Intranet intelligent operation and maintenance workflow container system based on LLM and embedded model

Abstract

The invention relates to an intranet intelligent operation and maintenance workflow container system based on LLM and an embedded model, and the system comprises an operation and maintenance knowledge base data embedding module which carries out the vectorization of high-dimensional discrete operation and maintenance knowledge base text data through the embedded model; the BMC state data crawling script module is used for performing data crawling on server state data under a server cluster BMC page IP (Internet Protocol) by a script and realizing automatic execution in the intranet container workflow platform module; the semantic large model module is used for analyzing user semantic data, operation and maintenance knowledge base vectorization data and BMC state data in an intranet environment; the intranet container workflow platform module is used for conveniently deploying an application platform by a container in an intranet environment to realize the integration of each module and the compilation of a work task process, adding specific conditions of intranet BMC state data on the basis, and comprehensively answering operation and maintenance related problems; the problems that the efficiency is low and the error rate is high when a large language model is adopted in intranet intelligent operation and maintenance are solved, and the problem solving is faster and more accurate.

Description

Technical Field

[0001] This invention relates to an information processing technology based on artificial intelligence for maintaining computer system operation, and particularly to an intranet intelligent operation and maintenance workflow container system based on LLM and embedded models. Background Technology

[0002] There are three technical bases for this approach: First, the development of domestically developed open-source large-scale models is promising, and related technologies are relatively mature, with DeepSeek and Qwen series being typical examples. Second, real-time technical support is lacking in intranet environments such as defense. Third, server clusters typically consist of a large number of servers, making traditional manual solutions for single-point problems inefficient. With the rapid development of artificial intelligence technology, its applications are gradually permeating enterprise offices and daily life, helping people solve a massive number of problems. As a core technology in the field of artificial intelligence, semantic large-scale models have been the subject of intense research in recent years, with widespread applications including customer service, financial advisory, and technical support.

[0003] Intranet intelligent operation and maintenance (O&M) is an application approach that fully leverages the capabilities of semantic big data models. Related patents include "Intelligent O&M System Based on AI Big Data Model and RPA," "Intelligent O&M Assisted Decision-Making Method, Device, and Medium Based on Large Language Model," and "An Intelligent O&M System, Method, Electronic Device, and Storage Medium." The "Intelligent O&M System Based on AI Big Data Model and RPA" patent uses a large language model to provide intelligent answers; however, it does not employ embedded models to vectorize the knowledge base, potentially leading to situations where keywords cannot be searched for complex queries. Both the "Intelligent O&M Assisted Decision-Making Method, Device, and Medium Based on Large Language Model" and "An Intelligent O&M System, Method, Electronic Device, and Storage Medium" patents require manual input of device status data. With a large number of devices, manual data transcription is inefficient and has a high error rate. Summary of the Invention

[0004] To address the issues of low efficiency and high error rate in using large language models for intelligent intranet operations and maintenance, a workflow container system for intelligent intranet operations and maintenance based on LLM and embedded models is proposed.

[0005] Compared to other similar patents, this invention innovatively proposes using workflow to control server cluster status data, and using models to propose solutions after problems occur, allowing users to make judgments and choices, and can solve: 1) Accurate feedback: It uses a combination of large language models and embedding models to understand and return the operation and maintenance information required by users in real time and accurately. At the same time, it can query knowledge base data more accurately when combining keyword and semantic matching.

[0006] 2) Status control: Select a workflow platform to crawl BMC status data according to the user's selected range, and provide the required server status data accurately and quickly. Compared with manual single-point query, it is much more efficient and can also provide a more comprehensive control over the server cluster status.

[0007] 3) Cluster analysis: Based on the operation and maintenance knowledge base, the system analyzes abnormal BMC status data and provides corresponding solutions. Users can then decide on the solution selection, which greatly improves the efficiency of problem solving while preserving the user's initiative.

