A method and system for intelligent operation and maintenance of a large data center model based on ME-RAG

Abstract

The application discloses a kind of based on ME-RAG's data center big model intelligent operation and maintenance method and system, the data center material is divided into multiple fields in the application, create an expert agent in each field, build the multimodal database including text, picture and table through unstructured document, as the external embedded knowledge base of expert agent;Based on ME-RAG framework, manager agent combines expert skill list and user problem dynamically selects the expert field involved, and is assigned to relevant expert agent to answer;Each expert agent receives answer instruction, searches knowledge base according to problem, answers problem with text, if there are relevant pictures and tables, then return together, finally by the summary officer, the answer of all expert agents is summarized.The application improves the correctness and comprehensiveness of the answer to user problem through multi-expert operation and maintenance mechanism, and provides more intuitive multimodal information for user through multimodal database search mechanism.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance of data centers, specifically involving a method and system for intelligent operation and maintenance of large-scale data center models based on Multi-Experts RAG (ME-RAG). Background Technology

[0002] In recent years, the demand for computing power has grown dramatically, making data centers (DCs) a global focus. The stability and security of data centers directly impact the operation of numerous industries, their importance being self-evident. Therefore, effective operation and maintenance management of data center facilities is crucial. Currently, data center facility operation and maintenance primarily relies on professional technicians providing 24 / 7 uninterrupted service. This not only places extremely high demands on the professional skills and dedication of maintenance personnel but also requires significant investment in talent development and operational resources.

[0003] In recent years, large conversational language models such as GPT, LLama, Mistral, and ChatGLM have offered hope for solving the aforementioned challenges. They possess powerful emergent and generalization capabilities, showing great potential in data center operations and maintenance. However, the data center field is highly specialized and rapidly evolving, and existing large language models, relying solely on their own knowledge reserves, struggle to provide intuitive and effective solutions for data centers. Retrieval-Augmented Generation (RAG) can deploy specialized large models for the data center domain at low cost. Current RAG methods typically treat all text data as a knowledge base, resulting in excessively large database sizes, significantly increased computational costs for retrieval modules, and unsatisfactory scalability. Furthermore, due to the lack of image-based information to aid understanding, these methods struggle to provide intuitive and clear guidance during data center equipment maintenance. The complex and large-scale nature of data center environments necessitates collaboration among multiple domain experts to efficiently solve problems. However, due to the context length limitations of large language models, when a user poses a question involving multiple knowledge domains, those domain fragments closely related to the question are prioritized, while fragments from less relevant domains may be ignored due to their lower ranking. As a result, the answers provided to users are often neither complete nor accurate enough.

[0004] Therefore, there is an urgent need for a more efficient and intelligent operation and maintenance solution that makes artificial intelligence technology easier to understand and apply, thereby improving operation and maintenance efficiency, ensuring operation and maintenance security, and ultimately achieving automated operation and maintenance of data centers. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a data center large-scale intelligent operation and maintenance method and system based on ME-RAG. Through a multi-expert operation and maintenance mechanism, it effectively improves the accuracy and comprehensiveness of answers to user questions, and through a multi-modal database retrieval mechanism, it provides users with more intuitive multi-modal information.

[0006] Technical solution: To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides an intelligent operation and maintenance method for large-scale data center models based on ME-RAG, comprising the following steps:

[0008] Divide the data center data into multiple domains, and create an expert agent for each domain;

[0009] For each domain, a multimodal database, including text, image, and tabular databases, is constructed using unstructured documents to serve as an embedded knowledge base for expert agents.

[0010] Based on the ME-RAG framework, the manager agent dynamically selects the relevant expert domains by combining the expert skills list with the user's question, and assigns the questions to the relevant expert agents for answering.

[0011] Upon receiving the response instruction, each expert agent searches their own knowledge base based on the question and answers it in text. If there are any relevant images or tables, they are also returned. Finally, the reporting specialist summarizes the answers from all the expert agents.

[0012] Preferably, the text database is constructed using a multi-vector approach, including: extracting the text portion of unstructured documents, dividing it into blocks according to sections, inputting each block into a large language model for summarization to obtain summary blocks, embedding the summary blocks, and linking them to the source text blocks via a multi-vector.

