A large model-based enterprise comprehensive management method, system, device and storage medium

Abstract

The application provides a kind of enterprise comprehensive management method, system, equipment and storage medium based on large model, belongs to the technical field of enterprise intelligent management, including based on original enterprise level comprehensive management procedure data, constructs vertical knowledge base;According to the vertical knowledge base constructed, a knowledge retrieval model Agent is established;Based on question and answer, an expert experience module MoE is constructed;According to the knowledge retrieval model Agent and the expert experience module MoE established, the large model answers the user's question, and outputs the answer.The application can effectively integrate the existing comprehensive management procedures and the experience accumulated by experts in the enterprise, and realize rapid response and accurate answer to the problem through intelligent means.

Description

Technical Field

[0001] This invention relates to the field of enterprise intelligent management technology, and in particular to an enterprise integrated management method, system, device and storage medium based on a large model. Background Technology

[0002] In modern enterprise management, the general management department plays a crucial role and is an indispensable key department within the enterprise. Its responsibilities encompass multiple aspects, including human resource management, general administrative affairs handling, resource procurement, and project management. The effective execution of these functions is essential for maintaining the company's daily operations and promoting the achievement of its strategic goals.

[0003] However, in practice, due to the wide range and complexity of the areas involved, comprehensive management experts often face a large number of professional challenges, which include not only routine business process consulting, but may also involve deeper strategic decision support.

[0004] Currently, when employees encounter such problems, they typically seek help directly from comprehensive management experts. In this scenario, business experts need to provide answers or guidance based on the company's comprehensive management procedures and their own experience, such as: which departments are responsible for R&D project initiation, and what the travel reimbursement process is. Answering such questions requires comprehensive management experts to consult knowledge, integrate information, and even consider past experience.

[0005] Despite the experts' extensive experience and expertise, the sheer volume of problems they handle daily not only consumes their time and energy significantly but also risks providing incomplete or inaccurate answers due to limitations in their individual experience. Furthermore, as businesses expand and diversify, the traditional method of manually resolving all issues is no longer sufficient to meet the demands for efficient and precise services.

[0006] The existing solution is to build an intelligent question-and-answer assistant based on historical question-and-answer pairs, which only considers expert experience. Specifically, it first collects past question-and-answer pairs from the comprehensive management department to build a database; second, it extracts keywords from user questions; based on the relevance of question-and-answer pairs in the database to the keywords, it retrieves highly similar questions; finally, it displays the questions and answers in descending order of similarity. The above solution obtains its solution by first matching keywords with questions in question-and-answer pairs to find similar or identical questions in past questions, then obtaining the corresponding answers and providing them as the solution for the user's question.

[0007] However, the above solutions rely heavily on keywords, and question-answering assistants cannot fully understand the user's semantics. The accuracy of the search depends primarily on the frequency of keyword occurrences, placing high demands on the keywords in the user's question. If the keywords change, it can easily lead to inaccurate searches and failure to obtain effective answers. Summary of the Invention

[0008] To address the above problems, this invention provides a comprehensive enterprise management method, system, equipment, and storage medium based on a large model, which assists comprehensive management personnel in solving enterprise management problems, improves the accuracy of obtaining answers to problems, and significantly improves the efficiency of comprehensive management work.

[0009] This invention provides a comprehensive enterprise management method based on a large model, comprising: Based on the original enterprise-level comprehensive management procedure data, a vertical domain knowledge base is constructed; Based on the constructed vertical domain knowledge base, a knowledge retrieval model Agent is established; Based on question-answer pairs, an expert experience module (MoE) is constructed. Based on the established knowledge retrieval model Agent and expert experience module MoE, the large model answers user questions and outputs the results.

[0010] As a further improvement to the present invention, the construction of a vertical domain knowledge base based on the original enterprise-level comprehensive management procedure data includes: Collect the original enterprise-level integrated management procedure data, and perform heterogeneous data parsing on the collected original enterprise-level integrated management procedure data; Data augmentation is performed on the parsed enterprise-level integrated management procedure data to obtain enterprise-level integrated management procedure augmented data, and the enterprise-level integrated management procedure augmented data is stored. The enhanced data of the enterprise-level integrated management procedures is vectorized to generate a vector matrix, and the vector matrix is ​​stored in a vector database.

