Multi-agent collaborative science and technology information generation method and system

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a multi-agent collaborative science and technology information generation method and system, and the method comprises the steps: obtaining a user demand; performing multi-information source retrieval according to the user demand to obtain database retrieval content; generating initial science and technology information according to the user demand and the retrieval content; and auditing and optimizing the initial science and technology intelligence according to the retrieval content. According to the technical scheme of the invention, the multi-agent collaborative architecture is constructed based on the large language model, and the capabilities of information retrieval, semantic analysis, knowledge fusion and content generation are combined, so that the automatic understanding, organization and expression of the science and technology information are realized, and therefore, the efficiency, quality and intelligent level of science and technology information generation are remarkably improved. And more valuable auxiliary decision support is provided for enterprises.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for generating scientific and technological intelligence through multi-agent collaboration. Background Art

[0002] As global technological competition intensifies, technological innovation is becoming a key source of core competitiveness for businesses. Against this backdrop, the demand for scientific and technological intelligence focused on tracking cutting-edge technologies, analyzing industry trends, and analysing the competitive landscape is growing exponentially. Its application has deeply penetrated core aspects of corporate strategic decision-making, R&D planning, and market insights.

[0003] In current practice, the acquisition and generation of scientific and technological intelligence is still primarily manual, relying on analysts to sort through and interpret vast quantities of scientific literature, patent information, industry reports, and more. This process is not only time-consuming and inefficient, but also limited by individual cognitive abilities and knowledge, making it prone to information omissions and misinterpretations. Furthermore, faced with massive amounts of multi-source, dynamically updated scientific and technological information, traditional manual processing methods are unable to meet businesses' demands for highly timely, comprehensive, and accurate intelligence.

[0004] Based on the above problems, there is an urgent need for an efficient, intelligent and scalable method for generating scientific and technological intelligence. Summary of the Invention

[0005] One of the purposes of the embodiments of the present invention is to provide a multi-agent collaborative scientific and technological intelligence generation method and system. This method and system utilize a large language model to obtain user needs, and combines information retrieval, semantic analysis, knowledge fusion and content generation capabilities to achieve automatic understanding, organization and expression of scientific and technological information, thereby significantly improving the efficiency, quality and intelligence level of scientific and technological intelligence generation, and providing enterprises with more valuable auxiliary decision support.

[0006] To achieve the above objectives, an embodiment of the present invention provides a method for generating scientific and technological intelligence through multi-agent collaboration, comprising:

[0007] Obtain user needs;

[0008] Perform multi-information source search according to the user's needs to obtain database search content;

[0009] Generate initial scientific and technological intelligence based on the user needs and search content;

[0010] The initial scientific and technological intelligence is reviewed and optimized based on the search content.

[0011] Optionally, obtaining user needs includes:

[0012] Respond to acquired user needs and communicate with users using a large language model;

[0013] Determine whether the currently input user requirements are complete and clear;

[0014] If the judgment is incomplete, ask guiding questions and return to the step of judging whether the currently input user requirements are complete and clear;

[0015] If the judgment is unclear, try to clarify the ambiguous expression and return to the step of determining whether the currently input user requirements are complete and clear;

[0016] When it is judged to be complete and clear, an identification label is added to the currently input user demand and structured.

[0017] Optionally, performing a multi-information source search according to the user's needs to obtain database search content includes:

[0018] Based on the user needs, generate initial search questions using a large language model;

[0019] Using thought chaining to split the initial search question into multiple sub-search questions;

[0020] Clarify the search target of each sub-search question;

[0021] Conduct internal vector database search and external search for each sub-search question;

[0022] Perform quality screening on the results retrieved for each sub-search question, including scoring the source authority and demand matching, and integrating the two scores for quality screening;

[0023] Structure effective search content.

[0024] Optionally, generating initial scientific and technological intelligence according to the user needs and search content includes:

[0025] Use the LlamaFactory framework to load the Llama3-8B-Chinese-Chat model as the intelligence writing model;

[0026] Build Alpaca format dataset;

[0027] Set model fine-tuning parameters;

[0028] Training the intelligence writing model using the Alpaca format dataset, including generating prediction results, calculating losses, and updating parameters through the QLoRA adapter;

[0029] Save the fine-tuned adapter weights to obtain the trained intelligence writing model;

[0030] The user needs and search content are input into the trained intelligence writing model to generate initial scientific and technological intelligence.

[0031] Optionally, proofreading and reviewing the initial scientific and technological information according to the search content includes:

[0032] Using a large language model to analyze the content of the initial scientific and technological intelligence and identify ambiguous information;

[0033] Arrange the fuzzy information into specific search questions;

[0034] Conducting a multi-source search on the specific search question to obtain supplementary data;

[0035] Based on the supplementary data, the initial scientific and technological intelligence is corrected and improved to generate final scientific and technological intelligence.

[0036] In another aspect, the present invention provides a multi-agent collaborative scientific and technological intelligence generation system, the intelligence generation system comprising:

[0037] The intelligence communication agent is built using a large language model and runs in a multi-agent task orchestration framework based on LangChain. It is used to conduct multiple rounds of dialogue with demand providers to obtain user needs.

[0038] An intelligence collection agent, comprising a vector database and an execution tool set, wherein the vector database is used to convert professional documents related to intelligence analysis into vector form, the execution tool set is used to provide the intelligence collection agent with tools to perform corresponding operations, and the intelligence collection agent is used to perform multi-information source retrieval according to the user's needs to obtain database retrieval content;

[0039] An intelligence writing agent, used to generate initial scientific and technological intelligence according to the user needs and search content;

[0040] An intelligence review agent is used to review and optimize the initial scientific and technological intelligence based on the search content.

[0041] Optionally, the intelligence communication agent is used to:

[0042] Respond to acquired user needs and communicate with users using a large language model;

[0043] Determine whether the currently input user requirements are complete and clear;

[0044] If the judgment is incomplete, ask guiding questions and return to the step of judging whether the currently input user requirements are complete and clear;

[0045] If the judgment is unclear, try to clarify the ambiguous expression and return to the step of determining whether the currently input user requirements are complete and clear;

[0046] When it is judged to be complete and clear, an identification label is added to the currently input user demand and structured.

