Scientific research task autonomous planning and execution method and system

Abstract

The invention provides a scientific research task autonomous planning and execution method and system. The method comprises the steps that natural language query is accepted in an interactive dialogue mode; identifying intent based on natural language query; decomposing problems based on intentions; and corresponding answers are generated and returned based on different intentions. The system is constructed by a host module and a server module based on a model context protocol; wherein the host module comprises a plant chat robot and a plant inference device; the server module comprises a knowledge agent, a data agent, a composite agent and a review agent; the host module is used for decomposing a problem into one or more sub-tasks, and is also used for defining an execution sequence and cooperation logic among the sub-tasks. According to the method and system provided by the invention, through basic training, incremental pre-training and supervised fine tuning, scientific research tasks in the fields of agriculture and the like are autonomously planned and executed, and the efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of information technology, specifically to the interdisciplinary field of intelligent collaboration and automated scientific research tasks, and particularly to a method and system for autonomous planning and execution of scientific research tasks. Background Technology

[0002] Research activities in fields such as biology, medicine, and agronomy involve many complex tasks. For example, in agricultural research, these include knowledge organization, data retrieval, bioinformatics analysis, and review writing. Traditionally, these tasks rely heavily on manual coordination. Researchers need to manually search literature, organize data, select analytical methods, and deploy tools. This process is not only time-consuming and labor-intensive but also prone to problems due to differences in professional knowledge. This can lead to inconsistent procedures and biased results.

[0003] Existing technologies either support only single knowledge retrieval or single data query. They cannot work together to complete the entire task process. Other technologies lack self-feedback and correction mechanisms. These technologies struggle to adapt to the dynamic needs of complex scenarios such as omics analysis. These problems result in low efficiency in scientific research task processing and high professional barriers. Ultimately, they fail to meet the needs of modern agricultural scientific research and cannot achieve efficient, autonomous, and precise processing of complex tasks. Summary of the Invention

[0004] The purpose of this invention is to provide a method for autonomous planning and execution of scientific research tasks, comprising the following steps: accepting natural language queries in the form of interactive dialogue; identifying intents based on the natural language queries; decomposing questions based on the intents; and generating corresponding answers based on different intents.

[0005] The intent identified based on the natural language query includes: confirming that the intent is a scientific research knowledge requirement, confirming that the intent is a scientific research data requirement, or confirming that the intent is a comprehensive review requirement.

[0006] The intention-based problem decomposition further includes: decomposing the problem into one or more sub-tasks; and clarifying the execution order and collaboration logic between the sub-tasks.

[0007] Specifically, confirming that the intent is a scientific research knowledge requirement, the step of generating corresponding answers based on different intents further includes: vectorizing the natural language query into semantic embedding, performing multi-strategy RAG retrieval from a vector knowledge base; reordering the retrieved content by relevance, selecting context fragments as knowledge context, and generating targeted answers by combining natural language interpretation with key knowledge source annotation.

[0008] Specifically, confirming that the intent is a scientific research data requirement, the step of generating corresponding responses based on different intents further includes: rewriting the natural language query in a form that conforms to the database structure, and generating a standard SQL statement; obtaining structured data from the scientific research database based on the standard SQL statement, and returning it in a format that combines natural language interpretation with the structured data carrier.

[0009] Specifically, confirming that the intent is a comprehensive review request, the step of generating corresponding responses based on different intents further includes: developing a structured review plan; iteratively retrieving multi-source information based on the structured review plan; integrating and rewriting the obtained multimodal information to form a research review, and returning it in the form of a report.

[0010] The vector knowledge base is built on specific documents during a specific period; the specific documents are preprocessed and vectorized to construct the vector knowledge base.

[0011] This invention also provides a research task autonomous planning and execution system, constructed from a host module and a server module based on a model context protocol. The host module includes: a plant chatbot for accepting natural language queries in an interactive dialogue format and for identifying intents based on the natural language queries; and a plant inference engine for decomposing questions based on the intents and, when the intents are for comprehensive review requests, formulating a structured review plan. The server module includes: a knowledge agent for performing multi-strategy RAG detection; and a data agent, driven by the plant inference engine, for converting the natural language queries into a form conforming to a database structure. The formula is rewritten and a standard SQL statement is generated. Structured data is obtained based on the standard SQL statement. A composite intelligent agent is used to combine the knowledge intelligent agent, the data intelligent agent, and the analysis intelligent agent into an intelligent agent with a specific application. A review intelligent agent, which is also provided in this invention, is used to formulate a structured review plan and coordinate the knowledge intelligent agent to iteratively retrieve multi-source information based on the structured review plan. The obtained multimodal information is integrated and rewritten to form a research review, which is returned in the form of a report. The host module is used to decompose the problem into one or more sub-tasks and clarify the execution order and collaboration logic between the sub-tasks.

[0012] The construction and training of the plant chatbot and the plant inferencer include basic model training, incremental pre-training, and supervised fine-tuning stages. In the basic model training stage, the basic dataset comes from multiple sources, focusing on data containing scientific research tasks. In the incremental pre-training stage, incremental pre-training is performed on a preset number of Pangu basic models for specific knowledge in a specific field. In the supervised fine-tuning stage, the plant chatbot and the plant inferencer are subjected to supervised fine-tuning to make their capabilities consistent with specific scientific research tasks and user interaction patterns.

[0013] The server module further includes the analytical agent, which adopts the PDCA framework. The analytical agent includes a memory module and a toolkit integration unit. The PDCA framework operates in a closed loop, supported by the memory module and the toolkit integration unit. The memory module stores process data generated during the analysis. In the check phase of the PDCA framework, it identifies deviations between the analysis results and the expected goals. In the action phase, it generates correction schemes based on these deviations. The toolkit integration unit integrates the omics analysis toolkits required for the analysis. Through an interface with the data agent, it retrieves the raw data required for the analysis from biological databases. In the execution phase of the PDCA framework, it calls the omics analysis toolkits to execute the generated analysis code. In the planning phase of the PDCA framework, the analytical agent retrieves relevant literature and summarizes analysis methods through the knowledge agent. Guided by the AI ​​model, it plans tasks and generates the analysis code for execution, wherein the analysis code is adapted to the syntax rules of the omics analysis toolkits.