[0011] As a further improvement of the present invention, the step of collecting the original enterprise-level comprehensive management procedure data and performing heterogeneous data parsing on the collected original enterprise-level comprehensive management procedure data includes using PaddleOCR to convert the table data and text data in the original .doc format file into .txt format.

[0012] As a further improvement of the present invention, the parsed enterprise-level comprehensive management procedure data is enhanced to obtain enhanced enterprise-level comprehensive management procedure data. The storage of the enhanced enterprise-level comprehensive management procedure data includes compressing excessively long paragraphs and adding paragraph summaries.

[0013] As a further improvement of the present invention, the parsed enterprise-level integrated management procedure data is enhanced to obtain enhanced enterprise-level integrated management procedure data. The storage of the enhanced enterprise-level integrated management procedure data includes naming each file with keywords and using business domain isolation folders.

[0014] As a further improvement of the present invention, the vectorization of the enterprise-level integrated management procedure enhancement data, the generation of a vector matrix, and the storage of the vector matrix into a vector database include, after storing the vector matrix into the vector database, slicing the vector data in the vector database.

[0015] As a further improvement of the present invention, the large model, based on the established knowledge retrieval model Agent and expert experience module MoE, answers user questions and outputs answers, including using langchain to construct a CoT system based on the large model.

[0016] This invention provides an enterprise integrated management system based on a large model, comprising a vertical domain knowledge base construction unit, a knowledge retrieval model agent establishment unit, an expert experience module (MoE) construction unit, and a large model solution unit, wherein... The vertical domain knowledge base construction unit is used to construct a vertical domain knowledge base based on the original enterprise-level comprehensive management procedure data. The knowledge retrieval model Agent building unit is used to build a knowledge retrieval model Agent based on the constructed vertical domain knowledge base; The Expert Experience Module (MoE) construction unit is used to construct the Expert Experience Module (MoE) based on question-answer pairs. The large model answering unit is used to answer user questions and output answers based on the established knowledge retrieval model Agent and expert experience module MoE.

[0017] The present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described enterprise integrated management method based on a large model.

[0018] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described enterprise integrated management method based on a large model.

[0019] This invention provides a comprehensive enterprise management method, system, device, and storage medium based on a large-scale model. It constructs a vertical domain knowledge base using original enterprise-level comprehensive management procedure data; establishes a knowledge retrieval model (Agent) based on the constructed vertical domain knowledge base; and constructs an expert experience module (MoE) based on question-and-answer pairs. Using the established knowledge retrieval model (Agent) and expert experience module (MoE), the large-scale model answers user questions and outputs responses, effectively integrating existing comprehensive management procedures and accumulated expert experience within the enterprise. Through intelligent means, it achieves rapid response and accurate answers to questions; and through natural language interaction, it enables comprehensive management personnel or enterprise employees to obtain relevant information and assistance more quickly and accurately, further improving work efficiency. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the enterprise integrated management method based on a large model, according to an embodiment of the present invention.

[0021] Figure 2 This is a detailed flowchart illustrating the question-and-answer implementation process of the enterprise integrated management method based on a large model, according to an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of the structure of the enterprise integrated management system based on a large model according to an embodiment of the present invention.

[0023] Figure labeling: 10, Vertical domain knowledge base construction unit; 20, Knowledge retrieval model agent establishment unit; 30, Expert experience module (MoE) construction unit; 40, Large model solution unit. Detailed Implementation

[0024] The following describes specific embodiments and appendices. Figure 1-3 The invention is described in detail so that those skilled in the art can more fully understand its purpose, features and effects.

[0025] Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. In the event of any discrepancy between the definitions of terms in this invention and their commonly understood meaning by one of ordinary skill in the art to which this invention pertains, the definitions provided in this invention shall prevail.

[0026] In large enterprises, experts in comprehensive management departments often need to answer questions from employees and departments regarding administration, human resources, project management, and other areas. Their work involves a lot of understanding of requirements, knowledge retrieval, literature reading, and summarizing, which often leads to questions not being answered in a timely manner or providing incomplete or incorrect answers, thus affecting the efficiency of enterprise management.

