Multi-agent system for astronomical field academic research and configuration method thereof

Abstract

The invention discloses a multi-agent system for astronomical field academic research and a configuration method thereof, and belongs to the technical field of artificial intelligence. The system comprises: an astronomical academic thinking agent layer which is used as a system brain, is driven by a large language model with a thinking chain, and is responsible for planning and reflecting scientific research tasks; the astronomical academic execution agent layer serves as four limbs of the system and completes tool calling through a lightweight model; the astronomy MCP tool library is used for providing tool services such as network search and literature retrieval through a standardized MCP protocol; and the astronomical discussion platform provides an interactive interface between the multiple agents and the user and a discussion mechanism based on mention. According to the invention, by constructing a hierarchical collaborative architecture and docking with a standardized protocol, the defects of an existing intelligent agent system in the aspects of scientific research cognition depth, task collaborative efficiency and group intelligence excitation are overcome, and the crossing from passive information processing to active scientific research discovery is realized.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and multi-agent system technology, specifically relating to a multi-agent system for academic research in the field of astronomy and its configuration method. Background Technology

[0002] With the rapid development of artificial intelligence technology, the application of large language models in academic research assistance is becoming increasingly profound. Existing technologies can already realize functions such as intelligent summarization, literature review, and information retrieval, and some intelligent agent systems designed for in-depth research have emerged. These systems help researchers efficiently process existing knowledge and improve the efficiency of scientific research information acquisition through enhanced retrieval and generation technologies.

[0003] However, the core of academic research in astronomy lies in "exploring the unknown" and "scientific discovery," a process that goes far beyond information retrieval and induction. It heavily relies on in-depth theoretical speculation, interdisciplinary perspectives, rigorous hypothesis formulation and verification, and dynamic, iterative collaborative discussions. Current technological solutions have the following key shortcomings in supporting this complex process: 1. Functional simplification and fragmented processes: Existing systems are mostly single-function "tool-type" intelligent agents, such as independent paper retrieval or data visualization tools. These tools lack effective coordination mechanisms and cannot form an organic, holistic research process. Researchers need to manually switch and integrate information between different tools, resulting in a fragmented research process that is difficult to simulate the coherent, goal-oriented research behavior of human scientists.

[0004] 2. Lack of Deep Thinking and Proactive Planning Capabilities: Most current research-assisted intelligent agents are essentially passive question-and-answer systems, their operation relying on direct responses to user questions. They generally lack the ability for autonomous task planning, reflection, and complex reasoning chains. In astronomical research, the complete "thinking chain" from observing phenomena to proposing theoretical hypotheses, designing verification schemes, interpreting data, and revising models is the core of scientific discovery, and existing systems struggle to autonomously achieve this high-order cognitive process.

[0005] 3. Failure to simulate a real scientific research collaboration ecosystem: Major astronomical discoveries often stem from teamwork and intellectual exchange among experts from different fields (such as theoretical physicists, observational astronomers, and computational experts). Current technologies lack support for "multi-role intelligent agents" and their structured, sequential interaction mechanisms, making it impossible to construct a dynamic discussion environment that can simulate the debate, complementarity, and integration of different academic viewpoints. Consequently, it is difficult for collective intelligence to emerge at the system level that transcends the capabilities of a single intelligent agent.

[0006] 4. Insufficient flexibility and standardization in tool integration and invocation: Although some systems support tool invocation, their tool management often lacks a unified protocol and standardized interface, leading to difficulties in tool expansion, poor compatibility, and a high risk of errors during the invocation process. For the multimodal and specialized toolsets required for astronomical research, existing technologies lack a stable and scalable tool integration framework to provide reliable support.

[0007] Addressing the following key challenges in using intelligent agent systems to support academic research in the field of astronomy: 1. The problem of insufficient depth of scientific research cognition: Existing systems are mostly limited to passive response and shallow information processing, lacking autonomous cognitive capabilities throughout the entire scientific research cycle. This invention introduces a dedicated "thinking agent" with a strong thought chain to simulate the complete cognitive process of human scientists from problem formulation, hypothesis generation, scheme planning to result reflection. It aims to solve the fundamental limitation of agents in scientific research activities that "can only execute, not think," and enhance their deep reasoning and autonomous scientific planning capabilities.

[0008] 2. The problem of low collaborative efficiency in complex scientific research tasks: In astronomical research, multiple tasks such as theoretical derivation, data analysis and literature review are highly coupled, while existing systems have fragmented functions and difficulty in collaboration. This invention constructs a hierarchical collaborative architecture of "thinking-execution" and uses standardized interface protocols for efficient connection. It aims to achieve close linkage between planning and execution in complex tasks, break down the collaboration barriers between intelligent agents, and form an efficient and automated scientific research task pipeline.

