CN120218079A

CN120218079A - Intelligence information integration method and device based on multi-agent task chain strategy large model

Info

Publication number: CN120218079A
Application number: CN202510292547.1A
Authority: CN
Inventors: 耿国桐; 张全; 吕雁飞; 严兆辉; 雷帅; 高倩; 汤丽群
Original assignee: Military Science Information Research Center Of Military Academy Of Chinese Pla
Current assignee: Military Science Information Research Center Of Military Academy Of Chinese Pla
Priority date: 2025-03-12
Filing date: 2025-03-12
Publication date: 2025-06-27

Abstract

The invention provides an information integration method and device based on a multi-agent task chain strategy big model, which are used for realizing automatic aggregation and sequencing of information by collecting a plurality of open-source dynamic scientific information of a target information source, utilizing a document analysis agent, an abstract induction agent, a content generation agent, a sequencing agent and a report generation agent which are constructed by a big language model to respectively extract key information, analyze and induce semantics, evaluate and update content, sequence information and record version change and report generation. The method effectively solves the defects in the information integration of the prior art, such as generalization, objectivity, cross-document information processing, flexibility, robustness, illusion rate, extrapolation capability, small sample learning capability and the like, improves the efficiency and effect of information work, and has important significance in enhancing the service quality benefit of dynamic information, supporting strategic decisions and policy formulation.

Description

Information integration method and device based on multi-agent task chain strategy big model

Technical Field

The invention relates to the technical field of information integration analysis, in particular to an information integration method and device based on a multi-agent task chain strategy big model.

Background

In the current globalization and informatization rapid development age background, the importance of technological information is more and more prominent, and the importance of the technological information plays a key role in auxiliary decision making and technological strategy planning. However, the current technical information has the characteristics of wide sources and various formats, so that the information management faces serious challenges, and the phenomenon of information scattering not only affects the quality and usability of the technical information, but also increases the difficulty of a decision maker in screening useful information from massive data. Although the traditional classical machine learning algorithm can assist researchers to understand and organize information, the performance is limited by the quality and relativity of input data, the generalization capability of a model is limited, the characteristic engineering depends on expert service level, subjectivity and unilateral performance exist, the complexity of the model is limited, and complex cross-document information is difficult to mine. Deep learning techniques, while exhibiting information integration potential, the decimated integration lacks consistency and flexibility, heuristic integration requires a large amount of training data and model robustness is limited, both of which lack extrapolation capability for different document types. The large language model technology has strong learning ability and generalization ability, but in the information integration task, zero sample prompt is difficult to accurately extract target key information, information redundancy is extracted, and phantom characteristics reduce model reliability, so that the application of the large language model technology in information analysis is prevented. Therefore, it is needed to provide an information integration method capable of effectively solving the above problems, so as to improve the efficiency and effect of information work and better assist in optimizing strategic decisions and policy making processes.

Disclosure of Invention

In view of this, the embodiment of the invention provides an information integration method and device based on a multi-agent task chain policy big model, so as 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 the existing information integration technology.

The invention provides an information integration method based on a multi-agent task chain strategy big model, which comprises the following steps:

collecting a plurality of open source dynamic scientific and technological information of a target information source, wherein the target information source is divided into a plurality of fields according to content attributes, propagation purposes, audience ranges or information forms;

Constructing a literature analysis agent based on a large language model, and extracting first target key information from the open-source dynamic science and technology information according to a preset first guide prompt text;

Constructing an abstract induction intelligent body based on the large language model, carrying out semantic analysis on the first target key information according to a second guide prompt text, and structurally outputting second target key information after abstract induction;

Constructing a content generation agent based on the large language model to acquire the latest standard document about the second target key information through a preset link, evaluating the target key information positioning difference part according to the standard document and updating to obtain third target key information;

Constructing a sequencing agent based on the large language model so as to sequence the third target key information according to a preset expression logic according to a third guide prompt text to obtain fourth target key information;

And constructing a report generating agent based on the large language model, comparing differences of the fourth target key information generated by each batch according to the fourth guiding prompt text, recording version changes and generating an update report.

In some embodiments, the method further comprises:

collecting a plurality of open source dynamic scientific and technological information of the target information source in a subscription pushing, manual collection and retrieval matching mode;

or, a customized crawler tool is adopted to regularly capture a plurality of open source dynamic scientific and technological information of the target information source;

Or integrating API interfaces provided by a plurality of target information sources, and uniformly managing and searching by means of an elastic search platform;

wherein the target information sources include news media, academic research, technical blogs, and social networks.