[0027] This invention provides a comprehensive enterprise management method, system, equipment, and storage medium based on a large model, which improves existing management methods and thereby enhances the accuracy of problem-solving.

[0028] Example 1 As a specific embodiment of the present invention, this embodiment provides an enterprise comprehensive management method based on a large model, referring to... Figure 1 For example, in step 2, the specific steps are as follows: S100: Based on the original enterprise-level comprehensive management procedure data, construct a vertical domain knowledge base; S200. Based on the constructed vertical domain knowledge base, establish a knowledge retrieval model Agent; S300, based on question-answer pairs, construct the expert experience module MoE; S400: Based on the established knowledge retrieval model Agent and expert experience module MoE, the large model answers the user's question and outputs the answer.

[0029] By constructing a vertical domain knowledge base, the original enterprise-level comprehensive management procedure data within the enterprise is transformed into structured and easily searchable knowledge information. Then, a knowledge retrieval model Agent is built for the vertical domain knowledge base, which can improve the efficiency and accuracy of knowledge retrieval, enabling users to quickly obtain the knowledge they need. The expert experience module MoE built based on question-answer pairs can effectively utilize historical experience, thereby improving the quality of knowledge retrieval.

[0030] The enterprise integrated management method based on a large model, as described in this invention, can effectively integrate existing integrated management procedures and expert experience within an enterprise, achieving rapid response and accurate solutions to problems through intelligent means. By using natural language interaction, integrated management personnel or employees can obtain relevant information and assistance more quickly and accurately, reducing decision-making difficulty and further improving work efficiency.

[0031] Based on the original enterprise-level comprehensive management procedure data, S100 constructs a vertical domain knowledge base, including: S110. Collect the original enterprise-level integrated management procedure data, and perform heterogeneous data parsing on the collected original enterprise-level integrated management procedure data. Specifically, in this embodiment, the collected original enterprise-level integrated management procedure data needs to be converted to obtain data in a set format.

[0032] In this embodiment, the collected raw enterprise-level comprehensive management procedure data is a .doc format file, which contains tabular data and text data. The raw enterprise-level comprehensive management procedure data can be various rules and regulations and manuals of the company, such as employee manuals for human resources and project initiation standard process manuals for project management.

[0033] In other feasible embodiments, the collected raw enterprise-level integrated management procedure data may also be in other formats, such as .pdf files.

[0034] The original enterprise-level integrated management procedure data collected from the .doc format files is then converted to obtain enterprise-level integrated management procedure data in a predetermined format. In this embodiment, the table data and text data in the original .doc format files are converted together into .txt format. Preferably, PaddleOCR (an OCR tool library based on PaddlePaddle) is used to convert the table data and text data in the original .doc format files together into .txt format.

[0035] Converting .doc format to .txt format can remove redundant formatting from enterprise-level integrated management procedure data, reduce the complexity of data processing, and make it easier to adjust and optimize.

[0036] By converting raw enterprise-level comprehensive management procedure data from .doc format files to .txt format and then using PaddleOCR technology to process the tables and text data, the data processing workflow can be simplified, and the accuracy and efficiency of data extraction can be improved. PaddleOCR, based on deep learning algorithms, has a high character recognition accuracy and can better handle complex layouts and handwritten text.

[0037] Since .txt files typically take up less storage space, converting tabular data along with text data from the original .doc file to .txt format can help save storage resources and facilitate better management and utilization of internal raw integrated management procedure data.

[0038] In other feasible embodiments, other conversion tools can also be used to convert the table data in the original .doc format file, along with the text data, into .txt format.

[0039] Further processing of the enterprise-level integrated management procedure data, converted to .txt format, includes: Remove special characters, spaces, abnormal indentation, and heading indentation, etc. Correct error messages such as section numbers and typos.

[0040] The above processing can improve the standardization of enterprise-level comprehensive management procedure data in .txt format files, resulting in higher data quality.

