Cross-platform operation method and device and first agent

Abstract

The invention provides a cross-platform operation method and device and a first agent, and relates to the technical field of computers, in particular to the fields of artificial intelligence, cloud computing and the like. According to the specific implementation scheme, the method comprises the steps of obtaining problem description information used for describing cross-platform job problems in a natural language; a first task needing to be executed for answering the cross-platform job question is planned and obtained based on the question description information by using a pre-trained first intelligent agent; target second intelligent agents corresponding to the first tasks are called to execute the first tasks through the first intelligent agents, execution results of the first tasks are obtained, the target second intelligent agents are intelligent agents capable of executing the corresponding tasks in multiple preset second intelligent agents, and the target second intelligent agents are intelligent agents capable of executing the corresponding tasks in multiple preset second intelligent agents. Different second intelligent agents have the capability of executing different types of tasks; and summarizing based on the execution result of each first task by utilizing the first intelligent agent to obtain an answer of the cross-platform homework question, and displaying the answer of the cross-platform homework question.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, in particular to the fields of artificial intelligence, cloud computing, and more particularly to a cross-platform operation method, device, and first intelligent agent. Background Art

[0002] Many enterprises present operation and maintenance information on the operation and maintenance platform to facilitate operation and maintenance personnel. Operation and maintenance personnel only need to visit specific pages in the operation and maintenance platform to obtain the operation and maintenance information they need and then perform operation and maintenance work. Summary of the invention

[0003] The present disclosure provides a method, an apparatus, a device and a storage medium for improving cross-platform operation efficiency.

[0004] According to a first aspect of the present disclosure, a cross-platform operation method is provided, comprising:

[0005] Obtain problem description information for describing the cross-platform operation problem in natural language;

[0006] Using a pre-trained first intelligent agent, based on the problem description information, a first task required to be performed to answer the cross-platform operation problem is planned;

[0007] Using the first agent, calling the target second agents corresponding to each of the first tasks to respectively execute each of the first tasks, and obtaining the execution results of each of the first tasks, wherein the target second agent is an agent that can execute the corresponding task among a plurality of preset second agents, and different second agents have the ability to execute tasks of different categories;

[0008] Utilizing the first agent, the answer to the cross-platform operation problem is summarized based on the execution results of each of the first tasks and displayed.

[0009] According to a second aspect of the present disclosure, there is provided an agent training method, comprising:

[0010] Acquire first sample description information for describing a sample cross-platform job problem in natural language, third tasks required to be performed to answer the sample cross-platform job problem, execution results of each of the third tasks, the answer to the sample cross-platform job problem, a first sample prompt word, and a second sample prompt word;

[0011] Based on the first sample description information, the first sample prompt word and the third task, training a first large language model, wherein the first large language model is used to plan the tasks required to answer the sample cross-platform operation question under the guidance of the first sample prompt word;

[0012] Based on the execution results of each of the third tasks, the second sample prompt words, and the answer to the sample cross-platform homework question, training a second large language model, wherein the second large language model is used to summarize the execution results of each of the third tasks under the guidance of the second sample prompt words to obtain the answer to the sample cross-platform homework question;

[0013] The trained intelligent agent includes the first largest language model and the second largest language model.

[0014] According to a third aspect of the present disclosure, a cross-platform operation device is provided, comprising:

[0015] An information acquisition module, used for acquiring problem description information for describing the cross-platform operation problem in natural language;

[0016] A planning module, configured to utilize a pre-trained first agent to plan, based on the problem description information, a first task required to be performed to answer the cross-platform operation problem;

[0017] an execution module, configured to use the first agent to call the target second agent corresponding to each of the first tasks to respectively execute each of the first tasks, and obtain the execution results of each of the first tasks, wherein the target second agent is an agent that can execute the corresponding task among a plurality of preset second agents, and different second agents have the ability to execute tasks of different categories;

[0018] The answer module is used to utilize the first intelligent agent to obtain and display the answer to the cross-platform operation problem based on the execution result summary of each of the first tasks.

[0019] According to a fourth aspect of the present disclosure, there is provided an intelligent agent training device, comprising:

[0020] A sample acquisition module, used to acquire first sample description information for describing a sample cross-platform operation problem in natural language, a third task required to be performed to answer the sample cross-platform operation problem, an execution result of each of the third tasks, an answer to the sample cross-platform operation problem, a first sample prompt word, and a second sample prompt word;

[0021] A first training module is used to train a first large language model based on the first sample description information, the first sample prompt word and the third task, wherein the first large language model is used to plan the tasks required to answer the sample cross-platform operation question under the guidance of the first sample prompt word;

[0022] A second training module is used to train a second large language model based on the execution results of each of the third tasks, the second sample prompt words, and the answer to the sample cross-platform homework question, wherein the second large language model is used to summarize the execution results of each of the third tasks under the guidance of the second sample prompt words to obtain the answer to the sample cross-platform homework question;

[0023] The trained intelligent agent includes the first largest language model and the second largest language model.

[0024] According to a fifth aspect of the present disclosure, there is provided a first intelligent agent, comprising:

[0025] An input module, used for receiving input information;

[0026] a processing module, configured to determine a target task based on the input information received by the input module, determine a large language model based on the target task, and obtain output information by calling the large language model to execute any one of the methods described in the first aspect;

[0027] An output module is used to output the output information obtained by the processing module.

[0028] According to a sixth aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to execute the method according to any one of the above-mentioned first aspects.

[0029] According to a seventh aspect of the present disclosure, a computer program product is provided, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements the method according to any one of the above-mentioned first aspects.

[0030] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure.

[0032] Figure 1 It is a flow chart of the cross-platform operation method provided by the present disclosure;

[0033] Figure 2a It is a diagram of a user entering descriptive information through a conversational front end;

[0034] Figure 2b It is a schematic diagram of a user inputting description information through a service number;

[0035] Figure 2c It is a schematic diagram of a user inputting description information through an intelligent assistant;

[0036] Figure 3 is another flow chart of the cross-platform operation method provided by the present disclosure;

[0037] Figure 4 It is another flow chart of the cross-platform operation method provided by the present disclosure;

[0038] Figure 5 This is another flow chart of the cross-platform operation method provided by the present disclosure;

[0039] Figure 6a It is a schematic diagram of a user inputting evaluation information through a conversational front end;

[0040] Figure 6b It is a schematic diagram of a user inputting evaluation information through a service account;

[0041] Figure 6c It is a schematic diagram of a user inputting evaluation information through an intelligent assistant;

[0042] Figure 7 is a flow chart of a first agent training method provided by the present disclosure;

[0043] Figure 8 It is a schematic diagram of an architecture for implementing a cross-platform operation method provided by the present disclosure;

[0044] Fig. 9 It is a flow chart of the fault location method provided by the present disclosure;

[0045] Fig.10 It is a flow chart of a training method for a fault location intelligent agent provided by the present disclosure;

[0046] Fig.11 is a schematic diagram of an architecture for implementing a fault location method provided by the present disclosure;

[0047] Fig.12 It is a flow chart of the edge resource management method provided by the present disclosure;

[0048] Fig.13 It is a flow chart of a training method for an edge resource management agent provided by the present disclosure;

[0049] Fig.14 is a schematic diagram of an architecture for implementing an edge resource management method provided by the present disclosure;

[0050] Fig.15 It is a flow chart of the alarm analysis method provided by the present disclosure;

[0051] Fig.16 It is a flow chart of a training method of an alarm analysis agent provided by the present disclosure;

[0052] Fig.17 It is a schematic diagram of an architecture for implementing an alarm analysis method provided by the present disclosure;

[0053] Fig.18 It is a flow chart of the data query method provided by the present disclosure;

[0054] Fig.19 It is a flow chart of a training method for a data query agent provided by the present disclosure;

[0055] Fig. 20 is a schematic diagram of the interaction between the first agent and the second agent provided by the present disclosure;

[0056] Fig.21 It is a flow chart of the intelligent agent training method provided by the present disclosure;

[0057] Fig. 22 It is a structural schematic diagram of a cross-platform operation device provided by the present disclosure;

[0058] Fig.23 is a structural schematic diagram of an intelligent agent training device provided by the present disclosure;

[0059] Fig.24 It is a block diagram of an electronic device used to implement the cross-platform operation method of the embodiment of the present disclosure. DETAILED DESCRIPTION

[0060] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0061] First, the relevant terms or concepts in the embodiments of the present disclosure are explained:

[0062] AI Agent: It is an intelligent agent system with a large language model as the core controller. Its essence is an agent system that controls the large language model to solve user problems. It relies on the large language model as its core decision-making and processing unit, and can use the logical reasoning ability of the large language model to solve user problems. In practical applications, the agent can be a software or hardware entity that can act autonomously.

[0063] LLM (Large Language Model) refers to a deep learning model trained on massive text data, which has strong expressive and generalization capabilities. It can not only generate natural language text, but also deeply understand the meaning of the text and handle various natural language tasks such as text summarization, question answering, translation, etc.

[0064] In order to more clearly illustrate the cross-platform operation method, device and first intelligent body provided by the present application, the following will take the scenario in which operation and maintenance personnel perform fault location as an example to exemplify the application scenario of the cross-platform operation method, device and first intelligent body provided by the present application. It can be understood that fault location is only one possible application scenario of the cross-platform operation method, device and first intelligent body provided by the present application. In other possible embodiments, the cross-platform operation method, device and first intelligent body provided by the present application can also be applied to other application scenarios besides fault location, and the following examples do not impose any limitation on this.

[0065] The operation and maintenance personnel find that a server cannot work properly. In order to determine the reason why the server cannot work properly, the operation and maintenance personnel need to obtain relevant information about the server operation, such as the number of visits, bandwidth usage, memory usage, etc.

[0066] For enterprises that have deployed intelligent operation and maintenance platforms, this information can often be obtained through querying the operation and maintenance platforms. However, for enterprises with more business lines, the operation and maintenance platforms need to display a wide variety of operation and maintenance-related information. In order to display information more clearly to operation and maintenance personnel, enterprises often choose to set up multiple operation and maintenance platforms, each of which is used to display part of the information, and set up multiple pages in each operation and maintenance platform, each of which is used to display one type of information.

[0067] Therefore, operation and maintenance personnel can query the information required for fault location through the operation and maintenance platform only if they know the page used to display the information they need to query. As the company's business lines increase, the number of operation and maintenance platforms and the number of pages in the operation and maintenance platforms also gradually increase, so that operation and maintenance personnel often need to remember the information displayed by a large number of pages and the operation and maintenance platforms to which these pages belong, so that they can accurately find the page used to display the information they need to query from a large number of pages, thereby obtaining the information they need. This brings great difficulties to the fault location work of operation and maintenance personnel, resulting in low fault location efficiency, and the same is true for other operation and maintenance operations.

[0068] Based on this, the present application provides a cross-platform operation method, device and first intelligent entity, which are used to improve the efficiency of operation and maintenance personnel in performing operation and maintenance operations in the scenario of the aforementioned multiple operation and maintenance platforms.

[0069] See also Figure 1 , Figure 1The figure shows a flow chart of a cross-platform operation method provided by the present application, including:

[0070] Step S101, obtaining problem description information for describing a cross-platform operation problem in a natural language;

[0071] Step S102, using a pre-trained first agent, planning based on the problem description information to obtain a first task required to answer the cross-platform operation problem;

[0072] Step S103, using the first agent, calling the target second agents corresponding to each first task to respectively execute each first task, and obtaining the execution results of each first task.

[0073] The target second agent is an agent that can perform the corresponding task among a plurality of preset second agents, and different second agents have the ability to perform tasks of different categories;

[0074] Step S104, using the first agent, based on the execution results of each first task, summarize and display the answer to the cross-platform operation problem.

[0075] This embodiment is selected to directly use the pre-trained first agent to plan the problem description information for the user feedback, and then call each second agent to execute each first task obtained by the plan, and summarize the execution results of each first task to obtain the answer to the cross-platform operation problem. During the whole process, the operation and maintenance personnel only need to interact with the first agent, and do not need to access the corresponding operation and maintenance platform and the corresponding page according to the needs. Therefore, even in a multi-platform scenario, the user does not need to remember the information displayed by a large number of pages and the operation and maintenance platforms to which these pages belong, and can perform the operation and maintenance operation, which effectively improves the efficiency of cross-platform operations.

[0076] On the other hand, in this embodiment, each second agent only needs to have the ability to execute some types of tasks, and then the coordination between the second agents is achieved through the scheduling of the first agent to solve the cross-platform operation problem. Therefore, each second agent does not need to learn the execution methods of too many tasks, which reduces the difficulty of building the second agent and thus reduces the implementation cost of the cross-platform operation method.

[0077] The above steps S101 to S104 will be described below respectively:

[0078] Among them, in step S101, the problem description information can be input by the operation and maintenance personnel in text form or in voice form. For example, the operation and maintenance personnel can input the text "the cause of the failure of device A" or input the voice "the cause of the failure of device A".

[0079] When the operation and maintenance personnel input text, the text can be directly used as the problem description information. When the operation and maintenance personnel input voice, the voice can be recognized and the recognized text can be used as the problem description information.

[0080] The execution subjects of the cross-platform operation method provided in this application include but are not limited to mobile phones, desktop computers, tablet computers, smart watches, smart speakers, etc. For mobile phones, desktop computers, platform computers and other devices that support text input, operation and maintenance personnel can input text or voice, while for smart speakers and other devices that do not support text input, operation and maintenance personnel can only input voice.

[0081] The operation and maintenance personnel may input text or voice by opening a specific page and inputting text or voice on the page. For example, Figure 2a It is also possible to open a specific service number through a specific application and input text or voice in the dialog window with the service number, for example, Figure 2b It is also possible to input text or voice by interacting with a specific group chat robot in the group chat window of the social software, for example, Figure 2c shown.

[0082] It is understandable that although Figure 2a , Figure 2b as well as Figure 2c The inputs of the operation and maintenance personnel are shown to be text, but this is only for the convenience of showing the problem description information input by the operation and maintenance personnel. The principle is the same for the case of voice input, which will not be repeated here.

[0083] In step S102, the first agent may perform planning based on pre-learned knowledge about how to perform planning. During planning, the first agent may be implemented based on any algorithm model with planning capabilities.

[0084] Exemplarily, in a possible embodiment, step S102 includes:

[0085] Step S1021, using the pre-trained first intelligent agent, calling the first language model, and inputting the question description information and the first prompt word into the first language model, obtaining the first task output by the first language model under the reference of the first prompt word and required to be performed to answer the cross-platform job question.

[0086] The first prompt word is used to guide the first language model to perform task planning. The first prompt word can be pre-set or obtained when training the first language model. For example, the first prompt word can be:

[0087] You are {agent_name}, {agent_bio}

[0088] {agent_instructions}

[0089] The current stage is the task planning stage. You will be given a goal or problem, and your decision will be made independently without relying on human help. Please take advantage of LLM and pursue efficient strategies for task planning.

[0090] 1. You have a short-term memory of ~4,000 words

[0091] 2. No user assistance required

[0092] 3. When planning, you can use the tools mentioned in the reference tool

[0093] 4. Internet search, information aggregation and authenticity identification capabilities

[0094] 5. Stay humble. For problems that you are not sure about, call the command as much as possible, but try to call it as little as possible and do not call it repeatedly.

[0095] 6. When you can draw a conclusion from your own knowledge or historical memory, please be smart and efficient to complete the task and draw the conclusion.

[0096] 7. Frequently and constructively self-criticize your overall behavior and reflect on past decisions and strategies to improve your approach

[0097] 8. You can only think in {max_iter_num} steps and plan {max_iter_num} tasks at most, so plan your tasks as efficiently as possible

[0098] 9. You have the ability to reflect. If the completed tasks and results cannot obtain the information needed to answer the question or cannot achieve the goal, you should continue to plan, but not repeat the previous tasks.

