A water conservancy technical report intelligent generation and auditing system and method

Abstract

The application discloses a kind of water conservancy technical report intelligent generation and auditing system and method, belong to wisdom water conservancy technical field.System uses six-layer architecture, including front-end interaction layer, back-end service layer, intelligent agent cooperation layer, model service layer, knowledge and data storage layer, process data asset layer.System is parsed user demand and constructs dynamic task directed acyclic graph by master control autonomous planning intelligent agent, coordinates demand analysis, outline generation, knowledge retrieval, content writing, content optimization, style alignment, compliance verification, consistency check, traceability labeling and other function intelligent agents, according to task dependency relationship, execution state and user feedback dynamically scheduling demand analysis, outline generation, knowledge retrieval, content writing, content optimization, style alignment, compliance verification, consistency check and traceability labeling and other function intelligent agents complete report preparation, and the report generated has full-link version traceability ability, and support the assetization of preparation process data and reuse.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary fields of smart water conservancy, water conservancy informatization and artificial intelligence. Specifically, it relates to a system and method for intelligent generation and review of water conservancy technical reports, and more particularly to a system and method for intelligent generation, compliance verification, content traceability and process knowledge reuse of water conservancy technical reports based on autonomous planning and multi-agent collaboration. Background Technology

[0002] Throughout the entire lifecycle of water conservancy projects, including planning and design, construction, operation and management, acceptance and evaluation, special studies, project application, and results summary, numerous technical reports need to be prepared. These reports include feasibility study reports, preliminary design reports, implementation plans, acceptance reports, special analysis reports, evaluation reports, information technology construction plans, and digital twin platform construction reports. These reports are generally characterized by their length, complex structure, high level of specialization, and stringent compliance requirements. The quality of their preparation directly affects project initiation, approval, construction, and operation, imposing strict requirements on compliance with industry standards, technical rationality, completeness of content, and professional expression.

[0003] Currently, the preparation of water conservancy technical reports is still primarily done manually. The conventional process typically involves the compiler first studying the requirements document, then searching through historical reports, industry standards, technical specifications, and business materials, followed by drafting the outline, writing the content, repeated revisions, and manual proofreading. This entire process encompasses multiple stages, including understanding requirements, data retrieval, chapter organization, professional analysis, formatting, compliance review, and feedback rectification. The long workflow, numerous participating roles, and high collaboration costs make it prone to problems such as misunderstandings of requirements, insufficient data retrieval, incomplete chapter coverage, inconsistencies, and difficulty in tracing the revision process. Ultimately, this results in low preparation efficiency and inconsistent report quality, failing to meet the water conservancy industry's demands for efficient, high-quality, and standardized report preparation.

[0004] With the implementation of large language model technology, existing tools on the market typically generate text based on fixed templates or single-round prompts, lacking the ability to automatically model chapter dependencies, validation feedback, and the impact of user modifications. Furthermore, these tools are not customized for the water conservancy engineering field, lacking support from water conservancy professional knowledge bases, industry standard libraries, and engineering case libraries. They cannot adapt to the compilation rules, professional terminology, and technical logic requirements of water conservancy technical reports, and their outputs are prone to professional deviations, logical gaps, misuse of terminology, and non-compliance issues, failing to meet the requirements for compiling and using formal water conservancy engineering reports. Another type of water conservancy technical report generation system relies on water conservancy report compilation guidelines and adopts a manually preset fixed workflow mode. It generates chapter content by manually configuring process nodes and calling upon a basic knowledge base, and supports manual prompting and optimization. However, in practical applications, it was found that the solution has obvious technical limitations: on the one hand, the process is rigid and has extremely poor flexibility. The process must be manually configured for each type of report and each chapter. When faced with non-standard requirements, new types of reports, or personalized requirements, it needs to be reconfigured and cannot automatically adapt to changing scenarios. On the other hand, the capabilities cannot be accumulated or iterated. The system only records simple operation logs and cannot capture or extract the expert experience, decision-making logic, and business judgment behind manual modifications. It cannot be transformed into reusable capabilities, resulting in repetitive work for each task and making it difficult to continuously optimize the system.

[0005] The existing water resources report generation scheme has the following problems: The report generation process relies on fixed templates or preset processes, lacking the ability to plan autonomously and adjust dynamically. Most existing solutions require pre-configuration of fixed processing links and workflows for specific report types, and cannot automatically identify task objectives and autonomously plan compilation steps after users upload requirement documents of any format and type. They also cannot dynamically adjust the execution path based on user feedback, intermediate execution results, or verification conclusions, resulting in insufficient system flexibility, versatility, and scenario adaptability.

[0006] There is a lack of in-depth vertical adaptation for the water conservancy industry, and insufficient capabilities for professional compliance verification and technical rationality verification. Existing general-purpose intelligent generation systems typically do not undergo in-depth adaptation to the water conservancy industry's knowledge system, technical specification system, and report preparation rules. They often only achieve basic text generation and superficial checks such as typos, formatting, and grammar, making it difficult to identify professional errors, non-compliance with specifications, missing chapters, incomplete logic, and deviations in business style in water conservancy technical solutions. The generated results still require extensive manual professional review and repeated modifications.

[0007] The sources and decision-making processes of generated content are not traceable, resulting in insufficient credibility and auditability of the results. Existing technologies typically focus more on the final report output, without recording the entire chain of content sources, modification basis, and generation decision-making process. They cannot clearly identify the normative clauses, historical data, knowledge entries, or user instructions corresponding to report chapters, paragraphs, or technical conclusions, nor can they reconstruct the modification trajectory and formation logic of the content. This is detrimental to result verification, responsibility determination, and audit traceability in the context of strong regulatory scenarios in the water conservancy industry.

[0008] The lack of systematic accumulation and assetization mechanisms for high-value data during the development process hinders the system's ability to achieve continuous self-iteration. While existing systems can record some operation logs, they typically store user interaction records, content modification trajectories, expert review opinions, and approval feedback information as ordinary process data. They lack mechanisms for automatic cleaning, deduplication, anonymization, structured decomposition, and classification of this data. Consequently, they cannot further transform this data into reusable prompt word samples, skill samples, decision chain samples, expert experience knowledge bases, expert preference knowledge bases, and business characteristic knowledge bases. This makes it difficult for the system to leverage historical process data for capability accumulation, performance optimization, and continuous evolution.

[0009] Therefore, there is an urgent need to provide an intelligent system and method for the preparation of technical reports in the water conservancy industry. This system should be able to automatically identify the content of the required documents, autonomously plan the preparation tasks, and organize multiple agents to collaboratively complete the outline generation, content writing, modification and optimization, professional verification, risk warning and source tracing without the need for pre-configured and fixed processes. Furthermore, it should be able to structure and reuse the interactive information, modification behavior and decision-making trajectory throughout the entire process, thereby improving the efficiency, quality, standardization, credibility, traceability and continuous optimization capabilities of water conservancy technical report preparation. Summary of the Invention

[0010] The purpose of this invention is to propose an intelligent generation and review system and method for water conservancy technical reports. This invention aims to solve the technical problems existing in the preparation of water conservancy technical reports, such as strong reliance on manual labor, poor process adaptability, insufficient professional compliance verification, untraceable content sources, and difficulty in accumulating and reusing process data.

