Project declaration generation method, system and equipment based on multiple agents and medium

Abstract

The invention relates to the technical field of servers, and particularly provides a multi-agent-based project declaration generation method, system and device and a medium, and the method comprises the steps: receiving a project theme, a keyword, a core innovation point and a declaration guide inputted by a user; analyzing the demand and generating a dynamic semantic map, and decomposing the declaration task into atomic sub-tasks according to the dynamic semantic map; then scheduling information collection and understanding, technical route design, innovation point extraction and writing, declaration synthesis and verification, and a plurality of functional agents cooperatively execute subtasks; and each agent generates a structured knowledge report, a technical route map and an innovation point text, and finally synthesizes a complete declaration draft to be output for a user to review and modify. Through multi-agent cooperation and dynamic semantic analysis, full-process intelligent generation of scientific research project declaration is realized, and the writing efficiency and quality are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a method, system, device, and medium for generating project application documents based on multiple agents. Background Technology

[0002] Research project applications are a crucial part of research activities, but the writing of these applications has long faced the dual challenges of efficiency and quality. Researchers must manually retrieve information from massive amounts of literature, rely heavily on personal experience to design technical routes, and spend considerable time organizing content and adjusting formatting—a complex process prone to blind spots in research. Existing text editing and literature management tools lack domain intelligence and cannot provide end-to-end support; while single large-scale language models often suffer from factual errors, logical inconsistencies, or insufficient innovation when handling such complex and rigorous tasks. Therefore, there is an urgent need in this field for an intelligent solution that can simulate expert team collaboration and achieve a fully intelligent process from understanding requirements to generating high-quality drafts, in order to systematically improve the efficiency of research applications. Summary of the Invention

[0003] In view of the above-mentioned shortcomings of the prior art, the present invention provides a method, system, device and medium for generating project application documents based on multiple agents, so as to solve the above-mentioned technical problems.

[0004] In a first aspect, the present invention provides a method for generating project proposals based on multi-agent systems, comprising: The system receives research project application requirements input by users, including project theme, keywords, a summary of core innovation points, and application guidelines. The requirement information is parsed to generate a dynamic semantic graph, and the application generation task is decomposed into multiple atomic subtasks based on the semantic graph. The atomic subtask is executed by scheduling multiple functional intelligent agents, wherein the multiple functional intelligent agents include at least an information collection and understanding agent, a technology route design agent, an innovation point extraction and writing agent, and an application synthesis and verification agent; Through the collaborative work of the aforementioned multiple functional intelligent agents, structured knowledge reports, technology roadmaps, and innovative point description texts are generated, and a draft of a scientific research project application is synthesized. Output the draft of the research project application to the user interaction module for users to review and modify.

[0005] In one optional implementation, the system receives user-inputted research project application requirements, including: It supports multimodal input methods, including text, document upload, voice input, and sketch upload; Multimodal models are used to convert unstructured inputs into structured demand information. The user interface proactively guides users to complete key information through a conversational interface.

[0006] In an optional implementation, parsing the demand information to generate a dynamic semantic graph includes: Use domain-pretrained language models to perform semantic parsing on user input information to extract entities, logical relationships between entities, and user intent. Based on the extracted entities and logical relationships, a dynamic semantic graph is constructed, wherein the entities serve as nodes of the graph and the logical relationships serve as edges connecting the nodes. Assign attribute information to the nodes and edges in the semantic graph, wherein the attribute information includes at least priority weights and execution dependency constraints; The user intent is used to determine the task type and execution sequence of atomic subtasks during task decomposition: when the user intent includes designing the system, the task flow of the technical route design is mapped and activated; when the user intent includes solving interpretability, the task of interpretability verification and analysis is mapped and associated.

[0007] In an optional implementation, the declaration generation task is decomposed into multiple atomic sub-tasks based on the semantic graph, including: The entities and relationships in the semantic graph are mapped to a preset atomic task template library to determine the required atomic subtask types; Based on the execution dependency constraints marked on nodes and edges in the semantic graph, construct a dependency graph between atomic subtasks; The dependency graph is topologically sorted to obtain an initial task sequence that satisfies all execution dependency constraints; Calculate the overall priority score of each atomic subtask in the initial task sequence, wherein the overall priority score of a subtask is calculated by weighted summation of the priority weights of all semantic graph entities associated with it; Provided that the dependency graph is satisfied, the task sequence is reordered according to the comprehensive priority score to generate the ordered atomic subtask execution sequence, wherein the subtask with the higher comprehensive priority score is arranged at an earlier execution position.

[0008] In one optional implementation, multiple functional agents are scheduled to execute the atomic subtasks, including: For each atomic subtask, its task features are extracted, and the task features are matched with the capability profiles of multiple functional agents in an agent resource pool. Based on the matching results, the optimal functional agent instance is called from the agent resource pool to execute the atomic subtask. During the execution of atomic subtasks, the status indicators of the called agent instance are monitored in real time. When the status indicators trigger the dynamic adjustment conditions in the preset rule base, the dynamic workflow engine executes a path switching operation, which includes starting downstream tasks in parallel or skipping non-critical dependent tasks.