[0013] Preferably, the image database is constructed using a multi-vector approach, including: extracting images from unstructured documents, inputting the image names and surrounding context into a large language model, summarizing the function of the images to obtain summary blocks, embedding the summary blocks, and linking the summary blocks with the image storage paths using a multi-vector.

[0014] Preferably, the tabular database is constructed using a multi-vector approach, which includes: extracting the table portion of unstructured documents to form a CSV document, then inputting the table context, table name, and table content into a large language model for summarization to obtain a summary block, and linking the summary block with the CSV file using a multi-vector.

[0015] Preferably, the manager agent consists of a large language model and a single classifier. Expert skill lists are fed into the large language model via prompts, allowing the model to understand the skills possessed by each expert. The large language model then dynamically selects the appropriate expert agent to answer the user's question, combining the expert skill list with the user's question. The single classifier is trained using business texts from various domains as input. The single classifier and the large language model jointly select the expert agent, and the union of the two is taken as the expert agent selected by the manager agent.

[0016] Preferably, the expert agent consists of retrieval modules corresponding to three modal databases: a text database, an image database, and a table database, as well as a large language model. The retrieval module calculates in the semantic space to obtain the summary block with the highest semantic similarity to the user's question and returns the source linked to the summary block to the large language model. The large language model combines the information returned by the retrieval module to answer the user's question and present relevant images and / or table information.

[0017] Preferably, based on the actual data center operation and maintenance process, the data center data is divided into multiple domains according to different devices, and the expert skill list for each domain is provided to the large language model in the form of a directory through prompt words.

[0018] Secondly, this invention provides a ME-RAG-based intelligent operation and maintenance system for large-scale data center models. Built on the ME-RAG framework, it includes a manager agent, multiple expert agents, and a reporting specialist. Each expert agent corresponds to a domain of data center data. A multimodal database, including text, image, and table databases, is constructed using unstructured documents as an embedded knowledge base for the expert agents. The manager agent dynamically selects relevant expert domains based on the expert skill list and user questions, and assigns the questions to the relevant expert agents for answering. Upon receiving an answer instruction, each expert agent retrieves its own knowledge base based on the question and answers it in text; if relevant images and tables are available, they are also returned. The reporting specialist summarizes the answers from all expert agents.

[0019] Preferably, the system is built using the AutoGen open-source framework and consists of a manager agent, multiple expert agents, and a reporting specialist.

[0020] Thirdly, the present invention provides a computer system including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when loaded onto the processor, implements the steps of the ME-RAG-based intelligent operation and maintenance method for large data center models.

[0021] Beneficial Effects: Compared with existing technologies, this invention has the following beneficial effects: 1. This invention integrates a multimodal database into multi-expert retrieval, enabling ordinary non-multimodal large models to process multimodal information, including text, tables, and images, resulting in more intuitive and understandable answers. 2. This invention fully utilizes the thinking ability of large language models. The large language model, assisted by a classifier, acts as a manager, coordinating domain experts and handling fuzzy queries, effectively reducing the risk of illusions. A series of expert models divide the database into multiple domains, improving the database's scalability. This avoids the resource consumption caused by traversing large-scale databases and the possibility that large language models may ignore weakly related but relevant domains due to contextual limitations, effectively improving the accuracy and comprehensiveness of the answers. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the ME-RAG framework and expert model in an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of the manager agent's operating mode in an embodiment of the present invention.

[0024] Figure 3 This is a schematic diagram of the thought process for constructing a question-and-answer dataset in an embodiment of the present invention.

[0025] Figure 4 This is a comparison chart of the manager agent's answer effects with different composition methods in the embodiments of the present invention.

[0026] Figure 5 This is a demonstration image of the effect of an example image-based answer in an embodiment of the present invention.

[0027] Figure 6 This is a diagram illustrating the effect of a table-based response in an embodiment of the present invention.

[0028] Figure 7 This is a diagram illustrating the effect of a comprehensive answer in an example from an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the embodiments of this invention are described in detail below with reference to the accompanying drawings. These embodiments are implemented based on the technical solutions of this invention, providing detailed implementation methods and specific operating procedures. It should be understood that the specific examples described herein are merely illustrative of this invention, but the scope of protection of this invention is not limited to the following embodiments.

