An engineering report generation and quality control method based on multi-agent collaboration and electronic equipment

Abstract

The application discloses a kind of engineering report generation and quality control method and electronic equipment based on multi-agent cooperation. The method comprises: receiving engineering source file and constructing benchmark fact atlas for read-only locking;Run the masking algorithm to replace historical report data into isolated vector database;Control the writing agent to retrieve the skeleton and fill in the true value to generate the target paragraph;Control the review agent to carry out triple extraction and entity attribute comparison and trigger reconstruction;In response to the save instruction, trigger gateway interception to mount data to sandbox memory;End detection and automatic revision in sandbox, record tamper-proof log. The application can completely solve the problem of historical parameter pollution, provide deterministic constraints for multi-agent text generation, avoid model illusion error, and ensure that the report is globally consistent with the chapter context.

Description

Technical Field

[0001] This application relates to the fields of data processing and natural language processing, and conforms to the classification direction of natural language processing, information retrieval and data management. Specifically, it refers to an engineering report generation and quality control method and electronic device based on multi-agent collaboration applied to the scenario of generating feasibility study reports for large-scale engineering construction projects. Background Technology

[0002] Feasibility study reports for large-scale engineering construction projects typically include several highly related professional chapters, such as project overview, site selection requirements, environmental assessment, and financial calculations. During the preparation of such reports, technical personnel often refer to high-quality, similar reports from previous projects as templates for modification, in order to improve preparation efficiency and reuse existing writing standards.

[0003] Currently, existing report preparation solutions based on conventional document processing software or traditional document generation systems are insufficient to meet the high-precision data control requirements of engineering projects, exhibiting significant shortcomings. Firstly, historical data parameter contamination is severe. When modifying imported historical templates, technicians are highly likely to overlook engineering parameters hidden within the long text of the template. This leads to invalid historical data being mixed into the generated new report, affecting the accuracy of the engineering report and potentially causing construction risks. Secondly, inconsistencies exist between data across chapters. Engineering feasibility study reports are lengthy, and existing systems lack a globally linked constraint mechanism when core project parameters change during iterative calculations. This results in conflicts between data from the same source cited in different chapters, and manually reviewing such implicit conflicts sentence by sentence is extremely costly, severely reducing the overall efficiency of engineering design and report delivery. Summary of the Invention

[0004] This application provides a method and electronic device for generating and controlling engineering reports based on multi-agent collaboration, which solves the technical defects of serious historical data pollution and inconsistencies in facts across chapters in the existing engineering report generation process.

[0005] This application provides a method for engineering report generation and quality control based on multi-agent collaboration, including the following steps: receiving engineering source files, extracting parameter triples to construct a baseline fact graph and locking it in read-only mode; running a masking algorithm on imported historical reference reports, replacing numerical entities and proper nouns in the historical reference reports with special placeholders, and storing them in an isolated vector database; controlling the writing agent to retrieve the syntactic skeleton from the isolated vector database, calling the truth values ​​in the baseline fact graph to fill the syntactic skeleton, and generating the target paragraph; controlling the review agent to perform reverse entity extraction on the target paragraph to obtain a generated fact subgraph, comparing the generated fact subgraph with the baseline fact graph for conflict nodes, and if a conflict is found, generating a conflict patch containing revision logic and triggering a refactoring instruction; responding to the document save instruction, triggering an input / output gateway interception mechanism to mount the document data to sandbox memory; performing final reverse extraction of cross-chapter entities in sandbox memory, and if cross-chapter entities are detected to be inconsistent with the baseline fact graph, using an automatic revision agent to generate differential replacement text to cover erroneous fields, removing the input / output gateway interception mechanism, and recording an anti-tampering hash log.

[0006] Specifically, the imported historical reference reports are subjected to a masking algorithm, which replaces the numerical entities and proper nouns in the historical reference reports with special placeholders. This includes: loading a preset engineering domain entity recognition model to perform sequence labeling on the historical reference reports, and outputting a continuous string containing numerical entities and proper nouns; generating a special placeholder sequence corresponding to the continuous string, which includes an identifier representing the entity type and an incrementing index number; and using the special placeholder sequence to cover the corresponding original characters according to the position information of the continuous string in the original text, generating parameter-free desensitized text, and storing the desensitized text as a syntactic skeleton in the isolated vector database.