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

In some embodiments, the first guidance prompt text, the second guidance prompt text, the third guidance prompt text, and the fourth guidance prompt text each include a context definition portion, a body description portion, an output constraint portion, and a flow and example heuristic portion;

The background definition part comprises a role summary and a background summary for executing tasks in the current problem scene;

the main description part comprises a task summary aiming at the current problem scene and a skill description for solving the problem;

The output constraint section includes an overview of constraints and constraints on the output format;

the flow and example heuristics section includes hints for processing flow steps and provided cases.

In some embodiments, the method further comprises:

According to the data updating frequency of the target information sources in the multiple fields, the collection frequency of the open source dynamic scientific and technological information in the corresponding fields is dynamically adjusted;

And configuring dynamic weights for the target information sources in different fields, and adjusting grabbing priorities based on timeliness, authority and user feedback in each field.

In some embodiments, the method further comprises:

And 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 query.

In some embodiments, the method further comprises:

And establishing an error return channel, and triggering the reprocessing of the upstream agent when the downstream agent detects the logic contradiction.

In another aspect, the present invention also provides an information integration device based on a multi-agent task chain policy big model, which includes a processor, a memory, and a computer program/instruction stored on the memory, wherein the processor is configured to execute the computer program/instruction, and the device implements the steps of the method when the computer program/instruction is executed.

In another aspect, the present invention also provides a computer readable storage medium having stored thereon a computer program/instruction which when executed by a processor performs the steps of the above method.

In another aspect, the invention also provides a computer program product comprising a computer program/instruction which, when executed by a processor, implements the steps of the above method.

The invention has the advantages that:

Additional advantages, objects, and features of the invention will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned from practice of the invention. The objectives and other advantages of the invention may be realized and attained by the structure particularly pointed out in the written description and drawings.

It will be appreciated by those skilled in the art that the objects and advantages that can be achieved with the present invention are not limited to the above-described specific ones, and that the above and other objects that can be achieved with the present invention will be more clearly understood from the following detailed description.

Drawings

The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate and together with the description serve to explain the application. In the drawings:

Fig. 1 is a flow chart of an information integration method based on a multi-agent task chain policy big model according to an embodiment of the invention.

Fig. 2 is a flow chart of an information integration method based on a multi-agent task chain policy big model according to another embodiment of the invention.

Fig. 3 is a management diagram of each agent in the information integration method based on the multi-agent task chain policy big model according to another embodiment of the invention.

Fig. 4 is a schematic diagram of a prompt engineering structure in an information integration method based on a multi-agent task chain policy big model according to another embodiment of the invention.

Detailed Description

The present invention will be described in further detail with reference to the following embodiments and the accompanying drawings, in order to make the objects, technical solutions and advantages of the present invention more apparent. The exemplary embodiments of the present invention and the descriptions thereof are used herein to explain the present invention, but are not intended to limit the invention.

It should be noted here that, in order to avoid obscuring the present invention due to unnecessary details, only structures and/or processing steps closely related to the solution according to the present invention are shown in the drawings, while other details not greatly related to the present invention are omitted.

It should be emphasized that the term "comprises/comprising" when used herein is taken to specify the presence of stated features, elements, steps or components, but does not preclude the presence or addition of one or more other features, elements, steps or components.

The integration process of intelligence information can be seen as a process of integrating key information across documents. However, the existing key information integration technology based on cross-document is mainly divided into two main types, namely extraction type integration and heuristic type integration. The extraction type integration is to directly extract target information from an original document and integrate the target information according to a certain rule, and the heuristic integration is to automatically generate the target information through a certain appointed sentence. While these existing approaches exhibit great information integration potential under the enablement of deep learning techniques, challenges remain. First, the integration information generated by the extraction method lacks consistency and flexibility, and second, the heuristic method requires a large amount of training data, the model performance depends on the data quality, and the robustness is limited. Third, both methods lack extrapolation capability for different document types because they require a large amount of domain-specific training data. For example, the model may be more suitable for integrating news-like documents, where it is difficult to achieve optimal processing performance.