[0030] This invention discloses a ME-RAG-based intelligent operation and maintenance method for large-scale data center models. The method mainly includes: dividing data center data into multiple domains; creating an expert agent for each domain; constructing a multimodal database (including text, image, and table databases) using unstructured documents as an embedded knowledge base for the expert agents; based on the ME-RAG framework, a manager agent dynamically selects the relevant expert domains by combining the expert skill list with user questions and assigns the answers to the relevant expert agents; upon receiving an answer instruction, each expert agent searches its own knowledge base based on the question and answers it in text, returning any relevant images and tables; finally, a reporting specialist summarizes the answers from all expert agents, thereby achieving a more comprehensive, intuitive, and intelligent response to data center user questions.

[0031] This invention establishes a manager agent and multiple expert agents, clearly defining responsibilities and effectively alleviating the contextual limitations and incomplete retrieval issues inherent in large language models. Through the ME-RAG framework, the manager agent assigns questions to experts in all relevant domains, considering the issues from multiple perspectives. Expert agents transform private unstructured data into a multimodal index database, enabling data center dialogue and question answering. This invention employs a multi-expert agent mechanism, subdividing the responsibilities of experts in each domain. This avoids the problem of a single searcher retrieving multiple similar knowledge blocks in highly relevant domains while neglecting knowledge in less relevant domains, thus preventing the retrieval of only one knowledge domain. It also provides scalability for subsequent feature introductions. Specifically, for data center operation and maintenance, a real data center operation and maintenance process can be simulated, and a ME-RAG intelligent operation and maintenance system can be built based on different equipment (such as air conditioning equipment and chiller equipment).

[0032] This embodiment also discloses a ME-RAG-based intelligent operation and maintenance system for large-scale data center models. The aim is to establish a private large-scale data center model, improve data center operation and maintenance efficiency, reduce manual operation and maintenance pressure, and ensure the comprehensiveness and accuracy of answers. The system is built on the ME-RAG framework and includes a manager agent, multiple expert agents, and a reporting specialist. Each expert agent corresponds to a domain of data center data. A multimodal database, including text, image, and table databases, is constructed using unstructured documents as an embedded knowledge base for the expert agents. The manager agent dynamically selects relevant expert domains based on the expert skill list and user questions, and assigns the answers to the relevant expert agents. Upon receiving an answer instruction, each expert agent searches its own knowledge base based on the question and answers it in text; if relevant images and tables are available, they are also returned. The reporting specialist summarizes the answers from all expert agents.

[0033] The following is combined Figure 1The ME-RAG framework shown further illustrates the detailed steps of this embodiment of the invention. For example... Figure 1 As shown, the ME-RAG framework is used for multi-expert retrieval and enhanced generation in data centers, consisting of three roles: manager agent, expert agent, and reporting specialist. In a data center scenario, the manager agent is responsible for assigning user questions to all experts in the relevant domains. Each expert represents a specialist in a specific domain, and upon receiving commands from the manager agent, they apply their domain-specific knowledge to answer the user's question. The reporting specialist then summarizes all the expert answers. Compared to traditional RAG, ME-RAG provides more comprehensive and accurate answers to questions involving multiple domains.

[0034] Traditional RAGs mainly consist of a retrieval module and a generation module, and are typically used to process text information, as shown below:

[0035]

[0036] Among them, the retrieval module In the document Search within the scope to extract the most relevant document fragments. Then, these fragments are compared with the questions. Combined, as a generation module The input will eventually generate the answer. .

[0037] Regarding the ME-RAG framework proposed in this embodiment, all text files It will be categorized into multiple text repositories based on its domain, i.e. ,in Number of representative fields. In managerial agency. Choose the question Related After the expert, the reporting specialist By combining the answers, the above formula can be rewritten as follows:

[0038]

[0039] in, , , Let represent the document, retrieval module, and generation module of the i-th domain, respectively. This method of storing the database separately effectively alleviates the scalability limitations of large-scale databases, significantly expanding the horizontal coverage across different domains while ensuring the maintenance of vertical retrieval depth. This multi-expert mechanism not only better aligns with real-world workflows but also effectively mitigates the contextual limitations of large language models through collaborative responses, thereby ensuring the comprehensiveness and accuracy of the answers.