[0014] This invention effectively addresses the functional limitations and efficiency bottlenecks of existing technologies in agricultural research task processing by constructing a multi-agent collaborative framework based on Model Context Protocol (MCP), providing comprehensive technical support for agricultural research. Existing technologies either can only independently achieve single knowledge retrieval or only support simple data queries, failing to coordinate and complete the entire research task process from requirement identification to result output. Furthermore, some technologies lack self-feedback correction mechanisms, making it difficult to adapt to the dynamic needs of complex scenarios such as omics analysis. In contrast, this invention, relying on a collaborative architecture of host and server modules, integrates knowledge agents, data agents, analysis agents, composite agents, and review agents. It achieves integrated processing of multiple types of needs, including research knowledge querying, data acquisition, bioinformatics analysis, and research review writing, eliminating the need for researchers to switch between multiple systems. This significantly improves the continuity and completeness of task processing, covering the entire agricultural research scenario from basic information acquisition to complex analysis and summarization.

[0015] In terms of task execution efficiency, this invention significantly reduces the cost of manual intervention through natural language interaction and automated process design. The plant chatbot in the host module can directly receive natural language queries in interactive dialogue and accurately identify the type of request; the plant inference engine can automatically break down the problem into sub-tasks, clarifying the execution logic and collaborative relationships; the agents in the server module complete automated operations according to preset mechanisms. Specifically, the knowledge agent quickly obtains relevant literature content through multi-strategy RAG retrieval, the data agent completes the conversion of natural language to SQL statements and data extraction under the drive of the plant inference engine, and the analysis agent realizes autonomous advancement and bias correction of bioinformatics analysis based on the PDCA framework. The entire process does not require manual literature retrieval, data organization, or deployment of analysis tools, effectively reducing a large number of manual coordination links in traditional scientific research tasks, shortening the overall processing cycle of scientific research tasks, and allowing researchers to focus more on key research innovations.

[0016] Meanwhile, this invention ensures the accuracy and reliability of scientific research task processing through specialized design. The vector knowledge base is built on specific documents from a specific period. After preprocessing and vectorization, it forms structured resources. Combined with the multi-strategy RAG retrieval of the knowledge agent, it ensures the comprehensiveness and relevance of knowledge acquisition. The plant chatbot and plant inference engine are built through three stages: basic model training, incremental pre-training, and supervised fine-tuning. They gradually adapt to the professional terminology, research logic, and user interaction patterns in the field of agricultural research, improving the accuracy of intent recognition and task decomposition. The analysis agent stores the analysis process data through a memory module. In the "check" stage of the PDCA framework, it accurately identifies result deviations and generates correction schemes in the "action" stage, forming a closed-loop optimization mechanism. This effectively avoids result deviations caused by operational errors or inappropriate method selection in traditional manual analysis, providing a stable guarantee for the efficient and accurate execution of agricultural research tasks. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This invention relates to a flowchart of a method for autonomous planning and execution of scientific research tasks.

[0019] Figure 2 This invention relates to a flowchart for generating corresponding responses based on different intentions.

[0020] Figure 3This is a block diagram of the self-service planning and execution system for scientific research tasks involved in this invention.

[0021] Figure 4 This invention relates to a hierarchical flowchart of the model construction and training of the plant chatbot and plant inference engine.

[0022] Explanation of reference numerals in the attached figures: 100. Autonomous Planning and Execution System for Scientific Research Tasks; 110. Main Unit Module; 111. Plant Chatbot; 112. Plant Inference Engine; 120. Server Module; 121. Knowledge Agent; 122. Data Agent; 123. Composite Agent; 124. Review Agent; 125. Analytical Agent. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] Below, in conjunction with Figures 1-4 This invention relates to a method and system for autonomous planning and execution of scientific research tasks.

[0025] Figure 1 This is a flowchart illustrating the autonomous planning and execution method for scientific research tasks involved in this invention. For example... Figure 1 As shown, the method for autonomous planning and execution of scientific research tasks involved in this invention includes: in step S110, accepting natural language queries in the form of interactive dialogue; in step S120, identifying intent based on natural language queries; in step S130, decomposing questions based on intent; and in step S140, generating corresponding answers based on different intents and returning them.

[0026] Specifically, in step S110, the process of accepting natural language queries in the form of interactive dialogue is completed by the plant chatbot (Phyto-Chatbot) in the host module (Phytomni-Hub). This chatbot supports multiple interaction methods. Users can directly use natural language to make various research requests, such as inquiring about research results on a certain crop gene or querying agricultural product yield data for a specific region.

[0027] In step S120, the intent recognition based on natural language queries is also executed by the plant chatbot. It relies on a pre-set intent recognition algorithm, combined with a database of professional terms and demand features from fields such as agriculture, biology, and medicine, to determine the type of user's demand. These demand types are mainly divided into three categories: first, knowledge-based demands such as acquiring theories, methods, or research progress related to a research topic; second, data-based demands such as acquiring experimental data and statistical data; and third, comprehensive demands requiring the system to integrate multi-source information to form a research review. The identified intent is synchronously transmitted to the plant inference engine (Phyto-Reasoner) in the host module.

[0028] In step S130, the process of decomposing the problem based on intent is led by the plant inference engine. It breaks down the original problem according to the complexity of the intent and clarifies the execution logic of the sub-tasks. For simple knowledge-based or data-based needs, such as inquiring about the application method of a certain technology or single yield data, it will be decomposed into one or two single sub-tasks, such as knowledge retrieval and organization or data query and transformation. For complex needs, such as analyzing the transcriptome data of a certain variety and summarizing gene functions, it will be broken down into multiple sub-tasks such as literature retrieval, data acquisition, data preprocessing, differential gene screening, functional enrichment analysis, and result organization. The execution order of the sub-tasks and the agents that need to collaborate are also marked; for example, data acquisition is handled by the data agent, and differential analysis is completed by the analysis agent.