The large language model technology (Large Language Models, LLMs) benefits from huge network scale and rich training corpus, shows strong learning ability, generalization ability and emerging ability, and endows new kinetic energy for breaking through the limitations. While the generic expansion capability and small sample learning capability presented by LLMs technology presents a great potential to address the above problems, challenges are presented. Compared with single document key information extraction, the information integration task is more focused on interaction and integration of cross-document knowledge, and LLMs has a longer context window, so that the accurate extraction capability of a model to target key information is difficult to excite due to classical zero sample prompt, and the extracted information often has more redundancy. Further, the redundant information further interferes with the information integration effect of the model. In addition, the illusion feature is one of the problems that limit LLMs application to information analysis, and the illusion feature inherent in LLMs reduces the reliability of the model, and prevents the LLMs technology from being organically embedded in information integration. Therefore, it is necessary to provide an information integration method with good generalization, strong objectivity, high flexibility, strong robustness, low illusion rate, extrapolation capability and small sample learning capability, so that the large model effectively enables the integration process of dynamic information, and the service quality benefit of the dynamic information is improved, and the efficiency and effect of information work are improved.

Specifically, the invention provides an information integration method based on a multi-agent task chain strategy big model, as shown in fig. 1, the method comprises the following steps S101-S106:

Step S101, collecting a plurality of open source dynamic scientific and technological information of a target information source, wherein the target information source divides a plurality of fields according to content attributes, propagation purposes, audience ranges or information forms.

Step S102, building a literature analysis agent based on a large language model, and extracting first target key information from open-source dynamic science and technology information according to a preset first guide prompt text.

And step S103, constructing an abstract induction intelligent agent based on the large language model, carrying out semantic analysis on the first target key information according to the second guide prompt text, and structuring and outputting the second target key information after abstract induction.

Step S104, constructing a content generation agent based on the large language model to acquire the latest standard document about the second target key information through a preset link, evaluating the target key information positioning difference part according to the standard document and updating to obtain third target key information.

And step 105, constructing a sequencing agent based on the large language model so as to sequence the third target key information according to the preset expression logic according to the third guide prompt text to obtain fourth target key information.

And S106, constructing a report based on the large language model to generate an agent, comparing differences of fourth target key information generated in batches according to the fourth guide prompt text, recording version changes and generating an update report.

In step S101, the method may include collecting a plurality of open source dynamic scientific and technological intelligence information of the target information source in a subscription pushing, manual collection and search matching manner. Or, a customized crawler tool is adopted to regularly capture a plurality of open source dynamic scientific and technological information of the target information source. Or integrating API interfaces provided by a plurality of target information sources, and uniformly managing and searching by means of an elastic search platform. Among other things, target information sources include news media, academic research, technical blogs, and social networks.

And introducing a multisource acquisition strategy, constructing a distributed crawler cluster by using Scrapy frames, configuring different crawler templates to adapt to each information source (such as RSS subscription grabbing and API polling requests), and realizing real-time data stream buffering through a Kafka message queue to prevent data loss in a high concurrency scene.

The goal of this step is to collect open source dynamic scientific and technological intelligence information from different sources. These sources of information may be divided into multiple domains, such as news, academic articles, technical blogs, social media, etc., based on content attributes, propagation purposes, audience scope, or forms of information.

In step S102, the objective is to construct a literature analysis agent, and extract first objective key information from the collected open-source dynamic scientific and technological information by using a large language model (such as Qwen-72B, llama). Based on the large language model, the functions of the intelligent agent are realized through prompt engineering (prompt engineering). The first guiding prompt text requirement is a prompt word with clear and accurate design, so that the model can accurately understand the task requirement.

Illustratively, the term "extract key information related to artificial intelligence from the following literature { input literature }", input literature is "2024, 1 month, a company has released a new product", and output is "key information: release of new product, time: 2024, 1 month".

In step S103, the originally extracted scattered first target key information is converted into structured second target key information with logicality and operability, which specifically comprises the steps of carrying out semantic understanding, breaking through literal matching, and mining implicit association of technical terms, such as inheritance relation of 'ViT' and 'transducer architecture'. An abstraction of knowledge is performed, and general rules are extracted from specific cases, such as three major technical routes of "small sample learning" are summarized from 10 papers. Structured representation converts free text into machine-resolvable formats, such as JSON Schema, supporting downstream automation processes. The method comprises the steps of carrying out key transition from a data stack to knowledge construction, eliminating noise through semantic understanding and structuring recombination, filtering irrelevant details, focusing core knowledge elements, establishing association, revealing a technical evolution path, enabling decision-making, and enabling structured data to directly drive advanced applications such as automatic reporting, trend prediction and the like.

In step S104, the objective of this step is to construct a content generation agent, acquire the latest standard document on the second target key information through the preset link using the large language model, evaluate the target key information and locate the difference part, and update to obtain the third target key information. The reference is that the novel aircraft adopts the latest structure, reduces weight, only uses 3 ten thousand screws, the input information is that the aircraft adopts 6 ten thousand screws and adopts a steam emission mode, and the output is that the aircraft adopts 3 ten thousand screws and adopts an electromagnetic ejection emission mode after updating.

