Intelligence information integration method and device based on multi-agent task chain strategy large model

Abstract

The invention provides an intelligence information integration method and device based on a multi-agent task chain strategy large model, and the method comprises the steps: collecting a plurality of pieces of open-source dynamic science and technology intelligence information of a target information source, and employing a literature analysis agent, an abstract induction agent, a content generation agent, a sequencing agent and a report generation agent which are constructed through a large language model; key information extraction, semantic analysis and induction, content evaluation and updating, information sequencing and version change recording and report generation are performed respectively, and automatic aggregation and sequencing of intelligence information are realized. According to the method, the defects of generalization, objectivity, cross-document information processing, flexibility, robustness, illusion rate, extrapolation ability, small sample learning ability and the like in intelligence information integration in the prior art are effectively overcome, the efficiency and effect of intelligence work are improved, and the intellectual property is improved. The method is of great significance in enhancing the service quality benefit of dynamic intelligence information and supporting strategic decision and policy making.

Description

Technical Field

[0001] The present invention relates to the technical field of information integration and analysis, and particularly to an intelligence information integration method and device based on a multi-agent task chain strategy large model. Background Art

[0002] In the context of the rapid development of globalization and informatization in today's era, the importance of scientific and technological intelligence has become increasingly prominent, playing a key role in assisting decision-making and scientific and technological strategic planning. However, the current scientific and technological intelligence information presents the characteristics of wide sources and diverse formats, resulting in severe challenges in information management. The phenomenon of "scattered" information not only affects the quality and usability of scientific and technological intelligence but also increases the difficulty for decision-makers to screen useful information from massive data. Although existing classical machine learning algorithms can assist researchers in understanding and organizing information, their performance is limited by the quality and relevance of input data, the generalization ability of the model is limited, feature engineering depends on the professional level of experts, there are subjectivity and one-sidedness, and the complexity of the model is limited, making it difficult to mine complex cross-document information. Although deep learning technology shows potential for information integration, extractive integration lacks consistency and flexibility, and heuristic integration requires a large amount of training data and the robustness of the model is limited. Both lack the extrapolation ability for different document types. Although large language model technology has powerful learning and generalization abilities, in information integration tasks, its zero-shot prompting is difficult to accurately extract target key information, the extracted information is redundant, and the hallucination characteristic reduces the reliability of the model, hindering its application in intelligence analysis. Therefore, there is an urgent need to propose an intelligence information integration method that can effectively solve the above problems to improve the efficiency and effectiveness of intelligence work and better assist in optimizing the strategic decision-making and policy-making processes. Summary of the Invention

[0003] In view of this, embodiments of the present invention provide an intelligence information integration method and device based on a multi-agent task chain strategy large model to eliminate or improve one or more defects existing in the prior art and solve the problems of lack of flexibility and performance dependence on data quality in existing intelligence information integration technologies.

[0004] One aspect of the present invention provides an intelligence information integration method based on a multi-agent task chain strategy large model, the method comprising the following steps:

[0005] Collect a plurality of open-source dynamic scientific and technological intelligence information of target information sources, and the target information sources are divided into multiple fields according to content attributes, dissemination purposes, audience ranges or information forms;

[0006] Construct a literature analysis agent based on a large language model, and extract first target key information from the open-source dynamic scientific and technological intelligence information according to a preset first guiding prompt text;

[0007] Construct an abstract induction agent based on the large language model, perform semantic analysis on the first target key information according to the second guiding prompt text, and structurally output the second target key information after abstract induction;

[0008] Construct a content generation agent based on the large language model to obtain the latest standard literature on the second target key information through a preset link, evaluate the positioning difference part of the target key information according to the standard literature and update it to obtain the third target key information;

[0009] Construct an ordering agent based on the large language model to order the third target key information according to the third guiding prompt text according to a preset expression logic to obtain the fourth target key information;

[0010] Construct a report generation agent based on the large language model to perform difference comparison on the fourth target key information generated in each batch according to the fourth guiding prompt text, record the version changes and generate an update report.

[0011] In some embodiments, the method further includes:

[0012] Collect multiple open-source dynamic scientific and technological intelligence information of the target information source by means of subscription push, manual collection and retrieval matching;

[0013] Or, use a customized crawler tool to regularly crawl multiple open-source dynamic scientific and technological intelligence information of the target information source;

[0014] Or, integrate the API interfaces provided by multiple target information sources and perform unified management and retrieval with the help of the Elasticsearch platform;

[0015] Wherein, the target information sources include news media, academic research, technology blogs and social networks.

[0016] In some embodiments, the large language model uses the Qwen-72B model, Qwen-7B model, Qwen-14B model, Qwen-32B model, open-source Llama model or generalized linear model as the base.

[0017] In some embodiments, the first guiding prompt text, the second guiding prompt text, the third guiding prompt text and the fourth guiding prompt text all include a background definition part, a main body description part, an output constraint part, and a process and example inspiration part;

[0018] The background definition part includes an overview of the role and background for performing tasks in the current problem scenario;

[0019] The main body description part includes a task overview for the current problem scenario and a skill description for solving the problem;

[0020] The output constraint part includes a limitation on the output format and an overview of the constraint conditions;

[0021] The process and example inspiration part includes hints on the steps of the processing flow and provided cases.

[0022] In some embodiments, the method further includes:

[0023] Dynamically adjust the collection frequency of the open-source dynamic technology intelligence information for the corresponding field according to the data update frequency of the target information sources in multiple fields;

[0024] And configure dynamic weights for the target information sources in different fields, and adjust the crawling priority based on the timeliness, authority, and user feedback of each field.

[0025] In some embodiments, the method further includes:

[0026] Establish an intermediate cache database for caching the first target key information, the second target key information, the third target key information, and the fourth target key information, and establish an index for repeated scheduling and querying.