[0007] Specifically, based on the position information of the continuous strings in the original text, a special placeholder sequence is used to overwrite the corresponding original characters. The configuration is as follows: extract the start byte offset and end byte offset of the continuous strings in the historical reference report; construct mask markers based on the entity type identifier; establish a mapping table that records the correspondence between mask markers, start byte offsets, and end byte offsets; traverse the mapping table, and for each record, delete the byte array in the historical reference report between the start byte offset and the end byte offset, and insert the bytecode sequence corresponding to the mask marker at the corresponding start position to form the updated character stream sequence.

[0008] Specifically, the generated fact subgraph and the baseline fact graph are compared to perform conflict node comparison operations, including: extracting the specific numerical attributes corresponding to each node in the generated fact subgraph; searching for baseline nodes with the same named entity as the node in the baseline fact graph; reading the baseline numerical attributes of the baseline node and the tolerance threshold range configured on the baseline node; calculating the difference between the specific numerical attributes and the baseline numerical attributes, and determining whether the difference falls within the tolerance threshold range; if the difference exceeds the tolerance threshold range, the current node is determined to be a conflict node.

[0009] Specifically, the process generates a conflict patch containing revision logic and triggers a refactoring instruction. The configuration is as follows: obtain the sentence-level context of the erroneous entity identified as a conflict node in the target paragraph; extract the baseline numerical attributes of the baseline node as the forced constraint truth value; construct a structured patch object containing the erroneous entity, the forced constraint truth value, and the sentence-level context; call the local rewrite interface of the writing agent using the structured patch object as input parameters; set a maximum retry counter; if the current rewrite count does not exceed the threshold of the maximum retry counter, generate the corrected paragraph text based on the structured patch object; if the threshold of the maximum retry counter is reached, freeze the generation state of the target paragraph.

[0010] Specifically, in response to a document save command, the input / output gateway interception mechanism is triggered to mount the document data to the sandbox memory, including: intercepting write operation request streams for persistent storage devices; redirecting all document data in the write operation request stream to the isolated sandbox memory space; locking the read and write permissions of all document data in the sandbox memory space to prevent the all document data from being accessed by external processes; and triggering the traversal module to read all document data sequentially according to chapter order, entering the final reverse extraction stage of cross-chapter entities.

[0011] Specifically, the automatic revision agent is used to generate differential replacement text to cover erroneous fields. The configuration is as follows: locate the first and second position indices corresponding to inconsistent data in cross-chapter entities; generate a differential string stream for replacement based on the truth values ​​in the baseline fact graph; convert the full document data into a character array, erase the memory block data between the first and second position indices in the sandbox memory space; insert the byte data corresponding to the differential string stream into the first position index; calculate the change in the length of the memory array, and update the starting address pointer of the remaining characters in the full document data after the second position index.

[0012] Specifically, recording tamper-proof hash logs includes: collecting the previous snapshot data stream, the next snapshot data stream, and the differential replacement text in the sandbox memory where the overwrite operation was performed; concatenating the previous snapshot data stream, the next snapshot data stream, and the differential replacement text and performing a hash operation to generate a fixed-length hash digest of the current operation; obtaining the historical hash digest of the previous operation log; combining the historical hash digest with the fixed-length hash digest of the current operation and performing another hash operation to generate a chained hash signature and appending it to the tamper-proof log file system.

[0013] This application also provides an electronic device, including: a memory for storing a computer program; and a processor connected to the memory, configured to implement the above-described method for generating and controlling engineering reports based on multi-agent collaboration when executing the computer program.

[0014] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described method for generating and controlling engineering reports based on multi-agent collaboration.