[0009] 3. Problems with poor standardization of professional tool integration and invocation: To address the challenges of diverse professional tools required for astronomical research, the chaotic integration methods of existing system tools, and the lack of unified standards leading to poor scalability and unreliable invocation, this invention constructs a standardized "Astronomical Academic MCP Tool Library" based on model context protocols. This library modularizes and unifies the management of tools through protocols, aiming to solve the problems of difficulty in integrating tool resources and insufficient invocation stability, and to provide a solid and reliable tool support foundation for scientific research activities.

[0010] 4. The problem of difficulty in stimulating collective wisdom in multi-agent systems: In view of the limitations of individual agents' perspectives and the disorder and divergence in multi-agent interactions, this invention establishes an "astronomical academic discussion mechanism" based on targeted @mentions and round-by-round control. This mechanism aims to build an orderly and focused collaborative environment, simulating an academic seminar, to solve the problems of difficulty in in-depth discussion and ineffective convergence of collective wisdom in multi-agent systems, thereby fostering innovative scientific insights. Summary of the Invention

[0011] The purpose of this invention is to overcome the above-mentioned shortcomings of the prior art and provide a multi-agent system and its configuration method for academic research in the field of astronomy.

[0012] The objective of this invention is achieved through the following technical solution: a multi-agent system configuration method for academic research in the field of astronomy, which realizes an astronomical academic thinking agent, an astronomical academic execution agent, an astronomical academic MCP tool library, and an astronomical academic discussion mechanism, including the following steps: (1) Realize an intelligent agent for astronomical academic thinking: Configure a large language model with thinking chain capabilities for task planning and reflection, and generate tool call requests; (2) Implement an astronomical academic execution agent: Configure a lightweight large language model to receive tool call requests from the thinking agent and interact with the MCP tool library to complete the tool call requests and return the results; (3) Construct an astronomical academic MCP tool library: Establish a tool set including web search, web crawling, data analysis, and paper retrieval, and provide services to the execution agent through the MCP protocol; (4) Establish an astronomical academic discussion platform and mechanism: including determining the order of speaking; users setting the number of times to speak in each round; and also including a memory management module to track and compress the discussion history to avoid context overflow.

[0013] Furthermore, in step (1), the intelligent agent for realizing astronomical academic thinking includes: (1.1) Role definition module: Based on the secondary development of the Langgraph framework, several thinking agent roles with different astronomical scientific research expertise are defined. Each thinking agent reflects its professional field differences through exclusive system prompts and knowledge bases. (1.2) Model configuration module: The core is a large language model that supports the thinking chain and is compatible with the OpenAI API, which can be deployed privately or accessed through the official API; the differentiated capabilities of the thinking agent are achieved by configuring system prompts that include its role definition and collaboration specifications. (1.3) Knowledge base configuration module: Configure a dedicated knowledge base based on RAG for each thinking agent; the knowledge base includes academic papers, professional journals, textbooks and patent materials, and performs vectorization processing and stores them in a vector database for the agent to retrieve and call; (1.4) Inference and Response Module: Its workflow follows the reAct agent model, specifically including: Upon receiving a question, the first step for the thinking agent is to determine the information needed to generate an answer. Then, it initiates a tool call request to the executing agent or performs a knowledge base search to obtain information; The judgment and information acquisition steps are executed repeatedly until the thinking agent determines on its own that it can generate a complete answer. It also includes setting a maximum number of loop iterations, which forces the current result to be output when this number of iterations is exceeded, in order to avoid infinite loops.

[0014] Furthermore, the thinking agent includes: Theoretical astrophysicist, focusing on theoretical models and mathematical derivations; Observational astronomers focus on observational data and empirical research. Computational astrophysicist, specializing in simulation and computational methods; Cosmology expert, specializing in large-scale cosmological structure and evolution; Astrobiologists who focus on the possibility of extraterrestrial life.

[0015] Furthermore, in step (2), the astronomical academic execution intelligent agent includes: (2.1) Model and Functional Modules: Driven by a lightweight, instruction-based large language model; (2.2) First docking module: The thinking agent and the execution agent dock through the A2A protocol; wherein, each execution agent declares its name, list of executable tools and related calling parameters through a predefined agent card; (2.3) Second docking module: The execution agent and the MCP tool library dock through the MCP protocol; wherein, each execution agent is limited to calling the tool subset corresponding to its function, including the document retrieval tool set, the web search tool set or the knowledge base tool set.

[0016] Furthermore, the executing intelligent agents are divided into four categories based on their functions: Data analytics agents are used to perform data statistics, analysis, and visualization tasks; A document retrieval agent used to retrieve academic documents; A web retrieval agent is used to obtain the latest scientific research information through a search engine; The knowledge base retrieval agent is used to retrieve private information uploaded by users.