[0029] In step S140, the step of generating corresponding answers based on different intentions is executed by the corresponding agent of the server module (Phytomni-Bot). For scientific knowledge-related needs, the knowledge agent first converts the user query into semantic embedding, retrieves candidate content from a vector knowledge base containing years of literature, abstracts, and patents through keyword, semantic, and hybrid multi-strategy retrieval, sorts the candidate content by relevance, publication time, and citation volume, filters out important fragments, and then the plant inference engine integrates these fragments into a natural language answer, annotates the knowledge source, and returns it to the user. For scientific data-related needs, the data agent, driven by the plant inference engine, first rewrites the natural language query into a statement adapted to the database structure, then converts it into a standard SQL statement through the NL2SQL module, executes the statement to extract structured data from the scientific database, and finally returns it in the form of a natural language explanation combined with tables or charts. For comprehensive review requests, the review intelligence agent first combines the structured review plan developed by the plant reasoner to clarify the review topic, time frame, main chapters, and information source types. Then, it coordinates the knowledge intelligence agent to retrieve information from multiple sources multiple times. If the information coverage is incomplete in the first retrieval, it will optimize keywords and retrieve again. Subsequently, it integrates and rewrites the obtained text, data, charts, and other information, organizes the content according to the structure of a scientific review, and finally generates a formal report to return.

[0030] Figure 2 This invention relates to a flowchart for generating corresponding responses based on different intentions.

[0031] According to an embodiment of the present invention, the intent identification based on natural language query includes: confirming that the intent is a research knowledge-related need, confirming that the intent is a research data-related need, or confirming that the intent is a comprehensive review need. This process corresponds to... Figure 2 Steps S1400, S1410, S1420 and S1430.

[0032] Specifically, in step S1400, the plant chatbot relies on a preset intent recognition algorithm, combined with a database of professional terms and a database of demand features in the relevant field, to determine the intent of the user's natural language query, clarify whether the demand belongs to the category of scientific knowledge, scientific data, or comprehensive review, and synchronize the recognition results to the plant inference engine.

[0033] In step S1410, the process of confirming that the intent is a scientific research knowledge-related need is achieved through an intent recognition algorithm mounted on the plant chatbot. This algorithm combines a database of professional terms and a database of demand features in fields such as agriculture, biology, and medicine to identify the user's request for information on theories, methods, research progress, or technological applications related to a specific scientific research topic. These needs typically revolve around explanations of specific scientific concepts, technical principles, experimental methods, or summaries of relevant research results. The algorithm determines that the user's request is for systematic knowledge explanations or information descriptions, rather than specific data or integrated reports, by capturing the expressive features related to professional knowledge in the query.

[0034] In step S1420, when the intent is confirmed to be a scientific research data request, the algorithm will focus on identifying features related to specific indicators and values ​​in the query, such as requests for quantitative information involving a specific time range, regional range, or experimental sample. The focus of this type of request is to obtain structured experimental data, statistical data, or observational data. The query often implicitly contains descriptions related to data dimensions. The algorithm combines the field features of the scientific research database with keyword mapping rules to determine whether the user needs to satisfy their needs through data retrieval, extraction, and organization, rather than simply knowledge explanation or multi-source information integration.

[0035] In step S1430, confirming the intent as a comprehensive review request involves identifying the user's desire to systematically integrate, analyze, and summarize information from multiple sources. Such queries typically include descriptive features indicating a comprehensive overview, in-depth analysis, or summary of a specific research topic. The algorithm, combined with the definition of review tasks in the requirement feature database, determines that the user needs to integrate various types of information, such as literature, data, and research conclusions, to form a complete and logically coherent systematic report, rather than a single-dimensional knowledge acquisition or data query. The identification of all three intents relies on a pre-set algorithm model and a domain-specific resource library to ensure the accuracy and relevance of intent judgment, providing a precise basis for subsequent problem decomposition and task execution.

[0036] According to an embodiment of the present invention, the intention-based problem decomposition further includes: decomposing the problem into one or more sub-tasks; and specifying the execution order and collaboration logic between the sub-tasks.

[0037] Specifically, the process of decomposing a problem into one or more subtasks is completed by the Plant Inference Engine using a domain-specific task decomposition model. The decomposition logic is highly adapted to the identified intent type and complexity. For simple scientific knowledge or data-related needs, the Plant Inference Engine directly breaks down the original problem into one or two closely related subtasks, each corresponding to a single functional goal, ensuring a concise and efficient execution path. For comprehensive review or complex analysis needs, the Plant Inference Engine adopts a layered decomposition approach, gradually refining the original problem into multiple independently executable and precisely assignable subtasks. These subtasks cover multiple stages such as information retrieval, data collection, analysis and processing, and result integration, forming a complete task execution chain. Each subtask has clear input and output standards, laying the foundation for subsequent collaboration.

[0038] When defining the execution order of subtasks, the plant inference engine prioritizes analyzing the dependencies between them to construct logical execution paths. For subtasks with sequential dependencies, the main execution chain is determined, and subtasks within the chain must proceed sequentially, with the output of the previous subtask serving as the input for the next, ensuring a coherent flow and accurate results. For subtasks without direct dependencies, priority is assigned based on their impact on the overall goal, with subtasks affecting key objectives executed first. Other subtasks can proceed in parallel or be interspersed at appropriate times, balancing efficiency and effectiveness. This sequential arrangement ensures smooth process flow while fully utilizing system resources and avoiding unnecessary waiting.

[0039] The collaboration logic of subtasks is built on the model context protocol. The plant inference engine precisely assigns each subtask to an agent with the corresponding capabilities based on its functional attributes. Knowledge-related subtasks are handled by the knowledge agent, data query and processing subtasks by the data agent, analysis subtasks by the analysis agent, and comprehensive integration subtasks by the composite agent. During collaboration, agents synchronize their states and share information through a pre-defined communication mechanism. When a subtask is completed, the corresponding agent promptly feeds back the result to the plant inference engine, which then triggers the initiation of the next related subtask or dynamically adjusts the parameters of subsequent subtasks based on the execution results. If an anomaly occurs during the execution of a subtask, the collaboration mechanism supports quickly switching the executing agent or initiating a retry process to ensure that the overall task is not affected by local problems and maintain the stability and continuity of the collaboration chain.

