Alarm treatment method based on multiple agents

Abstract

In order to solve the problems that in the prior art, a use method is simple and single, and the answer content is one-sided, wrong and inaccurate, the invention provides a multi-agent-based alarm treatment method which comprises the following steps: sequentially outputting answers of questions replied by a user to a historical agent, a tool agent, a detection agent, an execution agent and a knowledge base agent; and finally, the knowledge base intelligent agent outputs the associated text blocks to the large model for cue word output, and knowledge answers of the questions are obtained. According to the method, multiple prompt words can be configured, large models of different manufacturers can be adapted, different characteristics are provided according to factory business requirements, multiple agents are used for providing support assistance, the accuracy is high, and the problem solving is comprehensive.

Description

Technical Field

[0001] The present invention relates to the field of AI intelligent technology, and in particular to an alarm management method based on multi-agents. Background Art

[0002] Alarm management requires a lot of analysis and interpretation every day, including knowledge interpretation, position number import, statistical plan establishment, statistical plan analysis, etc., which requires a lot of complicated and repetitive operations. Modern technology has brought AI capabilities, but most of the intelligence we use is used to deal with some problems separately, and it is impossible to deal with complex tasks in a related manner, resulting in scattered use and inability to process coherently. Most processing solutions are simply hard-coded according to the problem and continuously process to get the answer to the problem. Once the problem is switched or the conditions are changed, the correct answer cannot be given or the answer is wrong. Here, a general multi-task collaborative assistant device based on multi-agents is needed. In the processing of complex problems, the problems are subdivided separately, and the historical correlation processing is carried out before and after to obtain a general assistant device with high stability and accurate answers, and thus displayed to the front desk to assist operators in completing tedious and repetitive work.

[0003] In order to solve the above problems, the Chinese invention patent with patent number CN116842133A proposes a problem processing method, device, equipment and storage medium based on artificial intelligence, which involves the field of artificial intelligence, specifically natural language processing, deep learning, and intelligent question-answering technology, and can be applied in smart cities and smart government affairs. The solution is: determine the first correlation between the problem to be processed and multiple standard answers in the answer library; determine the target answer from multiple standard answers based on the first correlation of multiple standard answers; and respond to the problem to be processed based on the target answer. However, the invention has no historical content to provide support, resulting in simple answers and incomplete functions. Summary of the invention

[0004] The present invention mainly solves the problem that the method used in the prior art is simple and single, which easily leads to one-sided answers and inaccurate errors, and provides an alarm management method based on multi-agents.

[0005] The above technical problems of the present invention are mainly solved by the following technical solutions: A multi-agent-based alarm management method comprises the following steps: S1. The answer to the question replied by the user is first output to the historical agent; S2, the tool agent then analyzes the problem through the big model and selects to use the specified tool in the tool warehouse; S3, the detection agent performs a large model analysis based on the confirmation information of the corresponding tool used by the user, and determines whether to cancel, modify, or confirm the use; S4, then the execution agent analyzes again whether to use the warehouse tool or the knowledge base agent based on the historical knowledge provided by the history agent and the tool information provided by the tool agent; S5. Finally, the knowledge base agent inputs the related text blocks to the large model for prompt word output to obtain the knowledge answer to the question.

[0006] As a preferred solution, in step S1, each time a question is asked, the historical agent will rewrite the question through the large model based on historical information and current information, and give questions with historical memory.

[0007] As a preferred solution, the tool agent also analyzes the specific parameters and information of tool usage, and responds to the user through a large model, allowing the user to choose whether to use or cancel the use.

[0008] As a preferred solution, in step S4, if the warehouse tool is used, the task is directly started, the answer obtained is delivered to the historical agent, and is output to the foreground through large model optimization.

[0009] As a preferred solution, if a warehouse tool is used, the tool parameters and tool mode are also taken as input to request remote services.

[0010] As a preferred solution, in step S5, the knowledge base agent uses the built-in knowledge base to select the document answer vector corresponding to the question according to the user question, and finds the associated text block with the highest matching degree from the vector database.

[0011] As a preferred solution, step S5 specifically includes the following steps: S5.1. Transfer the local file to the intelligent background. The intelligent background will put the extracted text into the text segmenter, segment it into text blocks, and then use the vector matrix to vectorize the text. S5.2, storing the vectorized text into a vector database; S5.3, use the vector matrix to vectorize the question raised by the user into a query vector, and then use similarity matching in the vector database to find the first few text blocks that best match; S5.4. Use proprietary prompt words, embed related text blocks, combine them into complete prompts and send them to the large model to give the answer.