[0015] In summary, this application establishes a globally read-only baseline fact graph to lock in core engineering parameters, providing absolutely accurate comparative benchmark data for subsequent agent actions. Simultaneously, by constructing a multi-agent collaborative architecture encompassing specialized responsibilities such as extraction, writing, and review, it enables independent writing at the paragraph level and reverse comparison verification at the numerical level. A parameter masking and isolation mechanism completely blocks the inflow of old template parameters into the generation stage. Furthermore, before document data is written to disk, a sandbox gateway interception mechanism is triggered to perform final verification and automated partial replacement, and version logs are recorded using chained hashing. Through the coordinated efforts of these steps, this application achieves a closed-loop logical barrier from underlying data inflow and content generation to final file storage, effectively preventing historical data parameter contamination and maintaining strict global data consistency during the dynamic generation and editing of multi-chapter long texts. Attached Figure Description

[0016] Figure 1 This is the main flowchart of the engineering report generation and quality control method provided in the embodiments of this application.

[0017] Figure 2 This is a flowchart of the placeholder masking, isolation, and processing logic provided in the embodiments of this application.

[0018] Figure 3 This is a flowchart of the sandbox verification interception and partial replacement processing provided in the embodiments of this application.

[0019] Figure 4 This is a logical structure block diagram of the multi-agent collaborative architecture provided in the embodiments of this application.

[0020] Figure 5 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application.

[0021] Figure labeling: 401 - Writing agent, 402 - Reviewing agent, 403 - Automatically revising agent, 404 - Isolated vector database, 405 - Baseline fact graph, 406 - Distributed message communication basic components, 407 - Sandbox memory, 408 - Input / output gateway, 501 - Memory, 502 - Processor, 503 - System bus. Detailed Implementation

[0022] The present application will now be described in further detail with reference to the accompanying drawings and embodiments.

[0023] The method provided in this application relies on a server cluster that includes processing units, memory, and a distributed database. The system's functional topology is divided into four core layers in terms of hardware entities: a fact-locking layer, an agent scheduling layer, a sandbox verification layer, and a persistent tracking layer. Based on this physical infrastructure, contamination prevention control is implemented in the engineering report generation process.

[0024] like Figure 1 As shown in the figure, the engineering report generation and quality control method based on multi-agent collaboration provided in this application includes the following detailed execution logic:

[0025] In step S101, processor 502 receives the project source file, extracts parameter triples to construct a baseline fact graph 405, and locks it in read-only mode. Specifically, the fact locking layer connects to the source data receiving interface and parses the file format and hierarchical directory of legally imported approval documents or original engineering designs. Using preset dependency parsing rules, it extracts factual information from the text and structures it into parameter triples. The structure of these parameter triples is limited to subject entity, predicate relation, and specific content. The extracted set of parameter triples is imported into the system's underlying graph database to generate a baseline fact graph 405 containing multi-node associations. The system applies write protection attributes to the node hierarchy of the baseline fact graph 405, forcibly changing its access permissions to a read-only locked state, setting it as a globally unique truth source for the system. By constructing a read-only baseline fact graph 405 and storing it in the graph database, a stable truth anchor can be established for all subsequent generation logic, preventing source data from being overwritten or tampered with by concurrent threads.

[0026] In step S102, processor 502 runs a masking algorithm on the imported historical reference reports, replacing numerical entities and proper nouns in the reports with special placeholders and storing them in the isolated vector database 404. The core purpose of this step is to cut off the path for old data to flow into the new report generation stage. Since historical templates inevitably contain outdated engineering data, the system needs to perform data anonymization before storing it in the database. Through the placeholder masking isolation mechanism, uncontaminated data containing only the syntactic structure skeleton is generated.

[0027] In step S103, the writing agent 401 retrieves the syntactic skeleton from the isolated vector database 404 and fills the syntactic skeleton with truth values ​​from the baseline fact graph 405 to generate the target paragraph. At this stage, the multi-agent collaborative system intervenes. The writing agent 401, configured on the application server, performs semantic similarity retrieval from the aforementioned parameter-contaminated isolated vector database 404 in a chapter-by-chapter mode to obtain the required reference text skeleton. During paragraph generation, since specific slots in the skeleton are replaced by placeholders, the writing agent 401 forcibly queries the read-only baseline fact graph 405, filling the empty slots with real values ​​from the baseline fact graph 405 and outputting the chapter-level target paragraph. By assembling the creative writing function of the large language model with strongly consistent baseline truth values ​​for slot filling, it is possible to retain high-fluency syntactic generation while preventing the parameter generation illusion caused by the autoregressive mechanism from the source.

