task identifiers of the target construction subtask and its coordinated construction subtask, and determine a search template based on the task identifiers; For each construction subtask, the corresponding search template is called to retrieve the respective historical similar task sets, wherein the historical similar task sets include historical construction subtasks having the same task identifier as each construction subtask, and the historical construction subtasks are independent construction subtasks or collaborative construction subtasks; Obtaining constraints of the target construction subtask and its coordinated construction subtasks, and performing component matching on a historical similar task set of each construction subtask based on the constraints to obtain an initial matching result; In response to determining that the similarity between the initial matching result and the design document is less than a second preset similarity threshold, adjusting at least one of the constraint condition and the target retrieval template, and performing retrieval matching again after the adjustment is completed; the target retrieval template is at least one retrieval template corresponding to the target construction subtask and its collaborative construction subtask; In response to determining that the similarity between the matching result and the design document is greater than or equal to a second preset similarity threshold, determining that the task in the historical similar task set corresponding to the target construction subtask is the search result of the target construction subtask.

5. The method according to claim 4, characterized in that The constraint condition includes at least one of the following: the priority of each initial construction subtask, the resource constraint information of each initial construction subtask, and the dependency relationship between different initial construction subtasks; Adjusting at least one of the constraint condition and the target search template includes adjusting at least one of the following: Modifying the priorities of the target construction subtask and its coordinated construction subtasks; The dependency relationship between the target construction subtask and its collaborative construction subtask is added to the target retrieval target.

6. The method according to claim 1, characterized in that The target subtask set is input into the reasoning model, and the construction period and construction path of each construction subtask are integrated based on the constraint conditions, including: Inputting the target subtask set into a reasoning model; The inference model obtains available resource information, wherein the available resource information includes total available resources within the estimated construction period of the target project and available resources within a preset time interval; The reasoning model calls a constraint solving algorithm to determine the key subtasks and bottleneck subtasks in the target subtask set based at least on the available resource information; The inference model calls a path planning algorithm to schedule each construction subtask in the target subtask set based on the constraint conditions and the key subtasks and bottleneck subtasks, and obtains the duration and construction path of each construction subtask.

7. The method according to any one of claims 1 to 6, characterized in that The intention parsing model and the reasoning model are both interactive models; When performing intent parsing on the code vector, the intent parsing model updates the code vector based on the first interaction information in response to receiving the first interaction information, and re-performs intent parsing on the updated code vector; When integrating the construction period and construction path of each construction subtask, the reasoning model updates the constraint conditions based on the second interaction information in response to receiving the second interaction information, and integrates the construction period and construction path of each construction subtask based on the updated constraint conditions to obtain the construction plan of the target project.

8. A construction plan generating device, characterized in that: include: An acquisition module is configured to acquire a construction design document of a target project and preprocess the construction design document to obtain a coding vector of the design document; The parsing module is configured to, in response to receiving an instruction to construct a construction plan for a target project, input the coding vector of the design document into an intent parsing model, perform intent parsing on the coding vector based on a knowledge graph of the project domain and preset parsing rules, and obtain an initial construction subtask set; the initial construction subtask set includes at least one initial construction subtask and constraints of each initial construction subtask; A retrieval module is configured to search each initial construction subtask in the intention parsing model to obtain a retrieval result of each initial construction subtask; The iteration module is configured to, in response to determining that the similarity between the task attributes and constraints of the target initial construction subtask and the search result is less than or equal to a first preset similarity threshold, re-parse the encoding vector of the design document to obtain an updated subtask set; the target initial construction subtask is at least one construction subtask in the initial construction subtasks; The iteration module is further configured to determine that the current subtask set is the target subtask set in response to determining that the similarity between the task type and the constraint condition of each subtask in the current construction subtask and the search result is greater than a first preset similarity threshold; The reasoning module is configured to input the target subtask set into the reasoning model, integrate the construction period and construction path of each construction subtask based on the constraint conditions, and obtain the construction plan of the target project.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images