In step S105, the objective is to construct a sequenced agent, and sequence the third objective key information according to the preset expression logic by using the large language model, so as to obtain the fourth objective key information. Based on the large language model, the function of the intelligent agent is realized through prompt engineering, prompt words are designed, and the instruction model is sequenced according to time, event logic and the like. Illustratively, the prompt is used to logically sequence the following messages by time, { input message } ", input message is" event A:2023, month 12, which is proprietary to the company. Event B, month 1 of 2024, the company issues a new product. The output is the ordering result, 2023, 12 months, company gets the patent, 2024, 1 month, company issues new products.

In step S106, the objective is to construct a report generating agent, compare the differences of the fourth objective key information generated by each lot with a large language model, record the version changes and generate an update report. Based on the large language model, the function of the intelligent agent is realized through prompt engineering. And designing a prompt word to guide the model to carry out difference comparison and report generation.

The method is characterized in that the prompt word is used for carrying out difference comparison on the information of the old version of the input information, namely the aircraft adopts 6 spare screws and adopts a steam type emission mode. The new version is that the aircraft adopts 3 tens of thousands of screws, and adopts an electromagnetic ejection emission mode, and the output is 1, namely, the difference point is the difference point, and the number of the screws is reduced. 2. The emission pattern varies. After updating, the aircraft adopts 3 tens of thousands of screws and adopts an electromagnetic ejection emission mode.

In some embodiments, the first guidance prompt text, the second guidance prompt text, the third guidance prompt text, and the fourth guidance prompt text each include a context definition portion, a body description portion, an output constraint portion, and a flow and example heuristics portion.

The context definition section includes a role overview and a context overview for performing tasks under the current problem scenario.

The main description section includes a task summary for the current problem scenario and a skill description for solving the problem.

The output constraint section includes an overview of constraints and constraints on the output format.

In some embodiments, the method further comprises step S201 and step S202:

Step S201, according to the data updating frequency of the target information sources in the multiple fields, the collection frequency of the dynamic scientific and technological information of the corresponding fields is dynamically adjusted.

Step S202, dynamic weights are configured for target information sources in different fields, and grabbing priorities are adjusted based on timeliness, authority and user feedback in each field.

In step S201, the frequency of collecting the dynamic scientific and technological information of the corresponding domain is dynamically adjusted according to the data update frequency of the target information source. The collection frequency is optimized by monitoring the update condition of the data source in real time, so that the timeliness and the effectiveness of the information are ensured.

And (3) monitoring the data updating condition of the information sources in each field in real time, counting the updating frequency of the information sources in each field, identifying the fields with frequent updating and the fields with less updating, and dynamically adjusting the collecting frequency according to the analysis result. For example, the collection frequency is increased for areas with frequent updates, and the collection frequency is decreased for areas with less updates. For example, if information sources in a news domain are updated multiple times per day and government reports are updated once per month, the collection frequency of the news domain may be set to multiple times per day and the collection frequency of the government reports may be set to once per month.

In step S202, dynamic weights are configured for target information sources in different domains, and grabbing priorities are adjusted based on timeliness, authority and user feedback in each domain. By the configuration of dynamic weights, the information of important fields is ensured to be acquired and processed preferentially. According to timeliness, authority and user feedback of each field, the importance of the field is evaluated, an initial weight value is set for each field, and the weight value is dynamically adjusted according to the field characteristic change monitored in real time. For example, if the timeliness of a certain field suddenly increases, the weight thereof is increased. For example, for the news domain, the weight may be set to a higher value due to its high timeliness, and for the academic database domain, the weight may also be set to a higher value due to its strong authority.

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 repeating the scheduling and querying.

In some embodiments, the method further includes establishing an error return channel that triggers reprocessing of the upstream agent when the downstream agent detects a logical conflict.

The invention is described below in connection with a specific embodiment:

the embodiment provides an information integration method based on a multi-agent task chain policy big model, as shown in fig. 2, the association relationship between different agents is shown in fig. 3, and the method includes:

s1, collecting open-source dynamic scientific and technological information of a target information source.

Specifically, the target information sources include, but are not limited to, news, academic articles, technical blogs, social media, and other texts in different fields.

Specifically, the collection process adopts an automatic pushing algorithm or a manual collection or retrieval matching algorithm.

S2, constructing a large model document analysis agent, and extracting target related key information from the framed information source in fine granularity.

Specifically, the fine granularity extraction target related structured key information may be timestamp information, key event, etc.