[0028] After the target paragraph is generated, the processor 502 controls the review agent 402 to perform reverse entity extraction on the target paragraph to obtain a generated fact subgraph. The generated fact subgraph is then compared with the baseline fact graph 405 to identify conflict nodes. If a conflict is found, a conflict patch containing revision logic is generated, and a refactoring instruction is triggered. In practice, the generated chapter output text is transferred via memory to the review agent 402, which is responsible for quality control. The review agent 402 incorporates a reverse extraction model from the engineering domain. It performs triple reverse extraction to extract the generated text fact subgraph that actually exists within the generated text from the chapter output text. The review agent 402 calculates the tolerance state between the value of each entity attribute in the generated fact subgraph and the value of the corresponding node in the baseline fact graph 405. If the two cannot perfectly match, a conflict patch object is generated and routed back to the writing agent 401 according to the specified conflict node, forcing a regeneration. By configuring the review agent 402, which is opposed to the writing role, deterministic mathematical verification is performed on the generated paragraph, thus replacing the unreliable probabilistic self-correction system within the large model.

[0029] When the document compilation is nearing completion, step S104 is executed. In response to the document save instruction, processor 502 triggers the I / O gateway 408 interception mechanism to mount the document data to sandbox memory 407. Because the project report has multiple chapters, the system intercepts this action from the underlying I / O gateway 408 when it receives the final save or export instruction. The entire document stream currently in memory is redirected and mounted in the temporarily partitioned sandbox memory 407 runtime environment, blocking the write process to the persistent disk and providing a computation window for the final consistency verification.

[0030] In step S105, the final reverse extraction of cross-chapter entities is performed in the sandbox memory 407. If a cross-chapter entity is detected to be inconsistent with the baseline fact graph 405, the automatic revision agent 403 generates differential replacement text to cover the erroneous fields, removes the input / output gateway 408 interception mechanism, and records an anti-tampering hash log. The system performs a global reverse extraction of the entire document in the sandbox memory 407 to obtain a global fact set. If a contradiction is found in the cross-chapter reference facts during comparison, accurate differential replacement text is directly generated based on the truth value of the baseline fact graph 405, and the memory address in the sandbox is covered without leaving a trace. After the repair is completed, the input / output gateway 408 interception mechanism is removed. The final content summary of the document, along with the revision traces, is recorded as an anti-tampering operation log after hashing. By setting the forward interception and automatic revision agent 403 modules at the underlying interface, omissions that may be caused by manual operation can be effectively prevented, and a closed-loop recording of global quality data can be achieved.

[0031] like Figure 2As shown, this embodiment of the application further extends the functionality and imposes action constraints on the placeholder masking and isolation logic in the main step S102. The process includes the following specific steps: Processor 502 executes step S201, loading a preset engineering domain entity recognition model to perform sequence annotation on historical reference reports, outputting a continuous string containing numerical entities and proper nouns. Specifically, the engineering domain entity recognition model is fine-tuned and trained based on a pre-trained model architecture of engineering cost and building material annotation corpus. It utilizes a self-attention mechanism to capture word vector association features in the report context, thereby outputting a continuous string containing numerical entities and proper nouns. The engineering domain entity recognition model here has undergone weight optimization for the construction and cost engineering corpus, enabling accurate identification of physical dimension words such as elevation and earthwork volume. Step S202 is executed to generate a special placeholder sequence corresponding to the continuous string. This special placeholder sequence includes an identifier representing the entity type and an incrementing index number. Step S203 involves extracting the start and end byte offsets of the continuous string from the historical reference report; constructing a mask marker based on the entity type identifier; and establishing a mapping table. This mapping table records the structured correspondence between the generated mask markers, start byte offsets, and end byte offsets. Step S204 involves the system traversing the mapping table. For each active record in the table, the byte array in the historical reference report between the start and end byte offsets is directly erased and deleted. The bytecode sequence corresponding to the mask marker is then directly inserted at the starting position of the original continuous string, forming an updated character stream sequence. The desensitized and parameter-free character stream sequence is stored as the syntactic skeleton in the isolation vector database 404. Through a precise memory erasure and point-to-point insertion strategy based on the underlying byte offsets, it is ensured that old data is completely physically isolated and cleared, and the erasure process does not disrupt the original paragraph formatting and line breaks of the project document.