[0040] Acting Manager The main structure is as follows Figure 1 As shown in the left part of the diagram, in this embodiment, the agent manager uses a large language model and a single classifier to analyze user queries and select experts in relevant fields to respond. In this process, the large language model acts as a cognitive tool, evaluating the question and identifying the relevant field. However, the large language model relies solely on the expert's name, lacking understanding of each expert's specific skills. This is because detailed information about the data center may not have been incorporated during the initial training phase of the large language model. To enhance the large language model's understanding of each expert's specific skills, this embodiment introduces a directory listing each expert's capabilities, which is input into the prompts in a List format. This structured presentation allows the large language model to better understand the relevant professional knowledge available within the system.

[0041] While large language models are powerful, they can occasionally produce inexplicable "illusions," generating incorrect or illogical content, which can negatively impact the system's normal operation. To mitigate the potential impact of these "illusions," this embodiment proposes a single classifier framework that combines machine learning techniques with large language models to support system operation and build a robust managerial agent. Specifically, the classifier uses business text from various domains as input to train a Support Vector Machine (SVM) model. User queries are processed by this classification model to determine the most relevant domain. The managerial agent uses the union of the large language model and the single classifier output as the selection result. By collaborating with the single classifier, the large language model enables the managerial agent to further identify and select a suitable list of experts. This ensemble approach significantly improves the accuracy and reliability of expert selection, thereby ensuring the smooth and efficient operation of the system.

[0042] One of the main challenges RAG faces is handling vague user queries. Vague descriptions often lead to incorrect search directions or incomplete retrieval of relevant information. To address this issue, in this embodiment, the agent manager selects all potentially relevant domain experts based on their understanding of the user's query and the expertise of each specialist. These experts collaboratively provide responses, ensuring a comprehensive answer from multiple perspectives. A specific workflow example is shown below. Figure 2 As shown, when a user asks, "What features should be considered when inspecting a battery pack?", they may actually also want information related to the uninterruptible power supply (UPS) system, even though this intent is not explicitly stated in the question. Traditional RAG (Related Information Group) might retrieve information related to battery devices based on semantic similarity, which could lead to misleading results if the question seems more relevant to battery devices. The ME-RAG framework proposed in this embodiment addresses this issue, where the agency manager selects both a battery expert and an UPS expert to jointly answer the user's query. This multi-expert collaboration allows the system to respond more comprehensively to user queries, ensuring that all relevant aspects of the problem are fully addressed. By appropriately soliciting insights from multiple experts, the quality of the information provided is improved, the problem-solving perspective is broadened, and ultimately, users are helped to make more informed decisions.

[0043] Due to the complexity and large number of components in data center equipment, textual descriptions alone are often insufficient to convey information intuitively during routine maintenance. In many cases, appropriate images and tables are needed to provide a clearer and more comprehensive explanation. Therefore, as shown in the right part of Figure 1, experts must possess the ability to process multimodal information. In this embodiment, each expert consists of three parts: a database, a retrieval module, and a generation module. We categorize knowledge into three modalities: tables, images, and text, and utilize a large language model to construct a comprehensive multi-vector. This method enables effective retrieval and response generation for different types of data. By integrating multiple modalities, the system can provide users with richer and more valuable answers, improving overall comprehensibility and usability. The system's ability to access and process multiple forms of information ensures that actual users have a comprehensive understanding of the current problems, thereby improving the efficiency and effectiveness of data center equipment operation and maintenance.

[0044] The specific method for constructing a multimodal external knowledge base is as follows: data center-related information is subdivided into multiple domains and then constructed into a text database. Image database and table database Each domain's database contains multiple document blocks, and each document block is then stored in a multi-vector format after undergoing some embedding processing. , , ,Right now:

[0045]

[0046]

[0047]

[0048] Where k represents the index of the vector.