[0008] The technical solution of this invention is as follows: This invention proposes an intranet intelligent operation and maintenance workflow container system based on the combination of large language model (LLM) and embedding model, which includes four main modules: operation and maintenance knowledge base data embedding module, BMC (Basement Management Controller) status data crawling script module, semantic large model module, and intranet container workflow platform module.

[0009] The first module is the data embedding module for the operations and maintenance (O&M) knowledge base. This module vectorizes the high-dimensional, discrete text data of the O&M knowledge base using an embedding model, facilitating semantic matching by the semantic big data model module. Traditional knowledge bases often rely on keyword-based text retrieval, while data embedding introduces an embedding model different from the semantic big data model, enabling the semantic big data model to understand the semantic information of the text and achieving a degree of "fuzzy search." Currently, a hybrid retrieval mode combining both methods is commonly used. This embedding model can be retrieved using Ollam, and commonly used embedding models such as bge-m3 can be selected for text vectorization. The O&M knowledge base text data is arranged by O&M engineers and needs to be adapted to the vectorization parameters of the embedding module to a certain extent. Furthermore, the embedding model is usually performed after text segmentation; therefore, the segmentation parameters also need to adapt to the context window constraints of the embedding model to preserve appropriate contextual dependencies and further ensure retrieval accuracy.

[0010] Secondly, there is the BMC status data crawling script module. This module uses scripts to crawl server status data (such as network card availability and power current values) under the BMC page IP of the server cluster, and automatically executes the data within the intranet container workflow platform module, providing it to the semantic big model module for analysis. As the core data acquisition unit connecting the BMC device and the upper-layer data application system, it is responsible for accurately and in real-time acquiring key status information during the BMC device's operation, providing underlying data support for device monitoring, fault early warning, and maintenance decisions. The BMC is essentially an independent microcomputer focused on device management, generally including hardware status monitoring, remote control, and security management functions. The BMC typically provides a web management interface. Crawling device status data can be done using Python scripts to simulate the process of "browser login to access the data interface, and then extracting values." The core tools are requests (for handling HTTP requests) and BeautifulSoup / lxml (for parsing HTML). If the data is dynamically loaded via AJAX, it is necessary to analyze the API interface to directly request JSON data, which is more efficient. In the absence of a web management interface, BMC interacts only through a dedicated protocol (such as IPMI). Python needs to call the protocol library to simulate sending commands and parse the protocol response to obtain values. The core tools are pyipmi (IPMI protocol) or pysnmp (SNMP protocol).

[0011] Thirdly, there is the semantic large-scale model module, used to analyze user semantic data, vectorized data from the operations and maintenance knowledge base, and BMC status data within the intranet environment. This module requires a certain amount of AI computing power to improve the analysis speed. In most cases, open-source large-scale models such as Deepseek can be used, and a parameter count of less than 7 bytes is sufficient to achieve relatively smooth technical question-and-answer functionality. There are several ways to deploy this module. A relatively simple method is to use an open-source large-scale model tool, such as Ollam, which allows you to directly pull model parameters via command. The specific selection and debugging of the large-scale model should be based on the user's actual needs. Generally, it is necessary to simultaneously change the model's prompt words and test the model output after the changes. Fine-tuning training can also be used to achieve better output results. Additionally, the output constraint functions built into some large-scale model tools can be used to impose certain constraints on the output text.

[0012] Fourth is the intranet container workflow platform module, which enables the convenient deployment of application platforms in the intranet environment to integrate various modules and orchestrate work processes. This allows for accurate analysis of user needs within the intranet environment, and the addition of specific intranet BMC status data to comprehensively answer operation and maintenance related questions.

[0013] This module contains two sub-modules: the intranet container sub-module and the workflow platform sub-module.