[0027] In some embodiments, the method further includes:

[0028] Establish an error feedback channel, and when a downstream intelligent agent detects a logical contradiction, trigger the reprocessing of the upstream intelligent agent.

[0029] On the other hand, the present invention also provides an intelligence information integration device based on a multi-agent task chain strategy large model, including a processor, a memory, and a computer program / instructions stored on the memory. It is characterized in that the processor is used to execute the computer program / instructions, and when the computer program / instructions are executed, the device implements the steps of the above method.

[0030] On the other hand, the present invention also provides a computer-readable storage medium, on which computer program / instructions are stored, and when the computer program / instructions are executed by a processor, the steps of the above method are implemented.

[0031] On the other hand, the present invention also provides a computer program product, including computer program / instructions, and when the computer program / instructions are executed by a processor, the steps of the above method are implemented.

[0032] The beneficial effects of the present invention are at least:

[0033] The present invention provides an intelligence information integration method and device based on a large model of multi-agent task chain strategy. By collecting multiple open-source dynamic scientific and technological intelligence information from target information sources, and using literature analysis agents, abstract induction agents, content generation agents, serialization agents, and report generation agents constructed by large language models to extract key information, conduct semantic analysis and induction, evaluate and update content, serialize information, and generate version change records and reports respectively, the automatic aggregation and serialization of intelligence information are achieved. This method effectively solves the deficiencies in intelligence information integration in the prior art, such as generalization, objectivity, cross-document information processing, flexibility, robustness, hallucination rate, extrapolation ability, and few-shot learning ability, etc., improves the efficiency and effectiveness of intelligence work, and is of great significance for enhancing the quality and benefits of dynamic intelligence information services, supporting strategic decision-making and policy formulation.

[0034] Additional advantages, objects, and features of the present invention will be partially described below and will become partially apparent to those of ordinary skill in the art after studying the following text, or may be learned from the practice of the present invention. The objects and other advantages of the present invention can be realized and obtained by the structure specifically pointed out in the specification and the drawings.

[0035] Those skilled in the art will understand that the objects and advantages that can be achieved by the present invention are not limited to the above specifically described, and the above and other objects that the present invention can achieve will be more clearly understood according to the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The drawings described herein are used to provide a further understanding of the present invention, form a part of this application, and do not limit the present invention. In the drawings:

[0037] Figure 1 It is a schematic flowchart of the intelligence information integration method based on the large model of multi-agent task chain strategy according to an embodiment of the present invention.

[0038] Figure 2 It is a schematic flowchart of the intelligence information integration method based on the large model of multi-agent task chain strategy according to another embodiment of the present invention.

[0039] Figure 3 It is a management diagram of each agent in the intelligence information integration method based on the large model of multi-agent task chain strategy according to another embodiment of the present invention.

[0040] Figure 4 It is a schematic diagram of the prompt engineering structure in the intelligence information integration method based on the large model of multi-agent task chain strategy according to another embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0041] To make the objectives, technical solutions and advantages of the present invention more clear and understandable, the present invention will be further described in detail below in conjunction with the embodiments and the drawings. Herein, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but do not limit the present invention.

[0042] Herein, it should also be noted that in order to avoid obscuring the present invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the present invention are shown in the drawings, while other details less relevant to the present invention are omitted.

[0043] It should be emphasized that the term "comprising / including" as used herein refers to the presence of features, elements, steps or components, but does not exclude the presence or addition of one or more other features, elements, steps or components.

[0044] The integration process of intelligence information can be regarded as a process of cross-document key information integration. However, the existing cross-document key information integration technologies are mainly two categories, namely extraction-based integration and heuristic integration. Extraction-based integration directly extracts target information from the original documents and integrates it according to certain rules; heuristic integration automatically generates target information through a specified sentence. Although these existing methods have shown great potential for information integration empowered by deep learning technologies, they still face challenges. First, the integrated information generated by the extraction-based method lacks consistency and flexibility; second, the heuristic method requires a large amount of training data, and the model performance depends on the data quality, with limited robustness. Third, due to the need for a large amount of domain-specific training data, both methods lack the extrapolation ability for different document types. For example, the model may be more suitable for integrating news documents, but it is difficult to achieve an ideal processing performance for research documents.

[0045] Large Language Models (LLMs) benefit from a large network scale and rich training corpora, demonstrating powerful learning, generalization, and emergence capabilities, which endow new impetus to break through the above limitations. Although the general expansion ability and few-shot learning ability of LLMs technology show great potential in solving the above problems, they also face challenges. Compared with single-document key information extraction, information integration tasks pay more attention to the interaction and integration of cross-document knowledge. Since LLMs have a relatively long context window, classical zero-shot prompts are difficult to stimulate the model's accurate extraction ability of target key information, and the extracted information often has a lot of redundancy. Furthermore, the redundant information will further interfere with the model's information integration effect. In addition, the hallucination feature is also one of the problems restricting the application of LLMs in intelligence analysis. The inherent hallucination feature of LLMs reduces the reliability of the model and hinders the organic embedding of LLMs technology in intelligence information integration. Therefore, an intelligence information integration method with good generalization, strong objectivity, the ability to analyze cross-document information, high flexibility, strong robustness, low hallucination rate, extrapolation ability, and few-shot learning ability is needed to enable large models to effectively empower the integration process of dynamic intelligence information, enhance the quality and efficiency of dynamic intelligence information services, and improve the efficiency and effectiveness of intelligence work.

[0046] Specifically, the present invention provides an intelligence information integration method based on a large model of a multi-agent task chain strategy, as Figure 1 shown, the method includes the following steps S101 to S106:

[0047] Step S101: Collect multiple open-source dynamic scientific and technological intelligence information from target information sources, and the target information sources are divided into multiple fields according to content attributes, dissemination purposes, audience ranges, or information forms.