[0009] In an optional implementation, for each atomic subtask, its task features are extracted, and the task features are matched with the capability profiles of multiple functional agents in an agent resource pool. Based on the matching results, the optimal functional agent instance is invoked from the agent resource pool to execute the atomic subtask, including: The atomic subtask is parsed into a structured task feature set, which includes at least: task type, technical field, priority level, and expected maximum time consumption; The task feature set is matched and calculated with the profile data of available instances in the agent resource pool; wherein, the profile data is dynamically updated and includes at least the agent's capability tags, historical performance indicators in a specific technical field, and real-time resource status; Based on a predefined weighted scoring function, calculate the matching score of each available agent instance for the current atomic subtask, and select the agent instance with the highest matching score as the optimal functional agent instance for invocation; The predefined weighted scoring function is calculated as follows: Match score = w1 × domain match degree + w2 × performance efficiency factor + w3 × success rate - w4 × current load factor Wherein: Domain matching degree represents the relevance between the capability label of the agent instance and the technical field to which the atomic subtask belongs; Performance efficiency factor is an inverse or reciprocal factor calculated based on the average speed of the agent instance in processing similar tasks; Success rate is the historical success rate of the agent instance in processing similar tasks; Current load factor is positively correlated with the current task queue length of the agent instance; w1, w2, w3, w4 are a set of predefined weighting coefficients for each calculation factor.

[0010] In one optional implementation, during the execution of the atomic subtask, the status indicators of the called agent instance are monitored in real time; when the status indicators trigger dynamic adjustment conditions in a preset rule base, the dynamic workflow engine executes a path switching operation, which includes starting downstream tasks in parallel or skipping non-critical dependent tasks, including: The status indicators of the intelligent agent instance are collected in real time, including at least the current task execution time, CPU / memory usage, and task queue length; the status indicators are compared with preset thresholds to determine whether the dynamic adjustment condition is triggered. If the timeout adjustment condition is triggered, the dynamic workflow engine analyzes the task dependency graph it maintains: if there is a parallelizable sub-path that does not depend on all the outputs of the timeout task, then the downstream task is started in parallel. If the failure adjustment condition is triggered, the dynamic workflow engine determines the task failure type based on the error code: if it is a non-critical dependency task failure, a manual completion flag is recorded and the task is skipped to continue the process; if it is a critical dependency task failure, the entire workflow is aborted.

[0011] Secondly, the present invention provides a project application generation system based on multi-agent systems, comprising: The requirement receiving module is used to receive scientific research project application requirement information input by users. The requirement information includes project theme, keywords, a summary of core innovation points, and application guide document. The requirement parsing module is used to parse the requirement information, generate a dynamic semantic graph, and decompose the application generation task into multiple atomic sub-tasks based on the semantic graph. The task scheduling module is used to schedule multiple functional agents to execute the atomic sub-tasks, wherein the multiple functional agents include at least an information collection and understanding agent, a technology route design agent, an innovation point extraction and writing agent, and an application synthesis and verification agent; The draft generation module is used to generate structured knowledge reports, technology roadmaps, and innovation point description texts through the collaborative work of the multiple functional intelligent agents, and to synthesize a draft of a scientific research project application. The draft output module is used to output the draft of the scientific research project application to the user interaction module for users to review and modify.

[0012] Thirdly, a device is provided, comprising: Memory, used to store the multi-agent-based project proposal generation program; A processor is configured to implement the steps of the multi-agent-based project proposal generation method as provided in the first aspect when executing the multi-agent-based project proposal generation program.

[0013] Fourthly, a computer-readable medium is provided, on which a multi-agent-based project proposal generation program is stored, wherein when the multi-agent-based project proposal generation program is executed by a processor, the multi-agent-based project proposal generation method as provided in the first aspect is implemented.

[0014] The beneficial effects of this invention are as follows: the project proposal generation method, system, device, and medium provided by this invention, based on multi-agent intelligence, achieve intelligent generation of the entire research project application process through multi-agent collaboration and dynamic semantic analysis. Its beneficial effects are: First, the system decomposes the complex application process into specialized tasks, which are processed in parallel by multiple agents, significantly shortening the application writing cycle and improving efficiency; second, based on a domain knowledge base and dynamic reasoning, the generated content is logically rigorous and highlights innovative points, effectively improving the quality and competitiveness of the application; third, the system has adaptive scheduling and fault-tolerance mechanisms, which can flexibly cope with uncertainties in task execution and ensure the robustness of the process; finally, the system significantly reduces the workload of researchers in data collection, technical planning, and text synthesis, allowing them to focus more on the research of core scientific issues. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic flowchart of a method according to an embodiment of the present invention.