Preferably, this step constructs a large model document analysis agent, which is implemented using LLMs based on the prompt engineering, that is, inputting the content to be processed and the prompt word of the prompt engineering together into the large model to obtain the target output.

Preferably, the LLMs uses a Qwen-72B model, other versions Qwen models, such as versions Qwen-7B, versions Qwen-14B and versions Qwen-32B, or an open source Llama model, such as versions Llama-8B, versions Llama-13B and versions Llama-70B, and a GLM model, can be used as alternatives to the base model in the technology.

Preferably, the prompt engineering content meets the following requirements of (1) clear and brief, no unnecessary or confusing terms are used, easy-to-understand language is used, the words are accurate and clear, the probability of LLMs misunderstanding is reduced, illusions are suppressed, (2) final expectations are clearly expressed, the expected results or targets of tasks are clearly illustrated in detail, and (3) focus on important details, noise interference is reduced, unnecessary information is avoided or details which are not important for the tasks are avoided to prevent confusion of description.

Preferably, the above-mentioned prompt engineering structure is shown in fig. 4, and the overall prompt engineering is composed of background definition, main body description, output constraint, flow and example elicitation.

Preferably, the context definition includes a role summary and a context summary.

Preferably, the content of the role summary in the above background definition is "literature analysis expert".

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

Preferably, the subject description includes a task summary and a skill description.

Preferably, the task summary content in the main body description is "you are a professional literature analysis expert, which has the ability to understand and analyze complex documents in depth, and can identify and extract key information and time marks thereof. The method can help the user to accurately extract the key information and the time information of the target related information from the multi-document information source.

Preferably, the skill description content in the subject description is "you need to have document retrieval, information extraction, time series analysis, key information identification skills".

Preferably, the output constraints include an output format and a constraint summary.

Preferably, the output format content in the output constraint is "structured text output, including key information and corresponding time stamp".

Preferably, the content of the constraint summary in the output constraint is that the process needs to be efficient and accurate and can adapt to documents in different fields and types.

Preferably, the above-described flow and example heuristics include workflow and example hints.

Preferably, the above process and the example startup operation process are:

S2-1, determining a document set and an analysis target;

s2-2, identifying key information and the context thereof in the document by applying a text analysis technology;

s2-3, extracting a time mark of key information and verifying;

s2-4, arranging and presenting the extracted information, and ensuring the accuracy and the readability of the information.

Preferably, the above-mentioned flow and example in-process example prompt content is:

"document 1:2024 1, a company has released a new product.

Key information is new product release time of 2024, 1 month

Document 2 this company obtained an important patent at month 12 of 2023.

Key information is obtaining patent, time is 2023 and 12 months. "

After determining a specific prompt project, the prompt content and the task instructions are combined together to form a final prompt content, wherein the final prompt content is obtained by asking you to extract corresponding information with { target subject } according to the literature { input }.

Specifically, the { input } represents information to be analyzed screened from the framed information sources, and the { target subject } represents a target subject of interest to the user.

And S3, constructing a large model information abstract induction intelligent body, abstract induction related key information and carrying out structural output.

Preferably, the step constructs a large model abstract induction agent, and is realized by using LLMs based on prompt engineering, namely, the content to be processed and prompt words of the prompt engineering are input into a large model together to obtain target output.

Preferably, the above-mentioned prompt engineering structure is also shown in fig. 4, and the overall prompt engineering is composed of a background definition, a main body description, an output constraint, a flow and an example heuristic.

Preferably, the summary content of the background in the background definition is that a user needs to perform document analysis, and the events in the plurality of pieces of fine-granularity key information are abstracted and summarized into a concise and refined result.

Preferably, the task summary content in the subject description is "you are an experienced literature analysis expert, and can identify and generalize the same or similar events and semantic points".

Preferably, the skill description content in the subject description is "you need to have semantic analysis and summarize skills".

Preferably, the output format content in the output constraint is "the result should be presented in a clear text form and in a striped manner, including semantic induction.

Preferably, the constraint summary content in the output constraint is that the process needs to be efficient and accurate, can process a large amount of fine-grained key information, can identify fine semantic differences and ensures that the information is not changed.

Preferably, the above process and the example startup operation process are:

s3-1, reading a plurality of pieces of fine granularity information and identifying key information;

s3-2, identifying the same or similar semantic points by using a semantic analysis technology;

S3-3, abstracting and summarizing key information to form striped and unified expression.

"literature A five-thousand stars embellish the night sky.

Document B, the five thousands of stars flash in the night sky.

Literature C global average air temperature increases.

Literature D, reported that global air temperature is generally raised by 1 degree celsius.