[0049] The specific source text to multi-vector transformation form is as follows:

[0050]

[0051]

[0052]

[0053] For the text content, this embodiment uses a large language model to analyze each text block. Summarize to obtain the summary block Then through a multi-vector converter Compare it with the corresponding Mapping and matching are performed. The retrieval module retrieves the modules with the highest similarity. and the corresponding The data is transmitted to the generation module; for image content, we utilize the image's text context. and its image name This prompts large language models to generate summary blocks that describe the intent of images. Then through a multi-vector converter Will and The retrieval module finds the corresponding image based on the image intent and then feeds it back to the generation module; for tables, a similar method to image processing is used. The text surrounding the table is... And the entire table (table content and table name). The input is fed into a large language model, which then generates tables and detailed descriptive summary blocks. Finally, through a multi-vector converter Perform matching. Considering database differences, the ME-RAG framework formula is rewritten as follows:

[0054]

[0055] Each vector database acts as a retrieval module, outputting fragments matching the user's question to the generation module, specifically the large language model, where experts ultimately arrive at the answer. Retrieval Module A combination of multi-vector retrieval and the Jaccard retrieval method was employed, which improved search quality at the semantic level while mitigating the impact of term frequency. The output results generated during the retrieval process are relevant to the user's question. Together, they were then input into the generation module. Finally, the reporting specialist summarizes the expert answers using prompts to generate the final answer. .

[0056] The effects of this invention will be illustrated below with a simulation experiment.

[0057] Simulation Experiment Setup: This experiment was built using the open-source framework Autogen and the ChatGLM3-6B model. Furthermore, the embedding model used was text2vec-base-chinese, which can effectively handle Chinese text representation. The dataset was processed based on Part 2 of "From Operations Novice to Expert: Data Center Facility Operations Guide," specifically the Key Equipment Operations Guide. To compare the experimental results, a question-and-answer dataset was constructed using a thought-chain approach, such as... Figure 3 As shown, the reasoning ability of a large language model is enhanced by outputting the intermediate steps of the reasoning process, thus making the reasoning process more transparent and comprehensive. The specific process is as follows:

[0058] First, the chapters in the book were divided into several non-overlapping paragraphs. Then, using the generative capabilities of a large language model, the key points of each paragraph were summarized and extracted.

[0059] Subsequently, for each identified key point, the large language model is prompted to raise a corresponding question and provide an answer based on information from the relevant paragraphs.

[0060] Finally, the generated content underwent rigorous review by professionals, who screened and improved it, resulting in a high-quality question-answering dataset. This dataset contains a total of 441 records, some of which contain poorly worded questions, serving as a test of the ME-RAG framework.

[0061] This systematic approach ensures the comprehensiveness and reliability of the question-and-answer dataset, thereby enhancing its practicality in data center operations.

[0062] Based on the constructed dataset, the number of experts in the ME-RAG framework is set to 10, including uninterruptible power supply (UPS) systems, data center infrastructure management (DCIM), data center operations management (DCOM), air-cooled air conditioning systems, structured cabling systems, batteries, backup generators, KVM systems, chilled water air conditioning systems, and high and low voltage systems.

[0063] To compare performance, the experiment used ROUGE and cosine similarity as the main evaluation metrics, focusing on the semantic space of text overlap and answer quality, respectively. ROUGE-1 highlights the matching of individual words, while ROUGE-L considers the preservation of word order in the text. On the other hand, cosine similarity measures the similarity between word vector pairs, effectively reflecting the semantic consistency between corresponding texts. Furthermore, to evaluate the decision-making accuracy of the managerial agent in the ME-RAG framework, the "Expert Accuracy" metric was introduced. Simultaneously, when the large language model "illusory" and did not select any expert, all metrics for the corresponding samples were set to 0 in the experiment. This comprehensive evaluation framework rigorously assesses the performance of the ME-RAG framework in generating relevant and accurate answers in a data center operational context.

[0064] Table 1 illustrates the advantages of the multi-expert mechanism proposed in this invention, comparing the results of the ME-RAG framework with those of a single-expert RAG that integrates all knowledge. The results show that the ME-RAG framework outperforms the single-expert agent across all evaluation metrics.