[0041] S120. Perform data augmentation on the parsed enterprise-level integrated management procedure data to obtain enterprise-level integrated management procedure augmented data, and store the enterprise-level integrated management procedure augmented data. The data enhancement of the parsed enterprise-level integrated management procedure data includes compressing excessively long paragraphs and adding paragraph summaries. Excessively long paragraphs often contain a large amount of information, making it easy to lose focus during retrieval. By compressing and extracting key information, paragraphs can be made more concise, improving readability.

[0042] By compressing excessively long paragraphs and adding paragraph summaries, the core content of the paragraphs can be summarized, which can enhance the accuracy of subsequent searches.

[0043] Preferably, a large model is used to compress extremely long paragraphs and summarize their content. The large model summarizes the paragraph content and fills it below the paragraph title. Using a large model to compress extremely long paragraphs and summarize their content improves the efficiency of data augmentation.

[0044] Furthermore, the obtained enterprise-level integrated management procedure enhancement data will be stored, including: Name each file using keywords; Use business domains to isolate folders, such as separating folders based on business domains like human resources, project management, and procurement.

[0045] Naming files using keywords improves the accuracy of large models when retrieving references based on questions. Using business domain-isolated folders ensures data isolation across different business domains from the source, preventing data clutter.

[0046] Retrieval accuracy is enhanced by using PaddleOCR and large models for text processing during the preprocessing stage.

[0047] S130. Vectorize the enhanced data of the enterprise-level integrated management procedures, generate a vector matrix, and store the vector matrix in a vector database. Specifically, in this embodiment, after obtaining the enhanced enterprise-level comprehensive management procedure data, it is segmented, cleaned, and vectorized to generate a vector matrix, which is then stored in a vector database. The vector database can be a lightweight Chroma database, a database system specifically designed for processing and storing large-scale vector data. Chroma can seamlessly integrate with machine learning and AI applications, supports various vector indexing technologies, and provides sub-second query response times even when processing millions or even billions of vector data points, significantly improving the query efficiency of large-scale vector data.

[0048] In other embodiments, the vector database may also be a Milvus database. This embodiment does not impose specific limitations on the vector database used.

[0049] Furthermore, the enterprise-level comprehensive management procedure enhancement data is segmented, cleaned, and vectorized based on a natural language embedding model to generate a vector matrix that can be directly stored in a vector database. The natural language embedding model is the BAAI General Embedding model, an open-source general-purpose vector model specifically designed to support various information retrieval and large language model retrieval enhancement applications. In other feasible embodiments, other types of natural language embedding models can also be used to segment, clean, and vectorize the enterprise-level comprehensive management procedure enhancement data to generate a vector matrix that can be directly stored in a vector database.

[0050] After storing the vector matrix in the vector database, the vector data in the database is sliced. The maximum length of a single text is limited by setting the batch size, and the length of repeated fragments in the context is limited by setting the overlap size. Using a reasonable slicing strategy can reduce the difficulty of retrieval while ensuring the relevance of the context, improving the continuity of retrieval results, and enhancing the understanding quality of the large model.

[0051] Furthermore, a vector database is generated for the knowledge of each business domain, and each vector database is stored using a single complete file as the smallest storage unit. The storage mode is continuous storage.

[0052] By creating independent vector databases for each business domain, effective data isolation can be achieved. On the one hand, this helps prevent data confusion between different business domains; on the other hand, it facilitates domain-specific retrieval during subsequent question-and-answer sessions. Query requests can directly locate the database of the relevant business domain, avoiding the additional overhead of cross-domain queries and thus improving query efficiency.

[0053] Furthermore, the original enterprise-level comprehensive management procedure data is collected and updated regularly, and then preprocessed and vectorized before being stored in a vector database. This ensures that the data in the vector database reflects the latest management procedures of the enterprise, thereby guaranteeing that the knowledge in the constructed vertical domain knowledge base is updated regularly, improving the accuracy of search results and reducing misleading information caused by outdated data.

[0054] Furthermore, a retrieval strategy for the vector database is determined. In this embodiment, the retrieval strategy for the vector database is based on the similarity between the user's question and the knowledge. Preferably, Euclidean distance is used to calculate the similarity between the user's question and the knowledge.

[0055] Furthermore, the retrieval strategy of the vector database, based on the similarity between user questions and knowledge, combines a keyword-based retrieval strategy to construct a hybrid retrieval mode to ensure retrieval accuracy.