Semantic induction:

1. The night sky has five tens of thousands of stars.

2. The global air temperature rises by 1 degree celsius. "

After determining a specific prompt project, the prompt content is combined with a task instruction to form a final prompt content, wherein the final prompt content is analyzed by the following pieces of fine-grained key information: { input }.

Specifically, the { input } represents the output in step S2.

S4, constructing a large model information evaluation content generation agent, evaluating the influence of the extracted and generalized information on the existing content based on a newer document, positioning the content to be updated, and updating the existing file content to obtain a series of refined and updated key target information.

Preferably, the step constructs a large model information evaluation content generation agent, and is realized by using LLMs based on prompt engineering, namely, the content to be processed and prompt words of the prompt engineering are input into a large model together to obtain target output.

Preferably, the content of the character summary in the above background definition is "information evaluation update expert".

Preferably, the content of the background summary in the above background definition is "the user needs to evaluate the existing file to determine whether to update according to the latest authority document, mark the content needing to be updated, and output the update result".

Preferably, the task summary content in the main body description is "you are a professional information evaluation update expert, which has the ability of deeply analyzing documents and evaluation information, and can quickly identify key update points and update related files.

Preferably, the skill description content in the subject description is "you need to have information analysis, content evaluation, key point identification, and mark skill skills".

Preferably, the output format content in the output constraint is "the result should include a list of updated points and a detailed description of each point, and a corresponding document reference, and the updated result is output.

Preferably, the constraint summary in the output constraint is "the assessment process needs to be based on the latest and authoritative documents, so that the accuracy and timeliness of the information are ensured.

Preferably, the above process and the example startup operation process are:

s4-1, determining an evaluation standard and a key field;

s4-2, searching and examining the latest authoritative document;

s4-3, comparing the existing file with the latest file, and identifying differences and update points;

S4-4, marking and updating the content to be updated, and providing literature support.

The existing file content is that the aircraft adopts 6 tens of thousands of screws and adopts a steam type emission mode.

Authoritative file content A. The novel aircraft adopts the latest structure, reduces the weight and only uses 3 ten thousand screws.

And the authoritative file content B is that the aircraft transmits in an electromagnetic ejection mode.

Authoritative file content C aircraft has increased support to helmet display screen.

Update point 1 aircraft construction, use screw quantity reduction.

Document reference authoritative File content A

Update point 2. Change of transmission mode.

Literature citation authoritative File content B

Updating 3. Support to the helmet display screen is increased.

Document reference, authoritative document content C

After updating, the aircraft adopts 3 tens of thousands of screws, adopts an electromagnetic ejection emission mode, and simultaneously increases the support for a helmet display screen. "

After determining a specific prompt project, the prompt content is combined with a task instruction to form a final prompt content, wherein the authority document { reference }, updated information { input }, is analyzed.

Specifically, the { reference } represents the latest authority in the related art, and the { input } represents the output of step S3.

S5, constructing a large-model information sequencing agent, adaptively refining the series of information according to a certain logic, and sequencing updated key target information to form a final information integration result.

Specifically, the self-adaptation refines the series of updated key target information according to a certain logic, and the updated key target information is sequenced to form a final information integration result, wherein the logic can comprise time logic, event logic, context logic and the like.

Preferably, the step constructs a large model information ordering intelligent agent, and is realized by LLMs based on prompt engineering, namely, the content to be processed and prompt words of the prompt engineering are input into a large model together to obtain target output.

Preferably, the above-mentioned prompt engineering structure is also shown in fig. 3, and the overall prompt engineering is composed of a background definition, a main body description, an output constraint, a flow and an example heuristic.

Preferably, the role summary in the above background definition is "literature analysis expert and information sequencing advisor".

Preferably, the content of the background summary in the above background definition is that "a user needs to perform deep analysis on a large number of documents, and can sequence according to a certain logic according to key information in the documents, so as to better understand and master the content of the documents.

Preferably, the task summary content in the above subject description is "you are an expert with rich experience in the field of document analysis and information sequencing, which is good at extracting key information from complex documents, and can effectively sequence information according to different logics.

Preferably, the skill descriptions in the subject description are "you have profound document retrieval, analysis and arrangement capabilities, and can be proficient in applying various analysis tools and methods, such as timeline analysis, logical reasoning, topic classification, etc., to ensure accurate ordering of information.

Preferably, the output format content in the output constraint is "providing the information abstract after sequencing, including key information points, time lines, logic relation diagrams and the like".