[0011] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: The intelligent generation and review system for water conservancy technical reports includes a front-end interaction layer, a back-end service layer, an intelligent agent collaboration layer, a model service layer, a knowledge and data storage layer, and a process data asset layer.

[0012] The system comprises four layers: a front-end interaction layer for receiving user input, displaying the report generation process, and providing feedback on execution results; a back-end service layer for handling business logic, task scheduling, and data management; a smart agent collaboration layer, the core decision-making brain of the system, for converting user-uploaded requirement documents and natural language instructions into executable task graphs and organizing multiple functional smart agents to work collaboratively; a model service layer for providing requirements understanding, content generation, content optimization, and verification assistance; a knowledge and data storage layer for storing report data, knowledge data, and process data; and a process data asset layer for collecting, cleaning, desensitizing, structurally decomposing, and classifying requirement trajectories, interaction records, content modification records, traceability annotation data, smart agent decision logs, and review feedback generated during report compilation, forming reusable data assets that support subsequent task reuse and system optimization. The front-end interaction layer, back-end service layer, smart agent collaboration layer, model service layer, and knowledge and data storage layer communicate asynchronously, loosely coupled, and maintain cross-module data consistency synchronization through standard RESTful interfaces and message queues.

[0013] Composition and functions of each module: Front-end interaction layer: The front-end interaction layer includes user login unit, report creation unit, requirement document upload unit, dialogue interaction unit, traditional editing unit, data retrieval unit, model selection unit, task progress display unit, and report export unit.

[0014] The system includes the following components: a user login unit for receiving user identity information and initiating a login request; a report creation unit for inputting the report name, report type, and compilation requirements for the current task; a requirements document upload unit for uploading basic data related to the task, including at least the project task book, construction requirements description, historical similar reports, information technology construction specifications, system architecture diagram, investment estimation table, and meeting minutes; a dialogue interaction unit for receiving supplementary requirements input by the user in natural language, such as generating or highlighting the data base, model platform, and business applications according to the implementation plan format, or supplementing the cybersecurity chapter; a traditional editing unit for manually modifying specific paragraphs and expressions after the system generates the initial draft; a data retrieval unit for browsing and retrieving existing materials from the database; a model selection unit for selecting different model strategies or generation modes; a task progress display unit for displaying the current task status, chapter generation progress, problem list, and rectification suggestions; and a report export unit for exporting the final deliverables.

[0015] The data generated by the front-end interaction layer mainly includes user identity data, report task data, attachment index data, natural language command data, text editing data, and model selection data, and all of the above data are sent to the back-end service layer.

[0016] Backend service layer: The backend service layer includes user management module, report management module, data source management module, task scheduling module, log management module, compliance management module, traceability management module, and process data processing module.

[0017] The module is divided into several parts: User Management, Report Management, Data Source Management, Task Scheduling, Log Log Management, Compliance Management, Traceability Management, and Process Data Management. The User Management module is used for authentication and access control; the Report Management module is used for creating report objects, maintaining report status, and managing report versions; the Data Source Management module is used for managing user-uploaded project materials, specification documents, and case files; the Task Scheduling module is used for generating task numbers, encapsulating task context, and issuing tasks to the agent collaboration layer; the Log Management module is used for recording request logs, operation logs, and model call logs; the Compliance Management module is used for receiving verification results and generating a problem list and rectification suggestions; the Traceability Management module is used for saving the mapping relationship between paragraph content and source evidence; and the Process Data Processing module is used for collecting requirement correction trajectories, content modification trajectories, and interaction process data, and then organizing and writing them into the knowledge and data storage layer.

[0018] After receiving a request from the front-end interaction layer, the back-end service layer authenticates, formats, and encapsulates the request to form a task context object, which is then sent to the agent collaboration layer. Simultaneously, the back-end service layer also receives intermediate results, final results, verification results, and traceability results returned by the agent collaboration layer, saves them to the knowledge and data storage layer, and then sends the corresponding status information and result summary back to the front-end interaction layer.

[0019] Intelligent Agent Collaboration Layer: The intelligent agent collaboration layer is the core decision-making and execution layer of the system, including a master autonomous planning intelligent agent and multiple functional execution intelligent agents. The functional execution intelligent agents include a requirements analysis intelligent agent, an outline generation intelligent agent, a knowledge retrieval intelligent agent, a content writing intelligent agent, a content optimization intelligent agent, a style alignment intelligent agent, a compliance verification intelligent agent, and a traceability annotation intelligent agent.

[0020] The master autonomous planning agent receives the task context from the backend service layer, identifies the report type of the task, plans the chapter structure and generation order, and coordinates other agents to execute collaboratively. The autonomous planning agent receives the task context from the backend service layer and, in conjunction with requirements documents, user instructions, historical interaction information, and preset rules, identifies the report type, compilation objectives, and chapter requirements. When it identifies the current task as a water resources information report, it breaks down the overall task into sub-tasks such as outline generation, data retrieval, chapter writing, content optimization, content verification, and traceability annotation, according to task decomposition rules and chapter dependencies. It then constructs a dynamic task directed acyclic graph (DAG), or dynamic task graph for short, to coordinate the execution of each agent according to priority. The construction of this graph uses a DAG analysis algorithm to determine chapter dependencies. Each sub-task node in the dynamic task DAG exists only upstream. The system establishes unidirectional dependencies pointing downstream, eliminating loops and allowing only local subtask rescheduling without forming a circular execution chain. The directed acyclic graph (DAG) analysis algorithm uses chapter dependency weights as the criterion, classifying task nodes into three categories: strong serial dependencies, weak dependencies, and no substantial dependencies. Strongly serially dependent nodes are executed sequentially according to chapter order; nodes with no substantial dependencies can be processed in parallel; weakly dependent nodes employ a semi-parallel scheduling mode. During system operation, four trigger scenarios are monitored in real-time: missing data, generation anomalies, verification failures, and user additions / modifications. If a corresponding event is detected, the relevant chapter dependencies are re-determined, and the dynamic task DAG is partially reconstructed and rescheduled. Strong serial dependency means that the entire content, technical indicators, and engineering quantities of a subsequent chapter depend entirely on the output of the preceding chapter. It can only start after the preceding chapter has been fully generated and verified. A typical chain is project background, construction necessity, current status analysis, and construction goals, which is a forced serial execution. Weak dependency means that multiple chapters share only a few basic engineering parameters, such as project construction scale and construction scope. The main writing is independent of each other and a semi-parallel scheduling is adopted: each weakly dependent chapter starts to generate the first draft simultaneously. After all the first drafts are produced, a unified consistency verification agent is scheduled to check the cross-chapter caliber. If there is a data conflict, the part is rewritten. A typical weak dependency combination is the functional requirements chapter, the data base chapter, and the security protection chapter. No substantial dependency means that there is no cross-reference between chapters in terms of data, specifications, and engineering boundaries. The entire process is generated independently and in parallel, such as investment estimation, implementation schedule, and operation and maintenance support plan. Semi-parallel scheduling is different from completely serial and completely parallel. It first generates the first drafts of each chapter in parallel, and then performs cross-chapter consistency verification and linkage correction in a unified manner, taking into account both generation efficiency and report logic consistency.During execution, if missing data, generation anomalies, content conflicts, verification failures, or user-added modification requests are detected, a rollback and rescheduling mechanism is triggered to make local adjustments to the task graph, thereby achieving autonomous planning and dynamic control of the report preparation process. In this embodiment, the autonomous planning agent further plans and generates chapters on project background, necessity of construction, current status of information system, design basis, construction goals, construction tasks, construction scope, construction principles, user analysis, business requirements, functional requirements, data requirements, performance requirements, and security protection requirements.