[0056] In this embodiment, the obtained enterprise-level integrated management procedure enhanced data vector database uses special data structures and algorithms to store and process vector data, making storage and retrieval operations highly efficient. In addition, the vector database can perform accurate matching based on the similarity between vectors, which is very important for knowledge question answering application scenarios.

[0057] In S100, raw enterprise-level comprehensive management procedure data is collected, and data preprocessing and data augmentation are performed. The processed data is then segmented, embedded, and other operations to build a vector database, thereby obtaining a vertical domain knowledge base.

[0058] Based on the constructed vertical domain knowledge base, S200 establishes a knowledge retrieval model Agent, including: By leveraging LangChain technology and mounting the vertical domain knowledge base, a knowledge retrieval model agent capable of natural language interaction with comprehensive management personnel is constructed. This agent performs knowledge retrieval and returns reference information. Preferably, the knowledge retrieval and reference functions are developed using LangChain's chains component. Utilizing LangChain technology enables rapid model construction, accelerating model building efficiency.

[0059] Specifically, the knowledge retrieval model agent can be used independently. Within the knowledge retrieval model agent, by querying specific comprehensive management questions, it returns direct knowledge results or guidance.

[0060] The prompt words are as follows: Accurately search for information and answer questions in the comprehensive management knowledge base.

[0061] They always end their answers with "Thank you for your question! Feel free to ask more questions!" Please refer to the following content: {context} Question: {question}.

[0062] Furthermore, after a user inputs a question, reference information is returned. Specifically, by traversing all knowledge base files and extracting the segment with the highest semantic relevance to the question, the model is returned. Simultaneously, the filename and paragraph information of the segment are returned, allowing users to directly view the reference segments in the original references. This improves the interpretability of the content generated by the model and provides users with a more intuitive and convenient fine-grained search result.

[0063] The knowledge retrieval model Agent enables knowledge retrieval and provides reference information, improving the interactive experience.

[0064] S300, based on question-answer pairs, constructs an Expert Experience Module (MoE) including: Based on expert experience, questions and answers are collected, and a question-and-answer pair database is constructed and stored in a structured manner. The question-and-answer pair database is a file system.

[0065] In real-world enterprise production environments, business experts frequently deal with occasional, low-frequency, and complex anomalies. The solutions and approaches to these problems are invaluable and represent expert experience that models need to reference and learn from. Therefore, by collecting these problems and answers, more comprehensive question-and-answer support can be provided for large models.

[0066] Furthermore, the expert experience module MoE adopts an NLP information retrieval strategy. By segmenting the question content, matching patterns, and weighting the average, it performs relevance matching based on the knowledge of the expert experience file system and returns the retrieved highly relevant expert experiences to the large model.

[0067] The MoE module matches expert experience with information retrieval strategies for reference by larger models. Preferably, the MoE module's information retrieval is based on keywords and utilizes the TD-IDF algorithm.

[0068] This embodiment utilizes langchain technology, combined with a vertical domain knowledge base, to create a large-scale model Agent that can interact with administrators using natural language, and constructs an expert module MoE to learn from historical question-and-answer experience, providing users with comprehensive question-and-answer data support.

[0069] Based on the established knowledge retrieval model (Agent) and expert experience module (MoE), the S400 uses a large model to answer user questions and outputs the answers. This includes utilizing langchain to construct a CoT (Cognition of Thought) system based on the large model, enabling intelligent scheduling of the Agent and MoE. A task routing (CoT) is constructed to identify user intent, specifically the business domain in which the user's question is located. This improves retrieval speed, reduces errors caused by global retrieval, and assigns the user's question to the corresponding knowledge retrieval model (Agent) to perform knowledge retrieval in the business domain in which the user's question is located.