[0040] like Figure 2 As shown, this flowchart constructs three independent and complete answer generation paths centered on intent classification. Each path corresponds to a dedicated execution stage and agent collaboration logic, ensuring that the output results accurately adapt to different types of research needs. For research knowledge-related needs, in step S1411, the process begins by converting the natural language query into semantic embedding. Subsequently, in step S1412, a multi-strategy retrieval is triggered to extract relevant content from the vector knowledge base. The retrieval results are reordered according to predetermined rules based on relevance, and contextual fragments that fit the needs are selected as knowledge support. Finally, these fragments are integrated through natural language and the knowledge sources are labeled to form a clear and well-founded answer. For research data-related needs, in step S1421, the process first rewrites the natural language query to conform to the database's structural requirements. Then, the NL2SQL module converts the rewritten query into a standard statement. In step S1422, the statement is executed to extract structured data from the research database. Finally, the results are presented in the form of natural language interpretation combined with structured carriers, balancing comprehensibility and data integrity. The process for comprehensive review requests begins with developing a structured review plan in step S1431. In step S1432, relevant agents are coordinated to iteratively retrieve multi-source information based on the plan. If the initial retrieval coverage is insufficient, the conditions are optimized for a second retrieval. In step S1433, the collected text, data, charts, and other multimodal information are integrated and rewritten, and the content is organized according to the standardized structure of a research review, ultimately generating a formal report with a standard format and logical coherence. These three paths are independent yet rely on a unified system collaboration framework, jointly achieving efficient and accurate response generation in different research scenarios.

[0041] According to an embodiment of the present invention, confirming that the intent is a scientific research knowledge-related need, and generating corresponding answers based on different intents, further includes: Natural language queries are vectorized into semantic embeddings, and multi-strategy RAG retrieval is performed from a vector knowledge base. The retrieved content is reordered based on relevance, and contextual fragments are selected as knowledge contexts. Targeted answers are generated by combining natural language interpretation with annotations of key knowledge sources.

[0042] Specifically, in the natural language query vectorization stage, the system employs tools adapted for capturing the semantics of scientific texts to convert user-submitted scientific knowledge queries into high-dimensional semantic vectors. These vectors must maintain the same dimension as the text vectors stored in the vector knowledge base to ensure accurate semantic matching in subsequent retrieval processes. This step is also... Figure 2 The initial stage of the knowledge-based answer generation path in Chinese scientific research lays the foundation for subsequent accurate retrieval. The construction of the vector knowledge base relies on literature resources within a specific period, covering academic articles, research abstracts, and related patents from multiple years. These documents need to be preprocessed to remove redundant formatting information, correct expression deviations, and complete structured decomposition. Then, they are converted into semantic vectors using the same vectorization tools and stored to form the basic resource base supporting retrieval.

[0043] The execution of multi-strategy RAG retrieval revolves around the vector knowledge base and strictly follows... Figure 2 The main retrieval logic of this path simultaneously employs three retrieval strategies. Keyword retrieval extracts scientific terminology from natural language queries, such as specific gene names and research method names, and precisely matches them with keywords in the knowledge base text, quickly filtering directly relevant candidate content. Semantic retrieval calculates the cosine similarity between the query semantic vector and the text vector in the knowledge base, capturing content with different expressions but similar semantics, compensating for potential omissions in keyword retrieval. Hybrid retrieval integrates the results of the first two strategies, eliminating duplicate content to form a preliminary candidate set, ensuring both retrieval accuracy and comprehensive coverage.

[0044] The relevance re-ranking process further filters and optimizes the initial candidate set, prioritizing content based on multiple dimensions. First, semantic relevance is the central criterion, prioritizing content highly relevant to the query topic. Second, publication date is considered, prioritizing recent publications to ensure timely information. Citation count is also used to select research findings with high recognition and influence within the field. This comprehensive evaluation across three dimensions forms the ranked candidate content list. Subsequently, when selecting contextual fragments from the ranked list, the focus is on extracting research background, main methods, experimental results, and conclusions directly relevant to the query, while redundant information with low topic relevance is removed, ensuring the selected fragments effectively support the answer.

[0045] Finally, in the answer generation stage, the system integrates the selected context fragments using natural language processing, organizing the scattered information into a clear and logical explanation according to the logic of scientific knowledge explanation, avoiding fragmented presentation. This step also corresponds to... Figure 2 The final output of the scientific research knowledge-based path in China is to form an answer that is both accurate and logical.

[0046] According to an embodiment of the present invention, confirming that the intent is a scientific research data request and generating a corresponding response based on different intents further includes: rewriting the natural language query into a form that conforms to the database structure and generating a standard SQL statement; obtaining structured data from the scientific research database based on the standard SQL statement and returning it in a format that combines natural language interpretation with the structured data carrier.

[0047] Specifically, in Figure 2 In the research data-related request response generation path shown, natural language query rewriting is the first step in the process. This step is completed by a data agent driven by a plant-based inference engine, with the goal of adapting the natural language query to the structural requirements of the research database. The data agent, relying on preset keyword mapping rules and combining them with the field characteristics of the research database, adjusts and optimizes the natural language query submitted by the user. By converting data-related expressions in the query into standardized expressions recognizable by the database, the ambiguity and arbitrariness of natural language are eliminated, ensuring that subsequent data retrieval steps can accurately locate the required information. This step also... Figure 2 This laid the foundation for the generation of subsequent SQL statements.

[0048] The next step is the standard SQL statement generation stage, executed by the NL2SQL module within the data intelligence agent. This module, fine-tuned and optimized with samples specific to the research field, accurately understands the main requirements of the rewritten query. It combines the table structure and field relationships of the research database to convert the rewritten query into a syntactically sound and logically rigorous standard SQL statement. Whether it's a simple single-condition data query or a complex query scenario involving multiple table joins and multi-dimensional filtering, this module can generate an appropriate SQL statement, ensuring that the query instructions can be accurately parsed and executed by the research database. Figure 2 The key transition link between query rewriting and data retrieval in the middle join.