[0065] Table 1 Comparison of RAG results from single-expert and multi-expert teams

[0066] method Rouge-1 Rouge-L Cosine Single Expert RAG 0.5990 0.5439 0.7926 ME-RAG 0.6534 0.5920 0.8150

[0067] Meanwhile, to demonstrate the superiority of the ME-RAG framework in decision-making capabilities, the experiment used different decision-making methods and conducted a comprehensive comparison across various evaluation indicators, such as... Figure 4 As shown in the diagram. In this experiment, "NC (No content, no table of contents)" means the agent manager only knows the expert's name and does not provide the expert's specific skill content in the prompt. "LC (Content generated by the large language model)" means that the expert content generated by the large language model is included in the agent manager's skill list. "MC (Manually created content)" refers to integrating manually created content (such as a book's table of contents) into the agent manager's skill list. "MC+Class" means adding a single classifier to handle the situation where the large language model produces "illusions".

[0068] from Figure 4The results clearly show that relying solely on the embedded knowledge of the large language model can only provide a rough classification of user queries, making it difficult to accurately grasp the specific skills of each expert. This highlights the limitations of relying solely on the understanding capabilities of the large language model and underscores the necessity of incorporating expert skill content. However, it is worth noting that using expert content generated by the large language model leads to even worse results. In the comparative experiments of the "LC" group, we observed that in many cases, the agent manager failed to effectively classify user queries. This is because the text directly input into the large language model, and the content generated by it, is overly detailed and verbose. Such lengthy context hinders the large language model's ability to effectively understand user queries, significantly increasing the likelihood of "illusions" and ultimately failing to generate accurate answers. In contrast, ME-RAG's decision mechanism performs exceptionally well in addressing the "illusion" problem. By using concise, manually curated, and more easily understood directory points, and with the support of a single classifier, it achieves optimal decision performance.

[0069] exist Figures 5 to 7 In the diagram, we simulate the actual results within a user interface to demonstrate the ME-RAG framework's capabilities in multimodal retrieval and multi-expert analysis. Due to space limitations, we have only included key information from the answers in the figures. Figure 5 The image retrieval capabilities of the ME-RAG framework were demonstrated, which can locate relevant images from a dataset based on user queries and provide an intuitive visual presentation. Figure 6 This demonstrates the ME-RAG framework's ability to retrieve answers from table content and respond to user queries. To provide a more intuitive understanding, the ME-RAG framework directly displays the retrieved table to the user. Figure 7 This demonstrates the comprehensive retrieval capabilities of the ME-RAG framework. When a user inquires about evaporators, evaporator devices in both air-cooled and water-cooled air conditioning systems are relevant to the question. Therefore, the ME-RAG framework selects experts in these two fields to provide relevant information about evaporators, retrieves relevant images for better visualization, and synthesizes the responses from both experts into a coherent summary.

[0070] Experimental results show that ME-RAG has good application prospects, strong scalability, and comprehensive answers. It addresses the problem of single retrieval in traditional RAG, alleviates the contextual limitations of large language models, introduces multimodal information to make data center problem processing more intuitive, and greatly reduces the operation and maintenance pressure on data center maintenance personnel.

[0071] This invention also discloses a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the steps of the aforementioned intelligent operation and maintenance method for a large data center model based on ME-RAG.

[0072] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A data center large-scale intelligent operation and maintenance method based on ME-RAG, characterized in that, Includes the following steps: Data center data is divided into multiple domains based on different devices, and an expert agent is created for each domain; For each domain, a multimodal database, including text, image, and tabular databases, is constructed using unstructured documents to serve as an embedded knowledge base for expert agents. The text database, image database, and table database are all constructed using a multi-vector approach. The image name and its surrounding context are fed into the large language model to summarize the function of the image and obtain a summary block. The table context, table name, and table content are fed into the large language model to form a summary block. Based on the ME-RAG framework, the manager agent dynamically selects the relevant expert domains by combining the expert skills list with the user's question, and assigns the questions to the relevant expert agents for answering. Upon receiving the response instruction, each expert agent searches their own knowledge base based on the question and answers it in text. If there are any relevant images or tables, they are returned as well. Finally, the reporting specialist summarizes the answers from all the expert agents. The manager agent consists of a large language model and a single classifier. Expert skill lists are fed into the large language model via prompt words, allowing the large language model to know the skills possessed by each expert. The large language model combines the expert skill list with the user's question to dynamically select the expert agent suitable for answering the user's question. The single classifier is trained using business texts from various domains as input. The single classifier and the large language model jointly select the expert agent, and the union of the two is taken as the expert agent selected by the manager agent. The expert agent consists of three modal databases: a text database, an image database, and a table database, as well as corresponding retrieval modules and a non-multimodal large language model. The retrieval module calculates the summary block with the highest semantic similarity to the user's question by performing calculations in the semantic space, and returns the source linked to the summary block to the non-multimodal large language model. The non-multimodal large language model combines the information returned by the retrieval module to answer the user's question and present relevant images and / or table information.