[0099]

[0100] Ensure that Task can be parsed by Python's json.loads

[0101] When the completed tasks can help answer this goal, generate a task to complete the task if possible, otherwise generate another task. A new task:

[0102] This embodiment is selected to utilize a pre-trained first intelligent agent, and to perform planning based on the problem description information through the trained first language model combined with the first prompt word. Due to the guidance of the first prompt word, the ability of the first language model to plan tasks can be fully utilized, thereby improving the accuracy of the planned first task, and further making the subsequent answers based on the first task more accurate.

[0103] In step S103, the number of preset second agents may be different depending on the application scenario. For example, in some application scenarios, the preset second agents are fault location agents and edge resource management agents, and in other application scenarios, the preset second agents are edge resource management agents, alarm analysis agents, and data query agents. In some other application scenarios, the preset second agents are fault location agents and data query agents. In some other application scenarios, the preset second agents are fault location agents, edge resource management agents, alarm analysis agents, and data query agents.

[0104] The fault location agent, edge resource management agent, alarm analysis agent and data query agent will be explained below and will not be repeated here.

[0105] In a possible embodiment, Figure 3 As shown, the target second agent corresponding to each first task may be determined in step S105, including:

[0106] Step S105 , using the first agent, based on the description information of each of the plurality of preset second agents, respectively determine the target second agent corresponding to each first task.

[0107] The description information is used to describe the capabilities of the second agent. By selecting this embodiment, the first agent can grasp the capabilities of each second agent based on the description information of each second agent, thereby more accurately calling the second agent with corresponding capabilities to perform the first task, improving the accuracy of the execution result, and further improving the accuracy of the answer obtained based on the subsequent summary of the execution result.

[0108] In step S104, the first agent may summarize based on the pre-learned knowledge on how to summarize the execution results of each task. When summarizing, the first agent may be implemented based on any algorithm model with summarizing capabilities.

[0109] Exemplarily, in a possible embodiment, step S104 includes:

[0110] Step S1041, using the pre-trained first intelligent agent, calling the second largest language model, and inputting the execution results of each first task and the second prompt word into the second largest language model, obtaining and displaying the answer to the cross-platform homework question that the second language model likes to output under the guidance of the second prompt word.

[0111] The second prompt word is used to guide the second language model to summarize. The second prompt word can be pre-set or obtained when training the first language model. The summary in this article refers to the process of refining the results of the task to obtain the answer to the question.

[0112] This embodiment is selected to utilize the pre-trained first intelligent agent to summarize through the trained second largest language model combined with the second prompt word. Due to the guidance of the second prompt word, the summarizing ability of the second largest language model can be fully utilized, thereby improving the accuracy of the summarized answer.

[0113] It is understandable that the planning ability of the first agent is limited, so the execution result of the first task obtained by planning may not be sufficient to answer the cross-platform operation problem. Based on this, in a possible embodiment, it can be as follows Figure 4 As shown, including:

[0114] Step S101, obtaining problem description information for describing a cross-platform operation problem in a natural language;

[0115] Step S102, using a pre-trained first agent, planning based on the problem description information to obtain a first task required to answer the cross-platform operation problem;

[0116] Step S103, using the first agent, calling the target second agents corresponding to each first task to respectively execute each first task, and obtaining the execution results of each first task.

[0117] The target second agent is an agent that can perform the corresponding task among a plurality of preset second agents, and different second agents have the ability to perform tasks of different categories;

[0118] Step S1042, using the first agent, determine whether the execution result of each first task contains all the information to answer the cross-platform operation question.

[0119] Step S1043: If yes, then use the first agent to summarize and display the answer to the cross-platform operation problem based on the execution results of each first task.

[0120] Step S106: If not, the pre-trained first agent is used to plan the second task that needs to be performed to answer the cross-platform operation problem based on the problem description information and the execution results of each first task.

[0121] Step S107, using the first agent, calling the target second agents corresponding to each second task to respectively execute each second task, and obtaining the execution results of each second task.

[0122] Step S108, using the first agent, based on the execution results of each first task and the execution results of the second task, an answer to the cross-platform operation problem is obtained and displayed.

[0123] Figure 4 The example shown is similar to Figure 1 The example shown is different only in that step S104 is refined into step S1042 and step S1043, and step S106, step S107 and step S108 are added, so only these different steps are described below.

[0124] Among them, in step S1042, the first agent can determine through reflection whether the execution results of each first task contain all the information to answer the cross-platform operation problem, that is, reflect on whether the execution results of each first task are sufficient to answer the cross-platform operation problem.

[0125] In step S106, the second task is planned in the same manner as the first task, except that when planning the second task, no task that overlaps with the first task will be planned, that is, any second task should be different from any first task.

[0126] In step S107, the method for determining the target second agent corresponding to each second task is the same as the method for determining the target agent corresponding to each first task. Please refer to the relevant description of the aforementioned step S105, which will not be repeated here.

[0127] In step S108, the summary method is the same as that of the aforementioned step S104, except that the summary of step S104 is based on the execution result of the first task, while the summary of step S108 is based on the execution result of the first task and the second task. Please refer to the relevant description of the aforementioned step S104 and step S1041, which will not be repeated here.

[0128] By selecting this embodiment, the first intelligent agent can automatically analyze whether the execution results obtained are sufficient to answer the questions of the operation and maintenance personnel. When it is still insufficient to answer the questions of the operation and maintenance personnel, the first intelligent agent can actively supplement the second task to further obtain execution results that are sufficient to answer the questions of the operation and maintenance personnel, thereby reducing the possibility of not being able to answer the questions of the operation and maintenance personnel, and improving the applicability of the solution.

[0129] In a possible embodiment, the cross-platform operation method provided by the present application can also be as follows: Figure 5 As shown, including:

[0130] Step S101, obtaining problem description information for describing a cross-platform operation problem in a natural language;

[0131] Step S102, using a pre-trained first agent, planning based on the problem description information to obtain a first task required to answer the cross-platform operation problem;

[0132] Step S103, using the first agent, calling the target second agents corresponding to each first task to respectively execute each first task, and obtaining the execution results of each first task.

[0133] The target second agent is an agent that can perform the corresponding task among a plurality of preset second agents, and different second agents have the ability to perform tasks of different categories;

[0134] Step S104, using the first agent, based on the execution results of each first task, summarize and display the answer to the cross-platform operation problem.

[0135] Step S109, obtaining evaluation information input for the displayed answer.

[0136] Evaluation information is used to determine the degree of match between Best Answers and cross-platform homework questions.

[0137] Step S110, correspondingly recording the question description information, answer and evaluation information.

[0138] Step S111, in response to the training instruction for the first agent, the first agent is optimized according to the stored question description information, answer and evaluation information.

[0139] Figure 5 The example shown is similar to Figure 1 The illustrated example differs only in that step S109, step S110 and step S111 are added, so only these different steps are described below.

[0140] Among them, in step S109, the way of obtaining evaluation information may be different according to different application scenarios. For example, the evaluation control may be displayed after the answer is displayed, so that the operation and maintenance personnel can interact with the evaluation control to input the evaluation information.

[0141] For example, corresponding to Figure 2a In the example shown, the evaluation control can be Figure 6a As shown, the evaluation control includes a good review sub-control and a bad review sub-control. In response to the operation and maintenance personnel clicking the good review sub-control, evaluation information indicating a high degree of match between the answer and the question is obtained, and in response to the operation and maintenance personnel clicking the bad review sub-control, evaluation information indicating a low degree of match between the answer and the question is obtained.

[0142] Similarly, corresponding to Figure 2b In the example shown, the evaluation control can be Figure 6b As shown, corresponding to Figure 2c In the example shown, the evaluation control can be Figure 6c shown.

[0143] By selecting this embodiment, the first intelligent agent can be continuously optimized through the evaluation information fed back by the operation and maintenance personnel, so that the first intelligent agent can more accurately answer the cross-platform operation questions raised by the operation and maintenance personnel.

[0144] The following will explain how to train the first agent. Figure 7 , Figure 7 The figure is a flow chart of the training method of the first agent provided by the present application, comprising:

[0145] Step S701, obtaining first sample description information for describing a sample cross-platform job problem in natural language, third tasks required to answer the sample cross-platform job problem, execution results of each third task, the answer to the sample cross-platform job problem, a first sample prompt word, and a second sample prompt word.

[0146] Step S702: training a first language model based on the first sample description information, the first sample prompt word and the third task.

[0147] The first language model is used to plan the tasks required to answer the sample cross-platform operation questions under the guidance of the first sample prompt word;

[0148] Step S703: training a second language model based on the execution results of each third task, the second sample prompt words, and the answers to the sample cross-platform homework questions.

[0149] The second largest language model is used to summarize the execution results of each of the third tasks under the guidance of the second sample prompt word to obtain the answer to the sample cross-platform operation problem.

[0150] The trained first intelligent agent includes a first large language model and a second large language model.

[0151] Among them, in step S702, the first sample description information and the first sample prompt word can be used as the input of the first large language model to obtain the output of the first large language model, and the third task can be used as the true value to calculate the loss between the output of the first large language model and the true value, and then the first large language model is adjusted according to the loss.

[0152] In step S703, the execution results of each third task and the second sample prompt word may be used as inputs of the second largest language model to obtain the output of the second largest language model, and the answer to the sample cross-platform homework question may be used as the true value to calculate the loss between the output of the second largest language model and the true value, and then the second largest language model may be adjusted according to the loss.

[0153] By selecting this embodiment, the cross-platform operation process of the operation and maintenance personnel can be integrated into the function of the first intelligent agent by training the first largest language model and the second largest language model, so that cross-platform operations can be performed through the first intelligent agent, thereby improving the accuracy of cross-platform operations.

[0154] In the above example, the process of implementing the cross-platform operation method based on the first agent is described. In order to more clearly describe the cross-platform operation method provided by the present application, the software architecture for implementing the cross-platform operation method will be described below. For example, see Figure 8 , Figure 8 The software architecture for implementing the cross-platform operation method provided by this application is shown.

[0155] Among them, the service number, super assistant and conversational front end of the user interface layer correspond to the aforementioned Figure 2b , Figure 2a as well as Figure 2c The interface in the example shown.

[0156] Fault location, resource management, alarm analysis and data question and answer in the intelligent assistant layer are four different second intelligent agents, namely the aforementioned fault location agent, edge resource management agent, alarm analysis agent and data query agent.

[0157] The perception, planning, execution and reflection of the engine layer are the four stages of the first intelligent agent, among which the perception stage and the planning stage correspond to the aforementioned step S102, the execution stage corresponds to the aforementioned step S103 and step S104, and the reflection stage corresponds to the aforementioned step S109, step S110 and step S111.

[0158] exist Figure 8 In the example shown, the perception phase includes problem reasoning, context analysis, and intent recognition. The planning phase includes step planning, tool parameter extraction, and intelligent questioning. The execution phase includes process execution, intelligent summary, and streaming answers. The reflection phase includes reinforcement learning, effect evaluation, and prompt word optimization.

[0159] The following will explain the four types of second agents mentioned above respectively:

[0160] As for the fault location agent, it is understandable that in some application scenarios, users want to locate the cause of the fault, and related technologies often require manual location. Manual location requires operation and maintenance personnel to have certain relevant knowledge, which is difficult for new operation and maintenance personnel, and may even lead to inaccurate fault location due to their limited relevant knowledge reserves.

[0161] Based on this, the present application provides a fault location agent, so that the first agent can be used to call the fault location agent to perform the fault location task in the first task and obtain the execution result of the fault location task.

[0162] The fault location agent can perform the fault location task in the following ways: Fig. 9 As shown, including:

[0163] Step S901: Acquire fault description information for describing the fault location task in natural language.

[0164] Step S902: utilizing the fault location agent to plan and obtain the first diagnostic steps required to perform the fault location task and the diagnostic sequence of each first diagnostic step based on the fault description information.

[0165] Step S903: utilizing the fault location agent, calling the target diagnostic tools corresponding to each first diagnostic step, respectively executing each first diagnostic step in a diagnostic sequence, and obtaining the diagnostic results of each first diagnostic step.

[0166] The target diagnostic tool is a tool among a plurality of preset diagnostic tools that can execute corresponding diagnostic steps.

[0167] Step S904: utilizing the fault localization agent to obtain the execution result of the fault localization task based on the diagnosis result summary of each first diagnosis step and feed it back to the first agent.

[0168] This embodiment is selected to directly use the pre-trained fault location agent to plan the fault description information, and then call each diagnostic tool to execute each first diagnostic step obtained by the plan, and summarize the diagnostic results of each first diagnostic step to obtain the execution result of the fault location task. In the whole process, the operation and maintenance personnel can realize fault location without manually locating the fault based on their own experience, which effectively improves the accuracy of fault location.

[0169] The above steps S901 to S904 will be described below respectively:

[0170] In step S901, the fault description information may be generated when the first agent plans and obtains the fault location task.

[0171] In step S902, the fault localization agent may perform planning based on the pre-learned knowledge about how to perform planning. During planning, the fault localization agent may be implemented based on any algorithm model with planning capability.

[0172] Exemplarily, in a possible embodiment, step S902 includes:

[0173] Step S9021, using the fault location intelligent agent, calling the third largest language model, and inputting the fault description information and the third prompt word into the third largest language model, obtaining the first diagnostic steps required to perform the fault location task and the diagnostic order of each first diagnostic step output by the third largest language model under the guidance of the third prompt word.

[0174] The third prompt word is used to guide the third language model to perform task planning. The third prompt word can be pre-set or obtained when training the third language model. For example, the third prompt word can be:

[0175] Your task is to come up with a short plan that helps me locate the user's problem in a few steps, using at most one of the provided tools for each step. You can create new steps with the help of the following tools.

[0176] Executing a tool can be understood as: using the tool to test the target, where the target is a sub-target of the original target. You will get a summary of the tool's output results.

[0177] You can use {tool_specification}

[0178] Note: log_retrieval needs to rely on the coverage node information to retrieve the log, so you need to execute the cover_check tool first

[0179] {current_date_and_time}

[0180] Construct data based on the planning results. The output can only be in the following JSON format. There should be no comments in the JSON result:

[0181]

[0182] Make sure the result can be parsed by Python's json.loads

[0183] Here are some examples.

[0184] Objective: Check the opencdnmt fault of XXX application, the request success rate of Henan Unicom has decreased, and the quality has deteriorated.

[0185] -Thinking: First of all, the request success rate in the target problem is an indicator provided by the customer and cannot be directly located, so there is no need to consider it. I need to first confirm whether there are any abnormalities in the basic information such as the overall bandwidth and number of requests of the opencdnmt service; at the same time, obtain the basic information of the fault reporting service, such as the changes in the scheduling coverage information of opencdnmt in Henan Unicom.

[0186] Step 1: Use the metric_check tool to check the bandwidth, number of requests, and status code indicator information of the domain name opencdnmt in Henan Unicom

[0187] Step 2: Use the cover_check tool to detect the current coverage changes of the domain name opencdnmt in Henan Unicom and the current bearer information

[0188] -Thinking: Now that I know the coverage and service node information of the opencdnmt domain name at the time of the failure (current), I need to query the basic indicators of the node, such as detection success rate, detection speed, node CPU pressure, etc., to see if there are any abnormalities; at the same time, obtain information on whether the node has abnormal changes at the same time

[0189] Step 3: Use the metric_check tool to detect the current detection success rate, detection speed, and node CPU pressure indicator information of the opencdnmt coverage node

[0190] Step 4: Use the change_check tool to detect recent abnormal changes or events in the opencdnmt coverage node.

[0191] -Thinking: In order to achieve the goal, I can execute [Step 1] and [Step 2] at the same time. After these two steps are completed, I can execute [Step 3]. So I can construct the operation logic like this:

[0192] Execution order: (Step1 AND Step2) -> (Step3 AND Step4)

[0193] Objective: To investigate why vd3.bdstatic.com Xinjiang Mobile's traffic increased by 502 yesterday.