[0017] Furthermore, the astronomical academic MCP toolkit includes: A set of literature retrieval tools for retrieving scientific research literature and results, including preprint platforms, academic databases and intellectual property databases. A suite of web search tools for retrieving authoritative news and journal information in the field of scientific research; A knowledge base toolset for mixed retrieval of user-owned documents, materials, and data; A set of data analysis tools to provide a professional data processing environment.

[0018] Furthermore, in step (4), the establishment of an astronomical academic discussion platform and mechanism includes: (4.1) Interactive interface module: Based on the Grado open source framework and the basic interface of Langgraph is developed in a secondary way, it provides functions such as chat interface, document upload, monitoring of the state of thinking agent, display of available tools list and setting the number of times to speak; (4.2) Discussion control module: used to receive the number of times to speak and the discussion topic set by the user, and control the discussion process, including randomly initializing the thinking agent to speak, determining the next speaker based on the content of the current speaking thinking agent and switching the speaking right based on the mention mechanism, and pausing the discussion when the speaking limit is reached. (4.3) Memory Management Module: Monitors the number of tokens in the discussion context in real time, and automatically triggers the history summary process when its occupancy rate reaches a preset threshold, compressing long discussion history into a summary, and forming a new context together with the most recent rounds of dialogue.

[0019] This invention also provides a multi-agent system for academic research in the field of astronomy, comprising: Astronomical Academic Thinking Agent: Equipped with a powerful thinking language model with thought chain capabilities, it is used for task planning and reflection, and to generate tool call requests; Astronomical academic execution agent: Configured with a lightweight large language model to receive tool call requests from the thinking agent and interact with the MCP tool library to complete the tool call requests and return the results; Astronomical Academic MCP Toolkit: This toolkit includes a collection of tools for web search, web scraping, data analysis, and paper retrieval, and provides services to the executing intelligent agent through the MCP protocol. The astronomical academic discussion platform and mechanism include: determining the order of speaking; allowing users to set the number of times they can speak in each round; and also includes a memory management module to track and compress the discussion history to avoid context overflow.

[0020] Furthermore, the thinking agent includes: Theoretical astrophysicists, used for theoretical models and mathematical derivations; Observational astronomers, used for observational data and empirical research; Computational astrophysicists, using simulation and computational methods; Cosmology experts specializing in large-scale cosmological structure and evolution; Astrobiologists use it to study the possibility of extraterrestrial life.

[0021] Furthermore, the executing intelligent agent includes: Data analysis agents can directly perform data statistics, analysis, and visualization tasks within the Jupyter sandbox environment; A literature retrieval agent used to retrieve academic literature from ArXiv, Wanfang, and the official websites of Science / Nature journals; A web retrieval agent is used to obtain the latest scientific research information through a search engine; The knowledge base retrieval agent is used to retrieve private information uploaded by users.

[0022] In summary, this invention aims to achieve a leap from passive information processing to proactive scientific discovery by constructing a multi-agent system with clear division of labor, high collaborative efficiency, and the ability to think deeply and engage in dynamic discussion. Specifically, the technical objectives achieved by this invention include: Achieving deep scientific research cognition: By introducing a dedicated thinking agent with strong thinking chain capabilities, the agent simulates the complete cognitive chain of human scientists from problem formulation, hypothesis generation, solution planning to result reflection, enabling the agent to have autonomous task planning, reflection and complex reasoning capabilities.

[0023] Building an efficient task collaboration system: By establishing a hierarchical "thinking-execution" architecture and a standardized docking mechanism, we can achieve efficient decoupling and collaboration between task planning and specific execution in complex astronomical research, forming an automated and integrated scientific research task workflow.

[0024] Provide standardized tool support: By building an astronomical academic MCP tool library based on a unified protocol, we can achieve standardized integration and efficient scheduling of special tools in the astronomical field, and provide a stable and scalable professional tool calling environment for intelligent agents.

[0025] Stimulating collective intelligence: By establishing a multi-agent discussion mechanism with directional speaking and turn control functions, an orderly and in-depth discussion environment can be constructed to promote the cross-integration of different professional perspectives, thereby stimulating collective intelligence and generating innovative scientific insights.

[0026] Compared with the prior art, the present invention has the following significant advantages: It achieves a deeper level of scientific understanding: by introducing a thinking agent with a strong thought chain, the system can perform multi-step reasoning, proactive planning, and self-reflection, elevating the agent's capabilities from the information processing level to the scientific reasoning level, thus solving the key problem of "only being able to execute, not to think".

[0027] It improves the automation and collaboration efficiency of complex task execution: Through the separation and collaboration of the thinking layer and the execution layer, as well as the standardized A2A docking protocol, the system can automatically decompose and complete the entire process of tasks from theoretical planning to data acquisition and analysis, forming an end-to-end automated scientific research workflow and overcoming the problems of task fragmentation and inefficient collaboration.