[0048] Step S102: Construct a literature analysis agent based on a large language model, and extract first target key information from the open-source dynamic scientific and technological intelligence information according to a preset first guiding prompt text.

[0049] Step S103: Construct an abstract induction agent based on a large language model, perform semantic analysis on the first target key information according to a second guiding prompt text, and structurally output second target key information after abstract induction.

[0050] Step S104: Construct a content generation agent based on a large language model to obtain the latest standard literature on the second target key information through a preset link, evaluate and update the different parts of the target key information positioning according to the standard literature, and obtain third target key information.

[0051] Step S105: Build an ordering agent based on the large language model to order the third target key information according to the third guiding prompt text according to the preset expression logic to obtain the fourth target key information.

[0052] Step S106: Build a report generation agent based on the large language model to perform differential comparison on the fourth target key information generated in each batch according to the fourth guiding prompt text, record the version changes and generate an update report.

[0053] In step S101, the method may include: collecting a plurality of open-source dynamic scientific and technological intelligence information of the target information source by means of subscription push, manual collection and retrieval matching. Or, using a customized crawler tool to regularly crawl a plurality of open-source dynamic scientific and technological intelligence information of the target information source. Or, integrating the API interfaces provided by multiple target information sources and performing unified management and retrieval with the help of the Elasticsearch platform. Among them, the target information sources include news media, academic research, technology blogs and social networks.

[0054] Introduce a multi-source collection strategy, use the Scrapy framework to build a distributed crawler cluster, configure different crawler templates to adapt to each information source (such as RSS subscription crawling, API polling request), and use the Kafka message queue to achieve real-time data stream buffering to prevent data loss in high-concurrency scenarios.

[0055] The goal of this step is to collect open-source dynamic scientific and technological intelligence information from different information sources. These information sources can be divided into multiple fields according to content attributes, communication purposes, audience scope or information form, such as news, academic articles, technology blogs, social media, etc.

[0056] In step S102, the goal is to build a literature analysis agent to extract the first target key information from the collected open-source dynamic scientific and technological intelligence information by using a large language model (such as Qwen-72B, Llama, etc.). Based on the large language model, the function of the agent is realized through prompt engineering. The first guiding prompt text is required to design clear and precise prompt words to ensure that the model can accurately understand the task requirements.

[0057] Exemplarily, use the prompt word "Extract the key information related to artificial intelligence from the following literature: {input literature}", the input literature is "In January 2024, a certain company released a new product", and the output is "Key information: Release a new product, Time: January 2024".

[0058] In step S103, the original extracted scattered "first target key information" is transformed into structured "second target key information" with logic and operability, which specifically includes: understanding semantics, breaking through literal matching, and mining the implicit associations of technical terms, such as the inheritance relationship between "ViT" and "Transformer architecture". Abstracting knowledge, extracting general rules from specific cases, such as summarizing "three major technical routes for few-shot learning" from 10 papers. Making a structured expression, converting free text into a machine-parsable format, such as JSON Schema, to support downstream automated processing. This step is a key leap from a data heap to knowledge construction. Through semantic understanding and structured reorganization, noise is eliminated, irrelevant details are filtered out, and core knowledge elements are focused on; associations are established to reveal the technical evolution path; decision-making is empowered, and structured data can directly drive advanced applications such as automated reports and trend predictions.

[0059] In step S104, the goal of this step is to construct a content generation agent that uses a large language model to obtain the latest standard literature on the second target key information through a preset link, evaluate the target key information, locate the different parts, and update them to obtain the third target key information. Exemplarily, use the prompt "Update the information according to the following standard literature: {reference}, update information: {input information}", the reference is "The new aircraft uses the latest structure, reduces weight, and only uses 30,000 screws", and the input information is "The aircraft uses more than 60,000 screws and uses a steam launch method", and the output is "Updated content: The aircraft uses more than 30,000 screws and uses an electromagnetic catapult launch method".

[0060] In step S105, the goal is to construct an ordering agent that uses a large language model to order the third target key information according to a preset expression logic to obtain the fourth target key information. Based on the large language model, the function of the agent is realized through prompt engineering, and prompt words are designed to guide the model to order according to time, event logic, etc. Exemplarily, use the prompt "Order the following information according to the time logic: {input information}", and the input information is "Event A: In December 2023, the company obtained a patent. Event B: In January 2024, the company released a new product.", and the output is "Ordering result: In December 2023, the company obtained a patent; in January 2024, the company released a new product".

[0061] In step S106, the goal is to construct a report generation agent that uses a large language model to compare the differences in the fourth target key information generated in each batch, record the version changes, and generate an update report. Based on the large language model, the function of the agent is realized through prompt engineering. Design prompt words to guide the model to perform difference comparison and report generation.

[0062] Exemplarily, using the prompt "Perform a differential comparison on the following information: {input information}", where the input information is "Old version: The aircraft used more than 60,000 screws and adopted a steam launch method. New version: The aircraft used more than 30,000 screws and adopted an electromagnetic catapult launch method", the output is "Differences: 1. The number of screws used has decreased. 2. The launch method has changed. Updated content: The aircraft used more than 30,000 screws and adopted an electromagnetic catapult launch method".

[0063] In some embodiments, the large language model uses the Qwen-72B model, Qwen-7B model, Qwen-14B model, Qwen-32B model, open-source Llama model, or generalized linear model as the base.

[0064] In some embodiments, the first guiding prompt text, the second guiding prompt text, the third guiding prompt text, and the fourth guiding prompt text all include a background definition part, a main body description part, an output constraint part, and a process and example inspiration part.