[0021] In the functional execution agent, the requirement analysis agent extracts structured information such as project name, construction unit, construction goals, system boundaries, current problems, business scenarios, and task requirements from task descriptions, user instructions, and uploaded materials; the outline generation agent generates chapter frameworks that conform to the characteristics of water conservancy informatization reports; the knowledge retrieval agent searches for relevant policy documents, industry standards, historical project cases, and terminology definitions around the current chapter; the content writing agent generates initial drafts for each chapter; the content optimization agent adds, deletes, and rewrites content based on user supplementary requirements and contextual constraints; the style alignment agent unifies the text to a formal report style; the verification assistance agent assists in judging completeness, consistency, and professional expression; and the traceability and annotation agent adds source identifiers to the generated content and establishes the correlation between content and evidence.

[0022] Model Service Layer: The model service layer includes a requirements understanding unit, an outline generation unit, a chapter writing unit, a content optimization unit, a style alignment unit, and a validation assistance unit.

[0023] The system includes the following components: a requirements understanding unit for understanding the professional semantics and domain concepts in water conservancy informatization reports, including at least terms such as data base, water conservancy knowledge platform, business middleware, digital twin scenario, four-prevention capabilities, and network security level protection, where the four-prevention capabilities include data prediction, engineering early warning, engineering rehearsal, and engineering contingency plan; an outline generation unit for generating chapter structures based on task types and requirements constraints; a chapter writing unit for outputting the initial draft of the chapter text; a content optimization unit for adding, deleting, and rewriting text; a style alignment unit for unifying writing style and expression; and a verification assistance unit for identifying inconsistencies, chapter omissions, and professional expression issues.

[0024] The model service layer does not directly face users, but serves as the underlying capability support layer of the intelligent agent collaboration layer, providing various intelligent agents with model capabilities such as semantic understanding, text generation, content adjustment, and quality judgment.

[0025] Knowledge and Data Storage Layer: The knowledge and data storage layer includes a user data storage module, a report data storage module, a task data storage module, a database storage module, a knowledge tag module, a traceability data storage module, and a process sedimentation data storage module.

[0026] The system includes the following modules: a user data storage module for storing user identity information, permission information, and operation records; a report data storage module for storing report text, chapter structure, version records, and exported results; a task data storage module for storing task numbers, task status, execution process, and scheduling information; and a database storage module for storing project-related materials, including at least historical reports of water conservancy informatization projects, relevant standards for smart water conservancy construction, requirements for digital twin water conservancy construction, network security specifications, system construction examples, and information system project management materials. A knowledge tagging module is used to attach search tags such as data resources, application architecture, security system, investment calculation, and acceptance evaluation to the materials to support subsequent knowledge retrieval and evidence location. A traceability data storage module stores the mapping relationship between generated paragraphs and standard basis, historical cases, and user instructions. A process-accumulated asset library module stores reusable assets after standardization, including at least a prompt word asset library, a skill asset library, a decision-making thought chain asset library, an expert experience library, an expert preference library, and a business characteristic library.

[0027] The knowledge and data storage layer is used to provide unified support for data carrying, data management, evidence traceability and process accumulation for the aforementioned layers.

[0028] Process data asset layer: The process data asset layer includes a process data acquisition module, a cleaning and desensitization module, a structured decomposition module, an asset classification and accumulation module, an asset quality evaluation module, and an asset retrieval and reuse module.

[0029] The process data acquisition module is used to collect user demand input, data retrieval records, content generation records, verification feedback records, manual modification records, traceability annotation records, and agent decision logs. The cleaning and desensitization module is used to perform data deduplication, noise removal, format standardization, and desensitization of sensitive information; The structured decomposition module is used to establish the relationship between user needs, input data, knowledge evidence, generated results, modification feedback, and final adopted results; The asset classification and accumulation module is used to transform process data into prompt word assets, skill assets, decision chain assets, expert experience assets, expert preference assets, and business characteristic assets; The asset quality evaluation module is used to evaluate asset quality based on manual adoption, verification pass rate, and subsequent reuse effect; The asset retrieval and reuse module is used to retrieve and call up accumulated assets in subsequent similar tasks, assisting in task planning, content generation, verification, and optimization.

[0030] Based on the above system, the present invention provides a method for compiling a water conservancy technical report, the method comprising the following steps: The system receives user-uploaded requirement documents, historical data, and related attachments, as well as user-inputted report names, dependent materials, template requirements, and natural language instructions. These are written into the knowledge and data storage layer to form a searchable task-level data index and knowledge index. The system then performs format parsing, text extraction, structure recognition, and metadata annotation on the input content to form standardized requirement inputs.

[0031] The autonomous planning agent performs semantic understanding on the standardized requirement input, identifies report type, business scenario, task objective, chapter requirements, constraints, and data gaps, and uses a directed acyclic graph (DAG) analysis algorithm to autonomously divide nodes into strongly sequential, weakly dependent, and non-substantially dependent nodes based on chapter dependency weights, constructing a dynamic task DAG without a preset fixed process; the dynamic task graph includes at least requirement parsing nodes, knowledge retrieval nodes, outline generation nodes, chapter generation nodes, content optimization nodes, compliance verification nodes, rectification and correction nodes, traceability annotation nodes, and result output nodes.

[0032] When the autonomous planning agent identifies incomplete requirement information, missing key data, or unclear task conditions, it initiates a clarification interaction with the user through dialogue and updates the dynamic task graph based on the information supplemented by the user.

[0033] The autonomous planning agent schedules the knowledge retrieval agent to retrieve industry standards, technical standards, historical reports, case materials, and terminology related to the current task from the water conservancy knowledge and standard support module, forming a knowledge evidence set; at the same time, the scheduling outline generation agent generates a report framework based on the standardized requirements input, the knowledge evidence set, and the task graph.