[0049] In the step of retrieving structured data using standard SQL statements, the data intelligence system establishes stable communication with the research database through a pre-defined database connection mechanism. It then submits the generated standard SQL statements to the research database to execute data retrieval operations. After the data is returned, the data intelligence system performs integrity verification on the retrieved structured data, removing data entries with missing key information or abnormal formats to ensure data quality. Simultaneously, it standardizes data from different sources and in different formats, unifying data units, field naming rules, etc., to ensure the consistency and readability of the final output data. This step corresponds to… Figure 2 The logic of data retrieval and organization in China.

[0050] Finally, we move on to the answer return section, which is also... Figure 2 The final output steps of the data-related requirements path for scientific research in China. The system will first provide natural language interpretation of the processed structured data, explaining the main meaning, trends, or key conclusions of the data in easy-to-understand terms, avoiding the difficulty for users to quickly understand the raw data. Simultaneously, to meet the requirements of scientific research scenarios for data accuracy and completeness, the system will present the raw data in conjunction with a structured data carrier. The structured data carrier can be selected in an appropriate form according to the data type, ensuring that the data is presented clearly and intuitively. This allows users to quickly grasp the main information and provides complete data support for subsequent in-depth analysis, resulting in an answer that is both easy to understand and practical.

[0051] According to an embodiment of the present invention, confirming the intent as a comprehensive review request and generating corresponding responses based on different intents further includes: developing a structured review plan; iteratively retrieving multi-source information based on the structured review plan; integrating and rewriting the obtained multimodal information to form a research review, and returning it in the form of a report.

[0052] Specifically, in Figure 2 In the comprehensive review request response generation path shown, formulating a structured review plan is the starting point of the process. This step is completed by the review agent driven by the plant inference engine. As a composite agent, the review agent first deeply analyzes the user's review requirements, clarifies the main theme and research boundaries of the review, and avoids content deviating from the direction. Subsequently, combined with the planning capabilities of the plant inference engine, it determines the time frame, key chapter framework, information coverage dimensions, and information source types of the review. The chapter framework follows the standard logic of scientific reviews, planning a content structure that includes research background explanation, important methodological overview, experimental results analysis, and conclusion summary. The information source types clearly cover multiple resources such as academic literature, patent technical documents, and public datasets, ensuring that the plan has a clear execution direction and providing a clear basis for subsequent iterative retrieval and information integration. This step also corresponds to... Figure 2 The starting point for planning a comprehensive path.

[0053] The next step involves iterative retrieval of multi-source information based on a structured review plan. This step is coordinated by the review agent and the knowledge agent, and strictly follows the... Figure 2 The retrieval logic for review articles is as follows: The review intelligence breaks down the structured review plan into specific retrieval instructions, specifying the keywords, information types, and time range for each round of retrieval, and then transmits these instructions to the knowledge intelligence agent. The knowledge intelligence agent then employs a multi-strategy RAG retrieval method, extracting relevant information from a vector knowledge base and multi-source scientific research resource databases. After the first round of retrieval, the review intelligence agent evaluates the coverage and relevance of the retrieval results. If it finds insufficient information or missing key research findings for a particular chapter, it optimizes the retrieval keywords or expands the range of information sources, initiating the next round of retrieval until the acquired information fully supports the content requirements of each chapter of the review, ensuring the completeness and relevance of multimodal information.

[0054] After the retrieval is completed, the process moves to the multimodal information integration and rewriting stage, a crucial step in forming a high-quality research review. The review intelligence first categorizes and organizes the acquired text, data, charts, and other multimodal information, eliminating duplicate and redundant content, while verifying the accuracy and consistency of the information, and annotating and reasonably analyzing contradictory research conclusions. Then, according to a pre-set chapter framework, the scattered information is logically reorganized: textual information is organized around the natural logic of the research content, first explaining the research background, then outlining the main methods, followed by analyzing the experimental results, and finally summarizing the research conclusions, forming coherent paragraphs; data information is transformed into standardized statistical expressions or chart descriptions; and chart information is formatted uniformly. During this process, the review intelligence relies on the logical reasoning capabilities of the plant-based inference engine to ensure natural transitions between chapters and a progressive overall content, forming a logically coherent and detailed initial draft of the review. This step also aligns with… Figure 2 The main process of information integration in China.

[0055] Finally, we move on to the research review generation and return stage. Figure 2 The final output steps of the comprehensive review request process are as follows: The review's intelligent system optimizes and polishes the initial draft, standardizing academic expression, unifying reference formats, and supplementing information source annotations and retrieval time descriptions to ensure the review meets the requirements of a scientific research report. The optimized content is then generated into a report. Finally, the generated research review report is delivered to the user via a plant-based chatbot, completing the closed loop of responding to and generating a comprehensive review request.

[0056] According to an embodiment of the present invention, a vector knowledge base is constructed on specific documents during a specific period; the specific documents are preprocessed and vectorized to construct a vector knowledge base.

[0057] Specifically, the specific period on which the vector knowledge base is built is clearly defined as 1900 to 2024. This time span covers more than a hundred years of scientific research achievements in the fields of agriculture and plant science. It includes relevant literature on classic theories and traditional research methods in the field, as well as the latest achievements of modern cutting-edge technologies and innovative research, ensuring that the knowledge base can provide information support for scientific research needs that has both historical depth and contemporary relevance.

[0058] The specific literature focuses on the fields of agriculture and plant science, forming a large and diverse collection, including 4 million relevant academic articles, 27 million research abstracts, and 200,000 related patents. The academic articles encompass bilingual publications in Chinese and English, including major journal articles and conference papers in sub-fields such as crop breeding, plant protection, soil and fertilizer, and agricultural ecology, as well as specialized research reports published by universities and research institutions. The research abstracts are derived from academic journals, dissertations, and research project completion reports in relevant fields worldwide, condensing the main viewpoints and key findings of various studies. The patent literature covers applied technological achievements such as crop variety breeding, agricultural technology inventions, and agricultural equipment innovation, comprehensively covering multiple information dimensions such as theory, methods, and technological applications that may be involved in scientific research knowledge needs. All literature has undergone rigorous screening, retaining only content highly relevant to agriculture and plant science, and eliminating literature with low relevance, incomplete data, or duplication, ensuring the high quality and high relevance of the literature collection.