[0014] The intranet container submodule is used to deploy virtual software environments, facilitating the construction of services required by workflow platforms. Container technology, as an operating system-level virtualization technology, essentially provides a "resource sandbox" for programs through the system kernel, achieving lightweight efficiency, strong portability, and secure isolation. This submodule can be deployed on Linux-based operating systems (such as the domestic Kylin and Euler systems), and testing has shown that it can be deployed on servers with ARM or x86 hardware architectures. Docker officially only provides image sources supporting x86 architecture for container deployment. To deploy ARM architecture containers, the pull source address and related configurations can be modified in the YAML file.

[0015] The workflow platform submodule is built upon the intranet container submodule. This submodule integrates the operations and maintenance knowledge base data embedding module, the BMC status data crawling script module, and the semantic big data model module as the main components of the platform's workflow. Its basic workflow is as follows: User initiates a question -> Semantic big data model performs semantic analysis -> Semantic matching and keyword-based retrieval of the operations and maintenance knowledge base by the embedded model -> Semantic big data model performs semantic analysis on the retrieval results and uses them as context for the user's question, providing feedback based on the user's question -> Feedback is divided into two scenarios: first, when BMC status data is not needed, the semantic big data model directly answers the relevant information in the user's question; second, when BMC status data is needed, it is crawled by the BMC status data crawling script module, and the semantic big data model provides an answer based on the data analysis. Figure 1 As shown. The workflow platform itself is a development platform for large-scale model-related applications. It is relatively easy to deploy and can combine simple code with large models.

[0016] The beneficial effects of this invention are as follows: Due to the adoption of the above technical solution and planning process, the beneficial effects of this invention are: 1) The combination of large language models and embedding models improves the efficiency of users in handling server cluster problems, and provides more accurate feedback, enabling faster and more accurate problem solving.

[0017] 2) The BMC status data crawling script module provides users with more comprehensive data about the server cluster, enabling them to efficiently control the status of the server cluster.

[0018] 3) The workflow platform sub-module analysis points out the possible causes of server cluster problem data, allowing users to choose solutions, and users and AI jointly conduct a comprehensive analysis of the problem. Attached Figure Description

[0019] Figure 1 This is a user operation and maintenance workflow service flowchart for the present invention; Figure 2 This is a diagram showing the running status and list of the large language model and embedding model of this invention; Figure 3 This is a confirmation image showing the installation of Docker and Docker Compose in this invention. Figure 4 Modify the mirror source image of the YAML file for this invention. Figure 5 This is a diagram showing the container pull image list of the present invention; Figure 6 This is a diagram confirming the container's operating status according to the present invention; Figure 7 This is a configuration diagram of the platform model list for this invention; Figure 8 This is a diagram showing the embedded index configuration of the operation and maintenance knowledge base of this invention; Figure 9 This is a workflow diagram for module integration in this invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0021] One server is used as the management server, while three others form a server cluster. All servers are based on the ARM64 architecture. The management server uses the domestically developed open-source Huawei Euler system. The operating system of the server cluster is not limited, but it must be equipped with a BMC chip for querying. This invention is an intranet intelligent operation and maintenance workflow container system based on a combination of large language models and embedded models. The deployment process and operation mode of each module in this system are shown below: Large language models and embedding models are pulled using the official Ollam method. Ollam is a commonly used open-source large model platform. First, download the Ollam software corresponding to your system, then search for the required large model on the official website, select the version number and model parameter count, and then enter the relevant command `ollama pull ...` to obtain the model. This method requires an internet connection. If an internet connection is not possible, models can also be created offline using gguf files. After pulling the model, you can enter `systemctl status ollama.service` to check its running status, and you can use the `ollama list` command to view the model list. Figure 2 As shown. After installing Docker and Docker Compose on the management server, you can use `docker -v` and `docker-compose -v` to verify the installation and use containers to run the workflow platform that will be installed later. The versions of Docker and Docker Compose must correspond to the requirements of the workflow platform. Here, the installed Docker version is v27.3.1, and the Docker Compose version is v2.30.3. Figure 3 As shown.