[0047] Optionally, the intelligence gathering agent is used to:

[0048] Based on the user needs, generate initial search questions using a large language model;

[0049] Using thought chaining to split the initial search question into multiple sub-search questions;

[0050] Clarify the search target of each sub-search question;

[0051] Conduct internal vector database search and external search for each sub-search question;

[0052] Perform quality screening on the results retrieved for each sub-search question, including scoring the source authority and demand matching, and integrating the two scores for quality screening;

[0053] Structure effective search content.

[0054] Optionally, the intelligence writing agent is used to:

[0055] Use the LlamaFactory framework to load the Llama3-8B-Chinese-Chat model as the intelligence writing model;

[0056] Build Alpaca format dataset;

[0057] Set model fine-tuning parameters;

[0058] Training the intelligence writing model using the Alpaca format dataset, including generating prediction results, calculating losses, and updating parameters through the QLoRA adapter;

[0059] Save the fine-tuned adapter weights to obtain the trained intelligence writing model;

[0060] The user needs and search content are input into the trained intelligence writing model to generate initial scientific and technological intelligence.

[0061] Optionally, the intelligence review agent is used to:

[0062] Using a large language model to analyze the content of the initial scientific and technological intelligence and identify ambiguous information;

[0063] Arrange the fuzzy information into specific search questions;

[0064] Conducting a multi-source search on the specific search question to obtain supplementary data;

[0065] Based on the supplementary data, the initial scientific and technological intelligence is corrected and improved to generate final scientific and technological intelligence.

[0066] Beneficial effects of the present invention:

[0067] The implementation method of the present invention constructs a multi-agent collaborative architecture based on a large language model, and combines information retrieval, semantic analysis, knowledge fusion and content generation capabilities to achieve automatic understanding, organization and expression of scientific and technological information, thereby significantly improving the efficiency, quality and intelligence level of scientific and technological intelligence generation, and providing enterprises with more valuable auxiliary decision support.

[0068] Other features and advantages of the embodiments of the present invention will be described in detail in the subsequent detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The accompanying drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following detailed description, they are used to explain the embodiments of the present invention, but do not constitute a limitation of the embodiments of the present invention. In the accompanying drawings:

[0070] Figure 1 A flowchart of a method for generating scientific and technological intelligence through multi-agent collaboration according to one embodiment of the present invention;

[0071] Figure 2 A flowchart of a method for obtaining user needs according to one embodiment of the present invention;

[0072] Figure 3 A flowchart of a method for obtaining search content according to one embodiment of the present invention;

[0073] Figure 4 A flowchart of a method for generating scientific and technological intelligence according to one embodiment of the present invention;

[0074] Figure 5 A flowchart of a method for reviewing and optimizing initial scientific and technological intelligence according to one embodiment of the present invention;

[0075] Figure 6 A structural block diagram of a multi-agent collaborative scientific and technological intelligence generation system according to one embodiment of the present invention;

[0076] Figure 7 Schematic diagram of the workflow of an intelligence collection agent according to one embodiment of the present invention.

[0077] Description of Reference Numerals

[0078] 1. Intelligence communication agent; 2. Intelligence collection agent; 3. Intelligence writing agent; 4. Intelligence review agent. DETAILED DESCRIPTION

[0079] The following describes the specific implementation of the embodiment of the present invention in detail with reference to the accompanying drawings. It should be understood that the specific implementation described herein is only used to illustrate and explain the embodiment of the present invention and is not used to limit the embodiment of the present invention.

[0080] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application are in compliance with the relevant provisions of laws and regulations. In the embodiments of this application, certain software, components, models, and other existing solutions in the industry may be mentioned. These should be considered as exemplary. Their purpose is only to illustrate the feasibility of implementing the technical solution of this application, but it does not mean that the applicant has or will necessarily use such solutions.

[0081] like Figure 1 The flowchart of the method for generating scientific and technological intelligence based on multi-agent collaboration according to one embodiment of the present invention is shown. Figure 1 In the method for generating intelligence, the method may include the following steps:

[0082] In step S10, user needs are obtained;

[0083] In step S11, a multi-information source search is performed according to user needs to obtain database search content;

[0084] In step S12, initial scientific and technological intelligence is generated according to user needs and search content;

[0085] In step S13, the initial scientific and technological information is reviewed and optimized according to the search content.

[0086] In this Figure 1 In the multi-agent collaborative scientific and technological intelligence generation method shown, step S10 is used to obtain user needs, and conduct multiple rounds of dialogue with the demand provider (i.e., the end user or the enterprise intelligence analyst) to complete the pre-demand collection and verification process of the scientific and technological intelligence writing task. This step S10 can automatically identify contextual inconsistencies or missing information in the user description, through a prompt-driven interactive question-and-answer process, and perform the next dialogue with the demand provider based on the returned content to communicate the needs, and repeat the cycle to gradually improve the needs. In this embodiment, the specific method for obtaining user needs in step S10 can be a variety of forms known to those skilled in the art. In one example of the present invention, this step S10 can be a method including: Figure 2 The steps shown in Figure 2 In the embodiment, step S10 may include:

[0087] In step S20, in response to the acquired user needs, a conversation is conducted with the user using the large language model;

[0088] In step S21, it is determined whether the currently input user requirements are complete and clear;

[0089] In step S22, if the input is incomplete, guiding questions are raised, and the process returns to the step of determining whether the currently input user requirements are complete and clear.

[0090] If the judgment is unclear, try to clarify the ambiguous expression and return to the step of determining whether the currently input user requirements are complete and clear;

[0091] When it is judged to be complete and clear, an identification label is added to the currently input user demand and structured.

[0092] In this Figure 2 In the illustrated method, step S20 is used to respond to the acquired user needs by engaging in a dialogue with the user using a large language model, guiding the user to clearly express their information needs, including but not limited to: technical topics of interest, analysis objectives, scenarios, desired report format and granularity, and the range of data sources to be retrieved. During each round of dialogue, an assessment is made as to whether an intelligence query agent needs to be mobilized for external search assistance.