[0028] It ensures the standardization and reliability of tool calls: Through a unified tool library built on the MCP protocol, it realizes the standardized integration and management of diverse tools in the astronomical field, greatly improves the accuracy of tool calls and the scalability of the system, and provides a solid technical foundation for scientific research activities.

[0029] This fosters the collective intelligence of multi-agent systems: by introducing a controllable discussion mechanism, agents from different disciplines can engage in orderly and in-depth debates and supplements around the same scientific problem, thereby generating innovative solutions that surpass the capabilities of individual agents, truly simulating the collaborative model of human scientific research teams. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is an overall architecture diagram of a multi-agent system provided in one embodiment of the present invention.

[0032] Figure 2 This is a configuration and workflow diagram of a thinking agent provided in one embodiment of the present invention.

[0033] Figure 3 This is a schematic diagram illustrating the invocation of the execution intelligence agent and tool library provided in one embodiment of the present invention.

[0034] Figure 4 This is a structural diagram of the MCP tool library provided in one embodiment of the present invention.

[0035] Figure 5 This is a timing diagram of the interaction and memory management of the discussion platform provided in one embodiment of the present invention. Detailed Implementation

[0036] The present invention will now be described in detail with reference to the accompanying drawings. Unless otherwise specified, the features of the following embodiments and implementations can be combined with each other.

[0037] Reference Figure 1 The system architecture diagram shown illustrates the configuration and coordination of the following components during the actual implementation of this system: The astronomical academic thinking intelligent agent layer acts as the "brain" of the system, responsible for the planning, reasoning, and reflection of scientific research tasks.

[0038] The astronomical academic execution intelligent agent layer acts as the "limbs" of the system, responsible for receiving instructions from the thinking layer and calling on specific tools to complete tasks.

[0039] The Astronomical Academic MCP Toolkit serves as the system's "toolbox," providing various research tools through standardized MCP protocols.

[0040] The astronomical academic discussion platform serves as the system's "collaboration field," providing an interface for multi-agent interaction with users and managing the discussion mechanism.

[0041] Specific implementation of key steps: Step 1: Configuration of the astronomical academic thinking agent, including: configuring a large language model with thought chain capabilities for task planning and reflection, and generating tool call requests; Specifically, it accesses large language models with thought chain capabilities through the OpenAI-compatible API protocol.

[0042] In one embodiment, the system instantiates multiple thinking agents representing different professional fields, based on the characteristics of astronomical scientific research. These agents include, but are not limited to: theoretical astrophysicists, observational astronomers, computational astrophysicists, cosmologists, and astrobiologists.

[0043] Model configuration: All thinking agents use a large language model that supports thought chain reasoning, such as the Qwen3-235B-A22B-Thinking-2507 model.

[0044] Role Definition: Each agent's expertise is defined through system prompts. For example, the system prompts configured for "Observational Astronomer" would emphasize their focus on interpreting and empirically analyzing observational data, and guide them to prioritize data acquisition and validation when planning tasks.

[0045] Knowledge base support: Each thinking agent is associated with a dedicated RAG-based knowledge base, which is constructed from materials such as astronomy textbooks, academic papers, professional journals, and patent documents. For example, after being vectorized using the Qwen3-Embedding-8B embedding model, it is stored in the Qdrant vector database for the agent to retrieve and refer to during reasoning. The embedding model and vector database can be selected by the agent.

[0046] Step 2: Configuration and tool invocation of the astronomical academic execution agent, including configuring a lightweight large language model to receive tool invocation requests from the thinking agent and interact with the MCP tool library to complete the tool invocation requests and return results; Specifically, a lightweight large language model is configured through the OpenAI compatible API, which interfaces with the thinking agent via the A2A protocol and interacts with the MCP tool library via the MCP protocol to complete tool call requests and return results; In one embodiment, the system decouples planning and execution by dividing the work between the thinking agent and the executing agent. The two communicate via the standardized Google A2A protocol.

[0047] Model configuration: The execution agent is driven by a more lightweight and efficient instruction fine-tuning large language model, such as the Qwen3-30B-A3B-Instruct-2507 model, whose core goal is to accurately parse instructions and call tools.

[0048] Functional division of labor: The system is configured with multiple dedicated execution agents, for example: Data analytics agents: used to directly perform data statistics, analysis, and visualization tasks within the Jupyter sandbox environment; Literature retrieval agent: Specialized in handling literature search requests across academic databases, such as searches from ArXiv, Wanfang, and the official websites of Science / Nature journals.

[0049] Web retrieval agent: Specifically designed to handle search requests for publicly available scientific research information on the Internet.