[0194] -Thinking: I need to first confirm whether the abnormal status code 502 indicator of the vd3.bdstatic.com domain name is abnormal; and confirm the coverage changes of the domain name in Xinjiang Mobile;

[0195] Step 1: Use the metric_check tool to check the 502 status code indicator information of the domain name vd3.bdstatic.com yesterday

[0196] Step 2: Use the cover_check tool to check the coverage change information of the domain name vd3.bdstatic.com in Xinjiang Mobile yesterday

[0197] -Thinking: Now I know that the 502 information generated by the domain name vd3.bdstatic.com yesterday has been covered by the scheduling node at that time; I need to query the basic indicators of the node, such as detection success rate, detection speed, node CPU pressure, etc., to see if there are any abnormalities; at the same time, obtain information on whether the node has abnormal changes at the same time; at the same time, I need to retrieve the domain name log on the node to summarize the log rules

[0198] Step 3: Use the metric_check tool to detect the detection success rate, detection speed, and node CPU pressure indicator information of the above nodes yesterday

[0199] Step 4: Use the log_retrieval tool to retrieve the log information of the domain name vd3.bdstatic.com on the above node yesterday

[0200] Step 5: Use the change_check tool to detect recent abnormal changes or events of vd3.bdstatic.com on the above nodes

[0201] -Thinking: In order to achieve the goal, I can execute [Step 1] and [Step 2] at the same time, and then execute [Step 3] and [Step 4] at the same time, so I can construct the operation logic like this:

[0202] Execution order: (Step1 AND Step2) -> (Step3 AND Step4 AND Step5)

[0203] This is the problem reported by the user: {goal}

[0204] This embodiment is selected to utilize a pre-trained fault location agent, and planning is performed based on the fault description information through the trained third language model combined with the third prompt word. Due to the guidance of the third prompt word, the ability of the third language model to plan tasks can be fully utilized, thereby improving the accuracy of the diagnostic steps obtained by the plan, and thus making the subsequent execution results based on the diagnostic steps more accurate.

[0205] The third prompt word is only a possible example, and in other possible embodiments, the third prompt word can also be set according to the needs of the user. In order to enable the third prompt word to better guide the third language model, the third prompt word can be used to describe the process of planning the diagnostic steps required to perform the example fault location task and the diagnostic order of each diagnostic step based on the example fault description information, wherein the example fault location task is the fault location task described by the example fault description information.

[0206] For example, the third prompt in the aforementioned example, "Goal: troubleshoot the opencdnmt fault reported by XXX application, the request success rate of Henan Unicom has decreased, and the quality has deteriorated" is the example fault description information, and the following content is the part used to describe the process of planning the diagnostic steps required to perform the example fault location task and the diagnostic order of each diagnostic step based on the example fault description information:

[0207] -Thinking: First of all, the request success rate in the target problem is an indicator provided by the customer and cannot be directly located, so there is no need to consider it. I need to first confirm whether there are any abnormalities in the basic information such as the overall bandwidth and number of requests of the opencdnmt service; at the same time, obtain the basic information of the fault reporting service, such as the changes in the scheduling coverage information of opencdnmt in Henan Unicom.

[0208] Step 1: Use the metric_check tool to check the bandwidth, number of requests, and status code indicator information of the domain name opencdnmt in Henan Unicom

[0209] Step 2: Use the cover_check tool to detect the current coverage changes of the domain name opencdnmt in Henan Unicom and the current bearer information

[0210] -Thinking: Now that I know the coverage and service node information of the opencdnmt domain name at the time of the failure (current), I need to query the basic indicators of the node, such as detection success rate, detection speed, node CPU pressure, etc., to see if there are any abnormalities; at the same time, obtain information on whether the node has abnormal changes at the same time

[0211] Step 3: Use the metric_check tool to detect the current detection success rate, detection speed, and node CPU pressure indicator information of the opencdnmt coverage node

[0212] Step 4: Use the change_check tool to detect recent abnormal changes or events in the opencdnmt coverage node.

[0213] -Thinking: In order to achieve the goal, I can execute [Step 1] and [Step 2] at the same time. After these two steps are completed, I can execute [Step 3]. So I can construct the operation logic like this:

[0214] Execution order: (Step1 AND Step2) -> (Step3 AND Step4)

[0215] Objective: To investigate why vd3.bdstatic.com Xinjiang Mobile's traffic increased by 502 yesterday.

[0216] -Thinking: I need to first confirm whether the abnormal status code 502 indicator of the vd3.bdstatic.com domain name is abnormal; and confirm the coverage changes of the domain name in Xinjiang Mobile;

[0217] Step 1: Use the metric_check tool to check the 502 status code indicator information of the domain name vd3.bdstatic.com yesterday

[0218] Step 2: Use the cover_check tool to check the coverage change information of the domain name vd3.bdstatic.com in Xinjiang Mobile yesterday

[0219] -Thinking: Now I know that the 502 information generated by the domain name vd3.bdstatic.com yesterday has been covered by the scheduling node at that time; I need to query the basic indicators of the node, such as detection success rate, detection speed, node CPU pressure, etc., to see if there are any abnormalities; at the same time, obtain information on whether the node has abnormal changes at the same time; at the same time, I need to retrieve the domain name log on the node to summarize the log rules

[0220] Step 3: Use the metric_check tool to detect the detection success rate, detection speed, and node CPU pressure indicator information of the above nodes yesterday

[0221] Step 4: Use the log_retrieval tool to retrieve the log information of the domain name vd3.bdstatic.com on the above node yesterday

[0222] Step 5: Use the change_check tool to detect recent abnormal changes or events of vd3.bdstatic.com on the above nodes

[0223] -Thinking: In order to achieve the goal, I can execute [Step 1] and [Step 2] at the same time, and then execute [Step 3] and [Step 4] at the same time, so I can construct the operation logic like this:

[0224] Execution order: (Step1 AND Step2) -> (Step3 AND Step4 AND Step5)

[0225] In step S903, the number and type of the preset diagnostic tools may vary according to different application scenarios. In addition, in a possible embodiment, the target diagnostic tools corresponding to each first diagnostic step may be obtained in the following manner:

[0226] The fault description information, the third prompt word, and the description information of each of the plurality of preset diagnostic tools are input into the third language model, and the first diagnostic steps required to perform the fault location task and the diagnostic sequence of each first diagnostic step output by the third language model under the guidance of the third prompt word are obtained, as well as the target diagnostic tool corresponding to each first diagnostic step output by the third language model under the guidance of the third prompt word.

[0227] By selecting this embodiment, the fault location intelligent agent can grasp the capabilities of each diagnostic tool based on the description information of each diagnostic tool, thereby more accurately calling the diagnostic tool with corresponding capabilities to perform the first diagnostic step, thereby improving the accuracy of the diagnostic results, and further improving the accuracy of the subsequent execution results obtained based on the summary of the diagnostic results.

[0228] Correspondingly, in this embodiment, in order to enable the third prompt word to better guide the third language model to accurately output the target diagnostic tools corresponding to each first diagnostic step, the third prompt word can be used to describe the precautions when calling each diagnostic tool. For example, the third prompt word in the aforementioned example "Note: log_retrieval log retrieval needs to rely on coverage node information, so the cover_check tool needs to be executed first" is used to describe the precautions when calling each diagnostic tool.

[0229] In step S904, the fault localization agent may summarize based on the pre-learned knowledge on how to summarize the diagnosis results of each diagnosis step. When summarizing, the fault localization agent may be implemented based on any algorithm model with summarization capability.

[0230] Exemplarily, in a possible embodiment, step S904 includes:

[0231] Step S9041, using the fault location agent, calling the fourth language model, and inputting the diagnosis results of each first diagnostic step and the fourth prompt word into the fourth language model, obtaining the answer to the fault location task output by the fourth language model under the guidance of the fourth prompt word and feeding it back to the first agent.

[0232] The fourth prompt word is used to guide the fourth language model to summarize. The fourth prompt word can be pre-set or obtained when training the fourth language model. The summary in this article refers to the process of refining the results of the task to obtain the answer to the question.

[0233] This embodiment is selected to utilize a pre-trained fault location agent and perform summarization through the trained fourth language model combined with the fourth prompt word. With the guidance of the fourth prompt word, the summarizing ability of the fourth language model can be fully utilized, thereby improving the accuracy of the execution result obtained by the summary.

[0234] The execution result obtained in step S904 can be used to represent the location result of the fault indicated by the fault location task, and can also be used to represent the location result and the solution to the fault. It is understandable that in some application scenarios, the operation and maintenance personnel are only concerned about the cause of the fault (i.e., the location result), while in other application scenarios, for inexperienced operation and maintenance personnel, they are also concerned about how to solve the fault. Therefore, by representing the location result through the execution result, or the location result plus the solution, the needs of different operation and maintenance personnel can be met in different situations.

[0235] The following will explain how to train a fault location agent. Fig.10 , Fig.10 The figure is a flow chart of the training method of the fault location agent provided by the present application, including:

[0236] Step S1001, obtaining second sample description information for describing a sample fault location problem in natural language, second diagnostic steps required to answer the sample fault location problem, diagnostic results of each second diagnostic step, an answer to the sample fault location problem, a third sample prompt word, and a fourth sample prompt word.

[0237] Step S1002: training a third language model based on the second sample description information, the third sample prompt word and the second diagnosis step.

[0238] The third language model is used to plan the diagnostic steps required to answer the sample fault location question under the guidance of the third sample prompt word.

[0239] Step S1003, training a fourth language model based on the diagnosis results of each second diagnosis step, the fourth sample prompt word, and the answer to the sample fault location question.

[0240] The fourth language model is used to summarize the diagnosis results of each second diagnosis step under the guidance of the fourth sample prompt word to obtain the answer to the sample fault location problem.

[0241] The trained fault localization agent includes the third largest language model and the fourth largest language model.

[0242] Among them, in step S1002, the second sample description information and the third sample prompt word can be used as the input of the third largest language model to obtain the output of the third largest language model, and the second diagnostic step can be used as the true value to calculate the loss between the output of the third largest language model and the true value, and then the third largest language model is adjusted according to the loss.

[0243] In step S1003, the diagnosis results of the second diagnostic steps and the fourth sample prompt word can be used as the input of the fourth language model to obtain the output of the fourth language model, and the answer to the sample fault location question is used as the true value to calculate the loss between the output of the fourth language model and the true value, and then the fourth language model is adjusted according to the loss.

[0244] By selecting this embodiment, the fault location process of the operation and maintenance personnel can be integrated into the function of the fault location intelligent agent by training the third language model and the fourth language model, so that fault location can be performed through the fault location intelligent agent, thereby improving the accuracy of fault location.

[0245] In the above example, the process of implementing fault location based on the fault location agent is described. In order to more clearly describe the fault location agent provided by the present application, the software architecture for implementing fault location is described below. For example, see Fig.11 , Fig.11 The figure shows the software architecture for implementing the fault location method provided in the present application.

[0246] Among them, the service number, super assistant and conversational front end of the user interface layer correspond to the aforementioned Figure 2b , Figure 2a as well as Figure 2c The interface in the example shown.

[0247] The fault location in the intelligent assistant layer is the aforementioned fault location agent.

[0248] The perception, planning, execution and reflection of the fault location layer are the four stages of the fault location agent, wherein the perception stage and the planning stage correspond to the aforementioned step S902, the execution stage corresponds to the aforementioned step S903 and step S904, and the reflection stage is used for the fault location agent to optimize itself based on its own output.

[0249] exist Fig.11In the example shown, the perception phase includes intent recognition, parameter extraction, and abnormal data perception, where abnormal data perception refers to the discovery of abnormal data. The planning phase includes execution chain generation, missing parameter tracing, and multi-round planning. The execution phase includes parameter acquisition, task scheduling execution, log retrieval through tools, indicator query, event query, etc. The reflection phase includes tool result summary, planning process summary, and intelligent suggestions.

[0250] based on Fig.11 Based on the framework shown, the fault localization agent can achieve the following functions:

[0251] Data-driven model training: Utilizes a large amount of historical failure data and log information for training, enabling the agent to identify different failure modes and characteristics.

[0252] Continuous learning and optimization: By continuously updating and optimizing the model, the latest fault data and solutions are introduced to maintain the advancement and accuracy of diagnostic capabilities.

[0253] Intelligent prompt generation: Combined with natural language processing technology, the prompt vocabulary is optimized to enable the intelligent agent to generate accurate diagnostic prompts based on the fault context.

[0254] Dynamic prompt adjustment: Dynamically adjust prompt words based on real-time feedback during troubleshooting to ensure the relevance and effectiveness of prompt information.

[0255] Data fusion and analysis: Integrate data from different monitoring systems, log files, and network traffic for comprehensive analysis to improve the accuracy of fault location.

[0256] Real-time data processing: Use streaming data processing technology to quickly analyze real-time data and promptly discover and locate faults.

[0257] Automated fault detection: The agent can automatically identify abnormal behaviors and fault signals, and quickly trigger the fault diagnosis process.

[0258] Decision support system: Provides detailed fault analysis reports and solution suggestions to assist operation and maintenance personnel in making decisions, and can automatically execute some solutions.

[0259] User-friendly interface design: Design an intuitive user interface so that operators can easily interact with the agent and obtain the required fault information and suggestions.

[0260] Automated tool chain integration: Seamlessly integrate with existing operation and maintenance tools and processes to automate the troubleshooting and resolution process.

[0261] For edge resource management agents, it is understandable that in some application scenarios, users want to plan and optimize edge resources, and related technologies often require manual query of edge resource conditions, and then manually formulate planning and optimization plans based on their own experience. Manual planning and optimization plans require operation and maintenance personnel to have certain relevant knowledge, which is difficult for new operation and maintenance personnel. They may even have limited relevant knowledge reserves, resulting in inaccurate manually formulated planning and optimization plans.

[0262] Based on this, the present application provides an edge resource management agent, so that the first agent can be used to call the edge resource management agent to execute the edge resource management task in the first task and obtain the execution result of the edge resource management task.

[0263] The edge resource management agent can perform edge resource management tasks in the following ways: Fig.12 As shown, including:

[0264] Step S1201: Obtain resource management description information for describing edge resource management tasks in natural language.

[0265] Step S1202, using the edge resource management agent, based on the resource management description information, plans to obtain the first resource analysis steps required to execute the edge resource management task and the resource analysis sequence of each first resource analysis step.

[0266] Step S1203, using the edge resource management intelligent body, calls the target resource analysis tool corresponding to each first resource analysis step to execute each first resource analysis step in the resource analysis order, and obtains the resource analysis results of each first resource analysis step.

[0267] The target resource analysis tool is a tool among a plurality of preset resource analysis tools that can execute corresponding resource analysis steps.

[0268] Step S1204, using the edge resource management positioning agent, summarizes the resource analysis results of each first resource analysis step to obtain the execution result of the edge resource management task and feeds it back to the first agent.

[0269] This embodiment is selected to directly use the pre-trained edge resource management agent to plan the resource management description information, and then call each resource analysis tool to execute the first resource analysis steps obtained by the plan, and summarize the resource analysis results of each first resource analysis step to obtain the execution result of the edge resource management task. During the whole process, the operation and maintenance personnel can formulate the planning and optimization plan without manually formulating the planning and optimization plan based on their own experience, which effectively improves the accuracy of edge resource management.

[0270] The above steps S1201 to S1204 are described below respectively:

[0271] Among them, in step S1201, the resource management description information may be generated when the first agent plans and obtains the edge resource management task.

[0272] In step S1202, the edge resource management agent may perform planning based on pre-learned knowledge about how to perform planning. During planning, the edge resource management agent may be implemented based on any algorithm model with planning capabilities.

[0273] Exemplarily, in a possible embodiment, step S1202 includes:

[0274] Step S12021, using the edge resource management agent, calls the fifth language model, and inputs the resource management description information and the fifth prompt word into the fifth language model, to obtain the first resource analysis steps output and executed by the fifth language model under the guidance of the fifth prompt word, and the resource analysis order of each first resource analysis step.

[0275] The fifth prompt word is used to guide the fifth language model to perform step planning. The fifth prompt word can be pre-set or obtained when training the fifth language model. For example, the fifth prompt word can be:

[0276] #Role

[0277] You are an edge construction expert who needs to be familiar with and analyze each link in the CDN or BEC construction process, including but not limited to demand analysis, construction process management, asset lifecycle management, construction plan output, and node capacity analysis.