[0093] Step S21 is used to determine whether the currently input user requirement is complete and clear. In this embodiment, the completeness determination can be performed by checking for missing key information dimensions (e.g., preset dimensions such as technical field, time range, geographic scope, and analysis objectives), and the clarity determination can be performed by detecting ambiguous expressions (e.g., unquantified or undefined terms such as "cutting-edge technology," "recent," and "effective"). Specifically, in this example, the completeness determination method may include: presetting key dimension templates, extracting entities using a pretrained model or rule matching, and detecting missing dimensions. The clarity determination method may also include constructing a fuzzy vocabulary, performing fuzziness scoring using a fuzziness scoring model, and performing contextual relevance verification (checking whether ambiguous terms are subsequently defined; for example, if a user says "cutting-edge technology, specifically referring to AI and blockchain," then the fuzzy mark for "cutting-edge technology" is removed).

[0094] Step S22 is used to perform corresponding operations based on the results of the judgment. When it is judged that the currently input user needs are incomplete, it identifies missing information points, actively asks guiding questions (such as "Do you need to focus on a specific technology track?", "Is there a target analysis cycle or region?"), and obtains the currently input user needs, and returns to step S21 for judgment. When it is judged that the currently input user needs are unclear, that is, they contain vague expressions, try to clarify the intention (such as "Which specific sub-sectors do you refer to as the 'cutting-edge technology'?"), obtain the currently input user needs, and return to step S21 for judgment. By repeatedly interacting with the user until their needs are judged to be complete and clear, add identification tags to the finally confirmed needs, and structure the complete needs.

[0095] Step S11 is used to perform a multi-information source search based on user needs to obtain database search content. Based on the acquired user needs, the large language model is used to generate an initial search question, the knowledge base and the network are used to retrieve the answer to the initial question, and the results are filtered to obtain the final search results. In this embodiment, the specific method for acquiring user needs in step S11 can be various forms known to those skilled in the art. In one example of the present invention, step S11 can include the following: Figure 3 The steps shown in Figure 3 In the step S11, the following steps may be included:

[0096] In step S30, based on user needs, an initial search question is generated using a large language model;

[0097] In step S31, the initial search question is split into multiple sub-search questions using a thinking chain;

[0098] In step S32, the search target of each sub-search question is clarified;

[0099] In step S33, internal vector database search and external search are performed for each sub-search question;

[0100] In step S34, the results retrieved for each sub-search question are quality screened, including: scoring the source authority and the demand matching degree, and fusing the two scores for quality screening;

[0101] In step S35, the effective search content is structured.

[0102] In this Figure 3In the method shown, step S30 is used to generate the initial search questions using the large language model for the acquired user needs. Step S31 is used to derive sub-search questions using the Chain of Thought (CoT) method, and to form fine-grained sub-questions by splitting complex questions. Specifically, in this example, each sub-question may include: a clear target object (such as "hydrogen fuel cell technology"), a clear focus (such as "development trends in the past five years"), and a quantifiable search direction (such as "changes in the number of domestic and foreign patent applications"). Step S32 is used to clarify the search target of each sub-search question (such as technology development trends, application scenarios, patent layout, etc.). Step S33 is used to search each sub-search question, giving priority to similarity search in the internal vector database (intelligence knowledge base built based on chromadb), and then calling external search tools (DuckDuckGo, full-text crawling tools) for incremental supplementation.

[0103] Step S34 is used to perform quality screening on the results retrieved for each sub-search question. Specifically, in this example, the retrieved results may be scored for authority and demand matching, and then the two scores may be combined for screening. In this embodiment, the specific method for performing authority scoring in step S34 may be various forms known to those skilled in the art. In one example of the present invention, it may include the following steps:

[0104] First, a scoring method is integrated to perform scoring, and the scoring method may include: source type identification, domain name / author matching and / or citation index. Among them, the source type identification scoring (SourceScore) method may be: if it is .gov, .org, or the official website of a well-known scientific research institution, the corresponding score SourceScore may be set to a value within the preset first interval; if it is a news media, forum, or personal blog, the corresponding score SourceScore may be set to a value within the preset second interval, and the minimum value of the first interval is greater than the maximum value of the second interval. The domain name / author matching scoring (AuthorScore) method may be: whether it appears in a trusted whitelist (such as CNIPA, IEEE, Nature); whether it matches a high-reputation author in the expert database. In this example, the domain name / author matching scoring may be The CitationScore method can be: if there are metadata such as impact factor, citation count, report rating, etc., then these data values ​​are normalized and averaged as the citation index. In this example, the document impact factor value can be By normalizing these metrics, the impact of document magnitude differences can be eliminated, ensuring fair comparison. If a document lacks metrics such as citation count or impact factor, a moderate citation score is assigned based on other authority metrics (such as source type and domain / author match). The values ​​of the selected citation metrics are averaged to form the CitationScore. The final score is calculated based on the obtained source type identification score (SourceScore), domain / author match score (AuthorScore), and citation metric score (CitationScore). In this example, the final score can be: AuthorityScore = α1 × SourceScore + α2 × AuthorScore + α3 × CitationScore, where α1, α2, and α3 are weighting coefficients. Each document is assigned an authority score of AuthorityScore ∈ [0, 1]. For example, high authority is defined as AuthorityScore ≥ 0.75.

[0105] In this embodiment, the specific method for performing the requirement matching score in step S34 may be various methods known to those skilled in the art. In one example of the present invention, the method may be to determine the semantic relevance between the search results and the sub-search questions, specifically including the following steps:

[0106] First, embedding semantic matching is performed, using BERT to generate semantic vectors for the query and document, and Cosine similarity is calculated. Next, keyword coverage is analyzed to determine whether the passage covers the target object, time dimension, and key keywords in the query. Finally, a RelevanceScore is output for each document's demand matching.

[0107] In this embodiment, the specific method for fusing the two types of scores for screening in step S34 may be various forms known to those skilled in the art. In one example of the present invention, the method may include the following steps:

[0108] First, the two scores are combined by calling final_score = α * AuthorityScore + β * RelevanceScore. In this example, the weights could be α = 0.4 and β = 0.6. Then, filtering logic is applied. This could be a dual-threshold filter (AuthorityScore ≥ 0.6 and RelevanceScore ≥ 0.7) or a Top-K filter (sort by final_score and retain the top K entries). In this example, the value of K could be 5.