[0070] In one possible embodiment, the prompt word is as follows: Your task is intent recognition. Based on the content of {order}, understand the meaning of {order} and select the intent that is semantically closest. Intents include: 1. Questions about <Human Resources>; 2. Questions about <Project Management>; 3. Questions about <General Administration>; 4. Questions about <Resource Procurement>. Only answer 1, 2, or 3. Example 1: Recognize the sentence <How to initiate a R&D project>, answer: 2. Example 2: Recognize the sentence <What should I do if I need to purchase a batch of monitors>, answer: 4. Your answer:

[0071] By using Task Routing (CoT), user issues can be accurately mapped to the relevant business domain, and searches can be performed in the knowledge base of the corresponding business domain, thus avoiding invalid searches.

[0072] Furthermore, it also includes building a Knowledge Integration Technology (CoT), which integrates knowledge from vertical domain knowledge bases and expert experience from file systems, enabling large models to return more comprehensive solutions.

[0073] In one possible embodiment, the prompt word is as follows: Your question is about knowledge integration, summarizing the information in {order1} and {order2}, and answering the {context}. Your answer:

[0074] The CoT system includes task routing and knowledge integration to automatically identify user intent and integrate relevant knowledge. By building the CoT system, intelligent scheduling of Agents and MoEs can be achieved.

[0075] Based on the established knowledge retrieval model (Agent) and expert experience module (MoE), the S400 employs a hybrid retrieval technique in answering user questions. This technique combines keyword matching and semantic understanding, increasing the likelihood of retrieving relevant documents and improving recall and precision. Furthermore, the large model merges the reference content provided by the knowledge retrieval model (Agent) and the expert experience module (MoE) to offer an integrated solution, enhancing the comprehensiveness of the solution.

[0076] The enterprise integrated management method based on a large model of the present invention leverages the semantic understanding capabilities of the large model to achieve large-scale dense retrieval based on semantics at the second level, as well as hybrid retrieval based on semantics and expert experience knowledge. In terms of form, the integrated management system, based on the large model, realizes the functions of literature retrieval, experience learning, and knowledge integration, directly improving the work efficiency of integrated management personnel.

[0077] Example 2 As a specific embodiment of the present invention, this embodiment provides an enterprise integrated management system based on a large model, referring to... Figure 3 It includes a vertical domain knowledge base construction unit 10, a knowledge retrieval model agent establishment unit 20, an expert experience module (MoE) construction unit 30, and a large model solution unit 40, among which, The vertical domain knowledge base construction unit 10 is used to construct a vertical domain knowledge base based on the original enterprise-level comprehensive management procedure data. The knowledge retrieval model Agent building unit 20 is used to build a knowledge retrieval model Agent based on the constructed vertical domain knowledge base; The expert experience module MoE construction unit 30 is used to construct the expert experience module MoE based on question-answer pairs; The large model answering unit 40 is used to answer user questions and output answers based on the established knowledge retrieval model Agent and expert experience module MoE.

[0078] The enterprise integrated management system based on a large model of the present invention uses four core units: a vertical domain knowledge base construction unit 10, a knowledge retrieval model agent establishment unit 20, an expert experience module (MoE) construction unit 30, and a large model answering unit 40. It improves enterprise management efficiency and service quality by integrating various enterprise-level management procedure data and expert experience. The vertical domain knowledge base construction unit 10 processes and transforms the original enterprise-level integrated management procedure data to build a vertical domain knowledge base, laying the foundation for subsequent knowledge retrieval and question answering. The knowledge retrieval model agent establishment unit 20, based on the vertical domain knowledge base, establishes a model agent specifically for retrieving relevant knowledge information, making information retrieval more accurate and efficient. The expert experience module (MoE) construction unit 30 uses a question-and-answer format to transform the expert experience accumulated within the enterprise into actionable knowledge data, supplementing the amount of searchable information, compensating for the deficiencies in the vertical domain knowledge base, and enhancing the system's intelligence and practicality. The large model answering unit 40, combining the resources provided by the first two units, utilizes the powerful processing capabilities of the large model to answer questions raised by enterprise users and output the answers.

[0079] The enterprise integrated management system based on a large model of the present invention digitizes the complex management procedures and expert experience data within the enterprise, which can quickly provide decision support for managers, greatly shorten the decision-making cycle, significantly reduce the information processing burden of integrated management personnel, and improve work efficiency and decision quality; through the effective organization and utilization of management procedure data and expert experience, it can provide users with more comprehensive decision-making methods.