2. The intelligent operation and maintenance method for large-scale data center models based on ME-RAG according to claim 1, characterized in that: The text database is constructed using a multi-vector approach, which includes: extracting the text portion of unstructured documents, dividing it into blocks according to sections, inputting each block into a large language model for summarization to obtain summary blocks, embedding the summary blocks, and linking them to the source text blocks via a multi-vector.

3. The intelligent operation and maintenance method for large-scale data center models based on ME-RAG according to claim 1, characterized in that: The image database is constructed using a multi-vector approach, which includes: extracting images from unstructured documents, inputting the image names and surrounding context into a large language model, summarizing the function of the images to obtain summary blocks, embedding the summary blocks, and linking the summary blocks with the image storage paths using a multi-vector.

4. The intelligent operation and maintenance method for large-scale data center models based on ME-RAG according to claim 1, characterized in that: The table database is constructed using a multi-vector approach, which includes: extracting tables from unstructured documents to form CSV documents, then inputting the table context, table name, and table content into a large language model for summarization to obtain summary blocks, and linking the summary blocks with the CSV files using a multi-vector approach.

5. The intelligent operation and maintenance method for large-scale data center models based on ME-RAG according to claim 1, characterized in that: Based on the actual data center operation and maintenance process, the data center data is divided into multiple domains according to different devices. The expert skill list for each domain is provided to the large language model in the form of a directory and through prompt words.

6. A data center large-scale intelligent operation and maintenance system based on ME-RAG, characterized in that: Built on the ME-RAG framework, it includes manager agents, multiple expert agents, and reporting specialists; data center data is divided into multiple domains according to different devices, and each expert agent corresponds to one domain of data center data. Through unstructured documents, a multimodal database including text database, image database, and table database is built as an embedded knowledge base for expert agents. The text database, image database, and table database are all constructed using a multi-vector approach. The image name and its surrounding context are fed into the large language model to summarize the function of the image and obtain a summary block. The table context, table name, and table content are fed into the large language model to form a summary block. The manager agent is used to dynamically select the relevant expert fields based on the expert skills list and the user's question, and assign them to the relevant expert agents for answering; each expert agent who receives the answer instruction searches its own knowledge base according to the question and answers the question in text, and returns relevant images and tables if applicable; the reporting specialist is used to summarize the answers of all expert agents. The manager agent consists of a large language model and a single classifier. Expert skill lists are fed into the large language model via prompt words, allowing the large language model to know the skills possessed by each expert. The large language model combines the expert skill list with the user's question to dynamically select the expert agent suitable for answering the user's question. The single classifier is trained using business texts from various domains as input. The single classifier and the large language model jointly select the expert agent, and the union of the two is taken as the expert agent selected by the manager agent. The expert agent consists of three modal databases: a text database, an image database, and a table database, as well as corresponding retrieval modules and a non-multimodal large language model. The retrieval module calculates the summary block with the highest semantic similarity to the user's question by performing calculations in the semantic space, and returns the source linked to the summary block to the non-multimodal large language model. The non-multimodal large language model combines the information returned by the retrieval module to answer the user's question and present relevant images and / or table information.

7. The intelligent operation and maintenance system for large-scale data center models based on ME-RAG according to claim 6, characterized in that: The system is built using the AutoGen open-source framework and consists of a manager agent, multiple expert agents, and a reporting specialist.

8. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is loaded into the processor, it implements the steps of the intelligent operation and maintenance method for large data center models based on ME-RAG as described in any one of claims 1-5.

Images