[0109] Step S35 is used to structure and organize the effective search content to generate a final search result package (including document fragments, source information, and search keywords).

[0110] Step S12 is used to generate a scientific and technological intelligence analysis report that meets professional standards based on the final user needs and search content, thereby realizing the intelligence and efficiency of the intelligence writing process. In this embodiment, the specific method for generating scientific and technological intelligence in step S12 can be various forms known to those skilled in the art. In one example of the present invention, step S12 can include the following: Figure 4 The steps shown in Figure 4 In the step S12, the following steps may be included:

[0111] In step S40, the Llama3-8B-Chinese-Chat model is loaded using the LlamaFactory framework as an intelligence writing model;

[0112] In step S41, construct an Alpaca format dataset;

[0113] In step S42, model fine-tuning parameters are set;

[0114] In step S43, the intelligence writing model is trained using the Alpaca format dataset, including: generating prediction results, calculating losses, and updating parameters through the QLoRA adapter;

[0115] In step S44, the fine-tuned adapter weights are saved to obtain a trained intelligence writing model;

[0116] In step S45, user needs and search content are input into the trained intelligence writing model to generate initial scientific and technological intelligence.

[0117] In this Figure 4In the method shown, step S40 is used to load the Llama3-8B-Chinese-Chat model using the LlamaFactory framework as an intelligence writing model. Step S41 is used to construct an Alpaca format dataset for training the intelligence writing model. Specifically, in this example, the Alpaca format dataset can include four parts: instruction, input, output, and system, where: instruction content is a clear writing task instruction (such as "write a new energy vehicle technology trend report"); input content is the search content or context background related to the task (such as the summary information provided by the intelligence collection agent); output content is the corresponding standard example report; system is used to define the output style and context setting system prompt words. Training data comes from multiple aspects. In this example, the source of training data can be real collected corporate science and technology intelligence demand cases, integrated science and technology research reports, industry analysis reports, white papers and other professional materials, and high-quality training samples hand-crafted based on the final version of the demand and search results. Step S42 is used to set model fine-tuning parameters. Model fine-tuning parameters can include learning rate, batch size, and training rounds. Specifically, in this example, the learning rate can be set to 2e-5, the batch size can be set to 32, and the number of training rounds can be set to 1-5. Step S43 trains the intelligence writing model using an Alpaca format dataset. Specifically, in this example, the steps of training the intelligence writing model may include generating prediction results, calculating losses, and updating parameters via the QLoRA adapter. Step S44 is used to save the fine-tuned adapter weights to obtain the trained intelligence writing model. Step S45 is used to input user requirements and search content into the trained intelligence writing model to generate initial scientific and technological intelligence.

[0118] Step S13 is used to review and optimize the initial scientific and technological information. By combining the search content, the fuzzy information in the initial scientific and technological information is reviewed and optimized. In this embodiment, the specific method for reviewing and optimizing the initial scientific and technological information in step S13 can be various forms known to those skilled in the art. In one example of the present invention, step S13 can include the following: Figure 5 The steps shown in Figure 5 In the step S13, the following steps may be included:

[0119] In step S50, the large language model is used to analyze the initial scientific and technological information text content to identify fuzzy information;

[0120] In step S51, the fuzzy information is organized into specific search questions;

[0121] In step S52, a multi-information source search is performed on the specific search question to obtain supplementary data;

[0122] In step S53, the initial scientific and technological information is corrected and improved based on the supplementary data to generate final scientific and technological information.

[0123] In this Figure 5 In the method shown, step S50 is used to identify ambiguous information in the initial scientific and technological intelligence. Specifically, in this example, a large language model can be used to analyze the text content of the initial scientific and technological intelligence to identify ambiguous expressions, unclear conceptual descriptions, insufficient data support or inferential content in the report. Step S51 is used to organize the identified ambiguous or unclear information points into specific search questions, such as: "What is the specific data basis for the XXX trend?", "What are the actual application cases of XXX technology?". Step S52 is used to perform multi-information source retrieval on specific search questions to obtain supplementary data. Multi-information source retrieval can be performed through external knowledge bases and the Internet to obtain the latest and authoritative supplementary materials or data basis. Step S53 is used to correct and improve the initial scientific and technological intelligence based on the supplementary data, clarify the original ambiguous expressions, supplement missing data or cases, improve the accuracy and completeness of the content, and generate the final scientific and technological intelligence.

[0124] On the other hand, the embodiment of the present invention also provides a multi-agent collaborative scientific and technological intelligence generation system. The structural block diagram of the system can be as follows: Figure 6 As shown. Figure 6 In the present invention, the detection system may include an intelligence communication agent 1, an intelligence collection agent 2, an intelligence writing agent 3 and an intelligence review agent 4.

[0125] The Intelligence Communication Agent 1 is built using a large language model and runs within a multi-agent task orchestration framework based on LangChain. It is used to conduct multiple rounds of dialogue with demand providers to obtain user requirements. Intelligence Communication Agent 1 is capable of extracting key information from user requirements and deciding whether to further confirm the requirements.

[0126] Intelligence collection agent 2 includes a vector database and an execution tool set, wherein the vector database is used to convert professional documents related to intelligence analysis into vector form, and the execution tool set is used to provide the intelligence collection agent with tools to perform corresponding operations. Professional documents may include research reports, analysis reports, and historical cases. Intelligence collection agent 2 is used to perform multi-information source retrieval based on user needs to obtain database retrieval content. In this example, intelligence collection agent 2 may also include network retrieval and knowledge base retrieval modules, which can automatically collect relevant information from the Internet and knowledge base and provide it to the intelligence writer agent to achieve multi-information source retrieval.

[0127] Intelligence Writing Agent 3 generates initial scientific and technological intelligence based on user needs and search results. Intelligence Writing Agent 3 is equipped with an intelligence writing model that is trained using proprietary intelligence information and can generate preliminary scientific and technological intelligence based on the intelligence content provided by the Intelligence Collection Agent and user needs.