[0022] The purpose of installing containers is to integrate a workflow platform. The official website provides an integration package for the x86 platform, which can be downloaded directly and followed according to the official tutorial. Since the official Docker images do not support the ARM64 architecture, the YAML file configuration for pulling the corresponding image source needs to be modified. For x86 architecture, simply use the official image source pull list; the YAML file does not need to be modified. Figure 4 As shown, after setting the image to the public domestic myhuaweicloud mirror source and modifying other related configuration items according to the mirror source, you can use the command `docker-compose up -d` to pull the mirror source in the docker folder. After the pull is complete, you can use the command `docker ps` to check if the mirror source list is complete. The complete mirror source list for the ARM64 architecture is as follows: Figure 5 As shown. Generally, after entering the command `docker-compose up -d`, the container status list and... Figure 6 Consistency means that the preliminary work for running the workflow platform has been completed.

[0023] The advantages of the large-scale model application development workflow platform lie in its high degree of editing freedom, broad task coverage, and support for local custom configuration, facilitating complex task execution within an intranet environment. Access to the workflow platform is granted via an IP address assigned to the local container, and this IP address must be on the same network segment as the server cluster's BMC IP address. In offline mode, the local model list needs to be configured manually, as shown in the configuration diagram below. Figure 7 As shown. After the operations and maintenance knowledge base is organized into a TXT file, the indexing method can be changed through the knowledge base index configuration provided by the platform. After simple testing, it is believed that for the operations and maintenance knowledge base, high-quality embedding, parent-child segmentation, and hybrid retrieval will yield more accurate and detailed indexing results. The specific configuration method still needs to be adjusted according to the knowledge base structure.

[0024] The final module integration workflow is as follows: Figure 8 As shown, this workflow is based on the technical solution and Figure 1The workflow, which involves executing tasks via a flowchart, can be integrated into a single, concise dialog interface by the platform's publishing function for user convenience. The main function of this workflow is to provide BMC status data for the server cluster when needed, based on potential or existing O&M-related issues raised by users on the management server. This data is then combined with an O&M knowledge base processed by a locally configured large language model (mistral) and an embedded model (nomic-embedded-text) to provide solutions to the O&M problems. The workflow is built by first introducing the large language model and the O&M knowledge base, then introducing the BMC status data crawling script, and finally adding conditional statements and error handling methods to connect the various modules. If needed, the modules can be expanded, such as using multiple large language models for problem analysis or multiple O&M knowledge bases with conditional statements to improve accuracy.

[0025] The technical comparison is as follows: Related search result 1: Application (Patent) Number: CN202510646108.6 Title: Intelligent Operation and Maintenance System Based on AI Large Model and RPA Abstract: This invention relates to an intelligent operation and maintenance system based on AI large-scale models and RPA, belonging to the field of data processing. It receives user input information in response to an RPA robot, performs semantic matching using an AI semantic model to obtain the inquiry question; processes the inquiry question using an answer material library, matching several historical answers; distributes the weights of the historical answers to obtain several centralized weights; processes the inquiry question using question-and-answer logs, distributes the answer frequency weights of the historical answers to obtain several frequency weights; and searches for the target answer based on the centralized weights and frequency weights, feeding it back to the RPA robot to reply to the user. This invention solves the technical problems of intelligent operation and maintenance systems struggling to effectively handle complex and non-standardized user questions, having low adaptability of answer matching algorithms, and limited multimedia information processing capabilities.

[0026] Comparison of key technical points: 1. This invention offers simpler and more accurate knowledge base retrieval: The aforementioned patent uses traditional statistical methods to centralize user-submitted questions and improve adaptability, which does indeed enhance the accuracy of answers to some extent. However, answering non-standardized questions using probabilistic methods always has limitations. The hybrid retrieval method using embedded models is simpler to deploy, requires no statistical methods, and the semantic matching knowledge base approach also results in higher knowledge base retrieval accuracy.