[0059] Before being incorporated into the vector knowledge base, specific documents undergo a multi-step preprocessing process. First, format parsing is performed. For different types of documents, such as academic articles, abstracts, and patents, corresponding parsing tools are used to extract the text content. PDF documents require specialized parsing technology to strip the text, ensuring complete extraction of the original information. Next, redundancy is removed, eliminating non-critical information such as headers, footers, figure and table annotations, formula symbols, and irrelevant formatting codes to avoid interfering with subsequent processing. Then, text cleaning is carried out, correcting typos and expression discrepancies, standardizing punctuation and terminology in both Chinese and English, and removing duplicate content to ensure accuracy and uniqueness. Finally, structured decomposition is performed, breaking down the complete document into independent segments such as title, abstract, research background, experimental methods, results analysis, and conclusions, according to the natural logic of scientific literature. Each segment has a clear information focus, laying the foundation for subsequent vectorization processing and precise retrieval.

[0060] After preprocessing, the system proceeds to the vectorization stage. Using the same tool as for natural language query vectorization, the system ensures consistent vector dimensions, guaranteeing the accuracy of subsequent semantic matching. This vectorization tool accurately captures the specialized semantic features of research texts in agriculture and plant science, independently converting each fragmented document into a high-dimensional semantic vector. It fully preserves the main information and logical connections contained within the fragment, ensuring the vector accurately reflects the text's research content. The vectorized semantic vectors are stored in a dedicated vector database. Simultaneously, a mapping is established between the semantic vectors and corresponding document fragments, as well as original source information including document title, publication date, author, journal name, and patent publication number, forming a clearly structured and highly efficient vector knowledge base. This knowledge base directly supports multi-strategy RAG retrieval by knowledge agents, ensuring the retrieval process can quickly locate document content related to the query semantics, providing stable and reliable underlying data support for accurate responses to research knowledge-related needs.

[0061] This invention also provides a research task autonomous planning and execution system, with a host module and a server module built based on a model context protocol. The host module includes: a plant chatbot for accepting natural language queries in interactive dialogue and for identifying intents based on the natural language queries; and a plant inference engine for decomposing questions based on intents and, when the intent is a comprehensive review requirement, formulating a structured review plan. The server module includes: a knowledge agent for performing multi-strategy RAG detection; a data agent for rewriting natural language queries into a database-structured format under the drive of the plant inference engine, generating standard SQL statements, and obtaining structured data based on the standard SQL statements; a composite agent for combining the knowledge agent, data agent, and analysis agent into an agent with a specific application; and a review agent, a composite agent for formulating a structured review plan, coordinating the knowledge agent to iteratively retrieve multi-source information based on the structured review plan, integrating and rewriting the obtained multimodal information to form a research review, and returning it in report form. The host module decomposes the problem into one or more sub-tasks, clarifying the execution order and collaboration logic between the sub-tasks.

[0062] Specifically, Figure 3 This is a block diagram of the autonomous planning and execution system for scientific research tasks involved in this invention. Figure 3 As shown, the autonomous planning and execution system for scientific research tasks 100 adopts a modular architecture, consisting of a host module 110 and a server module 120 built on the model context protocol. The two modules achieve real-time synchronization and smooth interaction of information through this protocol, ensuring that the logical connection between each link and the data transmission are coherent and accurate when the entire system processes scientific research tasks, and jointly supporting the autonomous operation of the entire process from receiving requirements to outputting results.

[0063] The host module 110, as a key unit for system-user interaction and task planning, contains two important components: the plant chatbot 111 and the plant inferencer 112. The plant chatbot 111 handles user interaction, accepting natural language queries from users in the form of interactive dialogue, efficiently handling both text input and speech-to-text requests. Simultaneously, it is responsible for identifying intent based on natural language queries. Relying on a pre-set intent recognition algorithm, combined with a database of professional terms and demand features in agriculture, biology, and medicine, it accurately determines whether the user's request is for scientific knowledge, scientific data, or a comprehensive review, and then transmits the recognition result to the plant inferencer 112. The plant inference engine 112 focuses on task decomposition and planning. Based on the intent conveyed by the plant chatbot 111, it decomposes the original problem into one or more independently executable sub-tasks, and clarifies the execution order and collaboration logic between the sub-tasks. For example, the execution order is such that sub-tasks with dependencies should be executed sequentially, while sub-tasks without direct dependencies can be executed in parallel according to priority. When the intent is a comprehensive review requirement, the plant inference engine 112 will also formulate an additional structured review plan, clarifying the topic boundaries, time range, chapter framework, and information source types of the review, providing clear guidance for the subsequent operations of the server module 120.

[0064] Server module 120, as the main unit for system function execution, integrates multiple specialized intelligent agents, each corresponding to the processing of different types of scientific research needs. Internally, it includes a knowledge agent 121, a data agent 122, a composite agent 123, and a review agent 124. The knowledge agent 121 is primarily responsible for responding to scientific knowledge-related needs, possessing multi-strategy RAG retrieval capabilities. It can retrieve relevant content from a vector knowledge base, providing information support for subsequent knowledge-based responses. The data agent 122 addresses scientific data-related needs. Driven by the plant inference engine 112, it converts natural language queries into a database-structured format, rewrites them, generates standard SQL statements, and then retrieves structured data from the scientific research database based on these standard SQL statements, ensuring the accuracy and completeness of the data acquisition.

[0065] The function of the composite intelligent agent 123 is to combine different intelligent agents. Based on the needs of specific application scenarios, it combines intelligent agents with single functions, such as the knowledge intelligent agent 121 and the data intelligent agent 122, into an integrated unit with specific application capabilities to address scientific research tasks requiring collaborative processing of multiple types of information. The review intelligent agent 124, as a composite intelligent agent, specifically handles comprehensive review needs. Based on the structured review plan formulated by the plant inference engine 112, it coordinates the knowledge intelligent agent 121 to iteratively retrieve multi-source information, integrates and rewrites the acquired text, data, charts, and other multimodal information, forming a research review that conforms to scientific research standards, and returns it to the user in the form of a report.