[0034] The autonomous planning agent breaks down the report into chapters according to the task graph and processes each chapter as an independent task node. For each chapter, the process of chapter task identification, knowledge retrieval, constraint construction and content generation are executed in sequence to form the initial draft of the chapter text. The generated content of the current chapter is jointly constrained by the results of the upstream chapter, the knowledge evidence set and user constraints.

[0035] The content writing agent calls the model service layer to generate initial drafts of each chapter's content. Then, the content optimization agent polishes, supplements, compresses, rewrites, or restructures the chapter content based on user instructions, historical examples, business style, and contextual relationships, resulting in an optimized chapter version.

[0036] If a user modifies the generated content through dialogue or traditional editing, the dialogue and editing synchronization module records the modification trajectory, update the version status and chapter status of the report, and feeds the modification information back to the autonomous planning agent. The autonomous planning agent determines whether to perform partial rewriting, chapter rewriting, linked chapter updates, or re-verification based on the modified content.

[0037] The water resources-specific compliance verification module is invoked to perform industry standard compliance verification, technical rationality verification, report integrity verification, basic compliance verification, and style consistency verification on the current report, and outputs problem items, risk levels, and rectification suggestions; among them, the basic compliance verification includes at least context consistency verification, typo verification, format verification, and plagiarism detection.

[0038] When the verification result fails, the autonomous planning agent reconstructs the subsequent task steps according to the problem type, problem location and rectification suggestions, and schedules relevant agents to perform targeted corrections. The targeted corrections include at least supplementary knowledge retrieval, partial chapter rewriting, cross-chapter content alignment and style correction, until the preset quality requirements are met.

[0039] During the content generation, optimization, and rectification process, the content source tracing module continuously records the normative basis, knowledge source, user instructions, model call records, and modification process corresponding to each chapter, paragraph, or key conclusion, forming a queryable content tracing chain. At the same time, the process data accumulation and assetization module automates the processing of requirement analysis trajectory, content modification trajectory, user interaction data, expert review opinions, approval feedback, and intelligent agent decision trajectory, and stores them according to business scenarios, value levels, and data types for subsequent knowledge reuse and system optimization.

[0040] Once the report passes verification, the report output module exports the report to the target format file and simultaneously saves version information, traceability information, and verification records, forming the final report output.

[0041] The beneficial effects of this invention include: 1. Improve report preparation efficiency and reduce reliance on manual labor. This invention achieves automatic requirement analysis and task decomposition through autonomous planning of intelligent agents, eliminating the need for manual pre-configuration of fixed workflows. Simultaneously, it utilizes multi-agent collaboration to complete the entire process of outline generation, data retrieval, content writing, modification and optimization, and verification and rectification, reducing the time spent by humans in requirement understanding, process configuration, chapter organization, and repeated communication. A comparative experiment was conducted using three different types of water conservancy reports—including an information technology implementation plan, a feasibility study report, and an acceptance report. Compared to existing fixed-workflow report generation schemes, under the same conditions of complete data, the initial draft generation time can be reduced by 30%–70%, and the overall compilation cycle by 20%–50%. Compared to purely manual compilation methods, compilation efficiency is improved by over 80%, reducing the reliance on the experience of professionals in water conservancy technical report compilation.

[0042] 2. Improve the quality and professional standardization of report content, and reduce the cost of manual review. This invention, through a water resources knowledge and standards support module and a water resources-specific compliance verification module, comprehensively reviews industry standard compliance, technical rationality, chapter completeness, basic text quality, and consistency of business style. This effectively reduces professional errors, chapter omissions, and non-standard expressions in reports. Compared to report generation methods without professional verification mechanisms, and compared to general generation schemes without vertical adaptation to water resources expertise, this invention increases chapter completeness to over 90%, improves the identification rate of standardization issues by over 30%, and reduces the number of manual reviews and revisions by 20%–40%, thus lowering the review costs for professionals.

[0043] 3. Achieve end-to-end traceability of reports to meet the stringent regulatory requirements of the water conservancy industry. This invention, through a content source tracing module, establishes a mapping relationship between report paragraphs, key conclusions, charts, and historical reports, regulatory clauses, knowledge entries, user instructions, and modification processes. This enables traceability of content sources, modification processes, and decision chains. Compared to methods that only output the final text result, this invention achieves a source labeling coverage rate of over 85% for key content, improves review and location efficiency by 30%–60%, and thus enhances the credibility, interpretability, and audit support capabilities of the report.

[0044] 4. A self-iterable closed loop for the system of business data assetization has been built, enabling continuous capability evolution. This invention utilizes a process data accumulation and assetization module to clean, deduplicatize, anonymize, structure, and classify data from requirement analysis trajectories, content modification trajectories, expert review opinions, approval feedback, and user interaction data. This results in sample prompt words, agent operation skills, decision chains, expert experience, and business characteristics. Through the continuous accumulation of high-value process data, the accuracy of generating subsequent similar reports can be improved by 10%–30%, and repetitive modifications reduced by 15%–35%. This truly realizes a positive flywheel effect: business generates data, data transforms into assets, assets enhance capabilities, and capabilities optimize business. The system can achieve continuous self-learning and self-optimization through business data. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating the overall process architecture of the present invention.

[0046] Figure 2 This is a flowchart illustrating the implementation of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only some, not all, of the embodiments of this invention, and are not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0048] like Figure 1 As shown, the intelligent generation and review system for water conservancy technical reports of the present invention includes a front-end interaction layer, a back-end service layer, an intelligent agent collaboration layer, a model service layer, a knowledge and data storage layer, and a process data asset layer.

[0049] In this embodiment, the front-end interaction layer is used to receive user operation requests and display system execution results. The front-end interaction layer includes a user login unit, a report creation unit, a data upload unit, a template selection unit, an agent selection unit, a model selection unit, a report editing unit, a task progress display unit, a problem prompting unit, and a report export unit. The user login unit receives username, password, and other identity information and initiates an identity authentication request to the back-end service layer; the report creation unit receives task information such as report name, report type, keywords, and compilation requirements; the data upload unit receives report dependency materials uploaded by the user; the template selection unit displays preset report templates and allows the user to select one; the agent selection unit selects a published report generation agent; the model selection unit selects different large language model services; the report editing unit displays and edits the generated report content; the task progress display unit shows the report generation task status; the problem prompting unit displays the results of content completeness, format standardization, and consistency checks; and the report export unit exports the final report to a specified format file.

[0050] In a preferred embodiment, the front-end interaction layer is implemented using a component-based approach. Different functional units are encapsulated as independent front-end components, and data is shared between components through a state management mechanism. When a user creates a report task, the front-end interaction layer encapsulates the report name, report type, template number, agent number, model number, uploaded data identifier, keywords, and compilation requirements into structured request data, and submits it to the back-end service layer through an interface. The front-end interaction layer also obtains the task status through polling or message push, and displays the stage status such as data parsing, content generation, report assembly, and export completion in real time.