[0278] #Target

[0279] 1.According to the user's question, accurately analyze and select the appropriate plug-in to answer according to the plug-in description.

[0280] 2. If the problem is complex, you need to call multiple different plug-ins, and call them in sequence to meet the needs of solving complex problems.

[0281] 3. Record and analyze historical conversations within three cycles to achieve the purpose of multi-round conversations.

[0282] 4. If there is no plug-in that can be called, the knowledge document is used to answer the question first.

[0283] 5. For conceptual questions, try to use knowledge documents to answer them, such as asking about processes, specifications, etc.

[0284] #limit

[0285] 1. You can only answer questions related to CDN or BEC construction. For non-related questions, you should clearly inform users that you cannot answer them.

[0286] 2. When calling a plug-in, ensure the applicability and accuracy of the plug-in to avoid providing incorrect information.

[0287] 3. Historical conversations within three cycles must be recorded and analyzed to maintain coherence and accuracy in multiple rounds of conversations.

[0288] 4. When conducting multiple rounds of conversations, you need to make full use of historical conversation records to avoid asking the same questions repeatedly and improve conversation efficiency.

[0289] This embodiment is selected to utilize the pre-trained edge resource management agent, and planning is performed based on the resource management description information through the trained fifth language model combined with the fifth prompt word. Due to the guidance of the fifth prompt word, the ability of the fifth language model planning step can be fully utilized, thereby improving the accuracy of the resource analysis step obtained by the plan, and thus making the subsequent execution results based on the resource analysis step more accurate.

[0290] The fifth prompt word is only a possible example. In other possible embodiments, the fifth prompt word can also be set according to user needs. In order to enable the fifth prompt word to better guide the fifth language model, the fifth prompt word can be used to describe the process of planning the resource analysis steps required to execute the example edge resource management task and the resource analysis order of each resource analysis step based on the example resource management description information. The example edge resource management task is the edge resource management task described by the example resource management description information;

[0291] And / or, the fifth prompt word is used to guide the fifth language model to use the first knowledge base, and the first knowledge base is used to record the health and / or construction cost of edge resources. The construction cost is used to represent the cost required for deploying some edge resources.

[0292] In step S1203, the number and type of preset resource analysis tools may vary according to different application scenarios. In addition, in a possible embodiment, the target resource analysis tools corresponding to each first resource analysis step may be obtained in the following manner:

[0293] The resource management description information, the fifth prompt word, and the description information of each of the multiple preset resource analysis tools are input into the fifth largest language model to obtain the first resource analysis steps required to execute the edge resource management task and the resource analysis order of each first resource analysis step output by the fifth largest language model under the guidance of the fifth prompt word, as well as the target resource analysis tools corresponding to each first resource analysis step output by the fifth largest language model under the guidance of the fifth prompt word.

[0294] By selecting this embodiment, the edge resource analysis agent can grasp the capabilities of each resource analysis tool based on the description information of each resource analysis tool, thereby more accurately calling the resource analysis tool with corresponding capabilities to perform the first resource analysis step, thereby improving the accuracy of the resource analysis results, and further improving the accuracy of subsequent execution results obtained based on the summary of the resource analysis results.

[0295] Correspondingly, in this embodiment, in order to enable the fifth prompt word to better guide the fifth language model to accurately output the target resource analysis tools corresponding to each first resource analysis step, the fifth prompt word can be used to describe the precautions when calling each resource analysis tool.

[0296] In step S1204, the edge resource analysis agent can summarize based on the pre-learned knowledge on how to summarize the resource analysis results of each resource analysis step. When summarizing, the edge resource analysis agent can be implemented based on any algorithm model with summarizing capabilities.

[0297] Exemplarily, in a possible embodiment, step S1204 includes:

[0298] Step S12041, using the edge resource management agent, calling the sixth language model, and inputting the resource analysis results of each first resource analysis step and the sixth prompt word into the sixth language model, obtaining the answer to the edge resource management task output by the sixth language model under the guidance of the sixth prompt word and feeding it back to the first agent.

[0299] The sixth prompt word is used to guide the sixth language model to summarize. The sixth prompt word can be pre-set or obtained when training the sixth language model. The summary in this article refers to the process of refining the results of the task to obtain the answer to the question.

[0300] This embodiment is selected to utilize the pre-trained edge resource management agent to perform summarization through the trained sixth language model combined with the sixth prompt word. Due to the guidance of the sixth prompt word, the summarizing ability of the sixth language model can be fully utilized, thereby improving the accuracy of the execution results obtained by the summary.

[0301] The execution result obtained in step S1204 can be used to indicate whether the edge resources can meet the requirements indicated by the edge resource management task, or an edge resource deployment scheme that can meet the requirements indicated by the edge resource management task. It is understandable that in some application scenarios, the operation and maintenance personnel have some requirements but are not clear whether the edge resources can meet the requirements, while in other application scenarios, the operation and maintenance personnel have some requirements but are not clear how to achieve the requirements. Therefore, the execution result indicates whether the edge resources can meet the requirements, or the deployment scheme that can meet the requirements, which can meet the needs of different operation and maintenance personnel in different situations.

[0302] The following will explain how to train the edge resource management agent. Fig.13 , Fig.13 The figure is a flow chart of the training method of the edge resource management agent provided by the present application, including:

[0303] Step S1301, obtain the third sample description information used to describe the sample edge resource management problem in natural language, the second resource analysis steps required to answer the sample edge resource management problem, the resource analysis results of each second resource analysis step, the answer to the sample edge resource management problem, the fifth sample prompt word and the sixth sample prompt word.

[0304] Step S1302: training a fifth language model based on the third sample description information, the fifth sample prompt word and the second resource analysis step.

[0305] Among them, the fifth language model is used to plan the resource analysis steps required to answer the sample edge resource management question under the guidance of the fifth sample prompt word.

[0306] Step S1303: training a sixth language model based on the resource analysis results of each second resource analysis step, the sixth sample prompt word, and the answer to the sample edge resource management question.

[0307] Among them, the sixth language model is used to summarize the resource analysis results of each second resource analysis step under the guidance of the sixth sample prompt word to obtain the answer to the sample edge resource management problem.

[0308] The trained edge resource management agents include the fifth language model and the sixth language model.

[0309] Among them, in step S1302, the third sample description information and the fifth sample prompt word can be used as the input of the fifth language model to obtain the output of the fifth language model, and the second resource analysis step is used as the true value to calculate the loss between the output of the fifth language model and the true value, and then the fifth language model is adjusted according to the loss.

[0310] In step S1303, the resource analysis results of the second resource analysis steps and the sixth sample prompt word may be used as inputs of the sixth language model to obtain the output of the sixth language model, and the answer to the sample edge resource management question may be used as the true value to calculate the loss between the output of the sixth language model and the true value, and then the sixth language model may be adjusted according to the loss.

[0311] By selecting this embodiment, the edge resource management process of the operation and maintenance personnel can be integrated into the functions of the edge resource management intelligent agent by training the fifth language model and the sixth language model, so that edge resource management can be performed through the edge resource management intelligent agent, thereby improving the accuracy of edge resource management.

[0312] In the above example, the process of implementing edge resource management based on the edge resource management agent is described. In order to more clearly describe the edge resource management agent provided by the present application, the software architecture for implementing edge resource management will be described below. For example, see Fig.14 , Fig.14 The figure shows the software architecture for implementing the edge resource management method provided in this application.

[0313] Among them, the service number, super assistant and conversational front end of the user interface layer correspond to the aforementioned Figure 2b , Figure 2a as well as Figure 2c The interface in the example shown.

[0314] Resource management in the intelligent assistant layer is the aforementioned edge resource management agent.

[0315] Perception, planning, execution and reflection of the resource management layer are the four stages of the fault location agent, among which the perception stage and the planning stage correspond to the aforementioned step S1202, the execution stage corresponds to the aforementioned step S1203 and step S1204, and the reflection stage is used for the edge resource management agent to optimize itself based on its own output.

[0316] exist Fig.14 In the example shown, the perception phase includes resource demand, resource health (i.e., obtaining the health of edge resources from the first knowledge base), construction cost (i.e., obtaining the construction cost from the first knowledge base), and model-based analysis. The planning phase includes tool chain, parameter parsing, and multiple rounds of planning. The execution phase includes relying on the enterprise large model platform (i.e., the aforementioned operation and maintenance platform), task scheduling execution, construction plan output achieved through tools, capacity assessment, construction process control, etc. The reflection phase includes result summary, model optimization, and intelligent suggestions.

[0317] based on Fig.14 Based on the framework shown, the edge resource management agent can achieve the following functions:

[0318] Natural language processing: The agent uses natural language processing technology to analyze resource construction requirements from operation and maintenance personnel or systems. These requirements may include the deployment of new nodes, expansion or optimization of existing resources, etc.

[0319] Dynamic demand assessment: Combining real-time data monitoring and historical data analysis, the intelligent agent can dynamically assess the urgency and impact of the demand, thereby prioritizing key tasks.

[0320] Intelligent reasoning ability: Using the reasoning ability of large models, intelligent agents can perform calculations in complex resource environments and derive the optimal resource allocation plan, including the location of computing nodes, optimization of network topology, and allocation of storage resources.

[0321] Multi-dimensional optimization: Considering multiple dimensions such as cost, performance, reliability and scalability, the intelligent agent can generate a construction plan that balances various factors to ensure maximum resource utilization.

[0322] Automation tool chain integration: The intelligent agent is seamlessly integrated with the automation deployment tool to achieve rapid deployment and configuration of resources. This includes automated script execution, configuration file generation, and real-time status monitoring.

[0323] Real-time adjustment and update: During the resource deployment process, the agent can make dynamic adjustments based on real-time feedback to ensure the accuracy and efficiency of resource construction.

[0324] Continuous monitoring: The agent continuously monitors deployed resources to detect and warn of potential problems in a timely manner. This can include abnormal fluctuations in resource usage, increased network latency, etc.

[0325] Closed-loop feedback mechanism: Through the feedback mechanism, the intelligent agent can collect resource usage data and feedback from operation and maintenance personnel, and continuously optimize and adjust resource management strategies.

[0326] Multi-agent collaboration: In large-scale CDN and edge computing networks, multiple agents can work together to distribute resource management tasks and improve the overall efficiency of the system.

[0327] Decision support system: The intelligent agent provides management with detailed resource status reports and optimization suggestions to assist decision makers in formulating strategic resource management plans.

[0328] As for the alarm analysis agent, it is understandable that in some application scenarios, users want to analyze the alarm, such as analyzing the cause of the alarm, filtering the alarm, etc., and related technologies often require combining their own experience to analyze the alarm. Manual alarm analysis requires operation and maintenance personnel to have certain relevant knowledge, which is difficult for new operation and maintenance personnel. It may even lead to inaccurate alarm analysis due to their limited relevant knowledge reserves, such as ignoring some alarms that should not be ignored.

[0329] Based on this, the present application provides an alarm analysis agent, so that the first agent can be used to call the alarm analysis agent to execute the alarm analysis task in the first task and obtain the execution result of the alarm analysis task.

[0330] The alarm analysis agent can perform alarm analysis tasks in the following ways: Fig.15 As shown, including:

[0331] Step S1501: Acquire alarm description information for describing the alarm analysis task in natural language.

[0332] Step S1502, using the alarm analysis agent, based on the alarm description information, plans to obtain the first alarm analysis steps required to execute the alarm analysis task and the alarm analysis sequence of each first alarm analysis step.

[0333] Step S1503, using the alarm analysis agent, calling the target alarm analysis tool corresponding to each first alarm analysis step to execute each first alarm analysis step in the alarm analysis order, and obtaining the alarm analysis results of each first alarm analysis step.

[0334] Among them, the target alarm analysis tool is a tool among multiple preset alarm analysis tools that can execute corresponding alarm analysis steps.

[0335] Step S1504, using the alarm analysis agent, based on the alarm analysis result summary of each first alarm analysis step, obtains the execution result of the alarm analysis task and feeds it back to the first agent.

[0336] This embodiment is selected to directly use the pre-trained alarm analysis agent to plan the alarm description information, and then call each alarm analysis tool to execute each first alarm analysis step obtained by the plan, and summarize the alarm analysis results of each first alarm analysis step to obtain the execution result of the alarm analysis task. In the whole process, the operation and maintenance personnel can realize the alarm analysis without manually performing the alarm analysis based on their own experience, which effectively improves the accuracy of the alarm analysis.

[0337] The above steps S1501 to S1504 are described below respectively:

[0338] Among them, in step S1501, the alarm description information can be generated when the first intelligent agent plans and obtains the alarm analysis task.

[0339] In step S1502, the alarm analysis agent may perform planning based on pre-learned knowledge about how to perform planning. During planning, the alarm analysis agent may be implemented based on any algorithm model with planning capabilities.

[0340] Exemplarily, in a possible embodiment, step S1502 includes:

[0341] Step S15021, using the alarm analysis agent, calling the seventh language model, and inputting the alarm description information and the seventh prompt word into the seventh language model, obtaining the first alarm analysis step output and executed by the seventh language model under the guidance of the seventh prompt word, and the alarm analysis order of each first alarm analysis step.

[0342] The seventh prompt word is used to guide the seventh language model to perform step planning. The seventh prompt word may be pre-set or obtained during training of the seventh language model.

[0343] This embodiment is selected to utilize a pre-trained alarm analysis agent, and planning is performed based on the alarm description information through the trained seventh language model combined with the seventh prompt word. Due to the guidance of the seventh prompt word, the ability of the seventh language model to plan steps can be fully utilized, thereby improving the accuracy of the alarm analysis steps obtained by the plan, and thus making the subsequent execution results based on the alarm analysis steps more accurate.

[0344] The seventh prompt word can be set according to user needs. In order to enable the seventh prompt word to better guide the seventh language model, the seventh prompt word can be used to describe the process of planning the alarm analysis steps required to perform the example alarm analysis task and the alarm analysis order of each alarm analysis step based on the example alarm description information, and the example alarm analysis task is the alarm analysis task described by the example alarm description information.

[0345] And / or, the seventh prompt word is used to guide the seventh language model to use the second knowledge base, and the second knowledge base is used to record the impact of the alarm on the business and the historical change information of the device that generates the alarm.

[0346] In step S1503, the number and type of preset alarm analysis tools may vary according to different application scenarios. In addition, in a possible embodiment, the target alarm analysis tools corresponding to each first alarm analysis step may be obtained in the following manner:

[0347] The alarm description information, the seventh prompt word, and the description information of multiple preset alarm analysis tools are input into the seventh language model to obtain the first alarm analysis steps required to execute the alarm analysis task and the alarm analysis order of each first alarm analysis step, which are output by the seventh language model under the guidance of the seventh prompt word, as well as the target alarm analysis tools corresponding to each first alarm analysis step, which are output by the seventh language model under the guidance of the seventh prompt word.

[0348] By selecting this embodiment, the alarm analysis intelligent agent can grasp the capabilities of each alarm analysis tool based on the description information of each alarm analysis tool, thereby more accurately calling the alarm analysis tool with corresponding capabilities to execute the first alarm analysis step, thereby improving the accuracy of the alarm analysis results, and further improving the accuracy of subsequent execution results summarized based on the alarm analysis results.

[0349] Correspondingly, in this embodiment, in order to enable the seventh prompt word to better guide the seventh language model to accurately output the target alarm analysis tool corresponding to each first alarm analysis step, the seventh prompt word can be used to describe the precautions when calling each alarm analysis tool.

[0350] In step S1504, the alarm analysis agent can summarize based on the pre-learned knowledge on how to summarize the alarm analysis results of each alarm analysis step. When summarizing, the alarm analysis agent can be implemented based on any algorithm model with summarizing capabilities.