[0017] Figure 2 This is a schematic block diagram of a system according to an embodiment of the present invention.

[0018] Figure 3 This is a schematic diagram of the structure of a device provided in an embodiment of the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0021] The project application generation method based on multi-agent provided in this embodiment of the invention is executed by a computer device, and correspondingly, the project application generation system based on multi-agent runs on the computer device.

[0022] Figure 1 This is a schematic flowchart illustrating a method according to an embodiment of the present invention. Wherein, Figure 1 The implementing entity can be a multi-agent-based project proposal generation system. Depending on different needs, the order of steps in this flowchart can be changed, and some steps can be omitted.

[0023] like Figure 1 As shown, the method includes: S1. Receive research project application requirements information input by the user, including project theme, keywords, core innovation point abstract, and application guide document; S2. Parse the requirement information, generate a dynamic semantic graph, and decompose the application generation task into multiple atomic sub-tasks based on the semantic graph; S3. Schedule multiple functional agents to execute the atomic sub-task, wherein the multiple functional agents include at least an information collection and understanding agent, a technology route design agent, an innovation point extraction and writing agent, and an application synthesis and verification agent; S4. Through the collaborative work of the multiple functional intelligent agents, a structured knowledge report, a technology roadmap, and an innovation point description text are generated, and a draft of a scientific research project application is synthesized. S5. Output the draft of the research project application to the user interaction module for user review and modification.

[0024] In one embodiment of the present invention, based on step S1, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0025] S101. Specific implementation of the multimodal input interface The system provides a graphical user interface and supports four main input methods: Text and document upload: Users can enter the project theme, keywords and a summary of the core innovations in the specified text box, and upload the application guide PDF document.

[0026] Voice Input: An integrated Automatic Speech Recognition (ASR) engine allows users to verbalize project ideas via microphone. The recognized text is automatically filled into the corresponding input box and can be edited.

[0027] Sketch Upload: Users can upload hand-drawn sketches of technical routes or system architectures. The system-integrated visual-language model (such as GPT-4V or similar open-source VLM) is responsible for recognizing and understanding the sketches, converting the graphic elements, connections, and text annotations into descriptive text.

[0028] S102. The process of converting unstructured data to structured data. The aforementioned multimodal input content is processed uniformly by a multimodal information fusion and understanding engine. The core of this engine is a finely tuned large multimodal model, and its workflow is as follows: For voice input, the text converted by ASR is directly sent to the engine.

[0029] For the uploaded sketch, the VLM model will generate a structured text describing its core content (e.g., "This sketch describes a process that includes three main steps: data preprocessing, model training, and result evaluation").

[0030] For all text information (including direct input, ASR results, VLM output, and OCR parsing results of uploaded documents), the engine uses domain pre-trained models (such as SciBERT) to perform deep semantic analysis, accurately extracting technical entities (such as "Transformer model" and "medical image"), method actions (such as "optimization" and "construction"), and performance metrics (such as "accuracy" and "latency"), and finally integrating them into a standardized structured requirement JSON template.

[0031] S103. Implementation of Actively Guided Interaction To address the issue of ambiguous or incomplete user input, the system incorporates a state machine-based dialogue management module. This module analyzes the received structured requirements information in real time. If it detects missing or insufficient confidence levels in key fields (such as "technical field segmentation" or "expected outcome"), it will proactively ask the user follow-up questions through the conversational UI on the user interface.

[0032] Example: When a user enters the topic "Research on next-generation battery materials" but does not specify the specific technical path, the system will pop up a prompt: "Please further explain whether you are interested in innovation in cathode materials, anode materials, or electrolyte systems?" Through this interaction, the system gradually guides the user to complete the information, ensuring the accuracy and completeness of the subsequent task decomposition of the intelligent agent.

[0033] In one embodiment of the present invention, based on step S2, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0034] S201. Parse the requirement information and generate a dynamic semantic graph, including: 1. Semantic parsing and element extraction: The semantic parsing engine loads a language model (e.g., SciBERT) pre-trained on a large amount of scientific literature. This model performs in-depth analysis on the structured requirement information processed in the S1 stage.

[0035] Entity extraction: The model identifies and extracts the core concepts from the text as entities. For example, given the user input "medical diagnosis decision support system based on causal reasoning", the engine will extract key entities such as "causal reasoning", "medical diagnosis", and "decision support system".

[0036] Relationship Extraction: The model further analyzes the sentence structure to identify the logical relationships between entities. In the example above, it will identify a "applied to" relationship between "decision support system" and "medical diagnosis," and a "based on" relationship between "decision support system" and "causal reasoning."

[0037] User intent recognition: The model integrates the entire input context to determine the user's macro-level goal. For example, it can identify two core intents from the input: "designing a system" and "solving interpretability issues".