[0032] In an optional implementation, this application quantitatively constrains the conflict node comparison operation in the main step S103 using mathematical rules. The processor 502 extracts the specific numerical attributes corresponding to each node in the generated fact subgraph. It extracts the numerical parameters of the current verification object and simultaneously retrieves the benchmark node with the same named entity from the benchmark fact graph 405, reading the benchmark numerical parameters of that benchmark node and the tolerance threshold range configured for that benchmark node. After unifying the data units of both, a tolerance comparison calculation is performed. The system performs an absolute value calculation operation on the difference between the first numerical parameter extracted from the generated text and the second numerical parameter of the benchmark fact, to determine whether the absolute value is zero. According to this determination logic, if the difference is equal to 0, the current data stream is determined to be in a consistent state. If the calculation finds that the difference is greater than 0, i.e., exceeding the zero tolerance threshold range set here, the system will perform a non-empty subtraction operation on the subgraph, subtracting the benchmark fact graph from the generated fact subgraph, and forcibly determine the current node with a non-empty subtraction result as a conflict node with parameter contamination or logical calculation errors.

[0033] Upon identifying a conflict node, the system triggers the generation of a conflict patch containing revision logic. The system obtains the sentence-level context fragment of the erroneous entity identified as the conflict node within the target paragraph and extracts the baseline numerical attribute of the baseline node as the mandatory constraint truth value. A structured patch object containing the erroneous entity, the mandatory constraint truth value, and the sentence-level context fragment is constructed. The review agent 402 uses this structured patch object as input and directly calls the local rewrite application interface of the writing agent 401. During this process, the system is configured with a maximum retry counter. If the number of regenerated modifications does not exceed the set threshold of the maximum retry counter, the writing agent 401 will generate the corrected paragraph text based on the structured patch object; if the maximum retry counter threshold is reached due to network or model degradation and the review comparison still fails, the generation process of the current target paragraph is frozen. This closed-loop rewrite mechanism, combining structured object packaging and transmission with counter truncation, achieves precise guidance for correcting abnormal fields while preventing the system from entering an infinite loop of deadlock.

[0034] In another embodiment of this application, such as Figure 3The diagram illustrates the execution steps of gateway interception and cross-chapter automatic revision at the document save boundary. When the front-end application triggers the save button, the operating system's input / output interface is activated, executing step S301 to intercept the write operation request stream to the persistent storage device and redirect the full document data in the request stream to the isolated sandbox memory space 407 allocated by the system. Step S302 locks the read / write permissions of the full document data in the sandbox memory space 407, preventing the full document data block from being accessed by other unrelated external processes or scheduled tasks in the background. Step S303 triggers the built-in traversal module, sequentially reading the full document data in the sandbox memory 407 according to the chapter order of the document structure, entering the final global reverse extraction stage of cross-chapter entities, and constructing a global fact set containing long-range relational facts for the entire document. Step S304, in the automatic revision stage, the processor 502 scans and locates the first and second position indices corresponding to inconsistent data in the cross-chapter entities. Stream information of the differential replacement text is generated based on the baseline truth value. The mounted full document data is converted into a basic character array. Within the contiguous address space of sandbox memory 407, the dirty data memory block between the first and second position indices is directly erased. Step S305 is executed to perform the insertion operation. The exact byte data corresponding to the differential string stream is sequentially inserted into the first position index. Since the length of the replacement character usually changes, the system internally recalculates the change in the length of the memory array and updates the starting address pointer of the remaining characters after the second position index in the full document data based on this offset. By introducing the movement calculation of the underlying memory pointer block and the direct erase and write operation, not only is the millisecond-level timeliness of the correction action guaranteed, but also the system resource consumption caused by replacing long-spaced documents is significantly reduced by avoiding a full rendering refresh.