[0080] Example 3 As a specific embodiment of the present invention, this embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps of the enterprise integrated management method based on a large model as described in Embodiment 1 above.

[0081] Example 4 As a specific embodiment of the present invention, this embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the enterprise integrated management method based on a large model as described in Embodiment 1 above.

[0082] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the enterprise integrated management method based on a large model as described in Embodiment 1.

[0083] The enterprise integrated management method, system, equipment and storage medium based on a large model provided by this invention can assist integrated management personnel in solving enterprise management problems, share a large amount of retrieval, reading and organization work, and thus greatly improve the efficiency of integrated management work.

[0084] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A comprehensive enterprise management method based on a large-scale model, characterized in that, The method includes: Based on the original enterprise-level comprehensive management procedure data, a vertical domain knowledge base is constructed. The vertical domain knowledge base includes multiple independent vector databases isolated by business domain, with each business domain corresponding to a vector database. Based on the constructed vertical domain knowledge base, a knowledge retrieval model Agent is established; Based on question-answer pairs, an expert experience module (MoE) is constructed. Based on the established knowledge retrieval model Agent and expert experience module MoE, the large model is used to answer user questions and output the answers; in, The construction of the vertical domain knowledge base based on the original enterprise-level comprehensive management procedure data includes: Collect the original enterprise-level integrated management procedure data, and perform heterogeneous data parsing on the collected original enterprise-level integrated management procedure data; Data augmentation is performed on the parsed enterprise-level integrated management procedure data to obtain enterprise-level integrated management procedure augmented data, and the enterprise-level integrated management procedure augmented data is stored. The enhanced data of the enterprise-level comprehensive management procedures is vectorized to generate a vector matrix, and the vector matrix is ​​stored in the corresponding business domain vector database. This includes segmenting, cleaning, and vectorizing the enhanced data of the enterprise-level comprehensive management procedures based on a natural language embedding model to generate a vector matrix, and storing the vector matrix in the corresponding business domain vector database. After storing the vector matrix in the vector database, the vector data in the vector database is sliced. The maximum length of a single text is limited by setting the batch size, and the length of repeated fragments in the context is limited by setting the overlap size. The step of establishing a knowledge retrieval model Agent based on the constructed vertical domain knowledge base includes: By using langchain technology, the vertical domain knowledge base is mounted to construct a knowledge retrieval model agent that can interact with the comprehensive management personnel in natural language. The knowledge retrieval model agent is used to perform knowledge retrieval and return reference information. The expert experience module (MoE) constructed based on question-answer pairs includes: Based on expert experience, questions and answers are collected, a question-and-answer pair database is constructed, and the data is stored in a structured manner. The process involves using the established knowledge retrieval model (Agent) and expert experience module (MoE) to answer user questions using a large model, and outputting the answers, including: Using LangChain, a CoT (Cooperation of Technology) system is constructed based on a large model to achieve intelligent scheduling of Agents and MoEs. The CoT system includes: A task routing (CoT) is constructed to identify the business domain to which the user's question belongs, and the user's question is assigned to the vector database of the corresponding business domain for knowledge retrieval, so as to reduce the error caused by the whole domain retrieval. A knowledge integration CoT is constructed, which integrates the knowledge in the vertical domain knowledge base with the expert experience in the expert experience module, and the large model returns a comprehensive solution.

2. The enterprise integrated management method based on a large model according to claim 1, characterized in that, The process of collecting the original enterprise-level integrated management procedure data and performing heterogeneous data parsing on the collected original enterprise-level integrated management procedure data includes using PaddleOCR to convert the table data and text data in the original .doc format file into .txt format. Then, the converted .txt format enterprise-level integrated management procedure data is processed, including: Remove special characters, spaces, abnormal indentation, and heading indentation errors; Correct the error section number, typos, and error messages.

3. The enterprise integrated management method based on a large model according to claim 2, characterized in that, The parsed enterprise-level comprehensive management procedure data is augmented to obtain enterprise-level comprehensive management procedure augmented data. The storage of the enterprise-level comprehensive management procedure augmented data includes compressing excessively long paragraphs and adding paragraph summaries. Specifically, a large model is used to compress excessively long paragraphs and add paragraph summaries.