[0128] Intelligence review agent 4 is used to review and optimize initial scientific and technological intelligence based on the search content to ensure the accuracy, completeness and consistency of the information.

[0129] In this embodiment, the specific method for the multi-agent collaboration can be various forms known to those skilled in the art. In one example of the present invention, the multi-agent collaboration method includes the following steps:

[0130] First, based on the multi-agent collaborative scientific and technological intelligence generation framework, roles are defined for intelligence communication agent 1, intelligence collection agent 2, intelligence writing agent 3, and intelligence review agent 4. Specifically, the role definition method can include the following steps:

[0131] The role of intelligence communication agent 1 is defined as:

[0132]

[0133] The role of intelligence gathering agent 2 is defined as:

[0134]

[0135]

[0136] The role of Intelligence Writing Agent 3 is defined as:

[0137]

[0138] The role of Intelligence Review Agent 4 is defined as:

[0139]

[0140] Then, the intelligence communicator agent 1 starts the intelligence generation process according to the user's needs and sends information collection instructions to the intelligence collection agent 2.

[0141] Then, the intelligence collection agent 2 performs network retrieval and knowledge base retrieval tasks according to the collection instructions, collects relevant data from multiple information sources, and returns the preliminary collected information to the demander. If the demand is confirmed, the content will be sent to the intelligence writing agent 3.

[0142] Then, the intelligence writing agent 3 writes preliminary scientific and technological intelligence based on the received information and the task requirements provided by the intelligence communicator agent 1.

[0143] Finally, Intelligence Review Agent 4 conducts a detailed review of the initially generated S&T intelligence to ensure the accuracy of key information. If any issues are found, Intelligence Collection Agent 2 is called upon to collect more information and generate an accurate report, ultimately forming the S&T intelligence report. This collaborative framework ensures the efficiency and accuracy of the S&T intelligence generation process. Through the division of labor and collaboration among agents, the quality and speed of intelligence generation can be greatly improved.

[0144] In this Figure 6 In the multi-agent collaborative scientific and technological intelligence generation system shown, the intelligence communication agent 1 can be used to obtain user needs. In this embodiment, the specific method for obtaining user needs can be various forms known to those skilled in the art. In one example of the present invention, the intelligence communication agent 1 can be used to perform the following steps: Figure 2 The method shown in Figure 2 In the example, the intelligence communication agent 1 can be used to perform the following steps:

[0145] In step S20, in response to the acquired user needs, a conversation is conducted with the user using the large language model;

[0146] In step S21, it is determined whether the currently input user requirements are complete and clear;

[0147] In step S22, if the input is incomplete, guiding questions are raised, and the process returns to the step of determining whether the currently input user requirements are complete and clear.

[0148] If the judgment is unclear, try to clarify the ambiguous expression and return to the step of determining whether the currently input user requirements are complete and clear;

[0149] When it is judged to be complete and clear, an identification label is added to the currently input user demand and structured.

[0150] In this Figure 2In the method shown, step S20 is used to respond to the acquired user needs, use the large language model to communicate with the user, and guide the user to clearly express their information needs, including but not limited to: technical topics of concern, analysis objectives, scenarios, expected report formats and granularity, the range of data sources that need to be retrieved, etc. In each round of dialogue, it is evaluated whether it is necessary to mobilize the intelligence query agent for external retrieval assistance. Step S21 is used to determine whether the user needs currently input are complete and clear. In this embodiment, the completeness judgment can be to check whether key information dimensions are missing (such as preset dimensions such as technical fields, time ranges, geographical ranges, and analysis targets), and the clarity judgment can be to detect ambiguous expressions (such as "cutting-edge technology", "recent", "good results", and other unquantified or undefined terms). Specifically, in this example, the specific steps of the completeness judgment method may include: presetting key dimension templates, extracting entities using pre-trained models or rule matching, and detecting missing dimensions. The specific steps of the clarity judgment method may include building a fuzzy word library, using a fuzziness scoring model to perform fuzziness scoring, and performing context relevance verification (checking whether fuzzy words are subsequently defined, such as when a user says "cutting-edge technology, specifically AI and blockchain", then clearing the fuzzy mark of "cutting-edge technology").

[0151] Step S22 is used to perform corresponding operations based on the results of the judgment. When it is judged that the currently input user needs are incomplete, it identifies missing information points, actively asks guiding questions (such as "Do you need to focus on a specific technology track?", "Is there a target analysis cycle or region?"), and obtains the currently input user needs, and feeds back to step S21 for judgment. When it is judged that the currently input user needs are unclear, that is, they contain vague expressions, try to clarify the intention (such as "Which specific sub-sectors do you refer to as the 'cutting-edge technology'?"), obtain the currently input user needs, and feed back to step S21 for judgment. By repeatedly interacting with the user until their needs are judged to be complete and clear, add identification tags to the finally confirmed needs, and structure the complete needs.

[0152] In this Figure 6 In the multi-agent collaborative scientific and technological intelligence generation system shown, the intelligence collection agent 2 can be used to perform multi-information source retrieval according to the user's needs to obtain database retrieval content. The workflow diagram of the intelligence collection agent is shown in FIG. Figure 7 In this embodiment, the specific method for obtaining the search content can be various forms known to those skilled in the art. In one example of the present invention, the intelligence collection agent 2 can be used to perform the following steps: Figure 3 The method shown in Figure 3 In the example, the intelligence gathering agent 2 can be used to perform the following steps:

[0153] In step S30, based on user needs, an initial search question is generated using a large language model;

[0154] In step S31, the initial search question is split into multiple sub-search questions using a thinking chain;

[0155] In step S32, the search target of each sub-search question is clarified;

[0156] In step S33, internal vector database search and external search are performed for each sub-search question;

[0157] In step S34, the results retrieved for each sub-search question are quality screened, including: scoring the source authority and the demand matching degree, and fusing the two scores for quality screening;

[0158] In step S35, the effective search content is structured.