[0012] As a preferred solution, the intelligent background uses an unstructured loader to load documents with various suffixes.

[0013] As a preferred solution, in step S5.2, a loading file record is also stored in the local database at the same time.

[0014] As a preferred solution, in step S3, if canceled, the user is directly notified of the cancellation; if modified, the tool is used to modify the prompt words, submitted to the large model, and output to the front desk; if confirmed for use, it is transferred to the execution agent for execution.

[0015] Therefore, the advantages of the present invention are: 1. The invention needs to be easy to understand, convenient to customize, and simple to maintain.

[0016] 2. The present invention can be configured with multiple prompt words, adapt to large models of different manufacturers, and provide different features according to factory business needs.

[0017] 3. The present invention uses a variety of intelligent agents to provide support and assistance, which has high accuracy and comprehensive problem solving.

[0018] 4. The construction method of the present invention is simple and easy to understand, and can be applied to a large number of similar irregular tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of the present invention.

[0020] Figure 2 It is the knowledge base intelligent agent architecture diagram of the present invention. DETAILED DESCRIPTION

[0021] The technical solution of the present invention is further specifically described below through embodiments and in conjunction with the accompanying drawings.

[0022] Embodiment 1: Alarm management requires a lot of analysis and interpretation every day, including knowledge interpretation, position number import, statistical plan establishment, statistical plan analysis, etc., which requires a lot of complicated and repetitive operations. Modern technology has brought AI capabilities, but most of the intelligence we use can only handle some problems individually and cannot handle complex tasks in a related manner, resulting in scattered use and inability to handle them coherently. Most processing solutions are simply hard-coded according to the problem and continuously process it to get the answer to the problem. Once the problem is switched or the conditions are changed, the correct answer cannot be given or the answer is wrong. Here, a general multi-task collaborative assistant device based on multi-agents is needed. In the processing of complex problems, the problems are subdivided separately, and the historical correlation processing is performed before and after to obtain a general assistant device with high stability and accurate answers, and thus displayed to the front desk to assist operators in completing tedious and repetitive work.

[0023] In order to deal with some problems in the prior art, such as the method of use is too simple and single, resulting in one-sided answers to questions, not comprehensive and in-depth enough, and in actual use, a customized method is often adopted, and no general and reusable solution is formed, which makes it easy to have errors and inaccuracies when switching problems and multi-tasking. In addition, the prior art often only uses a single tool or relies only on strategies based on retrieval enhancement generation (RAG), rather than comprehensively considering the problem through the collaborative work of multiple agents. In this case, when answering complex questions, the system often lacks the support of historical information, resulting in an incomplete understanding of the problem, simple answers, and incomplete functions. To this end, the present invention provides an alarm management method based on multi-agents to solve these technical deficiencies. The method can better utilize the advantages of historical data and multi-agent collaboration, improve the overall intelligence level of the system and its ability to deal with complex problems, and includes the following steps: S1. The answers to the questions replied by the user are output to the historical agent, tool agent, detection agent, execution agent and knowledge base agent in turn. Each agent has its specific functions and tasks when handling problems, ensuring that the problems can be analyzed from multiple angles and perspectives. They are first output to the historical agent. Each time a question is asked and answered, the historical agent will rewrite the question through a large model based on historical information and current information, and give questions with historical memory. This process ensures that the system can better understand the real needs of users and provide more accurate and detailed context information for subsequent processing stages of other agents. In this way, the system can not only understand the surface meaning of the current question, but also capture the deep meaning of the question in combination with the user's historical interaction information, avoiding one-sided answers or missing key information.

[0024] Through this multi-agent collaborative working mode, the present invention can effectively avoid the drawbacks of the existing technology that a single tool or method is difficult to fully solve complex problems. Each agent has its own unique functions and roles in the system, and cooperates with each other. Through the combination of historical information and current information, it ensures that the problem is handled more comprehensively and accurately, and improves the overall intelligence level of the system and its ability to handle complex problems. At the same time, this multi-agent design is also highly versatile and reusable, and can adapt to different scenarios and needs, avoiding the limitations of customized methods.

[0025] S2. The tool agent then analyzes the problem through the big model and chooses to use the specified tool in the tool warehouse. The tool agent also analyzes the specific parameters and information of the tool and responds to the user through the big model, allowing the user to choose whether to use or cancel the use.