4. The enterprise integrated management method based on a large model according to claim 3, characterized in that, The parsed enterprise-level integrated management procedure data is augmented to obtain enterprise-level integrated management procedure augmented data. The storage of the enterprise-level integrated management procedure augmented data includes naming each file with keywords and using business domain isolation folders.

5. The enterprise integrated management method based on a large model according to claim 1, characterized in that, The vector database is either the Chroma database or the Milvus database, and the natural language embedding model is the bge model.

6. The enterprise integrated management method based on a large model according to claim 5, characterized in that, Each vector database of the vertical domain knowledge base stores data using a single complete file as the smallest storage unit. The storage mode is continuous storage, and the original enterprise-level comprehensive management procedure data is collected and updated periodically, and then stored in the vector database after preprocessing and vectorization.

7. The enterprise integrated management method based on a large model according to claim 5, characterized in that, Based on the similarity between user questions and knowledge, a hybrid retrieval mode is constructed by combining keyword-based retrieval strategies to determine the retrieval strategy for the vector database. Euclidean distance is used to calculate the similarity between user questions and knowledge.

8. A comprehensive enterprise management system based on a large model, characterized in that, It includes a vertical domain knowledge base construction unit, a knowledge retrieval model agent establishment unit, an expert experience module (MoE) construction unit, and a large model solution unit. The vertical domain knowledge base construction unit is used to construct a vertical domain knowledge base based on the original enterprise-level comprehensive management procedure data. The vertical domain knowledge base includes multiple independent vector databases isolated by business domain, with each business domain corresponding to a vector database. The knowledge retrieval model Agent building unit is used to build a knowledge retrieval model Agent based on the constructed vertical domain knowledge base; The Expert Experience Module (MoE) construction unit is used to construct the Expert Experience Module (MoE) based on question-answer pairs. The large model answering unit is used to answer user questions and output answers based on the established knowledge retrieval model Agent and expert experience module MoE. in, The construction of the vertical domain knowledge base based on the original enterprise-level comprehensive management procedure data includes: Collect the original enterprise-level integrated management procedure data, and perform heterogeneous data parsing on the collected original enterprise-level integrated management procedure data; Data augmentation is performed on the parsed enterprise-level integrated management procedure data to obtain enterprise-level integrated management procedure augmented data, and the enterprise-level integrated management procedure augmented data is stored. The enhanced data of the enterprise-level comprehensive management procedures is vectorized to generate a vector matrix, and the vector matrix is ​​stored in the corresponding business domain vector database. This includes segmenting, cleaning, and vectorizing the enhanced data of the enterprise-level comprehensive management procedures based on a natural language embedding model to generate a vector matrix, and storing the vector matrix in the corresponding business domain vector database. After storing the vector matrix in the vector database, the vector data in the vector database is sliced. The maximum length of a single text is limited by setting the batch size, and the length of repeated fragments in the context is limited by setting the overlap size. The step of establishing a knowledge retrieval model Agent based on the constructed vertical domain knowledge base includes: By using langchain technology, the vertical domain knowledge base is mounted to construct a knowledge retrieval model agent that can interact with the comprehensive management personnel in natural language. The knowledge retrieval model agent is used to perform knowledge retrieval and return reference information. The expert experience module (MoE) constructed based on question-answer pairs includes: Based on expert experience, questions and answers are collected, a question-and-answer pair database is constructed, and the data is stored in a structured manner. The process involves using the established knowledge retrieval model (Agent) and expert experience module (MoE) to answer user questions using a large model, and outputting the answers, including: Using LangChain, a CoT (Cooperation of Technology) system is constructed based on a large model to achieve intelligent scheduling of Agents and MoEs. The CoT system includes: A task routing (CoT) is constructed to identify the business domain to which the user's question belongs, and the user's question is assigned to the vector database of the corresponding business domain for knowledge retrieval, so as to reduce the error caused by the whole domain retrieval. A knowledge integration CoT is constructed, which integrates the knowledge in the vertical domain knowledge base with the expert experience in the expert experience module, and the large model returns a comprehensive solution.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-7.

Images