[0065] The background definition part includes an overview of the role and background for performing tasks in the current problem scenario.

[0066] The main body description part includes an overview of the task for the current problem scenario and a description of the skills for solving the problem.

[0067] The output constraint part includes a limitation on the output format and an overview of the constraint conditions.

[0068] The process and example inspiration part includes hints on the steps of the processing flow and provided cases.

[0069] In some embodiments, the method further includes step S201 and step S202:

[0070] Step S201: Dynamically adjust the collection frequency of open-source dynamic scientific and technological intelligence information for the corresponding field according to the data update frequency of the target information sources in multiple fields.

[0071] Step S202: Configure dynamic weights for the target information sources in different fields and adjust the crawling priority based on the timeliness, authority, and user feedback of each field.

[0072] In step S201, according to the data update frequency of the target information sources, dynamically adjust the collection frequency of open-source dynamic scientific and technological intelligence information for the corresponding field. By real-time monitoring the update situation of the data source, optimize the collection frequency to ensure the timeliness and effectiveness of the information.

[0073] Monitor the data update status of information sources in various fields in real time, count the update frequencies of information sources in various fields, identify fields with frequent updates and fields with fewer updates, and dynamically adjust the collection frequency according to the analysis results. For example, for fields with frequent updates, increase the collection frequency; for fields with fewer updates, decrease the collection frequency. Exemplarily, if the information sources in the news field are updated multiple times a day, while government reports are updated once a month, the collection frequency for the news field can be set to multiple times a day, and the collection frequency for government reports can be set to once a month.

[0074] In step S202, configure dynamic weights for the target information sources in different fields, and adjust the crawling priorities based on the timeliness, authority, and user feedback of each field. Through the configuration of dynamic weights, ensure that information in important fields is obtained and processed first. Evaluate the importance of each field according to its timeliness, authority, and user feedback, set an initial weight value for each field, and dynamically adjust the weight value according to the changes in the field characteristics monitored in real time. For example, if the timeliness of a certain field suddenly increases, increase its weight. Exemplarily, for the news field, due to its high timeliness, the weight can be set to a relatively high value; for the academic database field, due to its strong authority, the weight can also be set to a relatively high value.

[0075] In some embodiments, the method further includes: establishing an intermediate cache database for caching the first target key information, the second target key information, the third target key information, and the fourth target key information, and establishing an index for repeated scheduling and querying.

[0076] In some embodiments, the method further includes: establishing an error feedback channel, and when a downstream agent detects a logical contradiction, triggering the reprocessing of an upstream agent.

[0077] On the other hand, the present invention also provides an intelligence information integration device based on a multi-agent task chain strategy large model, including a processor, a memory, and a computer program / instruction stored on the memory. The processor is used to execute the computer program / instruction, and when the computer program / instruction is executed, the device implements the steps of the above method.

[0078] The present invention is described below in conjunction with a specific embodiment:

[0079] This embodiment provides an intelligence information integration method based on a multi-agent task chain strategy large model, as Figure 2 shown, the association relationship between different agents is as Figure 3 shown, and the method includes:

[0080] S1. Collect open-source dynamic technology intelligence information from target information sources.

[0081] Specifically, the target information sources include, but are not limited to, original texts in different fields such as news, academic articles, technical blogs, and social media.

[0082] Specifically, the collection process uses an automatic push algorithm, manual collection, or a retrieval matching algorithm.

[0083] S2. Build a large model literature analysis agent to finely extract target-related key information from the framed information sources.

[0084] Specifically, the above-mentioned structured key information related to the target at a fine-grained level, and the target key information can be timestamp information, key events, etc.

[0085] Preferably, this step builds a large model literature analysis agent, which is implemented based on prompt engineering using LLMs, that is, the content to be processed and the prompt words of prompt engineering are jointly input into the large model to obtain the target output.

[0086] Preferably, the above-mentioned LLMs adopt the Qwen-72B model, or other versions of the Qwen model can also be used: various versions of Qwen-7B, various versions of Qwen-14B, various versions of Qwen-32B; or the open-source Llama model: various versions of llama-8b, various versions of llama-13b, various versions of llama-70b models; and generative interactive models such as the GLM model can all be used as alternatives to the base model in this technology.

[0087] Preferably, the content of prompt engineering should meet the following requirements: (1) Clear and concise, without using any unnecessary or confusing terms, using easy-to-understand language, with precise and clear wording to reduce the probability of LLMs misunderstanding and suppress hallucinations; (2) Clearly express the final expectation, and clearly explain the expected result or goal of the task in detail; (3) Focus on key details, reduce noise interference, and avoid adding unnecessary information or details that are unimportant to the task to prevent description chaos.

[0088] Preferably, the above-mentioned prompt engineering structure is as Figure 4 shown. The overall prompt engineering consists of background definition, main body description, output constraint, process, and example inspiration.

[0089] Preferably, the above-mentioned background definition includes role overview and background overview.

[0090] Preferably, the content of the role overview in the above-mentioned background definition is "literature analysis expert".

[0091] Preferably, the content of the background overview in the above-mentioned background definition is "The user needs to perform fine-grained information extraction from multiple documents, especially target-related key information and corresponding time information".

[0092] Preferably, the above-mentioned main body description includes a task overview and a skill description.

[0093] Preferably, the content of the task overview in the above-mentioned main body description is "You are a professional literature analysis expert, with the ability to deeply understand and analyze complex documents, and can identify and extract key information and its time markers. You can help users accurately extract key information related to the target and its time information from multi-document information sources".

[0094] Preferably, the content of the skill description in the above-mentioned main body description is "You need to possess skills in literature retrieval, information extraction, time series analysis, and key information identification".

[0095] Preferably, the above-mentioned output constraints include an output format and a constraint overview.