[0050] Knowledge base retrieval agent: Specifically designed to retrieve private databases uploaded by users.

[0051] Tool Invocation: After receiving a task, the executing agent invokes the corresponding toolset in the MCP tool library through the MCP protocol to complete the operation.

[0052] Step 3: Construction of the astronomical academic MCP toolkit, including the creation of a toolkit that includes web search, web crawling, data analysis, and paper retrieval, and the provision of services to the executing agent through the MCP protocol; The tool library encapsulates various astronomical research tools into standardized services using the MCP protocol. Preferably, the tool library is divided into the following tool sets based on function: Literature retrieval toolset: used to access various academic databases, such as Arxiv, ADS, Wanfang, etc.

[0053] Web search toolset: Used to access general search engines or science news websites, such as the Google Search API.

[0054] Knowledge base toolset: used for retrieving and analyzing user-owned private documents.

[0055] Data analytics toolsets: These provide professional data processing environments, such as Jupyter Sandbox.

[0056] Those skilled in the art will understand that the specific services listed in the above toolset are merely examples, and any service and tool that complies with the MCP protocol and can achieve the corresponding functions falls within the protection scope of this invention.

[0057] Step Four: The operation of the astronomical academic discussion platform and mechanism includes determining the speaking order; users setting the number of times they can speak in each round; and a memory management module for tracking and compressing discussion history to avoid context overflow. Determining the speaking order involves randomly selecting a thinking agent, then the currently speaking agent determines the next speaker based on its own content and switches speaking rights based on a mention mechanism.

[0058] In one embodiment, the user engages in discussions with the system through a web interface developed based on the Grado framework.

[0059] Discussion initiation: Users set the discussion topic (e.g., "Analysis of the atmospheric habitability of exoplanet K2-18b") and configure the number of times to speak per round.

[0060] Agent collaboration: The system randomly selects a thinking agent (e.g., an astrobiologist) to speak first. After the agent expresses its opinion, it designates the next speaker (e.g., @observational astronomer) through an @mention mechanism.

[0061] Discussion Process: The mentioned observational astronomer is activated and can supplement, question, or provide new evidence for the aforementioned viewpoint (e.g., responding, "More observational spectroscopic data from the James Webb Space Telescope is needed to confirm this potential biomarker"), and can continue to tag theoretical astrophysicists to request theoretical support. This process repeats until a complete answer is determined or the preset limit for the number of comments is reached.

[0062] History Compression: The system monitors token usage in the discussion context in real time. When usage reaches 90% of the threshold of the large model context window, the memory compression mechanism is automatically triggered: a strong inference model (e.g., Qwen3-235B-A22B-Instruct-2507 model) is called to generate a refined summary of the early discussion history, and then only the summary and the most recent 5 complete dialogues are retained as the new context, thereby effectively maintaining the continuity of long-term in-depth discussions.

[0063] Example 1: Taking "Investigating the Habitability of Exoplanets" as an example Suppose a user submits a research task: "Please analyze whether the exoplanet K2-18b possesses habitable conditions." The system's operation flow is as follows: 1. Specific Implementation of Astronomical Academic Thinking Intelligent Agent Thinking Agent Instantiation: The system randomly selects a thinking agent (such as an astrobiologist) and, as needed, mentions a theoretical astrophysicist (responsible for analyzing planetary formation theories) or an observational astronomer (responsible for verifying observational data) to participate in the collaboration. Each thinking agent is instantiated as an independent Agent object within the Langgraph framework, and its unique system prompts are loaded to define its role behavior.

[0064] The following are system prompts for the five thinking agents, mainly divided into "general guiding principles" and content for each thinking agent, with all agents sharing the content of the general guiding principles.

[0065] General guidelines: 1. Role Definition and Group Members: Clearly define the identity of each agent and the composition of the discussion group; 2. Interaction Mechanism (@Mention Mechanism): Supports directed dialogue with other experts by mentioning their names; 3. Requirements for dialogue continuity (5 items): It is mandatory to ask questions or invite other members after each response to maintain a smooth discussion; 4. Standardized Expression: Output the content directly, without using your own name, maintaining professionalism and ease of understanding; 5. Mandatory Tool Usage Requirements: When discussing cutting-edge topics, tools must be used to obtain the latest information; relying solely on existing knowledge is prohibited. 6. Strict requirements for citation: You must first use tools to search for the paper, and only cite existing literature. Fabricating references is prohibited. 7. List of interfacing intelligent agents and Agent Card (description) information: data analysis intelligent agent, document retrieval intelligent agent, web retrieval intelligent agent, knowledge base retrieval intelligent agent; 8. Citation and Reference Guidelines.