[0351] Exemplarily, in a possible embodiment, step S1504 includes:

[0352] Step S15041, using the alarm analysis agent, calling the eighth language model, and inputting the alarm analysis results of each first alarm analysis step and the eighth prompt word into the eighth language model, obtaining the answer to the alarm analysis task output by the eighth language model under the guidance of the eighth prompt word and feeding it back to the first agent.

[0353] The eighth prompt word is used to guide the eighth language model to summarize. The eighth prompt word can be pre-set or obtained when training the eighth language model. The summary in this article refers to the process of refining the results of the task to obtain the answer to the question.

[0354] This embodiment is selected to utilize a pre-trained alarm analysis agent, and summarizes through the trained eighth language model combined with the eighth prompt word. Due to the guidance of the eighth prompt word, the summarizing ability of the eighth language model can be fully utilized, thereby improving the accuracy of the execution result obtained by the summary.

[0355] The execution result obtained in step S1504 can be used to indicate the cause of the alarm indicated by the alarm analysis task, or the threshold adjustment suggestion of the alarm, or the processing method of the alarm. It is understandable that in some application scenarios, the operation and maintenance personnel are concerned about the cause of the alarm, and in other application scenarios, the operation and maintenance personnel are concerned about how to reasonably set the threshold of the alarm so as not to miss necessary alarms while reducing unnecessary alarms. In some application scenarios, the operation and maintenance personnel manage how to handle the alarms generated, such as whether they should be ignored, what measures should be taken to eliminate the alarms, etc. Therefore, by indicating the cause, threshold adjustment suggestion or processing method through the execution result, the needs of different operation and maintenance personnel can be met in different situations.

[0356] The following will explain how to train the alarm analysis agent. Fig.16 , Fig.16 The figure is a flow chart of the training method of the alarm analysis agent provided by the present application, including:

[0357] Step S1601, obtain the fourth sample description information used to describe the sample alarm analysis question in natural language, the second alarm analysis steps required to answer the sample alarm analysis question, the alarm analysis results of each second alarm analysis step, the answer to the sample alarm analysis question, the seventh sample prompt word and the eighth sample prompt word.

[0358] Step S1602, training a seventh language model based on the fourth sample description information, the seventh sample prompt word and the second alarm analysis step.

[0359] Among them, the seventh language model is used to plan the alarm analysis steps required to answer the sample alarm analysis question under the guidance of the seventh sample prompt word.

[0360] Step S1603, training an eighth language model based on the alarm analysis results of each second alarm analysis step, the eighth sample prompt word, and the answer to the sample alarm analysis question.

[0361] The eighth language model is used to summarize the alarm analysis results of each second alarm analysis step under the guidance of the eighth sample prompt word to obtain the answer to the sample alarm analysis question.

[0362] The trained alarm analysis agents include the seventh language model and the eighth language model.

[0363] Among them, in step S1602, the fourth sample description information and the seventh sample prompt word can be used as the input of the seventh language model to obtain the output of the seventh language model, and the second alarm analysis step is used as the true value to calculate the loss between the output of the seventh language model and the true value, and then the seventh language model is adjusted according to the loss.

[0364] In step S1603, the alarm analysis results of the second alarm analysis steps and the eighth sample prompt word may be used as inputs of the eighth language model to obtain the output of the eighth language model, and the answer to the sample alarm analysis question may be used as the true value to calculate the loss between the output of the eighth language model and the true value, and then the eighth language model may be adjusted according to the loss.

[0365] By selecting this embodiment, the alarm analysis process of the operation and maintenance personnel can be integrated into the function of the alarm analysis intelligent agent by training the seventh language model and the eighth language model, so that alarm analysis can be performed through the alarm analysis intelligent agent to improve the accuracy of the alarm analysis.

[0366] In the above example, the process of implementing alarm analysis based on the alarm analysis agent is described. In order to more clearly describe the alarm analysis agent provided by the present application, the software architecture for implementing alarm analysis will be described below. For example, see Fig.17 , Fig.17 The figure shows the software architecture for implementing the edge resource management method provided in this application.

[0367] Among them, the service number, super assistant and conversational front end of the user interface layer correspond to the aforementioned Figure 2b , Figure 2a as well as Figure 2c The interface in the example shown.

[0368] The alarm analysis in the intelligent assistant layer is the aforementioned alarm analysis agent.

[0369] Perception, planning, execution and reflection of the alarm analysis layer are the four stages of the fault location agent, among which the perception stage and the planning stage correspond to the aforementioned step S1502, the execution stage corresponds to the aforementioned step S1503 and step S1504, and the reflection stage is used for the alarm analysis agent to optimize itself based on its own output.

[0370] exist Fig.17 In the example shown, the perception phase includes alarm information, business impact (i.e., obtaining business impact from the second knowledge base), historical changes (i.e., obtaining historical changes from the second knowledge base), and model-based analysis. The planning phase includes historical alarms, intelligent filtering, and tool chains. The execution phase includes parameter acquisition, task execution, and result analysis. The reflection phase includes dynamic thresholds, alarm learning optimization, and intelligent positioning. Among them, the intelligent positioning in the reflection phase can be achieved with the help of the aforementioned intelligent positioning agent.

[0371] based on Fig.17 Based on the framework shown, the alarm analysis agent can achieve the following functions:

[0372] Preliminary analysis: The intelligent agent uses the large model capability to combine the alarm information with known cross-platform data, such as event centers and business management platforms, to make a preliminary alarm judgment and plan whether further analysis or positioning is needed.

[0373] Root cause location: With the help of intelligent reasoning and pattern recognition, the intelligent agent can quickly locate the root cause of the alarm, reducing the time and manpower cost of troubleshooting. It mainly uses the underlying rich data sources and tool plug-ins to learn historical data algorithms for different alarms to achieve the location effect.

[0374] Decision guidance: The intelligent agent automatically generates detailed analysis reports and processing suggestions to help operation and maintenance personnel make wise decisions and ensure rapid resolution of problems.

[0375] Intelligent filtering: By learning historical data and alarm patterns, the agent can effectively identify and filter false alarms and redundant alarms, reducing interference from operation and maintenance personnel.

[0376] Information integration and optimization: The intelligent agent integrates multi-source alarm information, removes duplicate and ambiguous data, and performs operations such as alarm merging to provide operation and maintenance personnel with a concise and clear alarm view.

[0377] Intelligent response measures: After identifying a problem, the intelligent agent can automatically execute response measures according to preset strategies, such as adjusting system configuration, restarting services, etc., reducing human intervention.

[0378] Continuous learning and optimization: The intelligent agent continuously optimizes its own analysis and response capabilities and adapts to system changes by constantly learning new alarm patterns and processing results.

[0379] For data query agents, it is understandable that in some application scenarios, operation and maintenance personnel want to query specific data. Then the operation and maintenance personnel need to access a specific page of a specific operation and maintenance platform to query the data. However, in cross-platform operation scenarios, the number of operation and maintenance platforms and the number of pages in each operation and maintenance platform are large, and it is difficult for operation and maintenance personnel to know which page of which operation and maintenance platform each data is located in, which leads to low efficiency of operation and maintenance personnel in querying data.

[0380] Based on this, the present application provides a data query agent, so that the first agent can be used to call the data query agent to execute the data query task in the first task and obtain the execution result of the data query task.

[0381] The data query agent can perform data query tasks in the following ways: Fig.18 As shown, including:

[0382] Step S1801, obtaining query description information for describing a data query task in natural language.

[0383] Step S1802, using a data query agent, based on the query description information, plans to obtain the first data location step required to execute the data query task and the data location sequence of each first data location step.

[0384] Step S1803, using the data query agent, calling the target data locating tool corresponding to each first data locating step to execute each first data locating step in the data locating order, and obtaining the data locating result of each first data locating step.

[0385] The target data locating tool is a tool among a plurality of preset data locating tools that can execute corresponding data locating steps.

[0386] Step S1804, using the data query agent, the execution result of the data query task is obtained based on the data location result summary of each first data location step and fed back to the first agent.

[0387] This embodiment is selected to directly use the pre-trained data query agent to plan the query description information for the query description information, and then call each data location tool to respectively execute each first data location step obtained by the plan, and summarize the data location results of each first data location step to obtain the execution result of the data query task. During the whole process, the operation and maintenance personnel can realize data query without manually performing data query based on their own accumulated experience, which effectively improves the efficiency of data query in cross-platform scenarios.

[0388] The above steps S1801 to S1804 are described below respectively:

[0389] Among them, in step S1801, the query description information can be generated when the first agent plans to obtain a data query task.

[0390] In step S1802, the data query agent may perform planning based on pre-learned knowledge about how to perform planning. During planning, the data query agent may be implemented based on any algorithm model with planning capabilities.

[0391] Exemplarily, in a possible embodiment, step S1802 includes:

[0392] Step S18021, using the data query agent, calling the ninth language model, and inputting the query description information and the ninth prompt word into the ninth language model, obtaining the first data location step output and executed by the ninth language model under the guidance of the ninth prompt word, and the data location order of each first data location step.

[0393] The ninth prompt word is used to guide the ninth language model to perform step planning. The ninth prompt word may be pre-set or obtained during training of the ninth language model.

[0394] This embodiment is selected to utilize a pre-trained data query agent, and planning is performed based on the query description information through the trained ninth language model combined with the ninth prompt word. Due to the guidance of the ninth prompt word, the ability of the ninth language model planning step can be fully utilized, thereby improving the accuracy of the data location step obtained by planning, and thus making the subsequent execution results based on the data location step more accurate.

[0395] The ninth prompt word can be set according to the needs of the user. In order to enable the ninth prompt word to better guide the ninth language model, the ninth prompt word can be used to describe the process of planning the data location steps required to perform the example data query task and the data location order of each data location step based on the example query description information, and the example data query task is the data query task described by the example query description information.

[0396] In step S1803, the number and type of preset data location tools may vary according to different application scenarios. In addition, in a possible embodiment, the target data location tools corresponding to each first data location step may be obtained in the following manner:

[0397] The query description information, the ninth prompt word, and the description information of each of the plurality of preset data locating tools are input into the ninth language model, and the first data locating steps required to perform the data locating task and the data locating sequence of each first data locating step, which are output by the ninth language model under the guidance of the ninth prompt word, are obtained, as well as the target data locating tools corresponding to each first data locating step, which are output by the ninth language model under the guidance of the ninth prompt word.

[0398] By selecting this embodiment, the data query agent can grasp the capabilities of each data location tool based on the description information of each data location tool, thereby more accurately calling the data location tool with corresponding capabilities to execute the first data location step, thereby improving the accuracy of the data location results, and further improving the accuracy of the subsequent execution results obtained based on the summary of the data location results.

[0399] Correspondingly, in this embodiment, in order to enable the ninth prompt word to better guide the ninth language model to accurately output the target data locating tools corresponding to each first data locating step, the ninth prompt word can be used to describe the precautions when calling each data locating tool.

[0400] In step S1804, the data query agent may summarize based on the pre-learned knowledge on how to summarize the data location results of each data location step. When summarizing, the data query agent may be implemented based on any algorithm model with summarizing capabilities.

[0401] Exemplarily, in a possible embodiment, step S1804 includes:

[0402] Step S18041, using the data query agent, calling the tenth language model, and inputting the data location results of each first data location step and the tenth prompt word into the tenth language model, obtaining the answer to the data query task output by the tenth language model under the guidance of the tenth prompt word and feeding it back to the first agent.

[0403] The tenth prompt word is used to guide the tenth language model to summarize. The tenth prompt word can be pre-set or obtained when training the tenth language model. The summary in this article refers to the process of refining the results of the task to obtain the answer to the question.

[0404] This embodiment is selected to use the pre-trained data query agent to summarize through the trained tenth language model combined with the tenth prompt word. Due to the guidance of the tenth prompt word, the summarizing ability of the tenth language model can be fully utilized, thereby improving the accuracy of the execution result obtained by the summary.

[0405] The execution result obtained in step S1804 can be used to represent a page that can display the data indicated by the data query task. For example, it can be a link to the page, so that the operation and maintenance personnel can directly jump to the corresponding page according to the execution result to obtain the required query data.

[0406] The following will explain how to train a data query agent. Fig.19 , Fig.19 The figure is a flow chart of the training method of the alarm analysis agent provided by the present application, including:

[0407] Step S1901, obtaining the fifth sample description information used to describe the sample data query question in natural language, the second data locating step required to answer the sample data query question, the data locating results of each second data locating step, the answer to the sample data query question, the ninth sample prompt word and the tenth sample prompt word.

[0408] Step S1902, training a ninth language model based on the fifth sample description information, the ninth sample prompt word and the second data positioning step.

[0409] The ninth language model is used to plan the data location steps required to answer the sample data query question under the guidance of the ninth sample prompt word.

[0410] Step S1903, training a tenth language model based on the reported data location results of each second data location step, the tenth sample prompt word, and the answer to the sample data query question.

[0411] The tenth language model is used to summarize the data location results of each second data location step under the guidance of the tenth sample prompt word to obtain the answer to the sample data query question.

[0412] The trained data query agents include the ninth language model and the tenth language model.

[0413] Among them, in step S1902, the fifth sample description information and the ninth sample prompt word can be used as the input of the ninth language model to obtain the output of the ninth language model, and the second data positioning step is used as the true value to calculate the loss between the output of the ninth language model and the true value, and then the ninth language model is adjusted according to the loss.

[0414] In step S1903, the data location results of the second data location steps and the tenth sample prompt word may be used as the input of the tenth language model to obtain the output of the tenth language model, and the answer to the sample data query question may be used as the true value to calculate the loss between the output of the tenth language model and the true value, and then the tenth language model may be adjusted according to the loss.

[0415] By selecting this embodiment, the data query process of the operation and maintenance personnel can be integrated into the function of the data query agent by training the ninth language model and the tenth language model, so that alarm analysis can be performed through the data query agent to improve the efficiency of data query.

[0416] The alarm analysis agent provided in this application can achieve the following functions:

[0417] Natural Language Processing (NLP): The intelligent agent can understand the user's natural language input and translate it into database query statements or platform interface parameters, thereby lowering the user's technical threshold.

[0418] Semantic understanding and parsing: Through semantic analysis, the agent can accurately understand the user's query intent, even in complex or ambiguous query requests.

[0419] Cross-platform data access: The agent supports queries from multiple data sources, including relational databases, NoSQL databases, data lakes, etc., and provides a unified query interface.

[0420] Data aggregation and integration: Intelligent agents can integrate data from different data sources, perform aggregation and calculations, and provide a global view and comprehensive analysis results.

[0421] Query optimization: The agent uses intelligent algorithms to optimize query paths and execution plans to improve query speed and efficiency.

[0422] Result caching: The agent can cache the results of frequently queried results to reduce repeated calculations and database accesses.

[0423] Query habit learning: The intelligent agent provides personalized query suggestions and auto-completion functions by learning the user's query habits and history.

[0424] Intelligent recommendation: Based on the user's query history and preferences, the agent can recommend relevant data sets and analysis dimensions to improve the user experience.

[0425] The first agent and the four second agents are described above. The following is an example of the interaction between the first agent and the four second agents. Fig. 20 .

[0426] The operation and maintenance personnel use the intelligent assistant (such as the aforementioned service number, page or intelligent robot) as the entrance to access the first intelligent agent. The first intelligent agent responds to the problem description information input by the user, plans each first task based on the problem description information, and dispatches the corresponding second intelligent agent to perform each first task respectively.

[0427] When each second intelligent agent performs its corresponding task, it will also plan based on the description information of the task, and schedule tools to execute the planned steps, obtain the execution results and feed them back to the first intelligent agent. The first intelligent agent will then summarize the answer and feed it back to the intelligent assistant, so that the intelligent assistant can display the answer to the user.

[0428] Corresponding to the aforementioned cross-platform operation method, the present application also provides an agent training method, such as Fig.21 As shown, including:

[0429] Step S2101, obtaining first sample description information for describing a sample cross-platform job problem in natural language, third tasks required to answer the sample cross-platform job problem, execution results of each third task, the answer to the sample cross-platform job problem, a first sample prompt word, and a second sample prompt word.