[0026] Tool agents are intelligent analysis systems based on large models. When faced with complex problems, they can select appropriate tools to handle them through in-depth understanding and analysis of the problems. Specifically, when the tool agent receives a question input by the user, it will first use the powerful natural language processing capabilities of the large model to perform semantic analysis and interpretation of the question. This analysis process is not limited to superficial word matching, but also involves a comprehensive understanding of the problem context, user intent, and potential needs. Through such analysis, the tool agent can clarify the core content of the current problem and possible solutions.

[0027] Next, the tool agent will query its internal or external tool warehouse according to the attributes and requirements of the problem. In the tool warehouse, there are various tools that can be used to solve problems. These tools cover multiple fields from data processing, information retrieval to complex calculations, image processing, etc. The tool agent can not only identify which tools may be suitable for solving the current problem, but also match the corresponding parameters and usage information for each tool based on the judgment of the big model. For example, if the problem involves data analysis, the tool agent may recommend a data analysis tool and specify specific parameters, such as the path of the data set, the analysis method, etc.

[0028] More importantly, the tool agent not only automatically selects tools and runs them directly, but also provides feedback to users on all possible options and details during the tool selection and parameter setting process. This means that users are always at the center of decision-making in this process. The tool agent presents the tool recommendations, usage parameters, expected effects, and other information it analyzes to users in a clear manner, ensuring that users can fully understand the functions of each tool and its possible impact.

[0029] After obtaining this information, users can choose whether to use the tool recommended by the tool agent based on their needs and judgment. Users can choose to directly execute the recommended tool, or adjust it according to the provided parameters, or even choose to cancel the operation. In this way, the tool agent not only embodies the intelligent automatic analysis capability, but also retains the user's ultimate control over the selection and use of the tool, ensuring a high degree of flexibility and user experience.

[0030] S3. The detection agent analyzes the big model through the confirmation information of the user's corresponding tool use to determine whether to cancel, modify, or confirm the use. If it is cancelled, the user is directly notified to cancel. If it is modified, the tool modification prompt words are constructed, submitted to the big model, and output to the front desk. If it is confirmed to be used, it is transferred to the execution agent for execution.

[0031] The detection agent is an intelligent system that is specifically used to make judgments and decisions during the process of user confirmation of the use of the tool. Its core function is to determine whether the next operation is to cancel, modify, or continue to execute the tool by analyzing user feedback or confirmation information. The entire process relies on the powerful analytical capabilities of the large model to ensure the accuracy and efficiency of the decision.

[0032] When a user chooses whether to use a tool based on the recommendation of the tool agent, the detection agent will receive confirmation information from the user. These confirmation messages may include explicit instructions, such as canceling, modifying parameters, and confirming use, or they may be selective feedback from the user based on the parameters provided by the tool agent. The primary task of the detection agent is to parse this information. Through the natural language processing capabilities of the large model, the detection agent can understand the user's true intentions, not just the superficial text instructions. It will combine the context and the user's historical behavior to ensure a deep understanding of the user's needs.

[0033] After analyzing the user's intention, the detection agent will make a corresponding judgment based on the result. If the user chooses to cancel the operation, the detection agent will immediately feedback the cancellation decision to the user and terminate the current tool usage process. In this case, there will be no subsequent tool execution steps, and the entire process ends directly under the user's cancellation instruction.

[0034] If the detection agent analyzes that the user wants to modify the parameters or certain details of the tool, it will enter the modification process. In this process, the detection agent will build a set of modification prompts based on the user's feedback. These prompts may include suggestions for adjusting tool parameters, guidance for modifying usage conditions, or reconfiguring certain steps. The detection agent will submit these prompts to the large model for further processing to ensure the accuracy and feasibility of the modification suggestions. Subsequently, the modified information will be output to the front desk in an appropriate format for further confirmation and selection by the user.

[0035] If the user confirms the use of the tool, it means that the user has no objection to the tool selection and parameter settings provided by the tool agent. Then the detection agent will pass this confirmation information to the execution agent. At this point, the task of the detection agent comes to an end, and the execution agent takes over the entire process and is responsible for the actual tool execution work. The execution agent will call the relevant tools to complete the task according to the established parameters and process.