[0066] Explanation of the prompts for each character: theoretical astrophysicist 1. General guiding principles; 2. Professional background and expertise: General relativity, quantum mechanics, statistical mechanics, mathematical modeling, computational astrophysics, theoretical prediction, dark matter / dark energy, cutting-edge developments; 3. Discussion methods: theoretical framework, physical principles, theoretical explanation, concept description, paper citation, connection between theory and observation, cutting-edge developments, and tool usage.

[0067] Observational astronomers 1. General guiding principles; 2. Professional background and expertise: Optical / infrared / radio observation, astronomical instruments, data processing, multi-band astronomy, observation and theoretical comparison; 3. Discussion Methods: Emphasize observational evidence, data analysis methods, and how observations support / challenge theories.

[0068] Computational astrophysicists 1. General guiding principles; 2. Professional background and expertise: Optical / infrared / radio observation, astronomical instruments, data processing, multi-band astronomy, observation and theoretical comparison; 3. Discussion Methods: Emphasis is placed on computational methods, simulation techniques, and numerical results.

[0069] cosmologist 1. General guiding principles; 2. Professional background and expertise: Cosmological models, cosmic microwave background, large-scale structure, cosmic expansion, parameter measurement; 3. Discussion methodology: Emphasis is placed on cosmological models, observational evidence, and timescales of cosmic evolution.

[0070] Astrobiologists 1. General guiding principles; 2. Professional background and expertise: Planetary habitability, life in extreme environments, biological characteristics, planetary formation, exoplanet characteristics; 3. Discussion methods: Emphasize the conditions for life, planetary environment, and biological characteristics.

[0071] Model and knowledge base configuration: Model: The thinking agent uniformly adopts the Qwen3-235B-A22B-Thinking-2507 model, which is deployed privately through SGLang or by calling third-party APIs to ensure the depth and coherence of the thought chain.

[0072] Knowledge Base: The system provides customized knowledge support for each thinking agent based on RAG technology. For example, when analyzing biomarkers, astrobiologists can use embedded search commands to query information such as the "Astrobiology Journal" and textbook chapters from the Qdrant vector database, and incorporate the search results into the reasoning process as context.

[0073] Answer process: Task Receiver: Astrobiologists receive user tasks.

[0074] Planning and Reflection: Initiate the ReAct loop and plan the analysis path: "Data such as mass, orbital parameters, and atmospheric composition of K2-18b need to be obtained. First, retrieve basic information from the knowledge base."

[0075] Knowledge retrieval: Retrieving basic information about K2-18b from the knowledge base revealed incomplete data (especially a lack of the latest atmospheric spectral data).

[0076] Tool Invocation Request: Generates a tool invocation request and sends a command to the web retrieval agent via the Google A2A protocol: "Search for the latest JWST observation data and research papers for K2-18b".

[0077] Generate a response: After obtaining the data returned by the execution layer, synthesize the information to write a preliminary habitability analysis report, share it on the discussion platform, and at the same time, @observational astronomers to request data reliability verification.

[0078] 2. Specific Implementation of Astronomical Academic Execution Agents Model and Division of Labor: The execution agent adopts the lightweight and efficient Qwen3-30B-A3B-Instruct-2507 model. For example, after receiving a request, the web retrieval agent needs to call the "web search" tool for parsing.

[0079] Protocol interoperability (corresponding to claims 3.2, 3.3): A2A Protocol: The thinking layer and the execution layer transmit structured requests (including tool names and input parameters such as search keywords "K2-18b JWST atmosphere 2024") through the Google A2A protocol.

[0080] MCP Protocol: The web retrieval agent uses the MCP client to call the search engine tools in the "MCP Web Search Toolset" to obtain web page links and summary results.

[0081] 3. Specific Implementation of the Astronomical Academic MCP Toolkit Tool call: Web search toolset: Tools that call up Google Search API or academic resource retrieval tools.

[0082] Literature retrieval toolset: If literature support is required, the literature retrieval agent can obtain preprint papers and other literature through the Arxiv API.

[0083] Knowledge Base Toolset: If a user uploads internal observation data, the knowledge base retrieval agent performs vectorized retrieval of the document.

[0084] 4. Specific Implementation of Astronomical Academic Discussion Platforms and Mechanisms Interactive Interface: Users interact through a web interface based on Grado, which displays the online thinking agent, discussion history, and discussion round control parameters in real time.

[0085] Discussion mechanism: The user sets the number of times they can speak to 5 and issues a task; the system then randomly selects an astrobiologist to speak first.

[0086] After an astrobiologist publishes their opinion, they tag an observational astronomer based on their own viewpoint; the latter, upon activation, may respond: "There are observational errors in citing JWST data, which need to be cross-validated with Hubble data," and then continue to tag a theoretical astrophysicist.