[0027] 2. This invention provides a more comprehensive answer: The aforementioned patent used statistical methods to select answers, but it did not include device status data, which may have resulted in incomplete answers. This invention uses a BMC status data crawling script module, and the Big Prophecy model will combine the data returned by this module to provide users with more comprehensive feedback.

[0028] Related search result 2: Application (Patent) Number: CN202510247549.9 Title: Intelligent Operation and Maintenance Auxiliary Decision-Making Method, Equipment and Medium Based on Large Language Model Abstract: This invention discloses an intelligent operation and maintenance auxiliary decision-making method, device, and medium based on a large language model. It includes an offline professional knowledge base and knowledge graph construction process, comprising a local knowledge base generation module, a knowledge graph construction module, and a generation logic rule module. The local knowledge base generation module and the knowledge graph construction module process input text data, while the generation logic rule module directly processes the results of the knowledge graph construction module. It also includes an online large model-based auxiliary decision-making process, comprising a text knowledge retrieval module, a rule query module, and a prompt template module. The prompt template module processes the results of the text knowledge retrieval module and the rule query module, which operate in parallel and process local knowledge base and knowledge graph data respectively. This invention combines a large model and a local knowledge base, significantly improving model performance and greatly reducing application costs.

[0029] Comparison of key technical points: 1. Both the knowledge base retrieval methods of this invention and the aforementioned patent have their advantages and disadvantages: The aforementioned patent uses a knowledge graph approach to optimize the knowledge base structure, resulting in higher retrieval efficiency and fewer errors. This invention uses an embedded model combined with a knowledge base to achieve a hybrid retrieval that combines vectorized semantic matching and keyword matching. Knowledge graphs are more standardized but require more resources and are more complex to process, while hybrid retrieval is relatively simpler but can still achieve a high retrieval accuracy.

[0030] 2. The data analysis of this invention is simpler and more comprehensive: The aforementioned patent requires users to manually input all relevant data, which may result in a small number of errors and is complex to operate. This invention uses a workflow platform equipped with a BMC status data script module, which can directly provide device status data in a workflow manner based on user queries, eliminating the need for manual input by the user, and can analyze the status based on this data, thus providing a more comprehensive answer.

[0031] Related search result 3: Application (Patent) Number: CN202311704609.2 Title: An Intelligent Operation and Maintenance System, Method, Electronic Device, and Storage Medium Abstract: This application provides an intelligent operation and maintenance system, method, electronic device, and storage medium. The system includes a natural language processing model to perform semantic analysis on first execution task information, determine the keyword information and task request type of the first execution task information, and send the keyword information and task request type to a knowledge base management module. If the task request type is an execution type, the system determines the task execution configuration information and target API interface parameters, and sends them to the operation and maintenance execution module. The knowledge base management module determines the target information corresponding to the first execution task information based on the keyword information and task request type. The operation and maintenance execution module determines the execution result corresponding to the first execution task information based on the task execution configuration information and target API interface parameters. This improves the processing efficiency and accuracy of operation and maintenance tasks and reduces the workload of manually configuring and managing operation and maintenance tasks.

[0032] Comparison of key technical points: 1. The execution method of this invention is more standardized: The aforementioned patents directly use the target API interface parameters for task execution without a pre-process of data crawling, which may lead to inaccurate problem analysis or difficulty in solving multiple problems in the server cluster. This invention combines a workflow platform to crawl BMC status data, and ultimately solves the problem through a collaborative approach between humans and models, giving full play to subjective initiative while also eliminating the problem of model judgment errors.

[0033] 2. This invention combines an embedded model to improve the accuracy of answers: The aforementioned patent uses a knowledge base management module to search for relevant text using keyword retrieval. However, in cases where users ask non-standardized questions, relevant text may not be found. This invention, on the other hand, employs a hybrid retrieval method, combining semantic matching based on embedded model vectorization with keyword matching. This approach yields more accurate retrieval results and is easier for large language models to analyze.