[0036] In general, the detection agent ensures that each step meets the user's actual needs through keen analysis of user feedback, and can respond quickly and accurately whether it is cancellation, modification or continuation. In this way, the detection agent plays an important coordination role in the entire intelligent tool system, improving the system's flexibility and user experience.

[0037] S4. The execution agent then uses the historical knowledge provided by the history agent and the tool information provided by the tool agent to analyze whether to use the warehouse tool or the knowledge base agent. If the warehouse tool is used, the task is started directly, and the answer is sent to the history agent and output to the front end through large model optimization. If the warehouse tool is used, the tool parameters and tool mode are also used as input to request remote service.

[0038] S5. Finally, the knowledge base agent inputs the associated text blocks to the large model for prompt word output to obtain the knowledge answer to the question. The knowledge base agent uses the built-in knowledge base to select the document answer vector corresponding to the question according to the user's question and finds the associated text block with the highest matching degree from the vector database.

[0039] The core function of the knowledge base agent is to provide accurate knowledge answers by processing questions raised by users, using its built-in knowledge base and powerful vector retrieval capabilities. When a user asks a question, the knowledge base agent will first parse the question to clarify the core content and information requirements of the question. Then, the knowledge base agent will call on its internal knowledge base resources, which usually include various documents, technical manuals, research reports, database records and other structured knowledge documents to ensure that users can be provided with answers based on reliable information.

[0040] In order to find the answer that best matches the user's question, the knowledge base agent relies on vectorization technology. It converts the user's question into a specific vector representation. This process usually relies on natural language processing technology and deep learning models to ensure that the question vector can accurately represent the semantics of the question. Next, the knowledge base agent will filter out the document answer vector that best matches the question vector from its built-in vector database. The vector database stores vectorized representations of a large number of documents, and each document or text paragraph is mapped to a vector to represent its semantics and content features. In this way, the knowledge base agent can quickly and efficiently find the associated text block that is closest to the question semantics.

[0041] Step S5 specifically includes the following steps: S5.1. Transfer local files to the intelligent background. The intelligent background will use the unstructured loader to load documents with various suffixes, and then put the extracted text into the text segmenter to segment and convert it into text blocks. Then, the vector matrix will be used to vectorize the text.

[0042] S5.2. The vectorized text is stored in the vector database, and a loading file record is stored in the local database at the same time.

[0043] S5.3. Use the vector matrix to vectorize the question raised by the user into a query vector, and then use similarity matching in the vector database to find the first few most matching text blocks.

[0044] S5.4. Use proprietary prompt words, embed related text blocks, combine them into complete prompts and send them to the large model to give the answer.

[0045] Compared with the prior art, the present invention has the following advantages: The existing technology uses a simple and single method with a high possibility of failure; the present invention uses multiple intelligent agents to coordinate processing, with diverse levels and efficient and accurate judgment.

[0046] The existing technology has a fixed solution and switches questions, which can easily lead to misjudgment; the present invention is flexible and customizable, realizes a universal and replicable mode, switches questions and answers accurately, and improves the success rate of responses.

[0047] The prior art uses only tools or only knowledge bases, which may result in one-sided answers. The present invention combines tools and knowledge base agents to conduct collaborative analysis and consider answers to questions together, thereby improving the accuracy of subsequent answers.

[0048] The prior art has no historical content to provide support, resulting in simple answers and inability to use past experience; the present invention utilizes historical information, extracts important question keywords, reorganizes new standard questions, efficiently uses historical experience data, and improves the overall problem-solving ability.

[0049] The existing technology has fixed parameters or no parameters, and cannot be customized and developed; the multi-parameter configuration method of the present invention is conducive to later modification and maintenance.

[0050] In the present invention, the user inputs the question into the history agent, and the question is rewritten through the large model to give a question with memory, so that the answer is more humane, avoiding short-term memory and one-sided response. Through the tool information in the tool warehouse, the tool agent analyzes the user's question and extracts the tool parameters and specific information. According to the user's choice, the tool is used directly or by asking, which improves the user's experience and enhances the versatility of the tool. Through the user's reply, the inspection agent uses the large model to detect the reply type and goes to the inaccessible path, which improves the diversity of the reply to the question.

[0051] Combined with historical knowledge and tool information provided by the tool selection agent, it re-analyzes whether to use warehouse tools or knowledge base agents, integrating the responses from task tools and knowledge base questions and answers, improving the richness of answers and the accuracy and effectiveness of answers. By using the existing knowledge vector library and combining it with user questions, it finds the most matching text answer block and obtains effective answers to questions, avoiding the non-featured expert library of the existing large model and expanding the answering capabilities of the proprietary alarm management knowledge base.