[0430] Step S2102: training a first language model based on the first sample description information, the first sample prompt word and the third task.

[0431] The first language model is used to plan the tasks required to answer the sample cross-platform job questions under the guidance of the first sample prompt word.

[0432] Step S2103: training a second language model based on the execution results of each third task, the second sample prompt words, and the answers to the sample cross-platform homework questions.

[0433] Among them, the second largest language model is used to summarize the execution results of each third task under the guidance of the second sample prompt word to obtain the answer to the sample cross-platform operation problem.

[0434] The trained intelligent agent includes the first largest language model and the second largest language model.

[0435] By selecting this embodiment, an intelligent agent can be trained so that the intelligent agent can plan the problem description information according to the problem description information fed back by the user, and then call each second intelligent agent to respectively execute each first task obtained by the plan, and summarize the execution results of each first task to obtain the answer to the cross-platform operation problem. During the whole process, the operation and maintenance personnel only need to interact with the first intelligent agent, and do not need to access the corresponding operation and maintenance platform and the corresponding page according to the needs. Therefore, even in a multi-platform scenario, the user does not need to remember the information displayed by each of the large number of pages and the operation and maintenance platforms to which these pages belong, and can perform the operation and maintenance operation, which effectively improves the efficiency of cross-platform operations.

[0436] Corresponding to the aforementioned cross-platform operation method, the present application also provides a cross-platform operation device, such as Fig. 22 As shown, including:

[0437] An information acquisition module 2201 is used to acquire problem description information for describing a cross-platform operation problem in a natural language;

[0438] A planning module 2202 is used to use a pre-trained first agent to plan a first task required to answer the cross-platform operation problem based on the problem description information;

[0439] The execution module 2203 is used to use the first agent to call the target second agent corresponding to each of the first tasks to respectively execute each of the first tasks, and obtain the execution results of each of the first tasks, wherein the target second agent is an agent that can execute the corresponding task among a plurality of preset second agents, and different second agents have the ability to execute tasks of different categories;

[0440] The answer module 2204 is used to utilize the first agent to obtain and display the answer to the cross-platform operation problem based on the execution result summary of each of the first tasks.

[0441] In a possible embodiment, the planning module uses a pre-trained first agent to plan, based on the problem description information, a first task required to answer the cross-platform operation problem, including:

[0442] Using a pre-trained first intelligent agent, calling a first large language model, and inputting the question description information and a first prompt word into the first large language model, obtaining a first task required to answer the cross-platform job question output by the first large language model under the guidance of the first prompt word.

[0443] In a possible embodiment, the planning module is further used for:

[0444] The first agent is used to determine target second agents corresponding to each first task based on respective description information of a plurality of preset second agents, wherein the description information is used to describe the capabilities of the second agent.

[0445] In a possible embodiment, the answer module uses the first agent to summarize and display the answer to the cross-platform operation question based on the execution results of each of the first tasks, including:

[0446] Using the first agent, determining whether the execution result of each of the first tasks contains all the information to answer the cross-platform operation question;

[0447] If yes, using the first agent, based on the execution results of each of the first tasks, the answer to the cross-platform operation problem is obtained and displayed;

[0448] The device also includes a reflection module, which is used to:

[0449] If not, then using the pre-trained first agent, based on the problem description information and the execution results of each of the first tasks, a second task that needs to be performed to answer the cross-platform operation problem is planned;

[0450] Using the first agent, calling the target second agents corresponding to each of the second tasks to respectively execute each of the second tasks, and obtaining execution results of each of the second tasks;

[0451] Utilizing the first agent, the answer to the cross-platform operation problem is obtained and displayed based on the execution results of each of the first tasks and the execution results of the second tasks.

[0452] In a possible embodiment, the answer module uses the first agent to summarize and display the answer to the cross-platform operation question based on the execution results of each of the first tasks, including:

[0453] Using the pre-trained first intelligent agent, the second largest language model is called, and the execution results of each of the first tasks and the second prompt word are input into the second largest language model, so as to obtain and display the answer to the cross-platform homework problem output by the second largest language model under the guidance of the second prompt word.

[0454] In a possible embodiment, the device further includes an optimization module, which is used to:

[0455] Obtaining evaluation information input for the displayed answer, wherein the evaluation information is used to indicate a matching degree between the answer and the cross-platform homework question;

[0456] Correspondingly recording the question description information, the answer, and the evaluation information;

[0457] In response to a training instruction for the first agent, the first agent is optimized according to the stored question description information, answers and evaluation information.

[0458] In a possible embodiment, the training process of the first agent includes:

[0459] Acquire first sample description information for describing a sample cross-platform job problem in natural language, third tasks required to be performed to answer the sample cross-platform job problem, execution results of each of the third tasks, the answer to the sample cross-platform job problem, a first sample prompt word, and a second sample prompt word;

[0460] Based on the first sample description information, the first sample prompt word and the third task, training a first large language model, wherein the first large language model is used to plan the tasks required to answer the sample cross-platform operation question under the guidance of the first sample prompt word;

[0461] Based on the execution results of each of the third tasks, the second sample prompt words, and the answer to the sample cross-platform homework question, training a second large language model, wherein the second large language model is used to summarize the execution results of each of the third tasks under the guidance of the second sample prompt words to obtain the answer to the sample cross-platform homework question;

[0462] The trained first intelligent agent includes the first largest language model and the second largest language model.

[0463] In a possible embodiment, the first task includes a fault location task, and the plurality of second agents include a fault location agent;

[0464] The execution module uses the first agent to call the target second agent corresponding to each of the first tasks to respectively execute each of the first tasks, and obtains the execution results of each of the first tasks, including:

[0465] Using the first agent, calling the fault location agent to execute the fault location task, and obtaining an execution result of the fault location task;

[0466] The process of the fault location agent performing the fault location task includes:

[0467] Acquire fault description information for describing the fault location task in natural language;

[0468] Using the fault location agent, based on the fault description information, a first diagnostic step required to perform the fault location task and a diagnostic order of each of the first diagnostic steps are planned;

[0469] Using the fault location agent, calling the target diagnostic tool corresponding to each of the first diagnostic steps to respectively execute each of the first diagnostic steps in the diagnostic order to obtain the diagnostic results of each of the first diagnostic steps, wherein the target diagnostic tool is a tool that can execute the corresponding diagnostic step among a plurality of preset diagnostic tools;

[0470] By using the fault location agent, the execution result of the fault location task is summarized based on the diagnosis results of each of the first diagnosis steps and fed back to the first agent.

[0471] In a possible embodiment, the execution result is used to indicate the result of locating the fault indicated by the fault locating task, or the execution result is used to indicate the result of locating the fault and a solution to the fault.

[0472] In a possible embodiment, the using the fault location agent to plan and obtain the first diagnostic steps required to perform the fault location task and the diagnostic order of each of the first diagnostic steps based on the fault description information includes:

[0473] Utilizing the fault location agent, the third largest language model is called, and the fault description information and the third prompt word are input into the third largest language model, so as to obtain the first diagnostic steps required to perform the fault location task and the diagnostic order of each of the first diagnostic steps output by the third largest language model under the guidance of the third prompt word.

[0474] In a possible embodiment, the third prompt word is used to describe the process of planning the diagnostic steps required to perform an example fault location task and the diagnostic order of each diagnostic step based on the example fault description information, and the example fault location task is the fault location task described by the example fault description information.

[0475] In a possible embodiment, the inputting the fault description information and the third prompt word into the third largest language model to obtain the first diagnostic steps required to perform the fault location task and the diagnostic order of each of the first diagnostic steps output by the third largest language model under the guidance of the third prompt word, including:

[0476] The fault description information, the third prompt word, and the description information of each of the multiple preset diagnostic tools are input into the third largest language model to obtain the first diagnostic steps required to perform the fault location task and the diagnostic order of each of the first diagnostic steps, which are output by the third largest language model under the guidance of the third prompt word, and the target diagnostic tools corresponding to each of the first diagnostic steps, which are output by the third largest language model under the guidance of the third prompt word.

[0477] In a possible embodiment, the third prompt word is used to describe the precautions when calling each of the diagnostic tools.

[0478] In a possible embodiment, the using the fault localization agent to obtain the execution result of the fault localization task based on the diagnosis result summary of each of the first diagnosis steps and feeding back the result to the first agent includes:

[0479] Utilizing the fault location agent, the fourth language model is called, and the diagnosis results of each of the first diagnostic steps and the fourth prompt word are input into the fourth language model, so as to obtain the answer to the fault location task output by the fourth language model under the guidance of the fourth prompt word and feed it back to the first agent.

[0480] In a possible embodiment, the training process of the fault location agent includes:

[0481] Acquire second sample description information for describing a sample fault location question in natural language, second diagnostic steps required to answer the sample fault location question, diagnostic results of each of the second diagnostic steps, an answer to the sample fault location question, a third sample prompt word, and a fourth sample prompt word;

[0482] Based on the second sample description information, the third sample prompt word and the second diagnostic step, training a third language model, wherein the third language model is used to plan the diagnostic steps required to answer the sample fault location question under the guidance of the third sample prompt word;

[0483] Based on the diagnosis results of each of the second diagnosis steps, the fourth sample prompt word, and the answer to the sample fault location question, a fourth language model is trained, wherein the fourth language model is used to summarize the diagnosis results of each of the second diagnosis steps under the guidance of the fourth sample prompt word to obtain the answer to the sample fault location question;

[0484] The trained fault location agent includes the third largest language model and the fourth largest language model.

[0485] In a possible embodiment, the first task includes an edge resource management task, and the plurality of the second agents include an edge resource management agent;

[0486] The execution module uses the first agent to call the target second agent corresponding to each of the first tasks to respectively execute each of the first tasks, and obtains the execution results of each of the first tasks, including:

[0487] Using the first agent, calling the edge resource management agent to execute the edge resource management task, and obtaining the execution result of the edge resource management task;

[0488] The process of the edge resource management agent performing the edge resource management task includes:

[0489] Acquire resource management description information for describing the edge resource management task in natural language;

[0490] Using the edge resource management agent, based on the resource management description information, a first resource analysis step required to execute the edge resource management task and a resource analysis sequence of each of the first resource analysis steps are planned;

[0491] Utilizing the edge resource management agent, calling the target resource analysis tool corresponding to each of the first resource analysis steps to respectively execute each of the first resource analysis steps in accordance with the resource analysis order, and obtaining the resource analysis results of each of the first resource analysis steps, wherein the target resource analysis tool is a tool among a plurality of preset resource analysis tools that can execute the corresponding resource analysis step;

[0492] The edge resource management positioning agent is utilized to obtain the execution result of the edge resource management task based on the summary of the resource analysis results of each of the first resource analysis steps and feed it back to the first agent.

[0493] In a possible embodiment, the execution result is used to indicate whether the edge resource can meet the requirement indicated by the edge resource management task, or an edge resource deployment solution that can meet the requirement indicated by the edge resource management task.

[0494] In a possible embodiment, the using the edge resource management agent to plan the first resource analysis steps required to execute the edge resource management task and the resource analysis sequence of each of the first resource analysis steps based on the resource management description information includes:

[0495] Utilizing the edge resource management agent, the fifth language model is called, and the resource management description information and the fifth prompt word are input into the fifth language model to obtain the first resource analysis steps output by the fifth language model under the guidance of the fifth prompt word and the resource analysis order of each of the first resource analysis steps.

[0496] In a possible embodiment, the fifth prompt word is used to describe the process of planning the resource analysis steps required to execute an example edge resource management task and the resource analysis order of each resource analysis step based on example resource management description information, and the example edge resource management task is the edge resource management task described by the example resource management description information; and / or, the fifth prompt word is used to guide the fifth language model to use a first knowledge base, and the first knowledge base is used to record the health and / or construction cost of edge resources.

[0497] In a possible embodiment, the inputting the resource management description information and the fifth prompt word into the fifth language model to obtain the first resource analysis steps output by the fifth language model under the guidance of the fifth prompt word and the resource analysis order of each of the first resource analysis steps includes:

[0498] The resource management description information, the fifth prompt word, and the description information of each of the multiple preset resource analysis tools are input into the fifth largest language model to obtain the first resource analysis steps required to execute the edge resource management task and the resource analysis order of each of the first resource analysis steps, which are output by the fifth largest language model under the guidance of the fifth prompt word, and the target resource analysis tools corresponding to each of the first resource analysis steps, which are output by the fifth largest language model under the guidance of the fifth prompt word.

[0499] In a possible embodiment, the utilizing the edge resource management positioning agent to obtain the execution result of the edge resource management task based on the resource analysis result summary of each of the first resource analysis steps and feeding back the result to the first agent includes:

[0500] Utilizing the edge resource management agent, the sixth language model is called, and the resource analysis results of each of the first resource analysis steps and the sixth prompt word are input into the sixth language model, so as to obtain the answer to the edge resource management task output by the sixth language model under the guidance of the sixth prompt word and feed it back to the first agent.

[0501] In a possible embodiment, the training process of the edge resource management agent includes:

[0502] Acquire third sample description information for describing a sample edge resource management problem in natural language, a second resource analysis step required to answer the sample edge resource management problem, resource analysis results of each of the second resource analysis steps, an answer to the sample edge resource management problem, a fifth sample prompt word, and a sixth sample prompt word;

[0503] Based on the third sample description information, the fifth sample prompt word and the second resource analysis step, training a fifth language model, wherein the fifth language model is used to plan the resource analysis steps required to answer the sample edge resource management question under the guidance of the fifth sample prompt word;

[0504] Based on the resource analysis results of each of the second resource analysis steps, the sixth sample prompt word, and the answer to the sample edge resource management question, a sixth language model is trained, wherein the sixth language model is used to summarize the resource analysis results of each of the second resource analysis steps under the guidance of the sixth sample prompt word to obtain the answer to the sample edge resource management question;

[0505] The trained edge resource management agent includes the fifth language model and the sixth language model.

[0506] In a possible embodiment, the first task includes an alarm analysis task, and the plurality of the second agents include an alarm analysis agent;

[0507] The execution module uses the first agent to call the target second agent corresponding to each of the first tasks to respectively execute each of the first tasks, and obtains the execution results of each of the first tasks, including:

[0508] Using the first agent, calling the alarm analysis agent to execute the alarm analysis task, and obtaining the execution result of the alarm analysis task;

[0509] The process of the alarm analysis agent performing the alarm analysis task includes:

[0510] Acquire alarm description information for describing the alarm analysis task in natural language;

[0511] Using the alarm analysis agent, based on the alarm description information, a first alarm analysis step required to execute the alarm analysis task and an alarm analysis sequence of each of the first alarm analysis steps are planned;

[0512] Utilizing the alarm analysis agent, calling the target alarm analysis tool corresponding to each of the first alarm analysis steps to respectively execute each of the first alarm analysis steps in accordance with the alarm analysis sequence, and obtaining the alarm analysis results of each of the first alarm analysis steps, wherein the target alarm analysis tool is a tool among a plurality of preset alarm analysis tools that can execute the corresponding alarm analysis step;

[0513] By utilizing the alarm analysis agent, the execution result of the alarm analysis task is obtained based on the alarm analysis result summary of each of the first alarm analysis steps and fed back to the first agent.

[0514] In a possible embodiment, the execution result is used to indicate the cause of the alarm indicated by the alarm analysis task, or a threshold adjustment suggestion for the alarm, or a processing method for the alarm.

[0515] In a possible embodiment, the using the alarm analysis agent to plan the first alarm analysis steps required to perform the alarm analysis task and the alarm analysis sequence of each of the first alarm analysis steps based on the alarm description information includes:

[0516] Utilizing the alarm analysis agent, the seventh language model is called, and the alarm description information and the seventh prompt word are input into the seventh language model to obtain the first alarm analysis step outputted and executed by the seventh language model under the guidance of the seventh prompt word and the alarm analysis order of each of the first alarm analysis steps.

