Fault root cause location method, device, storage medium and program product

Abstract

The present invention provides a method, device, storage medium and program product for locating the root cause of a fault, which belongs to the field of artificial intelligence. The method includes: according to the operation abnormality data corresponding to the target application system and the first prompt words corresponding to the multiple agent groups, according to the multiple query questions corresponding to the multiple agent groups, respectively retrieve the reference document sets corresponding to the multiple agent groups in the set document set; distribute the multiple reference documents contained in the reference document set corresponding to each agent group to the multiple agents in the corresponding agent group, and obtain the initial analysis results corresponding to the multiple agents; obtain the target analysis results corresponding to the multiple agent groups; and obtain the abnormal root cause information according to the target analysis results corresponding to the multiple agent groups. The entire fault root cause location process of the embodiment of the present invention does not need to accumulate a large amount of historical data in advance, does not require high data quality, and has a high accuracy in root cause location.

Description

Technical Field

[0001] The present invention relates to the field of Internet technology, and in particular to a fault root cause locating method, device, storage medium and program product. Background Art

[0002] Root cause location is a key step in the operation and maintenance of many application systems (such as cloud computing systems). In actual applications, if a user submits an unavailable work order for a cloud computing system service instance, or a monitoring module used to detect data anomalies in a cloud computing system detects that the cloud computing system has abnormal data, root cause location can be performed.

[0003] When performing root cause location work, the prior art generally compares the unavailable work orders or abnormal data with the previously accumulated historical similar information, and uses a large language model to analyze the comparison results to obtain the root cause location results. However, root cause location using only a single large language model cannot cope with complex cloud computing systems, and the accuracy of root cause location is low. Summary of the invention

[0004] The embodiments of the present invention provide a fault root cause location method, device, storage medium and program product to improve the accuracy of root cause location.

[0005] In a first aspect, an embodiment of the present invention provides a method for locating a root cause of a fault, the method comprising:

[0006] Generate multiple query questions corresponding to the multiple agent groups according to the operation abnormality data corresponding to the target application system and the first prompt words respectively corresponding to the multiple agent groups, wherein the multiple agent groups correspond to the multiple fault location fields of the target application system and the multiple query questions one by one, and the first prompt words are used to prompt the corresponding agent group to perform fault analysis in the corresponding fault location field;

[0007] Retrieving reference document sets corresponding to the multiple agent groups respectively from the set document set according to the multiple query questions;

[0008] Distributing a plurality of reference documents contained in a reference document set corresponding to each agent group to a plurality of agents in the corresponding agent group, so that the plurality of agents input the corresponding query question and the reference documents respectively distributed into a large language model, and obtain initial analysis results corresponding to the plurality of agents respectively, wherein the reference documents distributed to the plurality of agents are different;

[0009] Obtaining target analysis results corresponding to the plurality of agent groups, wherein the target analysis results corresponding to the target agent group are obtained from initial analysis results corresponding to the plurality of agents in the target agent group;

[0010] The abnormal root cause information is obtained according to the target analysis results corresponding to the multiple agent groups.

[0011] In a second aspect, an embodiment of the present invention provides a fault root cause location device, the device comprising:

[0012] A generating module, configured to generate a plurality of query questions corresponding to the plurality of agent groups according to the operation abnormality data corresponding to the target application system and the first prompt words respectively corresponding to the plurality of agent groups, wherein the plurality of agent groups correspond one-to-one to the plurality of fault location fields of the target application system and the plurality of query questions, and the first prompt words are used to prompt the corresponding agent group to perform fault analysis in the corresponding fault location field;

[0013] A retrieval module, configured to retrieve reference document sets corresponding to the plurality of agent groups respectively from a set document set according to the plurality of query questions;

[0014] A distribution module, configured to distribute a plurality of reference documents contained in a reference document set corresponding to each agent group to a plurality of agents in the corresponding agent group, so that the plurality of agents input the corresponding query questions and the reference documents distributed to them into a large language model, and obtain initial analysis results corresponding to the plurality of agents, wherein the reference documents distributed to the plurality of agents are different;

[0015] An acquisition module, used to acquire target analysis results corresponding to the plurality of agent groups, wherein the target analysis results corresponding to the target agent group are obtained from initial analysis results corresponding to the plurality of agents in the target agent group;

[0016] The summary module is used to obtain abnormal root cause information based on the target analysis results corresponding to the multiple agent groups.