[0035] Based on this, after the document data is officially written to disk, the system executes the action of recording a tamper-proof hash log. The security component collects the pre-overwrite snapshot data stream, the post-overwrite snapshot data stream, and the differential replacement text that plays a major corrective role, all stored in sandbox memory 407. These three parts are concatenated, and a hash operation is performed using a specified cryptographic algorithm to generate a fixed-length content digest for this finalization operation, which is recorded as a hash value. The historical hash digest retained from the previous operation log is retrieved. This historical hash digest is combined bitwise with the hash value of the current operation and subjected to a secondary hash operation to generate a chained hash signature, which is then appended to the tamper-proof file system dedicated to the project log.

[0036] Regarding the construction of the baseline fact graph 405, the system parses the file format and hierarchical directory of the project source file; extracts parameter triples containing subject, predicate, and object according to the preset syntax parsing rules; imports the parameter triples into the graph database to construct the baseline fact graph 405, and changes the access permission of the graph database to a locked state with only read permission.

[0037] To ensure the stable operation of the above logic and various modules, this application embodiment also provides a basic component configuration architecture to support the operation of the application server. For example... Figure 4 As shown, in the system logic architecture provided in this embodiment, a memory-based distributed message communication infrastructure component 406 is set up between the writing agent 401, which is responsible for the core text task, and the review agent 402, which handles the execution logic. For example, a message queue cluster can be deployed. By mounting such a message queue channel, the agent scheduling layer can serialize the generated paragraph objects and the conflict patch objects returned as feedback and send them into the asynchronous communication pipeline. By adding this type of middleware-level infrastructure bus, the deadlock coupling between the two computationally intensive agents at the memory level can be de-coupled, ensuring that various message structures can be reliably transmitted without out-of-order loss in large-scale, high-concurrency system throughput scenarios.

[0038] This application also provides an electronic device for running the aforementioned multi-agent cooperative program. For example... Figure 5 As shown, the electronic device mainly includes a memory 501 and a processor 502 connected via a system bus 503. The memory 501, specifically a non-volatile hard disk or a random access memory array, is primarily configured to store the computer program code supporting the aforementioned computational logic and related operating system firmware. The processor 502 typically includes a multi-core central processing unit or a tensor computation processing unit, and is electrically connected to the memory 501 to exchange data. The processor 502 is configured to, when reading and executing the computer program stored in the memory 501, invoke computational resources according to a predetermined timing sequence to fully implement any of the actions and processes of the multi-agent collaborative engineering report generation and quality control method described in the aforementioned embodiments. Through the high-speed computing power of these physical hardware modules, a solid hardware support is formed to ensure the algorithm's implementation.

Claims

1. A method for generating and controlling the quality of engineering reports based on multi-agent collaboration, characterized in that, include: Receive the project source file, extract parameter triples to construct a baseline fact graph and lock it in read-only mode; An occlusion algorithm is run on the imported historical reference report to replace numerical entities and proper nouns with special placeholders and store them in an isolated vector database. The writing agent is controlled to retrieve the syntactic skeleton from the isolated vector database, call the truth values ​​from the baseline fact graph to fill the syntactic skeleton, and generate the target paragraph. The review agent is controlled to perform reverse entity extraction on the target paragraph to obtain a generated fact subgraph, and compare the generated fact subgraph with the baseline fact graph to perform conflict node comparison. If a conflict is found, a conflict patch containing revision logic is generated and a refactoring instruction is triggered. In response to the document save instruction, the input / output gateway interception mechanism is triggered to mount the document data to the sandbox memory. In the sandbox memory, the final reverse extraction of cross-chapter entities is performed. If the cross-chapter entity is found to be inconsistent with the baseline fact graph, the automatic revision agent is used to generate differential replacement text to cover the erroneous field, the input / output gateway interception mechanism is removed, and the anti-tampering hash log is recorded.