[0038] 2. Construction and attribute assignment of dynamic semantic graph: Based on the above extraction results, the parsing engine constructs a graph data structure in memory, namely the dynamic semantic graph.

[0039] Graph Structure: Extracted entities (such as "causal reasoning" and "medical diagnosis") are used as nodes in the graph. Identified logical relationships (such as "based on" and "applied to") are used as edges connecting these nodes.

[0040] Attribute assignment: The system will calculate and assign key attribute information to each node and edge.

[0041] Priority weight: By analyzing the position and frequency of entities in the input, and combining this with explicit annotations by users on the interactive interface (such as dragging and sorting keywords), an initial priority weight is calculated for each entity node. For example, "causal reasoning" as the core method is assigned a weight of 0.9 (out of 1.0), while "decision support system" as the carrier is assigned a weight of 0.7.

[0042] Execution dependency constraints: Mark the logical order of task execution on the edges. For example, the edge "based on" implies a strong dependency constraint, meaning that "causal reasoning" must be understood before the "decision support system" can be designed.

[0043] 3. The mapping mechanism from user intent to task sequence: The completed semantic graph will be passed to the task decomposition module. Here, user intent directly determines the skeleton of the subsequent task flow.

[0044] When the system recognizes that the user intent includes "design system", it will map and activate the "technical route design task flow" from the task template library. This task flow will become the core of the atomic subtask sequence.

[0045] When the system simultaneously identifies user intent that includes "solving interpretability", it will associate and insert atomic subtasks such as "interpretability requirement analysis" and "interpretability verification method design" into the above task flow.

[0046] Through this mechanism, the system ensures that the final generated task execution sequence accurately corresponds to the user's in-depth scientific research goals.

[0047] S202. Based on the semantic graph, the application generation task is decomposed into multiple atomic sub-tasks, including: 1. Task mapping and dependency construction: The task planning engine first receives a dynamic semantic graph from upstream. Its built-in task mapper matches entities and relationships in the graph with a pre-defined library of atomic task templates.

[0048] Example: Identifying the entities "causal reasoning" and the relation "based on" will map to atomic subtask types such as "retrieve causal reasoning methodology" and "technical route design". Identifying the entity "medical diagnosis" will map to the subtask "investigate the current application status in the field of medical diagnosis".

[0049] Subsequently, the engine's dependency analyzer reads the execution dependency constraints marked on the edges of the semantic graph. For example, a constraint might specify that "background research" must be completed before "technology roadmap design" can begin. Based on these constraints, the engine constructs a task dependency graph in the form of a directed acyclic graph (DAG), where nodes are atomic subtasks and directed edges define the execution order.

[0050] 2. Initial sequence generation and priority calculation: The engine performs topological sorting on the above task dependency graph to generate one or more initial task sequences that do not violate any dependencies.

[0051] Next, the engine calculates the overall priority score for each atomic subtask in the sequence. The calculation rule is to sum the priority weights of all semantic graph entities associated with that subtask.

[0052] Example: The subtask "Retrieving Causal Reasoning Methodology" is associated with two entities, "Causal Reasoning" (weight 0.9) and "Methodology" (weight 0.5), so its combined priority score is 1.4. The subtask "Format Validation," however, is associated only with low-weight entities, hence its lower score.

[0053] 3. Priority-driven sequence optimization Finally, the engine reorders and optimizes the initial task sequence while strictly adhering to the constraints defined in the task dependency graph. Its core principle is to move subtasks with higher overall priority scores to the front of the sequence as much as possible, without causing dependency conflicts.

[0054] Optimization Results: This allows tasks that users are most concerned about and that are most relevant to the core innovation points (such as "technology roadmap design" and "innovation point extraction") to be prioritized and executed, even if some background research tasks are not yet fully completed. This priority-based intelligent scheduling significantly improves the system's response speed and user experience, ensuring that core content receives the most sufficient computing resources and is produced as early as possible.

[0055] In one embodiment of the present invention, based on step S3, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0056] S301. For each atomic subtask, extract its task features and match the task features with the capability profiles of multiple functional agents in an agent resource pool. Based on the matching results, call the optimal functional agent instance from the agent resource pool to execute the atomic subtask.

[0057] 1. Task Feature Extraction: When the scheduler receives an atomic subtask to be executed (e.g., "Retrieving the application of causal reasoning in biomedicine"), it first parses it to generate a structured task feature set. This feature set contains the following key fields: Task type: such as "information retrieval", "technical design" or "text generation"; Technical field: such as "biomedical"; Priority level: Inherited from the weights of associated entities in the semantic graph, such as "high"; Expected maximum time: preset based on task complexity, such as 120 seconds.