[0096] Preferably, the content of the output format in the above-mentioned output constraints is "Structured text output, including key information and corresponding time markers".

[0097] Preferably, the content of the constraint overview in the above-mentioned output constraints is "This process needs to be efficient and accurate, and can adapt to different fields and types of documents".

[0098] Preferably, the above-mentioned process and example inspiration include a work process and example tips.

[0099] Preferably, the work process in the above-mentioned process and example inspiration is as follows:

[0100] S2-1. Determine the document set and analysis target;

[0101] S2-2. Apply text analysis techniques to identify key information and its context in the documents;

[0102] S2-3. Extract the time markers of the key information and verify them;

[0103] S2-4. Organize and present the extracted information to ensure the accuracy and readability of the information.

[0104] Preferably, the content of the example tips in the above-mentioned process and example inspiration is as follows:

[0105] "Document 1: In January 2024, a certain company launched a new product.

[0106] Key information: Launch a new product, Time: January 2024

[0107] Document 2: The company obtained an important patent in December 2023.

[0108] Key information: Obtain a patent, Time: December 2023."

[0109] After determining the specific prompting engineering, the above prompting content and the following task instructions together constitute the final prompting content: Please extract the corresponding information related to {target topic} from the literature {input}.

[0110] Specifically, the above {input} represents the information to be analyzed screened from the defined information sources, and the above {target topic} represents the target topic that the user is interested in.

[0111] S3. Construct a large model information abstraction and induction agent to abstract and induce relevant key information and output it in a structured manner.

[0112] Preferably, this step constructs a large model abstraction and induction agent, which is implemented using LLMs based on prompting engineering, that is, the content to be processed and the prompting words of the prompting engineering are jointly input into the large model to obtain the target output.

[0113] Preferably, the above LLMs adopt the Qwen-72B model, and other versions of the Qwen model can also be used: various versions of Qwen-7B, various versions of Qwen-14B, various versions of Qwen-32B; or the open-source Llama model: various versions of llama-8b, various versions of llama-13b, various versions of llama-70b models; and generative interaction models such as the GLM model can all be used as alternatives to the base model in this technology.

[0114] Preferably, the above prompting engineering structure is also as Figure 4 shown. The overall prompting engineering consists of background definition, main body description, output constraints, process and example inspiration.

[0115] Preferably, the above background definition includes role overview and background overview.

[0116] Preferably, the role overview content in the above background definition is "literature analysis expert".

[0117] Preferably, the background overview content in the above background definition is "The user needs to conduct literature analysis and abstract and induce the events in multiple fine-grained key information into a concise and refined result".

[0118] Preferably, the above main body description includes task overview and skill description.

[0119] Preferably, the task overview content in the above main body description is "You are an experienced literature analysis expert and can identify and summarize the same or similar events and semantic points".

[0120] Preferably, the skill description content in the above main body description is "You need to possess semantic analysis and inductive summary skills".

[0121] Preferably, the above output constraints include output format and constraint overview.

[0122] Preferably, the output format content in the above output constraint is "The results should be presented item by item in a clear text form, including semantic induction."

[0123] Preferably, the constraint overview content in the above output constraint is "This process needs to be efficient and accurate, capable of processing a large amount of fine-grained key information, and able to identify subtle semantic differences to ensure that the information is not changed."

[0124] Preferably, the above process and example inspiration include a work process and example prompts.

[0125] Preferably, the work process in the above process and example inspiration is as follows:

[0126] S3-1. Read multiple pieces of fine-grained information and identify key information;

[0127] S3-2. Use semantic analysis techniques to identify the same or similar semantic points;

[0128] S3-3. Abstract and summarize the key information to form a unified and itemized expression.

[0129] Preferably, the example prompt content in the above process and example inspiration is as follows:

[0130] "Literature A: More than fifty thousand stars adorn the night sky.

[0131] Literature B: More than fifty thousand stars twinkle in the night sky.

[0132] Literature C: The global average temperature has risen.

[0133] Literature D: It is reported that the global temperature has generally risen by 1 degree Celsius.

[0134] Semantic induction:

[0135] 1. There are more than fifty thousand stars in the night sky.

[0136] 2. The global temperature has risen by 1 degree Celsius."

[0137] After determining the specific prompting engineering, the above prompt content and the following task instructions together form the final prompt content: Analyze the following fine-grained key information: {input}.

[0138] Specifically, the above {input} represents the output in step S2.

[0139] S4. Construct an agent for generating the content of the large model information evaluation, evaluate the impact of the above-extracted and summarized information on the existing content based on the latest literature, locate the content that needs to be updated, and update the existing file content to obtain a series of refined and updated key target information.

[0140] Preferably, this step constructs an intelligent agent for generating large model information evaluation content, which is implemented based on prompt engineering using LLMs, that is, the content to be processed and the prompt words of prompt engineering are jointly input into the large model to obtain the target output.

[0141] Preferably, the above LLMs adopt the Qwen-72B model, and other versions of the Qwen model can also be used: various versions of Qwen-7B, various versions of Qwen-14B, various versions of Qwen-32B; or the open-source Llama model: various versions of llama-8b, various versions of llama-13b, various versions of llama-70b models; and generative interaction models such as the GLM model can all be used as alternatives to the base model in this technology.

[0142] Preferably, the above prompt engineering structure is also as Figure 4 shown. The overall prompt engineering consists of background definition, main body description, output constraints, process and example inspiration.

[0143] Preferably, the above background definition includes role overview and background overview.

[0144] Preferably, the content of the role overview in the above background definition is "Information Evaluation and Update Expert".

[0145] Preferably, the content of the background overview in the above background definition is "The user needs to evaluate the existing documents to determine whether they need to be updated according to the latest authoritative literature, mark the content that needs to be updated, and output the update results".