[0051] The backend service layer is used to perform identity authentication, access control, task encapsulation, data parsing, task scheduling, result reception, status updating, log recording, and report data management. The backend service layer includes a user management module, a report management module, a data source management module, a template management module, an agent management module, a task scheduling module, a log management module, an audit management module, and an export management module. The user management module is used for user registration, login, permission verification, and session management; the report management module is used for creating report objects, maintaining report status, saving report content, and managing report versions; the data source management module is used for receiving and parsing user-uploaded data; the template management module is used for maintaining different types of report templates; the agent management module is used for maintaining agent configuration, publishing status, and calling permissions; the task scheduling module is used for creating report generation tasks and scheduling model services and the agent collaboration layer; the log management module is used for recording system operation logs, task logs, and exception logs; the audit management module is used for generating a problem list and rectification suggestions; and the export management module is used for generating the final report file.

[0052] In this embodiment, when the backend service layer receives a report generation request submitted by the frontend interaction layer, it first verifies the user's identity and permissions. Upon successful verification, the backend service layer generates a report number and a task number, and encapsulates the user number, report number, task number, report type, template number, agent number, model number, uploaded data number, database reference number, user compilation requirements, and task creation time into a task context object. This task context object serves as a unified data carrier for transmission between different layers within the system, ensuring that the data parsing, retrieval, generation, review, and export processes have a unified task identifier and status basis.

[0053] The agent collaboration layer is used to organize the report generation process based on report type and task context. For example... Figure 2 As shown, the intelligent agent collaboration layer includes a master intelligent agent and several functional intelligent agents. The master intelligent agent receives the task context from the backend service layer, identifies the report generation target, determines the task execution order, and schedules each functional intelligent agent to complete the corresponding processing. The functional intelligent agents include at least a requirements parsing intelligent agent, a template matching intelligent agent, a data retrieval intelligent agent, a content generation intelligent agent, a content optimization intelligent agent, a formatting intelligent agent, a review assistance intelligent agent, and an export assistance intelligent agent.

[0054] The requirement analysis agent extracts structured information such as report type, project name, system name, keywords, compilation requirements, document type, and output format from the task context. The template matching agent determines the report template based on the report type, keywords, and user selections. The document retrieval agent retrieves relevant reference content from uploaded documents and the database. The content generation agent calls the model service layer to generate the chapter text. The content optimization agent supplements, rewrites, or simplifies parts of the content based on user feedback or review issues. The formatting agent adjusts heading levels, paragraph formats, table formats, and numbering rules according to the template requirements. The review assistance agent checks the report's completeness, chapter consistency, and format compliance. The export assistance agent assists the export management module in generating the final document.

[0055] In this embodiment, the intelligent agent collaboration layer is not limited to a fixed process. The master intelligent agent determines the execution path based on the report type, template structure, data completeness, and generation status in the task context. For tasks with complete data and a clear template, the master intelligent agent executes the process in the order of "data parsing - template determination - data retrieval - chapter-by-chapter generation - review assistance - formatting - report export". For tasks with missing data or unclear templates, the master intelligent agent first triggers requirement supplementation or template recommendation, and then executes the subsequent generation tasks. For tasks where the user only requests modification of a partial chapter, the master intelligent agent only generates a partial rewrite task and does not repeat the full-text generation process.

[0056] The model service layer provides capabilities for calling pre-trained large language models, generating chapters, optimizing content, and providing semantic assistance. The model service layer includes a model access unit, a prompt word construction unit, a chapter generation unit, a content rewriting unit, and a model result return unit. The model access unit connects to different types of large language models; the prompt word construction unit organizes report type, chapter title, template requirements, user input, uploaded data fragments, and database search results into model input; the chapter generation unit calls the model to generate chapter content; the content rewriting unit rewrites partial content according to user modification requirements; and the model result return unit returns the model output to the backend service layer or the intelligent agent collaboration layer.

[0057] In one embodiment, chapter generation in the model service layer is implemented as follows: First, the master control agent determines the chapter to be generated based on the template structure; second, the data retrieval agent obtains reference fragments based on the current chapter title, report type, and keywords; third, the prompt word construction unit combines the chapter name, chapter writing requirements, references, output format, and constraints into the model input; then, the chapter generation unit calls the pre-trained large language model to generate the initial draft of the chapter; finally, the backend service layer receives the generation result, writes the chapter content into the report data table, and updates the chapter generation status.

[0058] The knowledge and data storage layer is used to store various types of data generated and accessed during system operation. This layer includes user data storage modules, report data storage modules, task data storage modules, data source storage modules, template data storage modules, agent configuration storage modules, database storage modules, retrieval index modules, and log data storage modules. The user data storage module stores user identity, permissions, and login records; the report data storage module stores report titles, report text, chapter structure, generation status, and exported file paths; the task data storage module stores task numbers, task status, task progress, creation time, completion time, and failure reasons; the data source storage module stores uploaded file information, parsed content, and file paths; the template data storage module stores template names, applicable report types, chapter structure, and format configurations; the agent configuration storage module stores agent names, applicable scenarios, default templates, database scope, model configurations, and prompt word rules; the database storage module stores industry data, historical reports, writing examples, and knowledge tags; the retrieval index module builds a full-text search index; and the log data storage module stores operation logs, task logs, model call logs, and exception logs.

[0059] In this embodiment, structured data can be stored using a relational database, task status and hot data can be stored using a cached database, and the main text of the database, historical reports, and uploaded data fragments can be indexed using a full-text search engine. This combination of relational database, cached database, and full-text search engine ensures the consistency of core data such as users, reports, and tasks while improving the efficiency of data retrieval and task status queries.

[0060] The process data asset layer is used for the unified collection, cleaning, desensitization, structured decomposition, classification, and reuse management of requirement input records, data retrieval records, content generation records, verification feedback records, manual modification records, traceability annotation records, and agent decision logs generated during the report preparation process. The process data asset layer includes a process data acquisition module, a cleaning and desensitization module, a structured decomposition module, an asset classification and retention module, an asset quality evaluation module, and an asset retrieval and reuse module. The process data acquisition module collects data from the entire process; the cleaning and desensitization module performs deduplication, noise removal, format standardization, and sensitive information desensitization; the structured decomposition module establishes the relationships between user requirements, input data, knowledge evidence, generated results, modification feedback, and the final adopted results; the asset classification and retention module transforms process data into prompt word assets, skill assets, decision chain assets, expert experience assets, expert preference assets, and business characteristic assets; the asset quality evaluation module evaluates asset quality based on manual adoption, verification pass rates, and subsequent reuse effects; and the asset retrieval and reuse module retrieves and reuses retained assets in subsequent similar tasks, assisting in task planning, content generation, verification, and optimization.

[0061] The following describes the implementation of the method of the present invention in conjunction with the system operation process.