2. The method as described in claim 1, characterized in that, The process of running a masking algorithm on the imported historical reference report to replace numerical entities and proper nouns in the historical reference report with special placeholders includes: loading a preset engineering domain entity recognition model to perform sequence labeling on the historical reference report, outputting a continuous string containing numerical entities and proper nouns; generating a special placeholder sequence corresponding to the continuous string, the special placeholder sequence including an identifier representing the entity type and an incrementing index number; according to the position information of the continuous string in the original text, using the special placeholder sequence to cover the corresponding original characters, generating parameter-free desensitized text, and storing the desensitized text as the syntactic skeleton in the isolation vector database.

3. The method as described in claim 2, characterized in that, The step of overwriting the corresponding original characters using the special placeholder sequence according to the position information of the continuous string in the original text is configured as follows: extracting the start byte offset and end byte offset of the continuous string in the historical reference report; constructing a mask marker according to the identifier of the entity type; establishing a mapping table, which records the correspondence between the mask marker, the start byte offset, and the end byte offset; traversing the mapping table, for each record, deleting the byte array in the historical reference report between the start byte offset and the end byte offset, and inserting the bytecode sequence corresponding to the mask marker at the corresponding starting position to form an updated character stream sequence.

4. The method as described in claim 1, characterized in that, The step of performing a conflict node comparison operation between the generated fact subgraph and the baseline fact graph includes: extracting the specific numerical attributes corresponding to each node in the generated fact subgraph; retrieving a baseline node with the same named entity as the node in the baseline fact graph; reading the baseline numerical attributes of the baseline node and the tolerance threshold range configured in the baseline node; calculating the difference between the specific numerical attributes and the baseline numerical attributes, and determining whether the difference falls within the tolerance threshold range; if the difference exceeds the tolerance threshold range, then determining the current node as a conflict node.

5. The method as described in claim 4, characterized in that, The step of generating a conflict patch containing revision logic and triggering a refactoring instruction is configured as follows: obtaining the sentence-level context of the erroneous entity identified as the conflict node in the target paragraph; and extracting the baseline numerical attribute of the baseline node as a forced constraint truth value. Construct a structured patch object containing the error entity, the forced constraint truth value, and the sentence-level context; The structured patch object is used as input parameters to call the local rewrite interface of the writing agent; Set a maximum retry counter. If the current number of rewrites does not exceed the threshold of the maximum retry counter, then generate the corrected paragraph text based on the structured patch object. If the threshold of the maximum retry counter is reached, then freeze the generation state of the target paragraph.

6. The method as described in claim 1, characterized in that, The step of responding to a document save command and triggering an input / output gateway interception mechanism to mount the document data to the sandbox memory includes: intercepting write operation request streams for persistent storage devices; redirecting all document data in the write operation request streams to the isolated sandbox memory space; locking the read and write permissions of the all document data in the sandbox memory space to prevent the all document data from being accessed by external processes; and triggering a traversal module to read the all document data sequentially according to chapter order, entering the final reverse extraction stage of cross-chapter entities.

7. The method as described in claim 6, characterized in that, The step of using an auto-revision agent to generate differential replacement text to cover error fields is configured as follows: locating the first and second position indices corresponding to inconsistent data in the cross-chapter entity; generating a differential string stream for replacement based on the truth values ​​in the baseline fact graph; converting the full document data into a character array; and erasing the memory block data between the first and second position indices in the sandbox memory space. Insert the byte data corresponding to the differential string stream into the first position index; calculate the change in the length of the memory array, and update the starting address pointer of the remaining characters in the full document data after the second position index.

8. The method as described in claim 1, characterized in that, The tamper-proof hash log recording includes: collecting the previous snapshot data stream, the subsequent snapshot data stream, and the differential replacement text in the sandbox memory where the overwrite operation was performed; concatenating the previous snapshot data stream, the subsequent snapshot data stream, and the differential replacement text and performing a hash operation to generate a fixed-length hash digest of the current operation; obtaining the historical hash digest of the previous operation log; combining the historical hash digest with the fixed-length hash digest of the current operation and performing another hash operation to generate a chained hash signature and appending it to the tamper-proof log file system.

9. An electronic device, characterized in that, It includes a memory and a processor, the memory being used to store a computer program, and the processor being connected to the memory and configured to implement the method as described in any one of claims 1 to 8 when executing the computer program.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 8.

Images