[0146] Preferably, the above main body description includes task overview and skill description.

[0147] Preferably, the content of the task overview in the above main body description is "You are a professional information evaluation and update expert, with the ability to deeply analyze literature and evaluate information, and can quickly identify key update points and update relevant documents".

[0148] Preferably, the content of the skill description in the above main body description is "You need to possess skills such as information analysis, content evaluation, key point identification, and marking skills".

[0149] Preferably, the above output constraints include output format and constraint overview.

[0150] Preferably, the content of the output format in the above output constraints is "The results should include a list of update points and a detailed description of each point, as well as the corresponding literature citations, and at the same time output the updated results".

[0151] Preferably, the content of the constraint overview in the above output constraints is "The evaluation process needs to be based on the latest and authoritative literature to ensure the accuracy and timeliness of information".

[0152] Preferably, the above process and example inspiration include a workflow and example prompts.

[0153] Preferably, the workflow in the above process and example inspiration is as follows:

[0154] S4-1. Determine evaluation criteria and key areas;

[0155] S4-2. Retrieve and review the latest authoritative literature;

[0156] S4-3. Compare existing documents with the latest literature to identify differences and update points;

[0157] S4-4. Mark and update the content that needs to be updated, and provide literature support at the same time.

[0158] Preferably, the content of the example prompts in the above process and example inspiration is as follows:

[0159] "Content of existing document: The aircraft uses more than 60,000 screws and adopts a steam launch method.

[0160] Content of authoritative document A: The new aircraft adopts the latest structure, reduces weight, and only uses 30,000 screws.

[0161] Content of authoritative document B: The aircraft uses an electromagnetic catapult for launch.

[0162] Content of authoritative document C: The aircraft has increased support for the helmet display.

[0163] Update point 1: Aircraft structure, reduction in the number of screws used.

[0164] Literature citation: Content of authoritative document A

[0165] Update point 2: Change in launch method.

[0166] Literature citation: Content of authoritative document B

[0167] Update point 3: Increased support for the helmet display.

[0168] Literature citation: Content of authoritative document C

[0169] Updated content: The aircraft uses more than 30,000 screws, adopts an electromagnetic catapult launch method, and has increased support for the helmet display."

[0170] After determining the specific prompt project, the above prompt content and the following task instructions together form the final prompt content: Analyze the following authoritative literature {References} and update the information {Input}.

[0171] Specifically, the above-mentioned {reference documents} represent the latest authoritative documents in the relevant field, and the above-mentioned {input} represents the output of step S3.

[0172] S5. Construct a large model information serialization agent to adaptively serialize the above series of refined and updated key target information according to a certain logic to form the final information integration result.

[0173] Specifically, the above-mentioned adaptive serialization of the above series of refined and updated key target information according to a certain logic to form the final information integration result, and the logic can include time logic, event logic, context logic, etc.

[0174] Preferably, this step constructs a large model information serialization agent, which is implemented based on prompt engineering and uses LLMs, that is, the content to be processed and the prompt words of prompt engineering are jointly input into the large model to obtain the target output.

[0175] Preferably, the above-mentioned LLMs adopt the Qwen-72B model, and other versions of the Qwen model can also be used: various versions of Qwen-7B, various versions of Qwen-14B, various versions of Qwen-32B; or the open-source Llama model: various versions of llama-8b, various versions of llama-13b, various versions of llama-70b models; and generative interaction models such as the GLM model can all be used as alternatives to the base model in this technology.

[0176] Preferably, the above-mentioned prompt engineering structure is also as Figure 3 shown. The overall prompt engineering consists of background definition, main body description, output constraint, process and example inspiration.

[0177] Preferably, the above-mentioned background definition includes role overview and background overview.

[0178] Preferably, the content of the role overview in the above-mentioned background definition is "literature analysis expert and information serialization consultant".

[0179] Preferably, the content of the background overview in the above-mentioned background definition is "The user needs to conduct in-depth analysis of a large number of documents and be able to serialize the key information in the documents according to a certain logic for better understanding and mastering the content of the documents".

[0180] Preferably, the above-mentioned main body description includes task overview and skill description.

[0181] Preferably, the content of the task overview in the above-mentioned main body description is "You are an expert with rich experience in the field of literature analysis and information serialization, good at extracting key information from complex documents and being able to effectively serialize the information according to different logics".

[0182] Preferably, the skill description content in the above main description is "You have profound abilities in literature retrieval, analysis, and organization, and are proficient in using various analysis tools and methods, such as timeline analysis, logical reasoning, theme classification, etc., to ensure the accurate serialization of information."

[0183] Preferably, the above output constraints include output format and constraint overview.

[0184] Preferably, the output format content in the above output constraints is "Provide a serialized information summary, including key information points, timeline, logical relationship diagram, etc."

[0185] Preferably, the constraint overview content in the above output constraints is "The serialization process should be objective and accurate, avoid subjective assumptions, ensure the integrity and logic of information, and at the same time only output the serialized original text results without excessive analysis."

[0186] Preferably, the above process and example inspiration include work process and example tips.

[0187] Preferably, the work process in the above process and example inspiration is as follows:

[0188] S5-1. Read and understand the literature content, and extract key information points;

[0189] S5-2. Classify and sort the information points according to the logic (time, event logic, context logic, etc.) of the content specified by the user;

[0190] S5-3. Construct the information serialization result, including information summary, timeline, logical relationship diagram, etc., and verify and adjust to ensure accuracy.

[0191] Preferably, the example tip content in the above process and example inspiration is as follows:

[0192] "Example 1: Serialize the research literature on a certain historical event

[0193] Key information points: Cause of the event, main participants, key turning points, result of the event.

[0194] Timeline: Arrange the key information points in the order of the time when the event occurred.