[0058] 2. Profile Matching and Score Calculation: The scheduler matches the task feature set with the dynamic profile data of all available instances in the agent resource pool. The profile data is stored in JSON format, updated in real time, and includes the following dimensions: Capability tags: such as ["Academic paper retrieval", "Patent analysis", "Policy interpretation"]; Historical performance metrics: including average processing time and success rate in a specific field (such as "biomedical"); Real-time resource status: including current task queue length, CPU load, etc.; The scheduler uses a predefined weighted scoring function to calculate the matching score for each agent instance: Match score = w1 × domain match degree + w2 × performance efficiency factor + w3 × success rate - w4 × current load factor Wherein: Domain matching degree represents the relevance between the capability label of the agent instance and the technical field to which the atomic subtask belongs; Performance efficiency factor is an inverse or reciprocal factor calculated based on the average speed of the agent instance in processing similar tasks; Success rate is the historical success rate of the agent instance in processing similar tasks; Current load factor is positively correlated with the current task queue length of the agent instance; w1, w2, w3, w4 are a set of predefined weighting coefficients for each calculation factor.

[0059] Domain matching degree: Calculated by comparing the task's technical field with the agent's capability labels. A perfect match is 1.0, a partial match is 0.5, and a no match is 0. Performance efficiency factor: Calculated as 1 / average processing time, encouraging the selection of instances with fast processing speed; Success rate: Historical success rate data is used directly; Current load factor: proportional to the current task queue length of the instance; Weighting coefficients w1-w4: Preset according to the system optimization target, for example, set to [0.4, 0.3, 0.2, 0.1].

[0060] 3. Optimal instance selection and invocation: The scheduler selects the agent instance with the highest matching score to execute the task. For example, an instance with excellent historical performance in the "biomedical" domain (success rate 0.98), fast processing speed (average time 110 seconds), and currently idle will be prioritized. This scheduling strategy based on multi-dimensional quantitative evaluation ensures that tasks are assigned to the most suitable agent, achieving optimal allocation of system resources and maximizing overall efficiency.

[0061] S302. During the execution of the atomic subtask, the status indicators of the called agent instance are monitored in real time; when the status indicators trigger the dynamic adjustment conditions in the preset rule base, the dynamic workflow engine executes a path switching operation, which includes starting downstream tasks in parallel or skipping non-critical dependent tasks.

[0062] 1. Status monitoring and condition judgment: The dynamic workflow engine achieves real-time monitoring through the following methods: Metrics collection: The engine collects status metrics in real time through the "heartbeat" data periodically reported by the intelligent agent, including: current task execution time, CPU / memory usage, and current task queue length. Threshold comparison: The engine compares the collected metrics with preset thresholds to make a judgment; Timeout adjustment conditions: When the task execution time exceeds 70% of the maximum allowed duration for this type of task (e.g., if the information gathering task has a time limit of 10 minutes, it will be triggered if it exceeds 7 minutes). Failure adjustment condition: Triggered immediately when the agent reports an unrecoverable error (such as data source connection failure, parsing exception, etc.).

[0063] 2. Dynamic path decision-making and execution: Based on different triggering conditions, the engine executes corresponding path switching operations. (1) Parallelization optimization under timeout conditions When the timeout adjustment condition is triggered, the engine accesses its maintained task dependency graph (DAG) for analysis: Dependency analysis: Identify which downstream tasks of the timed-out task can start first without relying on all of its output; Parallel startup: For branches that meet the conditions, the engine immediately starts the corresponding downstream tasks in parallel; Example: When the "information gathering agent" times out, since the core methodological parameters required by the "technology route design agent" have been provided by the semantic graph, the engine can start the technology route design task in parallel, allowing it to make preliminary plans based on the general methodology in the knowledge base.

[0064] (2) Process fault tolerance in case of failure When the failure adjustment condition is triggered, the engine performs tiered processing based on the error code: Non-critical dependency task failure: If a secondary data source is inaccessible, the engine records a "pending manual completion" flag, skips the task, and continues to execute subsequent processes; Critical dependency task failure: If the technical route design fails completely, the engine will immediately stop the entire workflow and send an interruption notification and a detailed error report to the user; This dynamic adjustment mechanism ensures that the system can still operate efficiently when faced with execution uncertainties. It improves overall efficiency through parallelization optimization and ensures the robustness of the process through hierarchical fault tolerance.

[0065] In one embodiment of the present invention, based on step S4, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0066] S401. Information Gathering and Understanding: The Execution Process of an Intelligent Agent The agent is first activated and, based on the "background research" subtask in the task sequence, performs retrieval in parallel through its built-in multiple data source interfaces. For example, for the project "Medical Diagnosis Decision Support System Based on Causal Reasoning," the agent will: retrieve relevant top conference papers from academic databases in the past three years; retrieve relevant technology patents from patent databases; obtain medical AI-related guidelines from policy databases; and use information extraction techniques to extract key data points from the retrieval results.