[0017] In a third aspect, an embodiment of the present invention provides an electronic device, comprising: a memory, a processor, and a communication interface; wherein the memory stores executable code, and when the executable code is executed by the processor, the processor executes the fault root cause locating method as described in the first aspect.

[0018] In a fourth aspect, an embodiment of the present invention provides a non-temporary machine-readable storage medium having executable code stored thereon. When the executable code is executed by a processor of an electronic device, the processor executes the fault root cause locating method as described in the first aspect.

[0019] In a fifth aspect, an embodiment of the present invention provides a computer program product, comprising: a computer program, which, when executed by a processor of an electronic device, causes the processor to execute the fault root cause locating method as described in the first aspect.

[0020] The method for locating the root cause of a fault provided by an embodiment of the present invention generates multiple query questions corresponding to multiple agent groups based on the abnormal operation data corresponding to the target application system and the first prompt words corresponding to the multiple agent groups. Since each agent group focuses on different fault location fields, the query questions corresponding to different agent groups actually tell the corresponding agent groups that the parts that need to be focused on in the abnormal operation data are different. Afterwards, the reference document sets corresponding to the multiple agent groups are retrieved from the set document set according to the multiple query questions, and the multiple reference documents contained in the reference document set corresponding to each agent group are distributed to the multiple agents in each agent group. During the distribution process, by making the multiple agents in the same agent group be distributed with different reference documents, it is ensured that the use of reference documents is richer, the analysis of the root cause of the fault is more comprehensive, and the accuracy of locating the root cause of the fault of the target application system is further ensured.

[0021] After obtaining their respective reference documents, multiple agents input the reference documents and target query questions into the large language model to obtain the initial analysis results corresponding to each of the multiple agents. Then, by integrating the multiple initial analysis results, the target analysis results corresponding to the target agent group that integrate multiple different reference documents can be obtained, and then accurate abnormal root cause information can be obtained based on the target analysis results corresponding to the multiple agent groups. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0023] Figure 1 A flowchart of a method for locating the root cause of a fault provided by an embodiment of the present invention;

[0024] Figure 2 A schematic diagram of an application of a fault root cause locating method provided by an embodiment of the present invention;

[0025] Figure 3 A flowchart of obtaining target analysis results corresponding to multiple agent groups respectively provided in an embodiment of the present invention;

[0026] Figure 4A schematic diagram of an application of a target agent group provided by an embodiment of the present invention;

[0027] Figure 5 A flowchart for retrieving reference document sets corresponding to multiple agent groups provided by an embodiment of the present invention;

[0028] Figure 6 A schematic diagram of the structure of a fault root cause locating device provided by an embodiment of the present invention;

[0029] Figure 7 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of the present invention are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0032] Some embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the case where there is no conflict between the embodiments, the following embodiments and the features in the embodiments can be combined with each other. In addition, the step sequence in the following method embodiments is only an example and not a strict limitation.

[0033] First, the terms or concepts involved in the embodiments of the present invention are explained:

[0034] 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 realize the decomposition of the solution steps of user problems, tool calls and summary of the solution results of multiple steps.

[0035] Large Language Model (LLM): 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.

[0036] The embodiment of the present invention provides a method for locating the root cause of a fault, and improves the accuracy of the root cause locating result through the following ideas: using multiple agent groups corresponding to multiple fault locating fields of the target application system, performing fault analysis in multiple fault locating fields on the operation abnormality data corresponding to the target application system, realizing the preliminary fault analysis of the operation abnormality data of the target application system, and laying the foundation for the subsequent accurate fault root cause locating. In the preliminary fault analysis process, according to the multiple query questions corresponding to the multiple agent groups, the reference document sets corresponding to the multiple agent groups are retrieved from the set document set, and the multiple reference documents contained in the reference document set corresponding to each agent group are distributed to the multiple agents in each agent group. In the distribution process, the reference documents distributed to the multiple agents in the same agent group are different, which ensures that the reference documents are used more abundantly and the fault root cause analysis is more comprehensive. Afterwards, the preliminary fault analysis results obtained by the multiple agents in each agent group are integrated to obtain the target fault analysis results corresponding to each agent group, and then the target fault analysis results corresponding to the multiple agent groups are summarized to obtain the final abnormal root cause information.