[0195] Logical relationship diagram: Show the logical connection between the cause, turning point and result of the event.

[0196] Example 2: Serialize the literature on the development of a certain scientific theory

[0197] Key information points: Proposal of the theory, main supporting evidence, development of the theory, related controversies.

[0198] Event logic: Arrange the key information points in the logical order of the development of the theory.

[0199] Context logic: Analyze the context connection between the theory proposed and the supporting evidence, as well as the development and disputes.

[0200] Example 3: Sequencing the literature on a certain policy change

[0201] Key information points: Policy introduction background, policy content, policy effects, policy adjustments.

[0202] Timeline: Arrange the key information points in the chronological order of the policy change.

[0203] Event logic: Analyze the logical connection among the policy introduction background, content, effects, and adjustments.

[0204] After determining the specific prompting engineering, the above prompting content and the following task instructions jointly constitute the final prompting content: Please sequence the following key information according to a certain logic: {input}.

[0205] Specifically, the above {input} represents the output of step S4.

[0206] S6. Build an agent for generating large model reports, compare the differences between the new and old literature before and after, record the version changes, generate a structured update report to ensure the traceability of the update process, and form a final archival record.

[0207] Specifically, this step is responsible for comparing the differences between the literature before and after the update, identifying and recording all difference points, outputting a structured update report to ensure the traceability of the above literature update process, and ensuring that the output result of the model for the main information content will not be changed again.

[0208] Preferably, this step builds an agent for generating large model reports, which is implemented based on prompting engineering using LLMs, that is, inputting the content to be processed and the prompting words of the prompting engineering into the large model to obtain the target output.

[0209] Preferably, the above LLMs adopt the Qwen-72B model, and other versions of Qwen models can also be used: various versions of Qwen-7B, various versions of Qwen-14B, various versions of Qwen-32B; or open-source Llama models: various versions of llama-8b, various versions of llama-13b, various versions of llama-70b models; and generative interactive models such as the GLM model can all be used as alternatives to the base model in this technology.

[0210] Preferably, the above prompting engineering structure is also as Figure 4 shown, and the overall prompting engineering consists of background definition, main body description, output constraints, process, and example inspiration.

[0211] Preferably, the above background definition includes a role overview and a background overview.

[0212] Preferably, the content of the role overview in the above background definition is "Literature analysis expert and version control consultant".

[0213] Preferably, the content of the background overview in the above background definition is "The user needs to accurately compare two documents, identify differences, and record version changes to ensure the traceability of the document update process and the output of a structured report."

[0214] Preferably, the above main body description includes a task overview and a skill description.

[0215] Preferably, the content of the task overview in the above main body description is "You are a professional literature analyst, proficient in using advanced text comparison techniques and version control methods, and able to accurately identify and record the subtle differences between documents."

[0216] Preferably, the content of the skill description in the above main body description is "You possess professional skills in text analysis, version control, data structuring, and report writing, and are able to efficiently process and analyze a large amount of literature data."

[0217] Preferably, the above output constraints include an output format and a constraint overview.

[0218] Preferably, the content of the output format in the above output constraints is "A structured update report, including a list of difference points and a record of version changes."

[0219] Preferably, the content of the constraint overview in the above output constraints is "The comparison process must be accurate and error-free, the report should be clear, structured, and easy to understand, while protecting the copyright and privacy of the documents."

[0220] Preferably, the above process and example inspiration include a work process and example tips.

[0221] Preferably, the work process in the above process and example inspiration is as follows:

[0222] S6-1. Import and preprocess the documents to prepare for comparison;

[0223] S6-2. Use a text comparison tool to identify the differences between the two documents;

[0224] S6-3. Analyze the difference points, classify and record the specific content and location of each difference;

[0225] S6-4. Generate a structured update report based on the difference points;

[0226] S6-5. Record the version changes to ensure that each update has a detailed record and traceability;

[0227] S6-6. Output the final update report.

[0228] After determining the specific prompting engineering, the above prompting content and the following task instructions together constitute the final prompting content: Please analyze the update situation of the following information: {input}.

[0229] Specifically, the above {input} represents the corresponding output of step S5 and the relevant content before the update.

[0230] Finally, an intelligence aggregation result on a specific topic and the corresponding update report can be obtained.

[0231] This embodiment proposes an intelligence information integration method based on a multi-agent task chain strategy large model. According to the process of information integration by human experts, the key information is organized and sequenced by decoupling complex tasks into multiple subtasks, so as to automatically aggregate and sequence the corresponding dynamic intelligence according to a specific topic. Combining with the embodiment, compared with the existing large model methods, the main cores of this embodiment are as follows: (1) Fine-grained task chain guidance and inspiration that conforms to the working paradigm of human experts. Compared with the classical method, the method of the present invention can guide the large model to decouple an abstract complex task into a reasonable work process that conforms to the thinking mode of human experts, construct multiple agents, and fully stimulate the reasoning ability of LLMs for different task requirements, effectively suppressing its hallucination rate. (2) Applying reasonable constraints to LLMs, taking into account the constraints on the target output and the generalization reasoning ability. Compared with the classical method, the method of the present invention does not forcibly use an absolute output format constraint or very specific and fixed example prompts, but provides an idea of a work process paradigm and an actual reasoning method. Compared with the classical LLMs, it takes into account the constraints on the target output and the generalization reasoning ability, can greatly stimulate the adaptive reasoning ability of the model while specifying the work process, and can better achieve better comprehensive reasoning performance. (3) Inspiring the adaptive logical reasoning ability of the model through methodology. Compared with the classical method, the method of the present invention inspires the adaptive logical reasoning ability of the model through methodology rather than specific examples, in order to improve the generalization ability of the model. The "methodology" brings higher "generalization", that is, teaching the model "the method of handling things" rather than "learning a single example". Therefore, compared with the classical method, the method of this embodiment has stronger out-of-domain extrapolation ability, that is, it can effectively process comprehensive information from different fields and different literature types. This embodiment is of great significance in enhancing the quality and efficiency of dynamic intelligence information services, improving the efficiency and effectiveness of intelligence work, and further assisting in optimizing the strategic decision-making and policy-making processes.