[0067] Generate structured knowledge reports containing research status, technological bottlenecks, and policy requirements using text summarization technology. S402. Technical Route Design: Execution Process of the Intelligent Agent This intelligent agent receives structured knowledge reports and original user requirements from upstream sources, and accesses the domain knowledge base to obtain relevant technical methodologies. Through its built-in planning algorithm, it analyzes the correspondence between current technical bottlenecks and user innovation points; retrieves methodological templates such as "causal discovery" and "counterfactual reasoning" from the knowledge base; generates a detailed technical roadmap containing key technical nodes, experimental steps, and dependencies; and outputs a visual roadmap and corresponding textual descriptions.

[0068] S403. Innovation Point Extraction and Writing of the Agent's Execution Process The agent simultaneously receives structured knowledge reports and technology roadmaps, and performs comparative analysis: it compares the project's technology roadmap with existing solutions to identify the project's technological innovation dimensions; it quantifies the advancement and technological value of each innovation point; and it generates a descriptive text of the core innovation points that conforms to academic standards.

[0069] For example, three major innovations were identified, including "an end-to-end causal reasoning framework that integrates domain knowledge".

[0070] S404. Execution process of application form synthesis and verification intelligent agent The intelligent agent acts as the final integrator: it calls the pre-set application form template; it intelligently fills the content modules output by each intelligent agent into the corresponding chapters; it fills in the background information in "Project Basis"; it fills in the technology roadmap in "Research Plan"; it fills in the description of innovation points in "Project Innovativeness"; it automatically generates abstracts and keywords; it performs compliance checks, including format review, word count, and verification of required fields; and it outputs a draft application form with a completeness of over 80%.

[0071] Ultimately, the generated draft application is presented to the user through an interactive module, supporting online editing and annotation, forming a closed-loop human-machine collaborative workflow. The entire process is completed within hours, significantly improving the efficiency of application writing.

[0072] In one embodiment of the present invention, based on step S5, a possible embodiment will be given below, and its specific implementation will be described in a non-limiting manner.

[0073] S501. Structured Results Presentation After receiving the complete draft generated by the application synthesis and verification agent, the user interaction module presents it to the user in the following ways: Multi-view display: The interface is divided into three main areas; Left-side outline navigation area: Displays the chapter structure of the application form, and supports clicking to jump to the next section; Central main editing area: Displays the full text of the application in rich text format, retaining all formatting styles; The annotation panel on the right displays the system's suggested modifications and compliance check results in real time. Content source tracing function: Each chapter and paragraph provides a "Source of Generation" floating prompt, which allows users to see which agent generated the content and which data sources it is based on, enhancing system transparency.

[0074] S502. Intelligent Interactive Editing Function The system provides the following core interactive functions to support user revisions: Intelligent annotation system: Users can select any text to add annotations, and the system automatically classifies annotation images through natural language understanding (such as "needs expansion", "insufficient evidence", "expression optimization"). One-click revision suggestions: Based on user annotations, the system calls the corresponding intelligent agent to provide revision suggestions. For example, when a user marks "the technical roadmap is not detailed enough," the technical roadmap design intelligent agent automatically generates supplementary content as alternatives. Real-time compliance checks: During the editing process, the system continuously runs validation rules in the background, providing real-time prompts for issues such as format violations, exceeding word limits, and missing required fields.

[0075] S503. Feedback Loop and Iterative Optimization The system establishes a complete human-machine collaborative workflow: Version management: Automatically saves each revision, supports version comparison and rollback; Feedback collection: User ratings and modification behaviors for the content generated by the agent are recorded and used to optimize the agent's performance; Rapid Regeneration: When users make significant structural adjustments, they can use the "Regenerate This Chapter" function to call on the corresponding intelligent agent to quickly generate new content.

[0076] Finally, users can export a satisfactory draft application form at any time into Word or PDF format that meets the requirements of the funding unit, completing the entire process from intelligent generation to final delivery.

[0077] In some embodiments, the multi-agent-based project proposal generation system may include multiple functional modules composed of computer program segments. The computer programs of each program segment in the multi-agent-based project proposal generation system may be stored in the memory of a computer device and executed by at least one processor to perform (see details). Figure 1 (Description) Functionality for generating project proposals based on multi-agent systems.

[0078] In this embodiment, the multi-agent-based project application generation system can be divided into multiple functional modules according to the functions it performs, such as... Figure 2 As shown. The module referred to in this invention is a series of computer program segments that can be executed by at least one processor and perform a fixed function, and is stored in memory. In this embodiment, the functions of each module will be described in detail in subsequent embodiments.

[0079] The requirement receiving module is used to receive scientific research project application requirement information input by users. The requirement information includes project theme, keywords, a summary of core innovation points, and application guide document. The requirement parsing module is used to parse the requirement information, generate a dynamic semantic graph, and decompose the application generation task into multiple atomic sub-tasks based on the semantic graph. The task scheduling module is used to schedule multiple functional agents to execute the atomic sub-tasks, wherein the multiple functional agents include at least an information collection and understanding agent, a technology route design agent, an innovation point extraction and writing agent, and an application synthesis and verification agent; The draft generation module is used to generate structured knowledge reports, technology roadmaps, and innovation point description texts through the collaborative work of the multiple functional intelligent agents, and to synthesize a draft of a scientific research project application. The draft output module is used to output the draft of the scientific research project application to the user interaction module for users to review and modify.