[0159] In this Figure 3 In the method shown, step S30 is used to generate the initial search questions using the large language model for the acquired user needs. Step S31 is used to derive sub-search questions using the Chain of Thought (CoT) method, and to form fine-grained sub-questions by splitting complex questions. Specifically, in this example, each sub-question may include: a clear target object (such as "hydrogen fuel cell technology"), a clear focus (such as "development trends in the past five years"), and a quantifiable search direction (such as "changes in the number of domestic and foreign patent applications"). Step S32 is used to clarify the search target of each sub-search question (such as technology development trends, application scenarios, patent layout, etc.). Step S33 is used to search each sub-search question, giving priority to similarity search in the internal vector database (intelligence knowledge base built based on chromadb), and then calling external search tools (DuckDuckGo, full-text crawling tools) for incremental supplementation.

[0160] Step S34 is used to perform quality screening on the results retrieved for each sub-search question. Specifically, in this example, the retrieved results may be scored for authority and demand matching, and then the two scores may be combined for screening. In this embodiment, the specific method for performing authority scoring in step S34 may be various forms known to those skilled in the art. In one example of the present invention, it may include the following steps:

[0161] First, a scoring method is integrated to perform scoring, and the scoring method may include: source type identification, domain name / author matching and / or citation index. Among them, the source type identification scoring (SourceScore) method may be: if it is .gov, .org, or the official website of a well-known scientific research institution, the corresponding score SourceScore may be set to a value within the preset first interval; if it is a news media, forum, or personal blog, the corresponding score SourceScore may be set to a value within the preset second interval, and the minimum value of the first interval is greater than the maximum value of the second interval. The domain name / author matching scoring (AuthorScore) method may be: whether it appears in a trusted whitelist (such as CNIPA, IEEE, Nature); whether it matches a high-reputation author in the expert database. In this example, the domain name / author matching scoring may be The CitationScore method can be: if there are metadata such as impact factor, citation count, report rating, etc., then these data values ​​are normalized and averaged as the citation index. In this example, the document impact factor value can be By normalizing these metrics, the impact of document magnitude differences can be eliminated, ensuring fair comparison. If a document lacks metrics such as citation count or impact factor, a moderate citation score is assigned based on other authority metrics (such as source type and domain / author match). The values ​​of the selected citation metrics are averaged to form the CitationScore. The final score is calculated based on the obtained source type identification score (SourceScore), domain / author match score (AuthorScore), and citation metric score (CitationScore). In this example, the final score can be: AuthorityScore = α1 × SourceScore + α2 × AuthorScore + α3 × CitationScore, where α1, α2, and α3 are weighting coefficients. Each document is assigned an authority score of AuthorityScore ∈ [0, 1]. For example, high authority is defined as AuthorityScore ≥ 0.75.

[0162] In this embodiment, the specific method for performing the requirement matching score in step S34 may be various methods known to those skilled in the art. In one example of the present invention, the method may be to determine the semantic relevance between the search results and the sub-search questions, specifically including the following steps:

[0163] First, embedding semantic matching is performed, using BERT to generate semantic vectors for the query and document, and Cosine similarity is calculated. Next, keyword coverage is analyzed to determine whether the passage covers the target object, time dimension, and key keywords in the query. Finally, a RelevanceScore is output for each document's demand matching.

[0164] In this embodiment, the specific method for fusing the two types of scores for screening in step S34 may be various forms known to those skilled in the art. In one example of the present invention, the method may include the following steps:

[0165] First, the two scores are combined by calling final_score = α * AuthorityScore + β * RelevanceScore. In this example, the weights could be α = 0.4 and β = 0.6. Then, filtering logic is applied. This could be a dual-threshold filter (AuthorityScore ≥ 0.6 and RelevanceScore ≥ 0.7) or a Top-K filter (sort by final_score and retain the top K entries). In this example, the value of K could be 5.

[0166] Step S35 is used to structure and organize the effective search content to generate a final search result package (including document fragments, source information, and search keywords).

[0167] In this Figure 6 In the illustrated multi-agent collaborative scientific and technological intelligence generation system, the vector database configuration of the intelligence collection agent 2 may include the following steps:

[0168] In step S60, the required libraries are imported; the required libraries include: chromadb function library, langchain_community.embeddings function library and langchain.text_splitter function library.

[0169] In step S61, the chromadb client is initialized and a new collection is created; the new collection is used to store vectorized document data, and the specific implementation is as follows:

[0170]

[0171] In step S62, a document vectorization function is defined to read the document content. The document vectorization function is used to convert the document content into a vector using the OpenAI embedding model:

[0172]

[0173] In step S63, a text embedding model is determined based on the document vectorization function and the OpenAI embedding model; the text embedding model is used to convert the text into a high-dimensional vector representation:

[0174]

[0175] In step S64, based on the text embedding model, the document is vectorized and the vectorized document is stored in the chromadb function library to obtain the chromadb vector database:

[0176]

[0177]

[0178] In step S65, the chromadb vector database is configured to the intelligence collection agent through the LangChain framework to realize intelligent retrieval and report generation functions. In this embodiment, the vector database is configured through RAG (Retrieval Augmented Generation), and professional documents related to intelligence analysis (such as research reports, analysis reports, historical cases, etc.) are converted into vector form to achieve long-term memory capabilities, which is convenient for the agent to query and learn. At the same time, chromadb is used as a vector database, and persistent storage (PersistentClient) is used to improve performance. The model used to convert text into vector representation is OpenAI's text-embedding-ada-002 model, which can convert text into 1536-dimensional vector representation and realize efficient similarity retrieval through cosine similarity calculation. The system supports multiple document formats (PDF, DOCX, TXT, URL, etc.) and improves retrieval efficiency through document segmentation.