[0066] Through the collaborative cooperation of the host module 110 and the server module 120, as well as the clear division of labor and efficient linkage among the components, the entire system achieves full-process autonomy of scientific research tasks from demand identification, planning and decomposition to execution output. It can accurately adapt to different types of scientific research needs and provide strong support for scientific research in agriculture, biology, medicine and other fields.

[0067] According to an embodiment of the present invention, the construction and training of the plant chatbot and the plant inferencer include a basic model training stage, an incremental pre-training stage, and a supervised fine-tuning stage. In the basic model training stage, the basic dataset comes from multiple sources, focusing on data containing scientific research tasks. In the incremental pre-training stage, incremental pre-training is performed on a preset number of Pangu basic models for specific knowledge in a specific domain. In the supervised fine-tuning stage, the plant chatbot and the plant inferencer are subjected to supervised fine-tuning to make their capabilities consistent with specific scientific research tasks and user interaction patterns.

[0068] Figure 4 This invention relates to a hierarchical flowchart of the model construction and training of a plant chatbot and a plant inference engine. (See attached diagram.) Figure 4 As shown, the flowchart presents the complete training chain in a hierarchical and progressive manner. From top to bottom, it consists of three stages: basic model training, incremental pre-training, and supervised fine-tuning. Each stage is closely connected, from building general capabilities to deepening domain knowledge, and then to precise adaptation to scenarios, gradually realizing the model's adaptation and support for agricultural scientific research tasks.

[0069] The basic model training phase focuses on the collection of multi-source scientific research data and the cultivation of basic capabilities. The data comes from major Chinese and English journal articles in the field of agricultural research, relevant patent documents from the State Intellectual Property Office, project completion reports from research institutions, and publicly available scientific research task execution records. The focus is on selecting content related to scientific research tasks. After preprocessing to remove redundant information, correct expression biases, convert non-text data, and classify it according to the type of scientific research task, the model undergoes multiple rounds of iterative training to master the ability of natural language understanding, basic logical deduction, and simple task decomposition, and initially establishes a cognitive framework for agricultural scientific research scenarios.

[0070] The incremental pre-training phase uses a pre-set number of Pangu basic models as carriers, specifically selecting the Pangu Chatting 38 billion parameter model (version 3.1.35.B010-128K) and the Pangu Inference 38 billion parameter model (version 3.0.1.1-128K). These two types of models are adapted to the interactive capabilities of the plant chatbot and the logical planning capabilities of the plant inference engine, respectively. At the same time, exclusive data from specific agricultural research areas are collected, covering professional literature, technical standards, field trial reports, and specialized terminology databases in fields such as crop genetics and breeding, plant disease and pest control, soil fertility assessment, and agricultural ecosystem surveys, to construct a systematic domain knowledge dataset. The training process adheres to the principle of preserving the basic capabilities of the model while supplementing domain knowledge, enabling the model to accurately identify specialized terms in the agricultural research field, understand the professional logical connections in research tasks, and complete the transformation from a general model to a domain-adapted model.

[0071] In the supervised fine-tuning phase, a labeled dataset was constructed around specific research tasks and user interaction patterns. The research task samples included labeled question-answer pairs covering scientific knowledge, scientific data, and comprehensive reviews. The user interaction samples included historical dialogue logs and contextual understanding cases. All samples were labeled with intent type, task decomposition logic, and optimal output results. During fine-tuning, the dataset was input into the incrementally pre-trained model. The key parameters of the model were adjusted using a small-batch iterative update method to improve the accuracy of intent recognition, the fluency of dialogue context understanding, the rationality of question decomposition in the plant inference engine, and the completeness of the review plan formulation. Subsequently, the model's performance was verified using a test set specifically for agricultural research and simulated user interactions. If issues such as intent misjudgment or task decomposition omissions were found, the model was returned to the fine-tuning dataset construction phase to supplement targeted samples and fine-tune the parameters again until the model's capabilities fully matched the specific research tasks and user interaction patterns in the agricultural field.

[0072] According to an embodiment of the present invention, the server module further includes an analytical agent that adopts a PDCA framework. The analytical agent includes a memory module and a toolkit integration unit. The PDCA framework operates in a closed loop under the support of the memory module and the toolkit integration unit. The memory module is used to store process data generated during the analysis process. In the inspection phase of the PDCA framework, it identifies the deviation between the analysis results and the expected goals. In the action phase, it generates a correction plan based on the deviation. The toolkit integration unit is used to integrate the omics analysis toolkits required for the analysis. Through the interface with the data agent, it obtains the raw data required for the analysis from the biological database. In the execution phase of the PDCA framework, it calls the omics analysis toolkit to execute the generated analysis code. In the planning phase of the PDCA framework, the analytical agent retrieves literature related to the requirements and summarizes the analysis methods through the knowledge agent. Under the guidance of the AI ​​model, it plans the task and generates analysis code for execution, wherein the analysis code is adapted to the syntax rules of the omics analysis toolkit.

[0073] like Figure 3 As shown, the server module also includes an analysis agent 125, which operates using the PDCA framework. It integrates a memory module and a toolkit integration unit. These two units work together to provide stable support for the closed-loop operation of the PDCA framework, ensuring that bioinformatics analysis tasks can be automated and precise.

[0074] In the planning phase of the PDCA framework, the analytical agent 125 first uses the knowledge agent 121 to retrieve relevant literature related to the current needs, extracting and summarizing suitable analytical methods from the literature to clarify the technical path for task execution. Subsequently, guided by the AI ​​model, and combining input information such as task objectives, data descriptions, and data paths, it plans the analytical task in detail, breaks down specific execution steps, and generates corresponding analytical code. The generated analytical code fully adapts to the syntax rules of the omics analysis toolkit in the toolkit integration unit, laying the foundation for smooth code execution. The preferred AI model is Deepseek. Deepseek was chosen as the preferred AI model because its pre-trained data covers knowledge in agricultural research and omics analysis, accurately understands professional terminology and analytical logic, and can combine task objectives, data descriptions, and data paths to complete analytical task planning and step decomposition, generating code that fully adapts to the syntax rules of the omics analysis toolkit in the toolkit integration unit, thus meeting the closed-loop operation requirements of the analytical agent's PDCA framework.