[0034] The above-described embodiments are merely one implementation of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A container system for intelligent intranet operation and maintenance workflow based on LLM and embedded models, characterized in that, include: The system includes a data embedding module for the operations and maintenance knowledge base, a BMC status data crawling script module, a semantic large model module, and an intranet container workflow platform module, among which: The operation and maintenance knowledge base data embedding module vectorizes high-dimensional discrete operation and maintenance knowledge base text data through an embedding model, so that the semantic large model module can perform semantic matching. The BMC status data crawling script module uses scripts to crawl server status data under the IP address of the BMC page in the server cluster, and automatically executes it in the intranet container workflow platform module to provide it to the semantic big model module for analysis. The semantic big model module is used to analyze user semantic data, vectorized data of the operation and maintenance knowledge base, and BMC status data in an intranet environment; The intranet container workflow platform module is designed to facilitate the deployment of applications in an intranet environment using containers, thereby integrating various modules and orchestrating workflows. This allows for precise analysis of user needs within the intranet environment, and by incorporating specific intranet BMC status data, it provides comprehensive answers to operation and maintenance-related questions. The intranet container workflow platform module comprises two sub-modules: the intranet container sub-module and the workflow platform sub-module. The intranet container sub-module deploys a virtual software environment to facilitate the setup of the services required by the workflow platform. The workflow platform sub-module, built upon the intranet container sub-module, integrates the operations and maintenance knowledge base data embedding module, the BMC status data crawling script module, and the semantic big model module as the main components of the platform's workflow. Its basic workflow is as follows: user initiates a question -> semantic big model performs semantic analysis -> semantic matching and keyword-based retrieval of the operations and maintenance knowledge base using the embedded model -> semantic big model performs semantic analysis on the retrieval results and uses them as context for the user's question, providing feedback based on the user's question -> feedback is divided into two scenarios: first, when BMC status data is not needed, the semantic big model directly answers the relevant information in the user's question; second, when BMC status data is needed, it is crawled by the BMC status data crawling script module, and the semantic big model then provides an answer based on the data analysis.

2. The intranet intelligent operation and maintenance workflow container system based on LLM and embedded model according to claim 1, characterized in that, The text data in the operations and maintenance knowledge base is arranged by operations and maintenance engineers and needs to be adapted to the vectorized parameters of the embedding module. In addition, the embedding model needs to be performed after the text is segmented, so the segmentation parameters also need to adapt to the context window constraints of the embedding model in order to retain appropriate contextual dependencies and further ensure the accuracy of retrieval.

3. The intranet intelligent operation and maintenance workflow container system based on LLM and embedded model according to claim 1, characterized in that, The BMC status data crawling script module serves as a data acquisition unit connecting BMC devices and upper-layer data application systems. It is responsible for accurately and in real-time acquiring key status information during the operation of BMC devices, providing underlying data support for equipment monitoring, fault early warning, and operation and maintenance decisions. BMC provides a web management interface. The process of crawling device status data uses Python scripts to simulate the flow from "browser login to accessing the data interface, and then extracting values." The core tools are requests and BeautifulSoup / lxml. If the data is dynamically loaded via AJAX, the API interface needs to be analyzed to directly request JSON data, which is more efficient. When BMC does not provide a web management interface, it interacts only through a dedicated protocol. Python needs to call the protocol library to simulate command sending and parse the protocol response to obtain values. The core tools are pyipmi or pysnmp.

4. The intranet intelligent operation and maintenance workflow container system based on LLM and embedded model according to claim 1, characterized in that, The semantic large model module requires AI computing power to improve the analysis speed. The specific selection and debugging of the large model should be based on the user's actual needs. It is necessary to change the model prompt words synchronously and test the model output after the change, or to use fine-tuning training to achieve better output results. It also includes using some output limitation functions built into some large model tools to impose certain constraints on the output text.

Images