[0179] In this Figure 6 In the illustrated multi-agent collaborative scientific and technological intelligence generation system, the execution tool configuration of the intelligence collection agent 2 may include a DuckDuckGo search tool, a full-text crawler tool, and a quality screening tool. DuckDuckGo is an internet search engine that fully utilizes the APIs of various network services to provide users with authoritative and accurate results. Specifically, in this example, calling the API may include the following steps:

[0180]

[0181] For links where it is necessary to obtain the full text content, the full text crawler is called. The full text crawler is defined as:

[0182]

[0183]

[0184] To improve the efficiency and accuracy of intelligence-gathering agents processing massive, multi-sourced results, the system integrates a filter module that can be called by the agent. This module, packaged as a modular API, can be called by each sub-question to perform a quality screening process, ensuring that the output results meet the quality assurance requirements of both source authority and semantic matching. Specifically, in this example, the specific method for configuring the filter module may include the following steps:

[0185] In step S70, the screening tool input is performed, and the screening tool receives the following structured input:

[0186] query: natural language expression of the sub-question;

[0187] candidates: a list of preliminary search results from the internal knowledge base or external network (in the format of document paragraph + meta information, such as source URL, publication time, etc.);

[0188] config: optional parameters (weight settings, source whitelist, etc.);

[0189] In step S71, a filter pipeline is defined. The filter tool determines the quality of candidate information and selects the best candidate information through the following steps:

[0190] Step 1: Authority Scoring

[0191] The authority_score(doc) function is called for each candidate result, and a score is calculated based on the following criteria: Source type identification: .gov, .org, or the official website of a well-known scientific research institution receives a higher score; news media, forums, or personal blogs receive a lower score. Domain / author matching: whether the document appears on trusted whitelists (such as CNIPA, IEEE, and Nature); and whether it matches a highly reputable author in the expert database. Citation metrics (optional): If metadata such as impact factor, citation count, and report rating are available, they are weighted and taken into account. The final output is an AuthorityScore∈[0,1] for each document, with high authority being an AuthorityScore ≥ 0.75.

[0192] Step 2: Relevance Scoring

[0193] Call the relevance_score(query, doc) function for each candidate passage to determine its semantic relevance to the subquestion: Embedding semantic matching: Use BERT to generate semantic vectors for the query and doc, and calculate cosine similarity. Keyword coverage analysis: Determine whether the passage covers the target object, time dimension, and focus keywords in the query.

[0194] Step 3: Quality score fusion and screening

[0195] Call final_score = α * AuthorityScore + β * RelevanceScore to combine the two scores, with default weights of α = 0.4 and β = 0.6 (adjustable). Then, perform filtering logic: dual threshold filtering (AuthorityScore ≥ 0.6 and RelevanceScore ≥ 0.7); or Top-K filtering (sort by FinalScore and retain the top K entries (K is set to 5 by default and is configurable).

[0196] The filter tool output is structured as follows:

[0197]

[0198] In the intelligence collection agent workflow, each sub-question will call the screening tool through the following interface:

[0199]

[0200]

[0201] In this Figure 6 In the multi-agent collaborative scientific and technological intelligence generation system shown, the intelligence writing agent 3 can be used to generate initial scientific and technological intelligence based on the user's needs and search content. In this embodiment, the specific method for generating initial scientific and technological intelligence can be various forms known to those skilled in the art. In one example of the present invention, the intelligence writing agent 3 can be used to perform the following steps: Figure 4 The method shown in Figure 4 In the example, the intelligence writing agent 3 can be used to perform the following steps:

[0202] In step S40, the Llama3-8B-Chinese-Chat model is loaded using the LlamaFactory framework as an intelligence writing model;

[0203] In step S41, construct an Alpaca format dataset;

[0204] In step S42, model fine-tuning parameters are set;

[0205] In step S43, the intelligence writing model is trained using the Alpaca format dataset, including: generating prediction results, calculating losses, and updating parameters through the QLoRA adapter;

[0206] In step S44, the fine-tuned adapter weights are saved to obtain a trained intelligence writing model;

[0207] In step S45, user needs and search content are input into the trained intelligence writing model to generate initial scientific and technological intelligence.

[0208] In this Figure 4 In the method shown, step S40 is used to load the Llama3-8B-Chinese-Chat model using the LlamaFactory framework as an intelligence writing model. Step S41 is used to construct an Alpaca format dataset for training the intelligence writing model. Specifically, in this example, the Alpaca format dataset can include four parts: instruction, input, output, and system, where: instruction content is a clear writing task instruction (such as "write a new energy vehicle technology trend report"); input content is the search content or context background related to the task (such as the summary information provided by the intelligence collection agent); output content is the corresponding standard example report; system is used to define the output style and context setting system prompt words. Training data comes from multiple aspects. In this example, the source of training data can be real collected corporate science and technology intelligence demand cases, integrated science and technology research reports, industry analysis reports, white papers and other professional materials, and high-quality training samples hand-crafted based on the final version of the demand and search results. Step S42 is used to set model fine-tuning parameters. Model fine-tuning parameters can include learning rate, batch size, and training rounds. Specifically, in this example, the learning rate can be set to 2e-5, the batch size can be set to 32, and the number of training rounds can be set to 1-5. Step S43 trains the intelligence writing model using an Alpaca format dataset. Specifically, in this example, the steps of training the intelligence writing model may include generating prediction results, calculating losses, and updating parameters via the QLoRA adapter. Step S44 is used to save the fine-tuned adapter weights to obtain the trained intelligence writing model. Step S45 is used to input user requirements and search content into the trained intelligence writing model to generate initial scientific and technological intelligence.

[0209] In this Figure 6In the multi-agent collaborative scientific and technological intelligence generation system shown, the intelligence review agent 4 can be used to review and optimize the initial scientific and technological intelligence based on the search content. In this embodiment, the specific methods of review and optimization can be various forms known to those skilled in the art. In one example of the present invention, the intelligence review agent 4 can be used to perform the following steps: Figure 5 The method shown in Figure 5 In the example, the intelligence review agent 4 can be used to perform the following steps:

[0210] In step S50, the large language model is used to analyze the initial scientific and technological information text content to identify fuzzy information;

[0211] In step S51, the fuzzy information is organized into specific search questions;

[0212] In step S52, a multi-information source search is performed on the specific search question to obtain supplementary data;

[0213] In step S53, the initial scientific and technological information is corrected and improved based on the supplementary data to generate final scientific and technological information.