[0087] Theoretical astrophysicists analyze planetary internal structures and the role of magnetic fields in maintaining the atmosphere using theoretical models. The discussion pauses after five replies or when a complete answer is deemed appropriate. Users can then enter "Continue Discussion" or provide guidance to start a new round.

[0088] Historical compression mechanism and memory module settings: When the number of discussion context tokens reaches 90% of the model window, the system automatically triggers compression: calling the Qwen3-235B-A22B-Thinking-2507 model to generate a summary of the early history (stored in JSON format, such as: "Summary: The first 20 rounds focused on the authenticity of the K2-18b atmospheric dimethyl sulfide signal, the observers questioned the confidence of the data, and the theorists proposed non-biological causes...").

[0089] The subsequent discussion will continue based on this summary and the last five rounds of dialogue, breaking through the limitations of context length while retaining the core arguments.

[0090] The core structure design of the memory digest. After compression is triggered, the memory module generates a structured digest that follows a predefined data pattern. An example of its core structure implementation is as follows: Structural design description: Structured summary: The compressed_history field refines lengthy discussion histories, extracting core issues, key arguments, consensus states, and unresolved issues, enabling the thinking agent to quickly grasp the overall picture of the discussion and avoid getting bogged down in details.

[0091] Role contribution record: The agent_contributions object is indexed by the thinking agent ID and independently records the core viewpoints and key actions of each role, ensuring that their professional contributions are highlighted and preserved, and can be accurately cited and inherited in subsequent discussions.

[0092] Context management: The recent_context field explicitly retains references to the last 5 rounds of dialogue, ensuring that the compressed memory can seamlessly connect with the most relevant recent context; at the same time, the next_speaker_scheduled field specifies the designated next speaker, ensuring that the continuity of the discussion process is not interrupted.

[0093] Traceability and queryability: The memory_id and trigger_condition fields provide traceability at the management level. This structured design allows the system to quickly locate and retrieve historical viewpoints or to-do items for specific roles, much like querying a database, significantly improving the availability and management efficiency of memory in long-term conversations.

[0094] In summary, through the specific embodiments described above, this invention constructs a multi-agent system capable of deeply simulating the entire process of astronomical scientific research, possessing autonomous thinking, efficient execution, and group collaboration capabilities. It is understood that any equivalent substitutions and reasonable modifications made by those skilled in the art to the selection of the large model, the specific content of the toolset, or the interactive interface framework, without departing from the core principles of this invention—namely, the "thinking-execution" hierarchical architecture, standardized protocol integration, and structured memory management—should be included within the scope of protection defined by the claims of this invention.

[0095] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A method for configuring a multi-agent system for academic research in the field of astronomy, characterized in that, To implement an intelligent agent for astronomical academic thinking, an intelligent agent for astronomical academic execution, an astronomical academic MCP toolkit, and an astronomical academic discussion mechanism, the following steps are included: (1) Realize an intelligent agent for astronomical academic thinking: Configure a large language model with thinking chain capabilities for task planning and reflection, and generate tool call requests; (2) Implement an astronomical academic execution agent: Configure a lightweight large language model to receive tool call requests from the thinking agent and interact with the MCP tool library to complete the tool call requests and return the results; (3) Construct an astronomical academic MCP tool library: Establish a tool set including web search, web crawling, data analysis, and paper retrieval, and provide services to the execution agent through the MCP protocol; (4) Establish an astronomical academic discussion platform and mechanism: including determining the order of speaking; users setting the number of times to speak in each round; and also including a memory management module to track and compress the discussion history to avoid context overflow.

2. The method for configuring a multi-agent system for academic research in the field of astronomy, as described in claim 1, is characterized in that... In step (1), the intelligent agent for realizing astronomical academic thinking includes: (1.1) Role definition module: Based on the secondary development of the Langgraph framework, several thinking agent roles with different astronomical scientific research expertise are defined. Each thinking agent reflects its professional field differences through exclusive system prompts and knowledge bases. (1.2) Model configuration module: The core is a large language model that supports the thinking chain and is compatible with the OpenAI API, which can be deployed privately or accessed through the official API; the differentiated capabilities of the thinking agent are achieved by configuring system prompts that include its role definition and collaboration specifications. (1.3) Knowledge base configuration module: Configure a dedicated knowledge base based on RAG for each thinking agent; the knowledge base includes academic papers, professional journals, textbooks and patent materials, and performs vectorization processing and stores them in a vector database for the agent to retrieve and call; (1.4) Inference and Response Module: Its workflow follows the reAct agent model, specifically including: Upon receiving a question, the first step for the thinking agent is to determine the information needed to generate an answer. Then, it initiates a tool call request to the executing agent or performs a knowledge base search to obtain information; The judgment and information acquisition steps are executed repeatedly until the thinking agent determines on its own that it can generate a complete answer. It also includes setting a maximum number of loop iterations, which forces the current result to be output when this number of iterations is exceeded, in order to avoid infinite loops.