[0232] In summary, for the intelligence information integration method and device based on the multi-agent task chain strategy large model of the present invention, by collecting multiple open-source dynamic scientific and technological intelligence information from target information sources, and using the literature analysis agent, abstract induction agent, content generation agent, sequencing agent, and report generation agent constructed by the large language model to extract key information, perform semantic analysis and induction, evaluate and update content, sequence information, and generate version change records and reports respectively, the automatic aggregation and sequencing of intelligence information are achieved. This method effectively solves the deficiencies in intelligence information integration of the prior art, such as generalization, objectivity, cross-document information processing, flexibility, robustness, hallucination rate, extrapolation ability, and few-shot learning ability, etc., improves the efficiency and effectiveness of intelligence work, and is of great significance for enhancing the quality and efficiency of dynamic intelligence information services, supporting strategic decision-making and policy formulation.

[0233] Those of ordinary skill in the art should understand that the various exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Specifically, whether to implement in hardware or software depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention. When implemented in hardware, it can be, for example, an electronic circuit, an application-specific integrated circuit (ASIC), appropriate firmware, a plug-in, a functional card, and so on. When implemented in software, the elements of the present invention are programs or code segments used to perform the required tasks. The programs or code segments can be stored in a machine-readable medium or transmitted via a data signal carried in a carrier wave over a transmission medium or a communication link.

[0234] It should be clear that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, the detailed description of known methods is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order between steps after understanding the spirit of the present invention.

[0235] In the present invention, the features described and / or illustrated for one embodiment can be used in the same way or in a similar way in one or more other embodiments, and / or combined with the features of other embodiments or replace the features of other embodiments.

[0236] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An intelligence information integration method based on a multi-agent task chain strategy model, characterized in that: The method comprises the following steps: Collect multiple open source dynamic science and technology intelligence information from target information sources, where the target information sources are divided into multiple fields according to content attributes, dissemination purposes, audience range or information form; Building a document analysis agent based on a large language model, extracting first target key information from the open source dynamic scientific and technological intelligence information according to a preset first guidance prompt text; Building an abstract induction agent based on the large language model, performing semantic analysis on the first target key information according to the second guidance prompt text, and outputting the second target key information in a structured manner after abstraction and induction; Building a content generation agent based on the large language model to obtain the latest standard document about the second target key information through a preset link, evaluating the target key information according to the standard document to locate the difference part and update it, so as to obtain the third target key information; Constructing a sequenced intelligent agent based on the large language model to sequence the third target key information according to the third guidance prompt text according to a preset expression logic to obtain fourth target key information; A report generation agent is constructed based on the large language model to perform a difference comparison on the fourth target key information generated in each batch according to the fourth guidance prompt text, record version changes and generate an update report.

2. The intelligence information integration method based on the multi-agent task chain strategy large model according to claim 1 is characterized in that: The method further comprises: Collect multiple open source dynamic science and technology intelligence information of the target information source by means of subscription push, manual collection and retrieval matching; Or, using a customized crawler tool to periodically crawl multiple open source dynamic science and technology intelligence information of the target information source; Or, integrate the API interfaces provided by multiple target information sources, and use the Elasticsearch platform for unified management and retrieval; The target information sources include news media, academic research, technical blogs and social networks.

3. The intelligence information integration method based on the multi-agent task chain strategy large model according to claim 1 is characterized in that: The large language model uses a Qwen-72B model, a Qwen-7B model, a Qwen-14B model, a Qwen-32B model, an open source Llama model or a generalized linear model as a base.

4. The intelligence information integration method based on the multi-agent task chain strategy large model according to claim 1 is characterized in that: The first guide prompt text, the second guide prompt text, the third guide prompt text and the fourth guide prompt text all include a background definition part, a main body description part, an output constraint part, and a process and example inspiration part; The background definition part includes an overview of roles and background used to perform tasks in the current problem scenario; The main description part includes a task overview for the current problem scenario and a description of the skills used to solve the problem; The output constraints section includes an overview of the restrictions and constraints on the output format; The process and example inspiration section includes hints on the process steps and provided examples.

5. The intelligence information integration method based on the multi-agent task chain strategy large model according to claim 1 is characterized in that: The method further comprises: According to the data update frequency of the target information sources in multiple fields, dynamically adjust the collection frequency of the open source dynamic scientific and technological intelligence information in the corresponding fields; Furthermore, dynamic weights are configured for the target information sources in different fields, and the crawling priority is adjusted based on the timeliness, authority and user feedback of each field.

6. The intelligence information integration method based on the multi-agent task chain strategy large model according to claim 1 is characterized in that: The method further comprises: An intermediate cache database is established to cache the first target key information, the second target key information, the third target key information and the fourth target key information, and an index is established for repeated scheduling and query.

7. The intelligence information integration method based on the multi-agent task chain strategy large model according to claim 1 is characterized in that: The method further comprises: Establish an error feedback channel, and when the downstream agent detects a logical contradiction, it triggers the reprocessing of the upstream agent.

8. An intelligence information integration device based on a multi-agent task chain strategy model, comprising a processor, a memory, and a computer program / instruction stored in the memory, characterized in that: The processor is used to execute the computer program / instructions. When the computer program / instructions are executed, the device implements the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method as claimed in any one of claims 1 to 7 are implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images