[0037] Figure 1 The present invention provides a flowchart of a method for locating the root cause of a fault. The method can be executed by a system for locating the root cause of a fault. The system for locating the root cause of a fault includes the following multiple agent groups. Figure 1 As shown, the method comprises the following steps:

[0038] 101. Based on the operation abnormality data corresponding to the target application system and the first prompt words corresponding to the multiple agent groups, multiple query questions corresponding to the multiple agent groups are generated, wherein the multiple agent groups correspond one-to-one to the multiple fault location fields and the multiple query questions of the target application system, and the first prompt word is used to prompt the corresponding agent group to perform fault analysis in the corresponding fault location field.

[0039] 102. According to multiple query questions, reference document sets corresponding to multiple agent groups are retrieved from the set document set respectively.

[0040] 103. Distribute the multiple reference documents contained in the reference document set corresponding to each agent group to the multiple agents in the corresponding agent group, so that the multiple agents input the corresponding query questions and the reference documents distributed to them into the large language model to obtain the initial analysis results corresponding to the multiple agents, wherein the reference documents distributed to the multiple agents are different.

[0041] 104. Obtain target analysis results corresponding to the plurality of agent groups respectively, wherein the target analysis results corresponding to the target agent group are obtained from initial analysis results corresponding to the plurality of agents in the target agent group respectively.

[0042] 105. Obtain abnormal root cause information based on target analysis results corresponding to multiple agent groups.

[0043] It should be noted that in the embodiments of the present invention, the target application system can be an application, such as a video application, an e-commerce application, etc., or a distributed system, such as a cloud computing system, a cloud storage system, a big data analysis system, etc., which is not limited here. The operation abnormality data corresponding to the target application system refers to the evaluation data that can define whether the operation status of the target application system is abnormal, such as CPU usage, memory usage, and the number of network connections, etc., which are not listed here one by one. Among them, the number of network connections can be represented by indicators such as the number of requests per second (Queries Per Seconds, referred to as QPS).

[0044] In actual applications, the operation abnormality data corresponding to the target application system can be obtained through the following situations: the user submits an unavailable work order for the target application system service instance, or the monitoring module used to detect data abnormalities in the target application system detects that the target application system has abnormal data. In actual applications, the sources of these abnormal data can be various, such as monitoring some predefined operating status indicators and finding that some indicators are abnormal over a period of time, or analyzing the operating status of a certain period of time from the log file and collecting some corresponding abnormal data, or performing traffic statistics analysis on the data packets transmitted over the network, and treating the distribution of the number of network connections within a period of time when the number of network connections suddenly increases or decreases as abnormal data, etc. The embodiment of the present invention does not limit the process of identifying and collecting abnormal data.

[0045] In the specific implementation, multiple query questions corresponding to the multiple agent groups are generated according to the acquired abnormal operation data and the first prompt words corresponding to the multiple agent groups. Figure 2 To illustrate: Figure 2There are three agent groups, namely: CPU agent group (the fault location field corresponding to this agent group is the CPU field), memory agent group (the fault location field corresponding to this agent group is the memory field) and IO traffic agent group (the fault location field corresponding to this agent group is the network connection field). Taking the memory agent group as an example, its corresponding first prompt word can be "You are an expert in the field of memory fault location. Please analyze the root cause of the fault from the perspective of the memory fault location field." Based on this, the above first prompt word and the operation abnormality data are combined to generate the query question corresponding to the memory agent group. It should be understood that the first prompt words corresponding to the CPU agent group and the IO traffic agent group are similar to the first prompt words corresponding to the memory agent group, and will not be repeated here, as long as they can prompt the corresponding agent group to perform fault analysis in the corresponding fault location field.

[0046] like Figure 2 As shown, after obtaining the query questions corresponding to each of the multiple agent groups in the above manner, based on each query question, a search is performed in the set document set to obtain a reference document set corresponding to each agent group (the reference document set includes reference document 1, reference document 2, reference document 3, etc.), wherein the set document set stores reference documents related to multiple fault location fields. The following takes any target agent group among the multiple agent groups as an example to specifically explain the application process of the reference document set:

[0047] It should be understood that the reference document set corresponding to the target agent group includes multiple reference documents, and the target agent group includes multiple agents. In practical applications, these multiple reference documents can be randomly sent to the multiple agents to ensure that the reference documents sent to the multiple agents are different. Afterwards, the multiple agents in the target agent group input the target query problem (the query problem corresponding to the target agent group is called the target query problem) and the reference documents distributed to each of them into the large language model, and the initial analysis results corresponding to each of the multiple agents can be obtained, and then the initial analysis results corresponding to each of the multiple agents are integrated (the specific integration process can be found in the subsequent embodiments, which will not be introduced here for the time being), and the target analysis results corresponding to the target agent group are obtained.