[0052] The prompt words of each agent are universal, and they have good answering capabilities for large models of different sizes, avoiding illusions. Different agents collaborate and utilize different model functions to integrate a set of powerful tool assistants, making the solution of complex problems more comprehensive and more accurate. In addition, the model parameters of the agent are customizable, which facilitates later maintenance and management. The processing results can also be stored in the database as a front-end display.

[0053] Embodiment 2: The intelligent agents in the present invention include the following: History Agent: After merging and assembling the previous question and answer information, it is combined with the rewrite prompt template (Rewrite Prompt) and submitted to the large model (LLM) to output a new question (new question) and historical knowledge abbreviation information (HistoryAbbreviation).

[0054] Rewrite the question prompts as follows: Given a chat history and the latest user question which might reference context in the chat history,formulate astandalone question which can be understood without the chat history.Do NOT answer the question, just reformulate it if needed and otherwise return it as is. Tool Agent: In the early stage, register API information to the tool warehouse (Tool Table) through Rest, integrate tool information into a common tool schema, combine with new questions (new question), build a tool analysis prompt template (Tool Prompt), submit it to the large model (LLM), and output the tool name (Tool Name) and tool parameters (Tool Input). Extract the corresponding warehouse information through the tool warehouse. If the tool is executed directly, use the execution agent directly to execute it. If confirmation execution is required, build confirmation prompts by integrating tool schema (Schema) and tool parameters (Tool Input), submit it to the large model (LLM), output the reply to the confirmation question and answer, and submit it to the front desk for display.

[0055] The tool analysis prompts are as follows: Assistant can ask the user to use tools to look up information that may be helpful in answering the users original question. The tools the human can use are: {tool_descs} RESPONSE FORMAT INSTRUCTIONS ----------------------------When responding to me,just output aresponse in one of two formats: **Option 1:** Use this if you want the human to use a tool. Markdown code snippet formatted in the following schema: ```json {{ "action":string,\The action to take.Must be one of{tool_names} "action_input": string\Tool parameters extracted from user input }} The tool confirmation prompt words are as follows: Fill in the tool information according to the tool name and parameter description. Do not falsify data and content, please fill in according to the actual situation. Here is a description of the tool: {tool_desc} The following is the collected tool information: {tool_info} RESPONSE FORMAT INSTRUCTIONS ----------------------------When responding to me,please output aresponse with format: ```json {{ "description": string,\The full description of the tool after filling }} ``` Check Agent: Through the tool warehouse (Tool Table), find the tool mode (Schema), combine the tool parameters (ToolInput) and user input (new question), build the detection prompt (Check Prompt), submit it to the large model (LLM), output confirmation (Confirm), cancel (Cancel), or modify (Edit). If canceled, directly notify the user to cancel. If modification is required, build the tool modification prompt (Edit Prompt), submit it to the large model (LLM), and output it to the front desk. If it is confirmed to be used, it will be transferred to the execution agent for execution.

[0056] The detection prompt words are as follows: I am a tool assistant, helping you use tools. Here is the previous tool description: {tool_desc} Here are the previous answers: {tool_info} I will analyze your reply this time and decide whether to call the tool based on the reply. The returned data can only be in JSON format. Please return it in the following format: ---------------------------- Please reply using one of the following 3 formats: **Option #1:** You clearly reply "confirm", do not forge data, and return the following json format: ```json {{ "check":"yes" }} ``` **Option#2:** You reply to cancel, and the following json format is returned: ```json {{ "check":"no" }} ``` **Option#3:** If you reply to modify, I will modify it according to the previous tool description and answer information. Do not forge parameters. If you have not modified it, do not modify the previous parameters. Return the json in the following format: ```json {{ "check":"edit", "action_input": string\Modified tool parameters }} ``` Modify the prompt words as follows: Modify the tool information according to the tool name and parameter description. Do not falsify data and content, please fill in according to the actual situation. Here is a description of the tool: {tool_desc} The following is the collected tool information: {tool_info} RESPONSE FORMAT INSTRUCTIONS ----------------------------When responding to me,please output aresponse with format: ```json {{ "description": string,\The full description of the tool after filling }} ``` Execute Agent: According to the tool schema, determine whether to use the knowledge base agent (RAG Agent) or the RestApi. If the knowledge base agent (RAG Agent) is used, it will be transferred to the knowledge base agent for execution. If the Rest Api is used, the tool parameters (Tool Input) and the tool schema (Tool Schema) will be used as input to request remote services. After obtaining the final answer, the reply prompt words are constructed and submitted to the large model (LLM), and the final result is given to the front desk.