Preferably, the constraint summary content in the output constraint is that the sequencing process should be objective and accurate, subjective assumption is avoided, the integrity and logic of information are ensured, and meanwhile, only the sequencing original text result is output, and excessive analysis is not needed.

Preferably, the above process and the example startup operation process are:

s5-1, reading and understanding literature content, and extracting key information points;

s5-2, classifying and ordering the information points according to logic (time, event logic, context logic and the like) of the content specified by the user;

s5-3, constructing information sequencing results, including information abstracts, timelines, logic relation diagrams and the like, and verifying and adjusting to ensure accuracy.

example 1 sequencing of research literature for certain historical events

Key information points are event cause, main participant, key turning point and event result.

And (5) a time line, namely arranging key information points according to the time sequence of the event.

Logical relationship diagram showing logical links between event causes, inflection points and results.

Example 2 ordering of the development literature of a scientific theory

Key information points are theoretical proposal, main support evidence, theoretical development and related disputes.

Event logic, namely arranging key information points according to the logic sequence of theoretical development.

Context logic analysis theory proposes context links with supporting evidence, development and disputes.

Example 3 ordering documents for policy transitions

Key information points are policy stage setting context, policy content, policy effect and policy adjustment.

And (5) a time line, namely arranging key information points according to the time sequence of policy transition.

Event logic, analyzing logical links between policy context, content, effects and adjustments. "

After determining a specific prompt project, the prompt content and the task instructions are combined to form a final prompt content, wherein the following key information is ordered according to a certain logic { input }.

Specifically, the { input } represents the output of step S4.

S6, constructing a large model report to generate an agent, comparing the differences between the previous and the next new and old documents, recording version changes, generating a structured update report, ensuring that the update process is traceable, and forming a final filing record.

Specifically, the step is responsible for comparing the differences of documents before and after updating, identifying and recording all the difference points, outputting a structured updating report, ensuring the traceability of the document updating process, and ensuring that the output result of the model on the main information content is not changed any more.

Preferably, the step constructs a large model report generating agent, and is realized by using LLMs based on prompt engineering, namely, the content to be processed and the prompt words of the prompt engineering are input into a large model together to obtain target output.

Preferably, the role summary in the above background definition is "literature analysis expert and version control advisor".

Preferably, the summary content of the background in the above background definition is that "the user needs to precisely compare two documents, identify the difference, and record the version change, so as to ensure traceability of the document update process and output of the structured report".

Preferably, the task summary in the above subject description is "you are a professional literature analyst, who is good at using advanced text comparison technology and version control methods, to accurately identify and record subtle differences between documents.

Preferably, the technical description content in the subject description is "you have the expertise of text analysis, version control, data structuring and report writing, and can efficiently process and analyze a large amount of literature data".

Preferably, the output format content in the output constraint is a "structured update report" including a difference point list and a version change record.

Preferably, the content of the constraint summary in the output constraint is that the comparison process must be accurate, the report needs to be clear and structured, and the report is easy to understand, and meanwhile, the copyright and the privacy of the document are protected.

Preferably, the above process and the example startup operation process are:

S6-1, importing and preprocessing documents, and preparing comparison;

S6-2, identifying the difference between the two documents by using a text comparison tool;

s6-3, analyzing the difference points, classifying and recording the specific content and position of each difference;

S6-4, generating a structured update report according to the difference points;

S6-5, recording version change, ensuring detailed record and traceability of each update;

S6-6, outputting a final update report.

After determining a specific prompt project, the prompt content and the task instructions are combined into a final prompt content, wherein the update condition of the following information is analyzed.

Specifically, the { input } represents the corresponding output of step S5 and the related content before update.

Finally, the information aggregation result of the specific subject can be obtained and the report can be updated correspondingly.