[0062] Step S1: Receive user report generation request. The user logs into the system through the front-end interaction layer and fills in the report name, report type, keywords, compilation requirements, report template, agent, and model on the report creation page, while also uploading the report's dependent materials. The front-end interaction layer encapsulates the user input and file upload results into a standardized request and sends it to the back-end service layer.

[0063] Step S2: Generate a task context object. After receiving the report generation request, the backend service layer verifies the user identity, report type, template number, agent number, model number, and uploaded materials. Upon successful verification, the backend service layer creates a report record and a task record, generates a report number and a task number, and encapsulates the relevant parameters into a task context object. The task context object includes at least the user number, report number, task number, report type, template number, agent number, model number, material number, keywords, compilation requirements, task status, and creation time.

[0064] Step S3: Parse uploaded data. The data source management module calls the corresponding parsing program based on the uploaded file type. For text files, the system extracts the main text and segments it into paragraphs; for Word files, the system extracts titles, main text, tables, and hierarchical structure; for Excel files, the system extracts worksheets, field names, and cell data; for PDF files, the system extracts parsable text and organizes it according to page numbers and paragraphs. After parsing, the system cleans, segments, removes duplicates, and annotates metadata to form a set of data fragments, and establishes associations between the data fragments and task numbers, report numbers, and file sources.

[0065] Step S4: Determine the report template. The template matching agent determines the report template based on the user's selection or the system's recommendation. If the user has specified a template, the system directly reads the template's chapter structure and format configuration; if the user has not specified a template, the system performs template matching based on the report type, keywords, and compilation requirements. The template matching process includes tag matching and keyword matching: the system first filters candidate templates based on the report type, then calculates the degree of matching between the user's keywords and template tags, template name, template description, and chapter titles, selecting the template with the highest matching degree as the recommended template. After user confirmation, the system writes the template's chapter structure into the task context.

[0066] Step S5: Construct the chapter generation task. The controlling agent, based on the chapter structure in the report template, breaks down the report generation task into multiple chapter generation sub-tasks. Each chapter generation sub-task includes a chapter number, chapter title, chapter level, chapter writing requirements, related data snippets, upstream chapter content, and output format requirements. For chapters with dependencies, the system determines the generation order according to the template order and content dependencies; for independent chapters, the system can generate them in parallel or sequentially.

[0067] Step S6: Retrieve Chapter References. The document retrieval agent retrieves reference content from uploaded documents and the database based on the chapter title, report type, user keywords, and template requirements. The retrieval process first filters candidate documents using document tags, then retrieves document fragments related to the current chapter through full-text search. The system sorts the search results based on keyword matching degree, tag relevance, document source priority, and update time, and selects the top-ranked document fragments as references for model generation. For duplicate document fragments, the system performs deduplication; for excessively long document fragments, the system retains a summary of content related to the chapter topic.

[0068] Step S7: Generate chapter content. After receiving the chapter generation subtask, the model service layer's prompt word construction unit generates model input based on the chapter title, template requirements, user-defined requirements, reference material excerpts, and output format requirements. Model input includes role descriptions, generation tasks, content constraints, reference materials, and output format requirements. The chapter generation unit calls a pre-trained large language model to generate a draft chapter. After the generation result is returned, the backend service layer writes the chapter content to the report data storage module and updates the corresponding chapter status to "generated."

[0069] Step S8: Summarize and form the initial report draft. After all chapters are generated, the backend service layer summarizes the content of each chapter according to the template chapter order to form the initial report draft. The formatting agent organizes the heading levels, paragraph formats, table formats, numbering rules, and document structure according to the template configuration. For duplicate headings, redundant descriptions, or inconsistent formatting content in the model output, the system performs preliminary corrections according to the template rules.

[0070] Step S9: Perform audit assistance processing. The audit assistance agent performs completeness checks, format compliance checks, and consistency checks on the initial draft report. The completeness check determines if any required chapters are missing; the format compliance check determines if heading levels, paragraph formats, table structures, and numbering rules conform to the template requirements; the consistency check identifies whether key elements such as the report name, project name, system name, timeline, technical approach, and data definitions are consistently expressed across different chapters. The system generates a structured problem list from the check results, including problem location, problem type, problem description, and suggested solutions, and returns it to the front-end interaction layer for display.

[0071] Step S10: Receive manual modification and partial rewriting requests. Users can view the initial report draft and issue list in the front-end report editing unit and manually modify the report content. For content requiring system-assisted modification, users can select a chapter or paragraph and input modification requirements. The front-end interaction layer submits the selected content, paragraph position, and modification requirements to the back-end service layer. The back-end service layer generates a partial rewriting task based on the paragraph position, calls the data retrieval agent to re-obtain relevant data, and then the model service layer generates replacement text. After user confirmation, the system writes the replacement text into the report content and records the content before modification, the content after modification, the modification time, and the user who performed the operation.

[0072] Step S11: Export the report file. After the user confirms the report content, the export management module reads the report text, chapter structure, template format, and export configuration to generate a target format file. During the export process, the system sets the title style, text style, table style, margins, paragraph spacing, and numbering format according to the template requirements. After the export is complete, the system writes the file path, export time, and report status to the report data storage module and returns the download address to the front-end interaction layer.

[0073] Step S12: Record process data and operation logs. The system records logs at key nodes such as user login, report creation, document upload, document parsing, template matching, document retrieval, model invocation, chapter generation, review assistance, manual modification, partial rewriting, and report export. Log content includes user ID, task ID, report ID, operation type, operation time, execution status, exception information, and result summary. Through log recording, the report generation process becomes traceable and auditable.

[0074] Step S13: Process Data Asset Accumulation and Reuse. After the report is exported, the process data asset layer reads the requirement input records, data parsing records, data retrieval records, agent call records, model call records, chapter generation records, review issue records, manual modification records, partial rewrite records, and report export records corresponding to this task. The process data cleaning and desensitization module performs deduplication, cleaning, format standardization, and sensitive information desensitization on the above data; the process data structuring module decomposes the data according to the relationship between user instructions, input data, generated results, modification feedback, and final adopted results; the asset classification and accumulation module accumulates the data into prompt word assets, template assets, skill assets, review rule assets, expert experience assets, user preference assets, data retrieval assets, and business characteristic assets based on data content and reuse value.

[0075] For the accumulated data assets, the asset quality evaluation module assesses them based on user adoption, the extent of manual modifications, approval status, and subsequent usage effects. Data assets that meet the preset quality requirements are added to the formal asset library; data assets that do not meet the requirements but have reference value are added to the candidate asset library; data assets with content errors, sensitive information, or unclear scope of application are not added to the library. When subsequent similar report generation tasks are initiated, the asset retrieval and reuse module retrieves reusable assets from the formal asset library based on report type, chapter theme, user preferences, and business scenario, and provides them to the template matching agent, document retrieval agent, content generation agent, review assistance agent, and content optimization agent for use, thereby improving the efficiency of subsequent report generation and review assistance.