[0057] Knowledge Base Agent (RAG Agent): The architecture diagram is as follows Figure 2 The data text processing architecture of the knowledge base is divided into the following stages: 1. Vectorization of local files (Local Embedding) First, the local document is transferred to the intelligent background, which will use the unstructured loader to load documents with various suffixes, such as .doc, .pdf, etc. The extracted text is then put into the text splitter, which is split and converted into text chunks according to the parameters of 250-word interval and 50-word overlap. The vector matrix is ​​then used to embed the text.

[0058] 2. Store in Vector Store The vectorized text (Embedding Text) is a string of numbers, which is directly input into the vector database (VectorStore). At the same time, a copy of the loaded file record is stored in the local sqlite database at the same time.

[0059] 3. Similarity Quetry When a user asks a question, the backend will use the embedding matrix to vectorize the question into a query vector. Then, the vector database will use the vector similarity matching to find the first few related text chunks. The similarity extraction parameter used here is 0.45.

[0060] 4. Give the answer based on LLM Use proprietary prompts and embed related text chunks to form a complete prompt and send it to the large model (LLM). Here, the large model (LLM) is requested in a stream manner to give the correct answer to the large model.

[0061] The processing results of the entire process are stored in the database for later historical data review.

[0062] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A multi-agent based alarm management method, characterized in that: The following steps are involved: S1. The answer to the question replied by the user is first output to the historical agent; S2, the tool agent then analyzes the problem through the big model and selects to use the specified tool in the tool warehouse; S3, the detection agent performs a large model analysis based on the confirmation information of the corresponding tool used by the user, and determines whether to cancel, modify, or confirm the use; S4, then the execution agent analyzes again whether to use the warehouse tool or the knowledge base agent based on the historical knowledge provided by the history agent and the tool information provided by the tool agent; S5. Finally, the knowledge base agent inputs the related text blocks to the large model for prompt word output to obtain the knowledge answer to the question.

2. According to the multi-agent based alarm management method of claim 1, it is characterized in that: In step S1, each time a question is asked, the historical agent will rewrite the question through the large model based on historical information and current information, and give questions with historical memory.

3. The multi-agent based alarm management method according to claim 1 is characterized in that: The tool agent also analyzes the specific parameters and information of tool usage, and responds to the user through a large model, allowing the user to choose whether to use or cancel the use of the tool.

4. The multi-agent based alarm management method according to claim 1 is characterized in that: In step S4, if the warehouse tool is used, the task is started directly, the answer obtained is delivered to the historical agent, and is output to the foreground through large model optimization.

5. The multi-agent based alarm management method according to claim 1 or 4, characterized in that: If a warehouse tool is used, the tool parameters and tool mode are also taken as input to request remote services.

6. The multi-agent based alarm management method according to claim 1, characterized in that: In step S5, the knowledge base agent uses the built-in knowledge base to select the document answer vector corresponding to the question according to the user's question, and finds the associated text block with the highest matching degree from the vector database.

7. The multi-agent based alarm management method according to claim 6 is characterized in that: The step S5 specifically comprises the following steps: S5.

1. Transfer the local file to the intelligent background. The intelligent background will put the extracted text into the text segmenter, segment it into text blocks, and then use the vector matrix to vectorize the text. S5.2, storing the vectorized text into a vector database; S5.3, use the vector matrix to vectorize the question raised by the user into a query vector, and then use similarity matching in the vector database to find the first few text blocks that best match; S5.

4. Use proprietary prompt words, embed related text blocks, combine them into complete prompts and send them to the large model to give the answer.

8. The multi-agent based alarm management method according to claim 7 is characterized in that: The intelligent backend will use an unstructured loader to load documents with various suffixes.

9. The multi-agent based alarm management method according to claim 7, characterized in that: In step S5.2, a load file record is also stored in the local database at the same time.

10. The multi-agent based alarm management method according to claim 1, characterized in that: In step S3, if canceled, the user is directly notified of the cancellation; if modified, the tool is used to modify the prompt words, submitted to the large model, and output to the front desk; if confirmed for use, it is transferred to the execution agent for execution.

Images