[0075] Once the execution phase begins, the toolkit integration unit plays a crucial role. This unit pre-integrates various omics analysis toolkits required for the analysis. Through an interface with the data agent 122, it accurately retrieves the raw data needed for the analysis from biological databases, ensuring the professionalism and completeness of the data source. Subsequently, following the analysis code generated in the planning phase, the corresponding omics analysis toolkits are invoked, automatically executing the analysis process without manual intervention in tool deployment, achieving seamless integration and efficient progress of the analysis process.

[0076] During the inspection phase, the memory module stores all types of process data generated during the analysis, including data processing details, code execution logs, and intermediate analysis results. Based on this stored information, the analysis agent 125 comprehensively verifies the final analysis results against the preset expected goals, accurately identifies the deviations between the results and the goals, clarifies the stages in which the deviations occurred and their possible causes, and provides a clear basis for subsequent corrective work.

[0077] In the action phase, the analytical agent 125 generates targeted correction plans based on the deviations identified in the inspection phase. These correction plans clearly define the direction of adjustment, such as optimizing analysis code parameters, supplementing specific raw data, and adjusting toolkit calling logic. Once generated, the plan is fed back to the planning phase, initiating a new round of task planning and execution. Through continuous iterative optimization, the deviations are gradually reduced until the analysis results fully meet the expected goals, forming a complete closed-loop working mode. This dynamic correction mechanism not only solves various problems in the automated analysis process in real time but also accumulates processing experience through a memory module, continuously improving the efficiency and accuracy of subsequent similar research tasks. It effectively adapts to the dynamic needs of complex scenarios such as omics analysis, providing researchers with stable and reliable bioinformatics analysis support.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for autonomous planning and execution of scientific research tasks, characterized in that, Includes the following steps: Accepts natural language queries in the form of interactive dialogue; Based on the natural language query, the intent is identified; Based on the aforementioned intent decomposition problem; The system generates corresponding answers based on different intentions.

2. The method according to claim 1, characterized in that, The intent identified based on the natural language query includes: The intent is confirmed to be a need for scientific research knowledge, a need for scientific research data, or a need for a comprehensive review.

3. The method according to claim 1, characterized in that, The intent-based decomposition problem further includes: Break down the problem into one or more sub-tasks; Define the execution order and collaboration logic between the subtasks.

4. The method according to claim 2, characterized in that, Confirming that the intent is a research knowledge-related need, the step of generating corresponding answers based on different intents further includes: The natural language query is vectorized into a semantic embedding, and multi-strategy RAG retrieval is performed from the vector knowledge base; The retrieved content is reordered based on relevance, and contextual fragments are selected as knowledge contexts. Targeted answers are then generated by combining natural language interpretation with annotations of key knowledge sources.

5. The method according to claim 2, characterized in that, Confirming that the intent is a research data-related request, the step of generating corresponding responses based on different intents further includes: The natural language query is converted into a form that conforms to the database structure and rewritten to generate standard SQL statements; Structured data is retrieved from scientific research databases based on the standard SQL statements and returned in a format that combines natural language interpretation with the structured data carrier.

6. The method according to claim 2, characterized in that, Confirming that the intent is a comprehensive overview request, the step of generating corresponding responses based on different intents further includes: Develop a structured review plan; Based on the aforementioned structured review plan, multi-source information is retrieved iteratively; The obtained multimodal information is integrated and rewritten to form a research review, which is then returned in the form of a report.

7. The method according to claim 4, characterized in that, The vector knowledge base is built upon specific documents from a specific period; The specific documents are preprocessed and vectorized to construct a vector knowledge base.

8. A research task autonomous planning and execution system, characterized in that, It is constructed based on the Model Context Protocol by a host module and a server module; wherein, the host module includes: A plant-based chatbot, configured to accept natural language queries in the form of interactive dialogue and to identify intent based on the natural language queries; and A plant-based inference engine is used to decompose the problem based on the intent and to develop a structured review plan when the intent is a comprehensive review requirement. The server module includes: A knowledge-based intelligent agent is used for multi-strategy RAG detection. A data intelligence agent, driven by the plant inference engine, is used to rewrite the natural language query into a form that conforms to the database structure, generate standard SQL statements, and obtain structured data based on the standard SQL statements. A composite intelligent agent is used to combine the knowledge intelligent agent, the data intelligent agent, and the analysis intelligent agent into an intelligent agent with a specific application. A review agent is a composite agent used to formulate a structured review plan, coordinate the knowledge agent to iteratively retrieve multi-source information based on the structured review plan, integrate and rewrite the obtained multimodal information to form a research review, and return it in the form of a report. The host module is used to decompose the problem into one or more sub-tasks and to define the execution order and collaboration logic between the sub-tasks.

9. The system according to claim 8, characterized in that, The construction and training of the plant chatbot and the plant inferencer include basic model training, incremental pre-training, and supervised fine-tuning stages. During the training phase of the basic model, the basic dataset comes from multiple sources, with a focus on data that includes research tasks. In the incremental pre-training phase, incremental pre-training is performed on a preset number of Pangu base models for specific knowledge in a specific domain. During the supervised fine-tuning phase, the plant chatbot and the plant inference engine are subjected to supervised fine-tuning to align their capabilities with specific research tasks and user interaction patterns.

10. The system according to claim 8, characterized in that, The server module also includes the analytical agent, which adopts the PDCA framework; The analytical agent includes a memory module and a toolkit integration unit, and the PDCA framework operates in a closed loop with the support of the memory module and the toolkit integration unit. The memory module is used to store process data generated during the analysis. In the inspection phase of the PDCA framework, deviations between the analysis results and the expected goals are identified. In the action phase, a correction plan is generated based on the deviations. The toolkit integration unit is used to integrate the omics analysis toolkit required for analysis. Through the interface linkage with the data agent, it obtains the raw data required for analysis from the biological database. In the execution phase of the PDCA framework, it calls the omics analysis toolkit to execute the generated analysis code. In the planning phase of the PDCA framework, the analytical agent retrieves relevant literature and summarizes analytical methods through the knowledge agent; under the guidance of the AI ​​model, it plans tasks and generates analytical code for execution, wherein the analytical code is adapted to the syntax rules of the omics analysis toolkit.

Images