[0080] Figure 3The multi-agent-based project proposal generation method provided in this application embodiment can be applied to devices. Those skilled in the art will understand that the device structures involved in the embodiments of this invention do not constitute a limitation on the device. A device may include more or fewer components than illustrated, or combine certain components, or have different component arrangements. In the embodiments of this invention, the device includes, but is not limited to, laptop computers, desktop computers, workbenches, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the embodiments of this application described and / or claimed herein.

[0081] The device 300 may include a processor 310, a memory 320, and a communication unit 330. These components communicate via one or more buses. Those skilled in the art will understand that the server structure shown in the figure does not constitute a limitation of the present invention. It may be a bus topology or a star topology, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0082] The memory 320 can be used to store execution instructions of the processor 310. The memory 320 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. When the execution instructions in the memory 320 are executed by the processor 310, the device 300 is able to perform some or all of the steps in the above method embodiments.

[0083] The processor 310 serves as the control center of the storage device, connecting various parts of the electronic device via various interfaces and lines. It executes software programs and / or modules stored in the memory 320, and calls data stored in the memory to perform various functions of the electronic device and / or process data. The processor can be composed of integrated circuits (ICs), such as a single packaged IC or multiple packaged ICs with the same or different functions connected together. For example, the processor 310 may consist only of a central processing unit (CPU). In this embodiment of the invention, the CPU may have a single processing core or include multiple processing cores.

[0084] The communication unit 330 is used to establish a communication channel, enabling the storage device to communicate with other devices. It can receive user data sent by other devices or send user data to other devices.

[0085] The present invention also provides a computer medium, wherein the computer medium may store a program, which, when executed, may include some or all of the steps provided in the embodiments of the present invention. The medium may be a magnetic disk, an optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0086] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a medium such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, or any other medium capable of storing program code. It includes several instructions to cause a computer device (which may be a personal computer, a server, or a second device, network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0087] The same or similar parts between the various embodiments in this specification can be referred to mutually. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple, and the relevant parts can be referred to the description in the method embodiments.

[0088] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or modules may be electrical, mechanical, or other forms.

[0089] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0090] In addition, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0091] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the present invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the present invention by those skilled in the art without departing from the spirit and essence of the invention, and such modifications or substitutions should all be within the scope of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should also be covered within the protection scope of the present invention.

Claims

1. A multi-agent based project proposal generation method, characterized by, include: The system receives research project application requirements input by users, including project theme, keywords, a summary of core innovation points, and application guidelines. The requirement information is parsed to generate a dynamic semantic graph, and the application generation task is decomposed into multiple atomic subtasks based on the semantic graph. The atomic subtask is executed by scheduling multiple functional intelligent agents, wherein the multiple functional intelligent agents include at least an information collection and understanding agent, a technology route design agent, an innovation point extraction and writing agent, and an application synthesis and verification agent; Through the collaborative work of the aforementioned multiple functional intelligent agents, structured knowledge reports, technology roadmaps, and innovative point description texts are generated, and a draft of a scientific research project application is synthesized. Output the draft of the research project application to the user interaction module for users to review and modify.

2. The method of claim 1, wherein, Receive research project application requirements input by users, including: It supports multimodal input methods, including text, document upload, voice input, and sketch upload; Multimodal models are used to convert unstructured inputs into structured demand information. The user interface proactively guides users to complete key information through a conversational interface.

3. The method of claim 1, wherein, Parsing the demand information and generating a dynamic semantic graph includes: Use domain-pretrained language models to perform semantic parsing on user input information to extract entities, logical relationships between entities, and user intent. Based on the extracted entities and logical relationships, a dynamic semantic graph is constructed, wherein the entities serve as nodes of the graph and the logical relationships serve as edges connecting the nodes. Assign attribute information to the nodes and edges in the semantic graph, wherein the attribute information includes at least priority weights and execution dependency constraints; The user intent is used to determine the task type and execution sequence of atomic subtasks during task decomposition: when the user intent includes designing the system, the task flow of the technical route design is mapped and activated; when the user intent includes solving interpretability, the task of interpretability verification and analysis is mapped and associated.