The embodiment provides an information integration method based on a multi-agent task chain strategy big model, which organizes and sequences key information in a mode of decoupling a complex task into a plurality of subtasks according to the information integration process of human experts, and achieves automatic aggregation and sequencing of corresponding dynamic information according to a specific theme. In combination with the embodiment, compared with the existing large model method, the main core of the embodiment has three points of (1) fine-grained task chain guidance and inspiring fitting the working paradigm of human expert. Compared with a classical method, the method can guide a large model to decouple an abstract complex task into a reasonable workflow conforming to the thinking mode of human experts, construct a plurality of agents, fully excite LLMs reasoning capability aiming at different task demands and effectively inhibit the illusion rate. (2) Reasonable constraint is applied to LLMs, and constraint and generalization reasoning capability on target output are considered. Compared with the classical method, the method does not strongly use absolute output format constraint or very specific and fixed example prompt, but provides a workflow paradigm idea and an actual reasoning method. Compared with classical LLMs, the method has the advantages that constraint and generalization reasoning capacity on target output are considered, the self-adaptive reasoning capacity of the model can be stimulated to a large extent while the workflow is specified, and better comprehensive reasoning performance can be achieved. (3) The adaptive logic reasoning capability of the model is inspired through methodology. Compared with the classical method, the method of the invention inspires the adaptive logic reasoning capability of the model through the methodology, rather than a specific example, so as to promote the generalization capability of the model, and the "methodology" brings higher "generalization", namely, teaching the model "a method for processing matters" rather than "learning a single example". Therefore, compared with the classical method, the method has stronger domain extrapolation capability, and can effectively process comprehensive information from different domains and different document types. The embodiment has important significance in enhancing the service quality benefit of the dynamic information, improving the efficiency and effect of the information work, further assisting in optimizing strategic decisions, policy making processes and the like.

In summary, according to the information integration method and device based on the multi-agent task chain strategy big model, through collecting a plurality of open source dynamic scientific information of a target information source and utilizing a document analysis agent, an abstract induction agent, a content generation agent, a sequencing agent and a report generation agent which are constructed by a big language model, key information extraction, semantic analysis and induction, content evaluation and update, information sequencing and version change record and report generation are respectively carried out, so that automatic aggregation and sequencing of information are realized. The method effectively solves the defects in the information integration of the prior art, such as generalization, objectivity, cross-document information processing, flexibility, robustness, illusion rate, extrapolation capability, small sample learning capability and the like, improves the efficiency and effect of information work, and has important significance in enhancing the service quality benefit of dynamic information, supporting strategic decisions and policy formulation.

Those of ordinary skill in the art will appreciate that the various illustrative components, systems, and methods described in connection with the embodiments disclosed herein can be implemented as hardware, software, or a combination of both. The particular implementation is hardware or software dependent on the specific application of the solution and the design constraints. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention. When implemented in hardware, it may be, for example, an electronic circuit, an Application Specific Integrated Circuit (ASIC), suitable firmware, a plug-in, a function card, or the like. When implemented in software, the elements of the invention are the programs or code segments used to perform the required tasks. The program or code segments may be stored in a machine readable medium or transmitted over transmission media or communication links by a data signal carried in a carrier wave.

It should be understood that the invention is not limited to the particular arrangements and instrumentality described above and shown in the drawings. For the sake of brevity, a detailed description of known methods is omitted here. In the above embodiments, several specific steps are described and shown as examples. The method processes of the present invention are not limited to the specific steps described and shown, but various changes, modifications and additions, or the order between steps may be made by those skilled in the art after appreciating the spirit of the present invention.

In this disclosure, features that are described and/or illustrated with respect to one embodiment may be used in the same way or in a similar way in one or more other embodiments and/or in combination with or instead of the features of the other embodiments.

The above description is only of the preferred embodiments of the present invention and is not intended to limit the present invention, and various modifications and variations can be made to the embodiments of the present invention by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. The information integration method based on the multi-agent task chain strategy big model is characterized by comprising the following steps:

2. The multi-agent task chain policy big model based intelligence information integration method according to claim 1, wherein the method further comprises:

3. The method for integrating information based on a multi-agent task chain policy big model according to claim 1, wherein the big language model adopts Qwen-72B model, qwen-7B model, qwen-14B model, qwen-32B model, open source ilama model or generalized linear model as a base.

4. The multi-agent task chain policy big model based intelligence information integration method according to claim 1, wherein the first guidance prompt text, the second guidance prompt text, the third guidance prompt text and the fourth guidance prompt text each include a background definition part, a main description part, an output constraint part, and a flow and example heuristic part;

5. The multi-agent task chain policy big model based intelligence information integration method according to claim 1, wherein the method further comprises:

6. The multi-agent task chain policy big model based intelligence information integration method according to claim 1, wherein the method further comprises:

7. The multi-agent task chain policy big model based intelligence information integration method according to claim 1, wherein the method further comprises:

8. An intelligence information integration device based on a multi-agent task chain policy big model, comprising a processor, a memory and computer programs/instructions stored on the memory, characterized in that the processor is adapted to execute the computer programs/instructions, which when executed implement the steps of the method according to any of claims 1 to 7.

9. A computer readable storage medium having stored thereon a computer program/instruction which when executed by a processor performs the steps of the method according to any of claims 1 to 7.

10. A computer program product comprising computer programs/instructions which, when executed by a processor, implement the steps of the method of any of claims 1 to 7.