[0517] In a possible embodiment, the seventh prompt word is used to describe the process of planning the alarm analysis steps required to execute an example alarm analysis task and the alarm analysis order of each alarm analysis step based on the example alarm description information, and the example alarm analysis task is the alarm analysis task described by the example alarm description information; and / or, the fifth prompt word is used to guide the fifth language model to use a second knowledge base, and the second knowledge base is used to record the impact of alarms on business and historical change information of the equipment that generates alarms.

[0518] In a possible embodiment, the step of inputting the alarm description information and the seventh prompt word into the seventh language model to obtain the first alarm analysis steps output by the seventh language model under the guidance of the seventh prompt word and the alarm analysis order of each of the first alarm analysis steps includes:

[0519] The alarm description information, the seventh prompt word, and the description information of each of the multiple preset alarm analysis tools are input into the seventh language model to obtain the first alarm analysis steps required to execute the alarm analysis task and the alarm analysis order of each of the first alarm analysis steps, which are output by the seventh language model under the guidance of the seventh prompt word, as well as the target alarm analysis tools corresponding to each of the first alarm analysis steps, which are output by the seventh language model under the guidance of the seventh prompt word.

[0520] In a possible embodiment, the using the alarm analysis agent to obtain the execution result of the alarm analysis task based on the alarm analysis result summary of each of the first alarm analysis steps and feeding back the result to the first agent includes:

[0521] Utilizing the alarm analysis agent, the eighth language model is called, and the alarm analysis results of each of the first alarm analysis steps and the eighth prompt word are input into the eighth language model, so as to obtain the answer to the alarm analysis task output by the eighth language model under the guidance of the eighth prompt word and feed it back to the first agent.

[0522] In a possible embodiment, the training process of the alarm analysis agent includes:

[0523] Obtaining fourth sample description information for describing a sample alarm analysis question in natural language, second alarm analysis steps required to be performed to answer the sample alarm analysis question, alarm analysis results of each of the second alarm analysis steps, an answer to the sample alarm analysis question, a seventh sample prompt word, and an eighth sample prompt word;

[0524] Based on the fourth sample description information, the seventh sample prompt word and the second alarm analysis step, training a seventh language model, wherein the seventh language model is used to plan the alarm analysis steps required to answer the sample alarm analysis question under the guidance of the seventh sample prompt word;

[0525] Based on the alarm analysis results of each of the second alarm analysis steps, the eighth sample prompt word, and the answer to the sample alarm analysis question, an eighth language model is trained, wherein the eighth language model is used to summarize the alarm analysis results of each of the second alarm analysis steps under the guidance of the eighth sample prompt word to obtain the answer to the sample alarm analysis question;

[0526] The trained alarm analysis agent includes the seventh language model and the eighth language model.

[0527] In a possible embodiment, the first task includes a data query task, and the plurality of the second agents include a data query agent;

[0528] The execution module uses the first agent to call the target second agent corresponding to each of the first tasks to respectively execute each of the first tasks, and obtains the execution results of each of the first tasks, including:

[0529] Using the first agent, calling the data query agent to execute the data query task, and obtaining the execution result of the data query task;

[0530] The process of the data query agent executing the data query task includes:

[0531] Acquire query description information for describing the data query task in natural language;

[0532] Using the data query agent, based on the query description information, a first data location step required to execute the data query task and a data location sequence of each of the first data location steps are planned;

[0533] Utilizing the data query agent, calling the target data locating tool corresponding to each of the first data locating steps to respectively execute each of the first data locating steps in the data locating order, and obtaining the data locating result of each of the first data locating steps, wherein the target data locating tool is a tool among a plurality of preset data locating tools that can execute the corresponding data locating step;

[0534] By using the data query agent, the execution result of the data query task is obtained based on the data positioning result summary of each of the first data positioning steps and fed back to the first agent.

[0535] In a possible embodiment, the execution result is used to represent a page that can display the data indicated by the data query task.

[0536] In a possible embodiment, the using the data query agent to plan based on the query description information to obtain the first data location steps required to execute the data query task and the data location order of each of the first data location steps includes:

[0537] The data query agent is used to call the ninth language model, and the query description information and the ninth prompt word are input into the ninth language model to obtain the first data location step output by the ninth language model under the guidance of the ninth prompt word and the data location order of each of the first data location steps.

[0538] In a possible embodiment, the ninth prompt word is used to describe the process of planning the data location steps required to execute an example data query task and the data location order of each data location step based on the example query description information. The example data query task is the data query task described by the example query description information.

[0539] In a possible embodiment, the step of inputting the query description information and the ninth prompt word into the ninth language model to obtain the first data locating steps output by the ninth language model under the guidance of the ninth prompt word and the data locating order of each of the first data locating steps includes:

[0540] The query description information, the ninth prompt word, and the description information of each of the plurality of preset data locating tools are input into the ninth language model to obtain the first data locating steps required to execute the data locating task and the data locating sequence of each of the first data locating steps, which are output by the ninth language model under the guidance of the ninth prompt word, and the target data locating tools corresponding to each of the first data locating steps, which are output by the ninth language model under the guidance of the ninth prompt word.

[0541] In a possible embodiment, the using the data query agent to obtain the execution result of the data query task based on the data location result of each of the first data location steps and feeding back the result to the first agent includes:

[0542] The data query agent is used to call the tenth language model, and the data location results of each of the first data location steps and the tenth prompt word are input into the tenth language model, so as to obtain the answer to the data query task output by the tenth language model under the guidance of the tenth prompt word and feed it back to the first agent.

[0543] In a possible embodiment, the training process of the data query agent includes:

[0544] Acquire fifth sample description information for describing the sample data query question in natural language, the second data location step required to answer the sample data query question, the data location results of each of the second data location steps, the answer to the sample data query question, the ninth sample prompt word, and the tenth sample prompt word;

[0545] Based on the fifth sample description information, the ninth sample prompt word and the second data location step, a ninth language model is trained, wherein the ninth language model is used to plan the data location steps required to answer the sample data query question under the guidance of the ninth sample prompt word;

[0546] Based on the data location results of each of the second data location steps, the tenth sample prompt word, and the answer to the sample data query question, a tenth language model is trained, wherein the tenth language model is used to summarize the data location results of each of the second data location steps under the guidance of the tenth sample prompt word to obtain the answer to the sample data query question;

[0547] The trained data query agent includes the ninth language model and the tenth language model.

[0548] Corresponding to the aforementioned intelligent agent training method, the present application also provides an intelligent agent training device, such as Fig.23 As shown, including:

[0549] The sample acquisition module 2301 is used to acquire first sample description information for describing a sample cross-platform operation problem in natural language, third tasks required to be performed to answer the sample cross-platform operation problem, execution results of each of the third tasks, answers to the sample cross-platform operation problem, first sample prompt words, and second sample prompt words;

[0550] A first training module 2302 is used to train a first large language model based on the first sample description information, the first sample prompt word and the third task, wherein the first large language model is used to plan the tasks required to answer the sample cross-platform operation question under the guidance of the first sample prompt word;

[0551] A second training module 2303 is used to train a second large language model based on the execution results of each of the third tasks, the second sample prompt words, and the answer to the sample cross-platform homework question, wherein the second large language model is used to summarize the execution results of each of the third tasks under the guidance of the second sample prompt words to obtain the answer to the sample cross-platform homework question;

[0552] The trained intelligent agent includes the first largest language model and the second largest language model.

[0553] The present application also provides a first intelligent agent, comprising:

[0554] An input module, used for receiving input information;

[0555] a processing module, configured to determine a target task based on the input information received by the input module, determine a large language model based on the target task, and obtain output information by calling the large language model to execute any of the above-mentioned cross-platform operation methods;

[0556] An output module is used to output the output information obtained by the processing module.

[0557] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.

[0558] Fig.24 A schematic block diagram of an example electronic device 2400 that can be used to implement an embodiment of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0559] like Fig.24 As shown, the device 2400 includes a computing unit 2401, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 2402 or a computer program loaded from a storage unit 2408 into a random access memory (RAM) 2403. In the RAM 2403, various programs and data required for the operation of the device 2400 can also be stored. The computing unit 2401, the ROM 2402, and the RAM 2403 are connected to each other via a bus 2404. An input / output (I / O) interface 2405 is also connected to the bus 2404.

[0560] A number of components in the device 2400 are connected to the I / O interface 2405, including: an input unit 2406, such as a keyboard, a mouse, etc.; an output unit 2407, such as various types of displays, speakers, etc.; a storage unit 2408, such as a disk, an optical disk, etc.; and a communication unit 2409, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 2409 allows the device 2400 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0561] The computing unit 2401 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 2401 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 2401 performs the various methods and processes described above, such as a cross-platform operation method or an agent training method. For example, in some embodiments, the cross-platform operation method or the agent training method may be implemented as a computer software program, which is tangibly included in a machine-readable medium, such as a storage unit 2408. In some embodiments, part or all of the computer program may be loaded and / or installed on the device 2400 via the ROM 2402 and / or the communication unit 2409. When the computer program is loaded into the RAM 2403 and executed by the computing unit 2401, one or more steps of the cross-platform operation method or the agent training method described above may be performed. Alternatively, in other embodiments, the computing unit 2401 may be configured to execute the cross-platform operation method or the intelligent agent training method in any other appropriate manner (e.g., by means of firmware).

[0562] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0563] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0564] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0565] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0566] The systems and techniques described herein may be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), and the Internet.

[0567] A computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship of client and server is generated by computer programs running on respective computers and having a client-server relationship with each other. The server may be a cloud server, a server of a distributed system, or a server combined with a blockchain.

[0568] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and this document does not limit this.

[0569] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A cross-platform operation method, comprising: Obtain problem description information for describing the cross-platform operation problem in natural language; Using a pre-trained first intelligent agent, based on the problem description information, a first task required to be performed to answer the cross-platform operation problem is planned; Using the first agent, calling the target second agents corresponding to each of the first tasks to respectively execute each of the first tasks, and obtaining the execution results of each of the first tasks, wherein the target second agent is an agent that can execute the corresponding task among a plurality of preset second agents, and different second agents have the ability to execute tasks of different categories; Utilizing the first agent, the answer to the cross-platform operation problem is summarized based on the execution results of each of the first tasks and displayed.

2. The method according to claim 1, wherein: The method of using the pre-trained first agent to plan a first task required to answer the cross-platform operation problem based on the problem description information includes: Using a pre-trained first intelligent agent, calling a first large language model, and inputting the question description information and a first prompt word into the first large language model, obtaining a first task required to answer the cross-platform job question output by the first large language model under the guidance of the first prompt word.

3. The method according to claim 1, further comprising: The first agent is used to determine target second agents corresponding to each first task based on respective description information of a plurality of preset second agents, wherein the description information is used to describe the capabilities of the second agent.

4. The method according to claim 1, wherein the using the first agent to obtain and display the answer to the cross-platform operation problem based on the execution results of each of the first tasks comprises: Using the first agent, determining whether the execution result of each of the first tasks contains all the information to answer the cross-platform operation question; If yes, using the first agent, based on the execution results of each of the first tasks, the answer to the cross-platform operation problem is obtained and displayed; The method further comprises: If not, then using the pre-trained first agent, based on the problem description information and the execution results of each of the first tasks, a second task that needs to be performed to answer the cross-platform operation problem is planned; Using the first agent, calling the target second agents corresponding to each of the second tasks to respectively execute each of the second tasks, and obtaining execution results of each of the second tasks; Utilizing the first agent, the answer to the cross-platform operation problem is obtained and displayed based on the execution results of each of the first tasks and the execution results of the second tasks.

5. The method according to claim 1, wherein: The using the first agent to obtain and display the answer to the cross-platform operation problem based on the execution results of each of the first tasks includes: Using the pre-trained first intelligent agent, the second largest language model is called, and the execution results of each of the first tasks and the second prompt word are input into the second largest language model, so as to obtain and display the answer to the cross-platform homework problem output by the second largest language model under the guidance of the second prompt word.

6. The method according to claim 1, further comprising: Obtaining evaluation information input for the displayed answer, wherein the evaluation information is used to indicate a matching degree between the answer and the cross-platform homework question; Correspondingly recording the question description information, the answer, and the evaluation information; In response to a training instruction for the first agent, the first agent is optimized according to the stored question description information, answers and evaluation information.

7. The method according to any one of claims 1 to 6, wherein: The training process of the first agent includes: Acquire first sample description information for describing a sample cross-platform job problem in natural language, third tasks required to be performed to answer the sample cross-platform job problem, execution results of each of the third tasks, an answer to the sample cross-platform job problem, a first sample prompt word, and a second sample prompt word; Based on the first sample description information, the first sample prompt word and the third task, training a first large language model, wherein the first large language model is used to plan the tasks required to answer the sample cross-platform operation question under the guidance of the first sample prompt word; Based on the execution results of each of the third tasks, the second sample prompt words, and the answer to the sample cross-platform homework question, training a second large language model, wherein the second large language model is used to summarize the execution results of each of the third tasks under the guidance of the second sample prompt words to obtain the answer to the sample cross-platform homework question; The trained first intelligent agent includes the first largest language model and the second largest language model.

8. The method according to claim 1, wherein: The first task includes a fault location task, and the plurality of second agents include a fault location agent; The using the first agent to call the target second agent corresponding to each of the first tasks to respectively execute each of the first tasks to obtain the execution results of each of the first tasks includes: Using the first agent, calling the fault location agent to execute the fault location task, and obtaining an execution result of the fault location task; The process of the fault location agent performing the fault location task includes: Acquire fault description information for describing the fault location task in natural language; Using the fault location agent, based on the fault description information, a first diagnostic step required to perform the fault location task and a diagnostic order of each of the first diagnostic steps are planned; Using the fault location agent, calling the target diagnostic tool corresponding to each of the first diagnostic steps to respectively execute each of the first diagnostic steps in the diagnostic order to obtain the diagnostic results of each of the first diagnostic steps, wherein the target diagnostic tool is a tool that can execute the corresponding diagnostic step among a plurality of preset diagnostic tools; By using the fault location agent, the execution result of the fault location task is summarized based on the diagnosis results of each of the first diagnosis steps and fed back to the first agent.

9. The method according to claim 8, wherein: The execution result is used to indicate the result of locating the fault indicated by the fault locating task, or the execution result is used to indicate the result of locating the fault and a solution to the fault.

10. The method according to claim 8, wherein: The method of using the fault location agent to plan and obtain the first diagnostic steps required to perform the fault location task and the diagnostic sequence of each of the first diagnostic steps based on the fault description information includes: Utilizing the fault location agent, the third largest language model is called, and the fault description information and the third prompt word are input into the third largest language model, so as to obtain the first diagnostic steps required to perform the fault location task and the diagnostic order of each of the first diagnostic steps output by the third largest language model under the guidance of the third prompt word.

11. The method according to claim 10, wherein: The third prompt word is used to describe the process of planning the diagnostic steps required to execute an example fault location task and the diagnostic order of each diagnostic step based on the example fault description information. The example fault location task is the fault location task described by the example fault description information.

12. The method according to claim 10, wherein: The step of inputting the fault description information and the third prompt word into the third language model to obtain the first diagnostic steps required to perform the fault location task and the diagnostic sequence of each of the first diagnostic steps output by the third language model under the guidance of the third prompt word includes: The fault description information, the third prompt word, and the description information of each of the multiple preset diagnostic tools are input into the third largest language model to obtain the first diagnostic steps required to perform the fault location task and the diagnostic order of each of the first diagnostic steps, which are output by the third largest language model under the guidance of the third prompt word, and the target diagnostic tools corresponding to each of the first diagnostic steps, which are output by the third largest language model under the guidance of the third prompt word.

13. The method according to claim 12, wherein: The third prompt word is used to describe the precautions when calling each of the diagnostic tools.

14. The method according to claim 8, wherein: The method of using the fault location agent to obtain the execution result of the fault location task based on the diagnosis result of each of the first diagnosis steps and feeding it back to the first agent includes: Utilizing the fault location agent, the fourth language model is called, and the diagnosis results of each of the first diagnostic steps and the fourth prompt word are input into the fourth language model, so as to obtain the answer to the fault location task output by the fourth language model under the guidance of the fourth prompt word and feed it back to the first agent.