4. The method of claim 3, wherein, Based on the semantic graph, the application generation task is decomposed into multiple atomic sub-tasks, including: The entities and relationships in the semantic graph are mapped to a preset atomic task template library to determine the required atomic subtask types; Based on the execution dependency constraints marked on nodes and edges in the semantic graph, construct a dependency graph between atomic subtasks; The dependency graph is topologically sorted to obtain an initial task sequence that satisfies all execution dependency constraints; Calculate the overall priority score of each atomic subtask in the initial task sequence, wherein the overall priority score of a subtask is calculated by weighted summation of the priority weights of all semantic graph entities associated with it; Provided that the dependency graph is satisfied, the task sequence is reordered according to the comprehensive priority score to generate the ordered atomic subtask execution sequence, wherein the subtask with the higher comprehensive priority score is arranged at an earlier execution position.

5. The method of claim 1, wherein, Scheduling multiple functional agents to execute the atomic subtasks includes: For each atomic subtask, its task features are extracted, and the task features are matched with the capability profiles of multiple functional agents in an agent resource pool. Based on the matching results, the optimal functional agent instance is called from the agent resource pool to execute the atomic subtask. During the execution of atomic subtasks, the status indicators of the called agent instance are monitored in real time. When the status indicators trigger the dynamic adjustment conditions in the preset rule base, the dynamic workflow engine executes a path switching operation, which includes starting downstream tasks in parallel or skipping non-critical dependent tasks.

6. The method of claim 5, wherein, For each atomic subtask, extract its task features. The task characteristics are then matched with the capability profiles of multiple functional agents in an agent resource pool. Based on the matching results, the optimal functional agent instance is invoked from the agent resource pool to execute the atomic subtask, including: The atomic subtask is parsed into a structured task feature set, which includes at least: task type, technical field, priority level, and expected maximum time consumption; The task feature set is matched and calculated with the profile data of available instances in the agent resource pool; wherein, the profile data is dynamically updated and includes at least the agent's capability tags, historical performance indicators in a specific technical field, and real-time resource status; Based on a predefined weighted scoring function, calculate the matching score of each available agent instance for the current atomic subtask, and select the agent instance with the highest matching score as the optimal functional agent instance for invocation; The predefined weighted scoring function is calculated as follows: Match score = w1 × domain match degree + w2 × performance efficiency factor + w3 × success rate - w4 × current load factor Wherein: Domain matching degree represents the relevance between the capability label of the agent instance and the technical field to which the atomic subtask belongs; Performance efficiency factor is an inverse or reciprocal factor calculated based on the average speed of the agent instance in processing similar tasks; Success rate is the historical success rate of the agent instance in processing similar tasks; Current load factor is positively correlated with the current task queue length of the agent instance; w1, w2, w3, w4 are a set of predefined weighting coefficients for each calculation factor.

7. The method of claim 5, wherein, During the execution of atomic subtasks, the status indicators of the called agent instance are monitored in real time. When the status indicators trigger dynamic adjustment conditions in the preset rule base, the dynamic workflow engine executes a path switching operation. The path switching operation includes starting downstream tasks in parallel or skipping non-critical dependent tasks, including: The status indicators of the intelligent agent instance are collected in real time, including at least the current task execution time, CPU / memory usage, and task queue length; the status indicators are compared with preset thresholds to determine whether the dynamic adjustment condition is triggered. If the timeout adjustment condition is triggered, the dynamic workflow engine analyzes the task dependency graph it maintains: if the downstream task has a parallelizable sub-path that does not depend on all the outputs of the timeout task, then the downstream task is started in parallel. If the failure adjustment condition is triggered, the dynamic workflow engine determines the task failure type based on the error code: if it is a non-critical dependency task failure, a manual completion flag is recorded and the task is skipped to continue the process; if it is a critical dependency task failure, the entire workflow is aborted.

8. A multi-agent based project proposal generation system, characterized by, include: The requirement receiving module is used to receive scientific research project application requirement information input by users. The requirement information includes project theme, keywords, a summary of core innovation points, and application guide document. The requirement parsing module is used to parse the requirement information, generate a dynamic semantic graph, and decompose the application generation task into multiple atomic sub-tasks based on the semantic graph. The task scheduling module is used to schedule multiple functional agents to execute the atomic sub-tasks, wherein the multiple functional agents include at least an information collection and understanding agent, a technology route design agent, an innovation point extraction and writing agent, and an application synthesis and verification agent; The draft generation module is used to generate structured knowledge reports, technology roadmaps, and innovation point description texts through the collaborative work of the multiple functional intelligent agents, and to synthesize a draft of a scientific research project application. The draft output module is used to output the draft of the scientific research project application to the user interaction module for users to review and modify.

9. A multi-agent based project proposal generation device, characterized by, include: Memory, used to store the multi-agent-based project proposal generation program; A processor is configured to implement the steps of the multi-agent-based project proposal generation method as described in any one of claims 1-7 when executing the multi-agent-based project proposal generation program.

10. A computer readable medium having stored thereon a computer program, characterized in that, The readable medium stores a project proposal generation program based on multi-agent intelligence, which, when executed by a processor, implements the steps of the project proposal generation method based on multi-agent intelligence as described in any one of claims 1-7.

Images