[0076] This invention targets the technical report preparation scenario in the water conservancy industry. Adaptable report types include, but are not limited to, water conservancy informatization construction plans, digital twin water conservancy platform construction reports, water conservancy project feasibility study reports, preliminary design reports, implementation plans, acceptance reports, special demonstration reports, and project applications. It supports intelligent generation, compliance verification, content traceability, and knowledge accumulation for various technical reports throughout their entire lifecycle, from planning, design, construction, operation, and acceptance. Driven by user needs, this invention focuses on autonomous planning without pre-defined processes, multi-agent collaborative division of labor, enhanced water conservancy knowledge generation, multi-dimensional compliance verification, full-link source traceability, and process data assetization. Through the coordinated operation of the front-end interaction layer, back-end service layer, agent collaboration layer, model service layer, knowledge and data storage layer, and process data asset layer, it achieves intelligent processing of the entire process of water conservancy technical reports, from requirements analysis, outline generation, content writing, interaction optimization, compliance review, risk warning, results export to knowledge accumulation.

[0077] The above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A smart system for generating and reviewing water conservancy technical reports, characterized in that: It includes a front-end interaction layer, a back-end service layer, an intelligent agent collaboration layer, a model service layer, a knowledge and data storage layer, and a process data asset layer; The front-end interaction layer provides interfaces for requirement input, report preview, manual intervention and modification feedback. It receives user operation instructions, requirement documents and supplementary materials, displays the report generation progress, execution results and problem list, and transmits various types of collected data to the back-end service layer. The backend service layer is used for task scheduling, agent lifecycle management and communication protocol adaptation between layers. It receives requests from the frontend interaction layer, completes identity authentication, task encapsulation and scheduling, receives various results returned by the agent collaboration layer and completes data persistence, and synchronizes the status and results to the frontend interaction layer. The intelligent agent collaboration layer includes a master autonomous planning intelligent agent and several functional execution intelligent agents. The master autonomous planning intelligent agent is used for requirement identification, construction of dynamic task directed acyclic graph and global exception handling. It schedules each functional intelligent agent to complete requirement parsing, outline generation, content writing, content optimization, compliance verification and traceability labeling. It also adjusts the task execution path according to execution exceptions and user modification instructions. The model service layer provides large language model inference services fine-tuned by LoRA in the water conservancy field. Based on water conservancy terminology, industry standards, historical report evidence, and chapter context, it provides constrained semantic understanding, chapter generation, and verification assistance for each agent. The knowledge and data storage layer includes a water conservancy standard knowledge base, a historical report sample library, and a vector index. Specifically, it includes user data, report data, task data, industry information, traceability data, and process-accumulated asset data, providing data retrieval, retrieval, and writing support for the other layers. The process data asset layer is used to collect, clean, de-identify, structure, decompose, classify, and precipitate the demand trajectory, content modification records, intelligent agent call logs, traceable and labeled data, expert review opinions, and user interaction data throughout the entire report compilation process. This forms reusable prompt word assets, skill assets, decision chain assets, expert experience assets, expert preference assets, and business characteristic assets, which are then used for subsequent task reuse and system optimization.

2. The intelligent generation and review system for water conservancy technical reports according to claim 1, characterized in that: The front-end interaction layer includes a user login unit, a report creation unit, a requirement document upload unit, a dialogue interaction unit, a traditional editing unit, a data retrieval unit, a model selection unit, a task progress display unit, and a report export unit; The user login unit is used to complete user identity verification and login; the report creation unit is used to enter the report name, report type and compilation requirements; the requirement document upload unit is used to upload basic project data; the dialogue interaction unit is used to receive supplementary compilation requirements from users in natural language; the traditional editing unit is used to manually modify and edit the content of the system-generated report; the data retrieval unit is used to retrieve and call existing business materials in the database; the model selection unit is used to select different model strategies and generation modes; the task progress display unit is used to display the task status, chapter progress, problems and rectification suggestions in real time, as well as display the chapter node status, verify the location of problems and trace the source relationship, and support users to trigger partial rewriting or linkage updates for problem items; the report export unit is used to export the final report to a specified format file; the business data generated by the front-end interaction layer is transmitted to the back-end service layer.

3. The intelligent generation and review system for water conservancy technical reports according to claim 1, characterized in that: The backend service layer includes a user management module, a report management module, a data source management module, a task scheduling module, a log management module, a compliance management module, a traceability management module, and a process data processing module; The user management module is used for authentication, permission allocation, and user information management; the report management module is used for creating report objects and controlling report status and version; the data source management module is used for parsing, storing, and indexing various types of data files uploaded by users. The task scheduling module is used to encapsulate the task context, including report type, chapter status, evidence set, verification result and user modification record, and update the task status according to the results returned by the agent; The log management module is used to uniformly collect, store, and retrieve system logs; the compliance management module is used to receive verification results and generate a structured list of issues and tiered rectification suggestions. The traceability management module is used to maintain the mapping relationship between report content and source evidence and modification trajectory; The process data processing module is used for the entire process of data collection, cleaning, desensitization, and structuring, as well as data modification, and synchronizes it to the knowledge and data storage layer. After the backend service layer completes the request processing, it sends the task to the intelligent agent collaboration layer and sends the execution result of the intelligent agent collaboration layer back to the frontend interaction layer.

4. The intelligent generation and review system for water conservancy technical reports according to claim 1, characterized in that: The intelligent agent collaboration layer includes a master-controlled autonomous planning intelligent agent and several functional execution intelligent agents; The main control autonomous planning agent is used to receive the task context of the backend service layer, identify the report type and compilation goal, construct a dynamic task graph without fixed process by combining chapter dependencies, and schedule the execution agents of each function. The functional execution agent includes: The demand parsing agent is used to extract structured elements such as report type, project scale, and applicable standards from user input; An outline generation agent is used to generate report chapter outlines that conform to water conservancy industry standards based on the results of requirements analysis. The knowledge retrieval agent is used to retrieve relevant normative clauses and historical cases from the knowledge and data storage layer; A content writing agent is used to generate initial drafts of each chapter based on the outline structure and search results; A content optimization agent is used to enhance the technical rationality and complete the content of the initial draft. The style alignment agent is used to perform three processes on chapter text: unifying language style, standardizing and replacing water conservancy professional terms, and verifying the consistency of cross-chapter data references. The compliance verification intelligent agent is used to verify the completeness, standardization, and technical rationality of reports and determine the level of problems. A consistency verification agent is used to perform logical consistency verification at the chapter level and across chapters. The verification-assisted intelligent agent is used to provide three types of underlying operators for the compliance verification intelligent agent: chapter omission detection, contextual conflict identification, and technical terminology error matching. A traceability annotation agent is used to generate globally unique version identifiers for each paragraph and key conclusion and maintain a linked list of modification history.