15. The method according to any one of claims 8 to 14, wherein: The training process of the fault location agent includes: Acquire second sample description information for describing a sample fault location question in natural language, second diagnostic steps required to answer the sample fault location question, diagnostic results of each of the second diagnostic steps, an answer to the sample fault location question, a third sample prompt word, and a fourth sample prompt word; Based on the second sample description information, the third sample prompt word and the second diagnostic step, training a third language model, wherein the third language model is used to plan the diagnostic steps required to answer the sample fault location question under the guidance of the third sample prompt word; Based on the diagnosis results of each of the second diagnosis steps, the fourth sample prompt word, and the answer to the sample fault location question, a fourth language model is trained, wherein the fourth language model is used to summarize the diagnosis results of each of the second diagnosis steps under the guidance of the fourth sample prompt word to obtain the answer to the sample fault location question; The trained fault location agent includes the third largest language model and the fourth largest language model.

16. The method according to claim 1, wherein: The first task includes an edge resource management task, and the plurality of second agents include an edge resource management agent; The using the first agent to call the target second agent corresponding to each of the first tasks to respectively execute each of the first tasks to obtain the execution results of each of the first tasks includes: Using the first agent, calling the edge resource management agent to execute the edge resource management task, and obtaining the execution result of the edge resource management task; The process of the edge resource management agent performing the edge resource management task includes: Acquire resource management description information for describing the edge resource management task in natural language; Using the edge resource management agent, based on the resource management description information, a first resource analysis step required to execute the edge resource management task and a resource analysis sequence of each of the first resource analysis steps are planned; Utilizing the edge resource management agent, calling the target resource analysis tool corresponding to each of the first resource analysis steps to respectively execute each of the first resource analysis steps in accordance with the resource analysis order, and obtaining the resource analysis results of each of the first resource analysis steps, wherein the target resource analysis tool is a tool among a plurality of preset resource analysis tools that can execute the corresponding resource analysis step; The edge resource management positioning agent is utilized to obtain the execution result of the edge resource management task based on the summary of the resource analysis results of each of the first resource analysis steps and feed it back to the first agent.

17. The method according to claim 16, wherein: The execution result is used to indicate whether the edge resource can meet the requirement indicated by the edge resource management task, or an edge resource deployment solution that can meet the requirement indicated by the edge resource management task.

18. The method according to claim 16, wherein: The using the edge resource management agent to plan the first resource analysis steps required to execute the edge resource management task and the resource analysis sequence of each of the first resource analysis steps based on the resource management description information includes: Utilizing the edge resource management agent, the fifth language model is called, and the resource management description information and the fifth prompt word are input into the fifth language model to obtain the first resource analysis steps output by the fifth language model under the guidance of the fifth prompt word and the resource analysis order of each of the first resource analysis steps.

19. The method according to claim 18, wherein: The fifth prompt word is used to describe the process of planning the resource analysis steps required to execute an example edge resource management task and the resource analysis order of each resource analysis step based on the example resource management description information, and the example edge resource management task is the edge resource management task described by the example resource management description information; and / or, the fifth prompt word is used to guide the fifth language model to use the first knowledge base, and the first knowledge base is used to record the health and / or construction cost of edge resources.

20. The method according to claim 18, wherein: The step of inputting the resource management description information and the fifth prompt word into the fifth language model to obtain the first resource analysis steps output by the fifth language model under the guidance of the fifth prompt word and the resource analysis sequence of each of the first resource analysis steps, including: The resource management description information, the fifth prompt word, and the description information of each of the multiple preset resource analysis tools are input into the fifth largest language model to obtain the first resource analysis steps required to execute the edge resource management task and the resource analysis order of each of the first resource analysis steps, which are output by the fifth largest language model under the guidance of the fifth prompt word, and the target resource analysis tools corresponding to each of the first resource analysis steps, which are output by the fifth largest language model under the guidance of the fifth prompt word.

21. The method according to claim 16, wherein: The utilizing the edge resource management positioning agent to obtain the execution result of the edge resource management task based on the resource analysis result of each of the first resource analysis steps and feeding it back to the first agent includes: Utilizing the edge resource management agent, the sixth language model is called, and the resource analysis results of each of the first resource analysis steps and the sixth prompt word are input into the sixth language model, so as to obtain the answer to the edge resource management task output by the sixth language model under the guidance of the sixth prompt word and feed it back to the first agent.

22. The method according to any one of claims 16 to 21, wherein: The training process of the edge resource management agent includes: Acquire third sample description information for describing a sample edge resource management problem in natural language, a second resource analysis step required to answer the sample edge resource management problem, resource analysis results of each of the second resource analysis steps, an answer to the sample edge resource management problem, a fifth sample prompt word, and a sixth sample prompt word; Based on the third sample description information, the fifth sample prompt word and the second resource analysis step, training a fifth language model, wherein the fifth language model is used to plan the resource analysis steps required to answer the sample edge resource management question under the guidance of the fifth sample prompt word; Based on the resource analysis results of each of the second resource analysis steps, the sixth sample prompt word, and the answer to the sample edge resource management question, a sixth language model is trained, wherein the sixth language model is used to summarize the resource analysis results of each of the second resource analysis steps under the guidance of the sixth sample prompt word to obtain the answer to the sample edge resource management question; The trained edge resource management agent includes the fifth language model and the sixth language model.

23. The method according to claim 1, wherein: The first task includes an alarm analysis task, and the plurality of the second agents include an alarm analysis agent; The using the first agent to call the target second agent corresponding to each of the first tasks to respectively execute each of the first tasks to obtain the execution results of each of the first tasks includes: Using the first agent, calling the alarm analysis agent to execute the alarm analysis task, and obtaining the execution result of the alarm analysis task; The process of the alarm analysis agent performing the alarm analysis task includes: Acquire alarm description information for describing the alarm analysis task in natural language; Using the alarm analysis agent, based on the alarm description information, a first alarm analysis step required to execute the alarm analysis task and an alarm analysis sequence of each of the first alarm analysis steps are planned; Utilizing the alarm analysis agent, calling the target alarm analysis tool corresponding to each of the first alarm analysis steps to respectively execute each of the first alarm analysis steps in accordance with the alarm analysis sequence, and obtaining the alarm analysis results of each of the first alarm analysis steps, wherein the target alarm analysis tool is a tool among a plurality of preset alarm analysis tools that can execute the corresponding alarm analysis step; By utilizing the alarm analysis agent, the execution result of the alarm analysis task is obtained based on the alarm analysis result summary of each of the first alarm analysis steps and fed back to the first agent.

24. The method according to claim 23, wherein: The execution result is used to indicate the cause of the alarm indicated by the alarm analysis task, or the threshold adjustment suggestion for the alarm, or the processing method of the alarm.

25. The method according to claim 23, wherein: The method of using the alarm analysis agent to plan the first alarm analysis steps required to perform the alarm analysis task and the alarm analysis sequence of each of the first alarm analysis steps based on the alarm description information includes: Utilizing the alarm analysis agent, the seventh language model is called, and the alarm description information and the seventh prompt word are input into the seventh language model to obtain the first alarm analysis step outputted and executed by the seventh language model under the guidance of the seventh prompt word and the alarm analysis order of each of the first alarm analysis steps.

26. The method according to claim 25, wherein: The seventh prompt word is used to describe the process of planning the alarm analysis steps required to execute the example alarm analysis task and the alarm analysis order of each alarm analysis step based on the example alarm description information, and the example alarm analysis task is the alarm analysis task described by the example alarm description information; and / or, the fifth prompt word is used to guide the fifth language model to use the second knowledge base, and the second knowledge base is used to record the impact of the alarm on the business and the historical change information of the equipment that generates the alarm.

27. The method according to claim 25, wherein: The step of inputting the alarm description information and the seventh prompt word into the seventh language model to obtain the first alarm analysis steps output by the seventh language model under the guidance of the seventh prompt word and the alarm analysis sequence of each of the first alarm analysis steps includes: The alarm description information, the seventh prompt word, and the description information of each of the multiple preset alarm analysis tools are input into the seventh language model to obtain the first alarm analysis steps required to execute the alarm analysis task and the alarm analysis order of each of the first alarm analysis steps, which are output by the seventh language model under the guidance of the seventh prompt word, as well as the target alarm analysis tools corresponding to each of the first alarm analysis steps, which are output by the seventh language model under the guidance of the seventh prompt word.

28. The method of claim 23, wherein: The utilizing the alarm analysis agent to obtain the execution result of the alarm analysis task based on the alarm analysis result summary of each of the first alarm analysis steps and feeding back the result to the first agent includes: Utilizing the alarm analysis agent, the eighth language model is called, and the alarm analysis results of each of the first alarm analysis steps and the eighth prompt word are input into the eighth language model, so as to obtain the answer to the alarm analysis task output by the eighth language model under the guidance of the eighth prompt word and feed it back to the first agent.

29. The method according to any one of claims 23 to 28, wherein: The training process of the alarm analysis agent includes: Obtaining fourth sample description information for describing a sample alarm analysis question in natural language, second alarm analysis steps required to be performed to answer the sample alarm analysis question, alarm analysis results of each of the second alarm analysis steps, an answer to the sample alarm analysis question, a seventh sample prompt word, and an eighth sample prompt word; Based on the fourth sample description information, the seventh sample prompt word and the second alarm analysis step, training a seventh language model, wherein the seventh language model is used to plan the alarm analysis steps required to answer the sample alarm analysis question under the guidance of the seventh sample prompt word; Based on the alarm analysis results of each of the second alarm analysis steps, the eighth sample prompt word, and the answer to the sample alarm analysis question, an eighth language model is trained, wherein the eighth language model is used to summarize the alarm analysis results of each of the second alarm analysis steps under the guidance of the eighth sample prompt word to obtain the answer to the sample alarm analysis question; The trained alarm analysis agent includes the seventh language model and the eighth language model.

30. The method of claim 1, wherein: The first task includes a data query task, and the plurality of the second agents include a data query agent; The using the first agent to call the target second agent corresponding to each of the first tasks to respectively execute each of the first tasks to obtain the execution results of each of the first tasks includes: Using the first agent, calling the data query agent to execute the data query task, and obtaining the execution result of the data query task; The process of the data query agent executing the data query task includes: Acquire query description information for describing the data query task in natural language; Using the data query agent, based on the query description information, a first data location step required to execute the data query task and a data location sequence of each of the first data location steps are planned; Utilizing the data query agent, calling the target data locating tool corresponding to each of the first data locating steps to respectively execute each of the first data locating steps in the data locating order, and obtaining the data locating result of each of the first data locating steps, wherein the target data locating tool is a tool among a plurality of preset data locating tools that can execute the corresponding data locating step; By using the data query agent, the execution result of the data query task is obtained based on the data positioning result summary of each of the first data positioning steps and fed back to the first agent.

31. The method according to claim 30, wherein: The execution result is used to represent a page that can display the data indicated by the data query task.

32. The method of claim 30, wherein: The method of using the data query agent to plan and obtain the first data location steps required to execute the data query task and the data location sequence of each of the first data location steps based on the query description information includes: The data query agent is used to call the ninth language model, and the query description information and the ninth prompt word are input into the ninth language model to obtain the first data location step output by the ninth language model under the guidance of the ninth prompt word and the data location order of each of the first data location steps.

33. The method of claim 32, wherein: The ninth prompt word is used to describe the process of planning the data location steps required to execute an example data query task and the data location order of each data location step based on the example query description information. The example data query task is the data query task described by the example query description information.

34. The method of claim 32, wherein: The step of inputting the query description information and the ninth prompt word into the ninth language model to obtain the first data locating step output by the ninth language model under the guidance of the ninth prompt word and the data locating sequence of each of the first data locating steps includes: The query description information, the ninth prompt word, and the description information of each of the plurality of preset data locating tools are input into the ninth language model to obtain the first data locating steps required to execute the data locating task and the data locating sequence of each of the first data locating steps, which are output by the ninth language model under the guidance of the ninth prompt word, and the target data locating tools corresponding to each of the first data locating steps, which are output by the ninth language model under the guidance of the ninth prompt word.

35. The method of claim 30, wherein: The method of using the data query agent to obtain the execution result of the data query task based on the data location results of each of the first data location steps and feeding back the result to the first agent includes: The data query agent is used to call the tenth language model, and the data location results of each of the first data location steps and the tenth prompt word are input into the tenth language model, so as to obtain the answer to the data query task output by the tenth language model under the guidance of the tenth prompt word and feed it back to the first agent.

36. The method according to any one of claims 30 to 35, wherein: The training process of the data query agent includes: Acquire fifth sample description information for describing the sample data query question in natural language, the second data location step required to answer the sample data query question, the data location results of each of the second data location steps, the answer to the sample data query question, the ninth sample prompt word, and the tenth sample prompt word; Based on the fifth sample description information, the ninth sample prompt word and the second data location step, a ninth language model is trained, wherein the ninth language model is used to plan the data location steps required to answer the sample data query question under the guidance of the ninth sample prompt word; Based on the data location results of each of the second data location steps, the tenth sample prompt word, and the answer to the sample data query question, a tenth language model is trained, wherein the tenth language model is used to summarize the data location results of each of the second data location steps under the guidance of the tenth sample prompt word to obtain the answer to the sample data query question; The trained data query agent includes the ninth language model and the tenth language model.

37. A method for training an intelligent agent, comprising: Acquire first sample description information for describing a sample cross-platform job problem in natural language, third tasks required to be performed to answer the sample cross-platform job problem, execution results of each of the third tasks, the answer to the sample cross-platform job problem, a first sample prompt word, and a second sample prompt word; Based on the first sample description information, the first sample prompt word and the third task, training a first large language model, wherein the first large language model is used to plan the tasks required to answer the sample cross-platform operation question under the guidance of the first sample prompt word; Based on the execution results of each of the third tasks, the second sample prompt words, and the answer to the sample cross-platform homework question, training a second large language model, wherein the second large language model is used to summarize the execution results of each of the third tasks under the guidance of the second sample prompt words to obtain the answer to the sample cross-platform homework question; The trained intelligent agent includes the first largest language model and the second largest language model.

38. A cross-platform operation device, comprising: An information acquisition module, used for acquiring problem description information for describing the cross-platform operation problem in natural language; A planning module, configured to utilize a pre-trained first agent to plan, based on the problem description information, a first task required to be performed to answer the cross-platform operation problem; an execution module, configured to use the first agent to call the target second agent corresponding to each of the first tasks to respectively execute each of the first tasks, and obtain the execution results of each of the first tasks, wherein the target second agent is an agent that can execute the corresponding task among a plurality of preset second agents, and different second agents have the ability to execute tasks of different categories; The answer module is used to utilize the first intelligent agent to obtain and display the answer to the cross-platform operation problem based on the execution result summary of each of the first tasks.

39. An intelligent agent training device, comprising: A sample acquisition module, used to acquire first sample description information for describing a sample cross-platform operation problem in natural language, a third task required to be performed to answer the sample cross-platform operation problem, an execution result of each of the third tasks, an answer to the sample cross-platform operation problem, a first sample prompt word, and a second sample prompt word; A first training module is used to train a first large language model based on the first sample description information, the first sample prompt word and the third task, wherein the first large language model is used to plan the tasks required to answer the sample cross-platform operation question under the guidance of the first sample prompt word; A second training module is used to train a second large language model based on the execution results of each of the third tasks, the second sample prompt words, and the answer to the sample cross-platform homework question, wherein the second large language model is used to summarize the execution results of each of the third tasks under the guidance of the second sample prompt words to obtain the answer to the sample cross-platform homework question; The trained intelligent agent includes the first largest language model and the second largest language model.

40. A first intelligent agent, comprising: An input module, used for receiving input information; a processing module, configured to determine a target task based on the input information received by the input module, determine a large language model based on the target task, and obtain output information by executing any one of the methods of claims 1 to 36 by calling the large language model; An output module is used to output the output information obtained by the processing module.

41. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1-36.

42. A computer program product comprising a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 36.

Images