3. The method for configuring a multi-agent system for academic research in the field of astronomy, as described in claim 1, is characterized in that... The thinking agent includes: Theoretical astrophysicist, focusing on theoretical models and mathematical derivations; Observational astronomers focus on observational data and empirical research. Computational astrophysicist, specializing in simulation and computational methods; Cosmology expert, specializing in large-scale cosmological structure and evolution; Astrobiologists who focus on the possibility of extraterrestrial life.

4. The method for configuring a multi-agent system for academic research in the field of astronomy, as described in claim 1, is characterized in that... In step (2), the implementation of the astronomical academic execution intelligent agent includes: (2.1) Model and Functional Modules: Driven by a lightweight, instruction-based large language model; (2.2) First docking module: The thinking agent and the execution agent dock through the A2A protocol; wherein, each execution agent declares its name, list of executable tools and related calling parameters through a predefined agent card; (2.3) Second docking module: The execution agent and the MCP tool library dock through the MCP protocol; wherein, each execution agent is limited to calling the tool subset corresponding to its function, including the document retrieval tool set, the web search tool set or the knowledge base tool set.

5. The method for configuring a multi-agent system for academic research in the field of astronomy according to claim 1, characterized in that, The executing agents are divided into four categories based on their functions: Data analytics agents are used to perform data statistics, analysis, and visualization tasks; A document retrieval agent used to retrieve academic documents; A web retrieval agent is used to obtain the latest scientific research information through a search engine; The knowledge base retrieval agent is used to retrieve private information uploaded by users.

6. The method for configuring a multi-agent system for academic research in the field of astronomy, as described in claim 1, is characterized in that... The astronomical academic MCP toolkit includes: A set of literature retrieval tools for retrieving scientific research literature and results, including preprint platforms, academic databases and intellectual property databases. A suite of web search tools for retrieving authoritative news and journal information in the field of scientific research; A knowledge base toolset for mixed retrieval of user-owned documents, materials, and data; A set of data analysis tools to provide a professional data processing environment.

7. The method for configuring a multi-agent system for academic research in the field of astronomy according to claim 1, characterized in that, In step (4), the establishment of an astronomical academic discussion platform and mechanism includes: (4.1) Interactive interface module: Based on the Grado open source framework and the basic interface of Langgraph is developed in a secondary way, it provides functions such as chat interface, document upload, monitoring of the state of thinking agent, display of available tools list and setting the number of times to speak; (4.2) Discussion control module: used to receive the number of times to speak and the discussion topic set by the user, and control the discussion process, including randomly initializing the thinking agent to speak, determining the next speaker based on the content of the current speaking thinking agent and switching the speaking right based on the mention mechanism, and pausing the discussion when the speaking limit is reached. (4.3) Memory Management Module: Monitors the number of tokens in the discussion context in real time, and automatically triggers the history summary process when its occupancy rate reaches a preset threshold, compressing long discussion history into a summary, and forming a new context together with the most recent rounds of dialogue.

8. A multi-agent system for academic research in the field of astronomy, characterized in that, include: Astronomical Academic Thinking Agent: Equipped with a powerful thinking language model with thought chain capabilities, it is used for task planning and reflection, and to generate tool call requests; Astronomical academic execution agent: Configured with a lightweight large language model to receive tool call requests from the thinking agent and interact with the MCP tool library to complete the tool call requests and return the results; Astronomical Academic MCP Toolkit: This toolkit includes a collection of tools for web search, web scraping, data analysis, and paper retrieval, and provides services to the executing intelligent agent through the MCP protocol. Astronomical academic discussion platform and mechanism: including determining the order of speaking; users setting the number of times they can speak in each round; It also includes a memory management module for tracking and compressing discussion history to avoid context overflow.

9. A multi-agent system for academic research in the field of astronomy according to claim 8, characterized in that, The thinking agent includes: Theoretical astrophysicists, used for theoretical models and mathematical derivations; Observational astronomers, used for observational data and empirical research; Computational astrophysicists, using simulation and computational methods; Cosmology experts specializing in large-scale cosmological structure and evolution; Astrobiologists use it to study the possibility of extraterrestrial life.

10. A multi-agent system for academic research in the field of astronomy according to claim 8, characterized in that, The executing intelligent agent includes: Data analysis agents can directly perform data statistics, analysis, and visualization tasks within the Jupyter sandbox environment; A literature retrieval agent used to retrieve academic literature from ArXiv, Wanfang, and the official websites of Science / Nature journals; A web retrieval agent is used to obtain the latest scientific research information through a search engine; The knowledge base retrieval agent is used to retrieve private information uploaded by users.

Images