5. The intelligent generation and review system for water conservancy technical reports according to claim 1, characterized in that: The model service layer includes a requirement understanding unit, an outline generation unit, a chapter writing unit, a content optimization unit, a style alignment unit, and a verification assistance unit. The model service layer is the capability support layer of the intelligent agent collaboration layer and is called by the intelligent agents executing various functions. It does not directly form a report task process. The requirement understanding unit is used to identify water conservancy professional terms, domain concepts, and business semantics; the outline generation unit is used to generate a hierarchical chapter structure based on report type and constraints; the chapter writing unit is used to generate logically coherent chapter text; the content optimization unit is used to rewrite text, reduce redundancy, and emphasize key points; and the style alignment unit is used to unify report format and expression. The verification assistance unit is used to identify chapter omissions, inconsistencies in context, and errors in professional terminology.

6. The intelligent generation and review system for water conservancy technical reports according to claim 1, characterized in that: The knowledge and data storage layer adopts a hybrid architecture of relational database, vector database, graph database and file system, including user data storage module, report data storage module, task data storage module, database storage module, knowledge tag module, traceability data storage module and process sedimentation data storage module. The traceability data storage module is used to store the relationship between content source, modification process and verification basis with report paragraphs, evidence basis and decision trajectory as nodes or triples. The user data storage module is used to encrypt and store user identity, permissions, and operation records; the report data storage module is used to store report text, chapter structure, version records, and exported files; the task data storage module is used to record task numbers, status, execution processes, and scheduling information; the database storage module is used to store water conservancy industry standards, historical reports, and technical case studies; the knowledge tagging module is used to add multi-dimensional search tags to various types of data; and the traceability data storage module is used to store the mapping relationship between report content, evidence, and decision-making trajectory in the form of triples. The process sedimentation data storage module is used to store standardized prompt word assets, skill assets, decision-making chain data, expert experience, user preferences, and reusable data assets with business characteristics.

7. The intelligent generation and review system for water conservancy technical reports according to claim 1, characterized in that: The process data asset layer includes a process data acquisition module, a cleaning and desensitization module, a structured decomposition module, an asset classification and accumulation module, an asset quality evaluation module, and an asset retrieval and reuse module. The process data acquisition module is used to collect process data for the entire process, including requirement input, data retrieval, content generation, verification feedback, manual modification, and traceability annotation. The cleaning and desensitization module is used to perform deduplication, format standardization, sensitive information desensitization, and outlier removal on process data. The structured decomposition module is used to establish the relationship between user needs, knowledge evidence, generated results, modification feedback and final adoption results; The asset classification and accumulation module is used to transform process data into prompt word assets, skill assets, decision chain assets, expert experience assets, expert preference assets, and business characteristic assets. The asset quality evaluation module is used to evaluate asset quality based on manual adoption, verification pass rate, and reuse effect; The asset retrieval and reuse module is used to retrieve previously stored assets in subsequent similar tasks.

8. A method for intelligent generation and review of water conservancy technical reports, applied to the intelligent generation and review system for water conservancy technical reports as described in any one of claims 1-7, characterized in that, Includes the following steps: Task reception and standardization processing: The front-end interaction layer receives user-input report information, natural language commands, requirement documents and supplementary materials, completes data formatting and metadata annotation, generates standardized task requests and sends them to the back-end service layer; the back-end service layer completes identity authentication, data indexing, task number generation, encapsulates the task context and sends it to the intelligent agent collaboration layer. Requirements analysis and dynamic task graph construction: The master autonomous planning agent calls the requirements analysis agent to extract structured requirements data, identify report types, compilation objectives, data gaps and constraints; If the requirement information is missing, the interaction will prompt the user to clarify and update the requirement data; a dynamic task graph without a preset fixed process will be built based on the chapter dependency relationship; Knowledge retrieval and report outline generation: The master control autonomously plans and schedules the knowledge retrieval agent to retrieve industry standards, historical cases, and technical standards and form a set of knowledge evidence; at the same time, it schedules the outline generation agent to generate the report chapter framework by combining structured requirements and knowledge evidence. Chapter-by-chapter content generation and optimization: The main control autonomous planning agent breaks down the tasks of each chapter according to the execution order and dependencies of the dynamic task graph; For each chapter, integrate the chapter task description, knowledge evidence, upstream chapter content, and user constraints to generate the chapter generation context; The content writing agent generates the first draft of the chapter, and then the content optimization agent and style alignment agent complete the structural adjustment and format unification. Layered compliance verification and retrospective correction: The compliance verification intelligent agent conducts layered verification of completeness, standardization, technical rationality, format and deduplication of each chapter and the entire report, classifies the problem level and outputs rectification suggestions; If the verification fails, the main controller autonomously plans the intelligent agent to reconstruct the task steps according to the problem type, and schedules the corresponding intelligent agent to perform partial rewriting, cross-chapter alignment, and style correction until the verification is successful. Content traceability annotation and real-time data recording: In the entire process of content generation, optimization and rectification, the traceability annotation intelligent agent binds the corresponding standard basis, data source, user instructions and modification trajectory to each paragraph and key conclusion, establishes a full-link traceability relationship and stores it in the traceability data storage module; Manual Interaction Revision and Linked Updates: If a user submits manual modifications or supplementary instructions, the system identifies the affected chapter scope and relationships, triggers a single-chapter partial update or a cross-chapter linked update, and simultaneously updates the task context and report version. Results assembly, export and full verification: After all chapters pass verification, the backend service layer integrates all chapter content, traceability data and verification records to complete report assembly, generates the final report according to the user-specified format and supports export; Process data assetization and system self-iteration: The process data asset layer cleans, deduplicates, desensitizes, and structurally decomposes the entire process requirement trajectory, modification records, agent decision logs, and review opinion data, transforming them into reusable data assets such as prompt words, skills, decision chains, and expert experience, and storing them in the database; the system reuses the accumulated assets to optimize the execution logic of subsequent tasks, and incrementally fine-tunes the model based on the asset data to achieve continuous self-optimization of the system.

9. The intelligent generation and review method for water conservancy technical reports according to claim 8, characterized in that: The method of constructing the dynamic task graph includes using a directed acyclic graph analysis algorithm to determine chapter dependencies, and distinguishing strong serial dependencies, weak dependencies, and no substantial dependencies based on dependency weights; nodes with strong serial dependencies are executed sequentially, nodes with no substantial dependencies are executed in parallel, and nodes with weak dependencies are scheduled in a semi-parallel manner; during task execution, four scenarios are monitored in real time: missing data, generation anomalies, verification failures, and user additions and modifications, triggering local reconstruction and rescheduling of the task graph.

10. The intelligent generation and review method for water conservancy technical reports according to claim 8, characterized in that: The process data assetization includes screening effective interaction samples to generate prompt word assets, extracting target task processes to generate skill assets, storing agent thinking logs to form decision-making thought chain assets, classifying review opinions to form expert experience assets, statistically analyzing user modification habits to form expert preference assets, and clustering scenario constraints to form business characteristic assets; all assets undergo automated verification and manual review before being put into the database.

Images