[0214] In this Figure 5 In the method shown, step S50 is used to identify ambiguous information in the initial scientific and technological intelligence. Specifically, in this example, a large language model can be used to analyze the text content of the initial scientific and technological intelligence to identify ambiguous expressions, unclear conceptual descriptions, insufficient data support or inferential content in the report. Step S51 is used to organize the identified ambiguous or unclear information points into specific search questions, such as: "What is the specific data basis for the XXX trend?", "What are the actual application cases of XXX technology?". Step S52 is used to perform multi-information source retrieval on specific search questions to obtain supplementary data. Multi-information source retrieval can be performed through external knowledge bases and the Internet to obtain the latest and authoritative supplementary materials or data basis. Step S53 is used to correct and improve the initial scientific and technological intelligence based on the supplementary data, clarify the original ambiguous expressions, supplement missing data or cases, improve the accuracy and completeness of the content, and generate the final scientific and technological intelligence.

[0215] Beneficial effects of the present invention:

[0216] The implementation method of the present invention constructs a multi-agent collaborative architecture based on a large language model, and combines information retrieval, semantic analysis, knowledge fusion and content generation capabilities to achieve automatic understanding, organization and expression of scientific and technological information, thereby significantly improving the efficiency, quality and intelligence level of scientific and technological intelligence generation, and providing enterprises with more valuable auxiliary decision support.

[0217] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0218] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0219] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0220] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0221] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0222] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0223] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0224] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0225] The above are merely embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. A multi-agent collaborative scientific and technological intelligence generation method, characterized in that: The method comprises: Obtain user needs; Perform multi-information source search according to the user's needs to obtain database search content; Generate initial scientific and technological intelligence based on the user needs and search content; The initial scientific and technological intelligence is reviewed and optimized based on the search content.

2. The method according to claim 1, characterized in that The obtaining of user needs includes: Respond to acquired user needs and communicate with users using a large language model; Determine whether the currently input user requirements are complete and clear; If the judgment is incomplete, ask guiding questions and return to the step of judging whether the currently input user requirements are complete and clear; If the judgment is unclear, try to clarify the ambiguous expression and return to the step of determining whether the currently input user requirements are complete and clear; When it is judged to be complete and clear, an identification label is added to the currently input user demand and structured.

3. The method according to claim 1, characterized in that Perform multi-information source search based on the user's needs to obtain database search content including: Based on the user needs, generate initial search questions using a large language model; Using thought chaining to split the initial search question into multiple sub-search questions; Clarify the search target of each sub-search question; Conduct internal vector database search and external search for each sub-search question; Perform quality screening on the results retrieved for each sub-search question, including scoring the source authority and demand matching, and integrating the two scores for quality screening; Structure effective search content.

4. The method according to claim 1, wherein Generating initial scientific and technological intelligence based on the user needs and search content includes: Use the LlamaFactory framework to load the Llama3-8B-Chinese-Chat model as the intelligence writing model; Build Alpaca format dataset; Set model fine-tuning parameters; Training the intelligence writing model using the Alpaca format dataset, including generating prediction results, calculating losses, and updating parameters through the QLoRA adapter; Save the fine-tuned adapter weights to obtain the trained intelligence writing model; The user needs and search content are input into the trained intelligence writing model to generate initial scientific and technological intelligence.

5. The method according to claim 1, wherein Reviewing and optimizing the initial scientific and technological intelligence based on the search content includes: Using a large language model to analyze the content of the initial scientific and technological intelligence and identify ambiguous information; Arrange the fuzzy information into specific search questions; Conducting a multi-source search on the specific search question to obtain supplementary data; Based on the supplementary data, the initial scientific and technological intelligence is corrected and improved to generate final scientific and technological intelligence.

6. A multi-agent collaborative scientific and technological intelligence generation system, characterized by: include: The intelligence communication agent is built using a large language model and runs in a multi-agent task orchestration framework based on LangChain. It is used to conduct multiple rounds of dialogue with demand providers to obtain user needs. An intelligence collection agent, comprising a vector database and an execution tool set, wherein the vector database is used to convert professional documents related to intelligence analysis into vector form, the execution tool set is used to provide the intelligence collection agent with tools to perform corresponding operations, and the intelligence collection agent is used to perform multi-information source retrieval according to the user's needs to obtain database retrieval content; An intelligence writing agent, used to generate initial scientific and technological intelligence according to the user needs and search content; An intelligence review agent is used to review and optimize the initial scientific and technological intelligence based on the search content.

7. The system according to claim 6, characterized in that The intelligence communication agent is used to: Respond to acquired user needs and communicate with users using a large language model; Determine whether the currently input user requirements are complete and clear; If the judgment is incomplete, ask guiding questions and return to the step of judging whether the currently input user requirements are complete and clear; If the judgment is unclear, try to clarify the ambiguous expression and return to the step of determining whether the currently input user requirements are complete and clear; When it is judged to be complete and clear, an identification label is added to the currently input user demand and structured.

8. The system according to claim 6, wherein: The intelligence gathering agent is used to: Based on the user needs, generate initial search questions using a large language model; Using thought chaining to split the initial search question into multiple sub-search questions; Clarify the search target of each sub-search question; Conduct internal vector database search and external search for each sub-search question; Perform quality screening on the results retrieved for each sub-search question, including scoring the source authority and demand matching, and integrating the two scores for quality screening; Structure effective search content.

9. The system according to claim 6, wherein: The intelligence writing agent is used to: Use the LlamaFactory framework to load the Llama3-8B-Chinese-Chat model as the intelligence writing model; Build Alpaca format dataset; Set model fine-tuning parameters; Training the intelligence writing model using the Alpaca format dataset, including generating prediction results, calculating losses, and updating parameters through the QLoRA adapter; Save the fine-tuned adapter weights to obtain the trained intelligence writing model; The user needs and search content are input into the trained intelligence writing model to generate initial scientific and technological intelligence.

10. The system according to claim 6, wherein: The intelligence review agent is used to: Using a large language model to analyze the content of the initial scientific and technological intelligence and identify ambiguous information; Arrange the fuzzy information into specific search questions; Conducting a multi-source search on the specific search question to obtain supplementary data; Based on the supplementary data, the initial scientific and technological intelligence is corrected and improved to generate final scientific and technological intelligence.

Images