[0048] By adopting the same method as above, the target analysis results corresponding to all agent groups are obtained, and these target analysis results are summarized by the summary agent to obtain the final abnormal root cause information. Optionally, the abnormal root cause information indicates the root cause object and the specific abnormal root cause that cause the abnormal operation of the target application system. The root cause object can be the target application system or the user, and for the specific abnormal root cause, for example, the abnormal root cause can be "the target application system service instance is closed due to insufficient memory, and insufficient memory causes the node to fail over."

[0049] In addition, it should be noted that, since multiple agent groups are used in the above-mentioned root cause location process, if a corresponding large language model is equipped for each agent group, resource consumption will increase and the cost will be high. In view of this, the embodiment of the present invention enables multiple agent groups to share a large language model. By enabling multiple agent groups to share a large language model, resource consumption is reduced and the cost is low. It can be understood that for different agent groups, since the first prompt words contained in the corresponding query questions are different, the reference document sets obtained are different, so even if the same large language model is used, the output results obtained by each agent group are different. For different agents in the same agent group, since the reference documents distributed to these agents are different, even if the same large language model is used, the output initial analysis results are different, so different agents can obtain different initial analysis results based on the use of different reference documents, which increases the diversity of the analysis results and is conducive to a more comprehensive analysis of the root cause of the fault.

[0050] Based on the above, the fault root cause location method provided by the embodiment of the present invention, since multiple agent groups correspond to multiple fault location fields one by one, therefore, by using multiple agent groups to perform fault root cause location operations, fault root cause analysis can be performed from multiple fault location fields, thereby improving the accuracy of fault root cause location. In the specific fault root cause analysis process, the reference document sets corresponding to the multiple agent groups are retrieved from the set document set according to the multiple query questions corresponding to the multiple agent groups, and the multiple reference documents contained in the reference document set corresponding to each agent group are distributed to the multiple agents in each agent group. In the distribution process, the reference documents distributed to the multiple agents in the same agent group are different, ensuring that the use of reference documents is more abundant, the fault root cause analysis is more comprehensive, and the accuracy of the subsequent target application system root cause location is further guaranteed. Afterwards, the preliminary fault analysis obtained based on the multiple agents in each agent group can be summarized and processed accordingly to obtain the final abnormal root cause information. The entire fault root cause location process is automatically carried out, with high work efficiency, and there is no need to accumulate a large amount of historical data in advance, and the data quality requirements are not high.

[0051] Figure 3 A flowchart of obtaining target analysis results corresponding to multiple agent groups provided by an embodiment of the present invention, such as Figure 3 As shown, obtaining the target analysis results corresponding to the multiple agent groups specifically includes the following steps:

[0052] 301. According to the similarity between the initial analysis results corresponding to the multiple agents in the target agent group, the multiple agents are divided into groups to obtain multiple groups, wherein the similarity between the initial analysis results corresponding to the agents in the same group is greater than a set threshold.

[0053] 302. Control the agents in the same group to summarize the initial analysis results corresponding to the agents in the same group through a cooperative integration method to obtain a first intermediate analysis result corresponding to each group.

[0054] 303. Control the intelligent agents in different groups to identify error information of the first intermediate analysis results corresponding to different groups through a debate-style integration method, so as to obtain the second intermediate analysis result corresponding to each group.

[0055] 304. Determine the target analysis result corresponding to the target agent group based on the second intermediate analysis results corresponding to each of the multiple groups.

[0056] In actual applications, since the reference documents distributed to multiple agents in the target agent group are different, the initial analysis results corresponding to the multiple agents obtained based on the reference documents distributed to each agent are also different. By integrating these different initial analysis results, the target analysis results corresponding to the target agent group can be obtained. The specific steps for obtaining the target analysis results corresponding to the target agent group are as follows:

[0057] S1. Determine the similarity between the initial analysis results corresponding to the multiple agents, and divide the multiple agents into groups according to the similarity to obtain multiple groups, so that the similarity between the initial analysis results corresponding to the agents in the same group is greater than a set threshold. Figure 4 Let's take an example:

[0058] exist Figure 4In the example, it is assumed that the target agent group is an IO flow agent group, which contains 4 IO flow agents, namely: the first IO flow agent, the second IO flow agent, the third IO flow agent, and the fourth IO flow agent. After obtaining the similarity between the initial analysis results corresponding to the above 4 IO flow agents, it is found that the similarity between the initial analysis results corresponding to the first IO flow agent and the second IO flow agent is greater than the set threshold (assuming that the set threshold is 80%), and the similarity between the initial analysis results corresponding to the third IO flow agent and the fourth IO flow agent is also greater than 80%, and the similarity between the initial analysis results corresponding to the first IO flow agent and the third IO flow agent and the fourth IO flow agent is less than 80%, and the similarity between the initial analysis results corresponding to the second IO flow agent and the third IO flow agent and the fourth IO flow agent is also less than 80%. At this time, the first IO flow agent and the second IO flow agent can be divided into the same group, which is called group A here, and the third IO flow agent and the fourth IO flow agent can be divided into the same group, which is called group B here. Among them, the method of calculating the similarity between the two initial analysis results can be to use an encoding model to encode the two initial analysis results into word vectors respectively, calculate the distance between the word vector encoding results (such as cosine distance, etc.), and use the distance calculation result as the similarity, but not limited to this.

[0059] S2. Control the agents in the same group to summarize the initial analysis results corresponding to each agent in the same group through a cooperative integration method to obtain the first intermediate analysis result corresponding to each group. Since the initial analysis results corresponding to the agents in the same group are highly similar, a cooperative integration method can be used to summarize the initial analysis results corresponding to each agent in the same group to obtain the first intermediate analysis result corresponding to each group. Figure 4 For example, the initial analysis results corresponding to the first IO traffic agent and the second IO traffic agent can be summarized to obtain the first intermediate analysis result corresponding to group A; the initial analysis results corresponding to the third IO traffic agent and the fourth IO traffic agent can be summarized to obtain the first intermediate analysis result corresponding to group B.

[0060] In a specific implementation, the second prompt word and the initial analysis results corresponding to other agents in the first group are input to any agent in the first group, so that any agent outputs the first intermediate analysis result corresponding to the first group through the large language model. The first group is any one of the multiple groups, and the second prompt word is used to prompt any agent to summarize the initial analysis results corresponding to each agent in the first group.

[0061] For example, assuming that the first group is the above-mentioned group A, the second prompt word and the initial analysis results corresponding to other agents in group A are input to any agent in group A (assuming the first IO traffic agent). The second prompt word can be "Your initial analysis results are similar to the initial analysis results corresponding to other agents in this group. Please combine your initial analysis results and summarize the initial analysis results corresponding to each agent in this group to generate the first intermediate analysis result corresponding to this group."

[0062] By inputting the second prompt word and the initial analysis results corresponding to other intelligent agents in the first group to any intelligent agent in the first group, it is possible to summarize the initial analysis results corresponding to multiple intelligent agents in the first group, obtain more comprehensive initial analysis results within the group, and lay the foundation for subsequent inter-group interactions.

[0063] S3, control the agents in different groups to identify the error information of the first intermediate analysis results corresponding to different groups through a debate-style integration method, so as to obtain the second intermediate analysis results corresponding to each group. Since the similarity between the initial analysis results corresponding to the agents in different groups is low, a debate-style integration method can be used to identify the difference information of the first intermediate analysis results corresponding to different groups, so as to obtain the second intermediate analysis results corresponding to each group. For ease of understanding, continue with Figure 4 For example, after obtaining the first intermediate analysis result corresponding to group A and the first intermediate analysis result corresponding to group B, any agent in group A performs difference information identification on the first intermediate analysis results corresponding to group A and group B respectively (at this time, mainly based on the first intermediate analysis result of group A, the error information in the first intermediate analysis result of group B is determined), and the second intermediate analysis result corresponding to group A is obtained; and any agent in group B performs difference information identification on the first intermediate analysis results corresponding to group A and group B respectively, and the second intermediate analysis result corresponding to group B is obtained (at this time, mainly based on the first intermediate analysis result of group B, the error information in the first intermediate analysis result of group A is determined).

[0064] In a specific implementation, a third prompt word and a first intermediate analysis result corresponding to the second group are input to any agent in the first group, so that any agent outputs a second intermediate analysis result corresponding to the first group through a large language model. The third prompt word is used to prompt any agent to determine error information in the first intermediate analysis result of the second group based on the first intermediate analysis result of the first group, and generate a second intermediate analysis result corresponding to the first group based on the error information, and the first group and the second group are any two of the multiple groups.

[0065] For example, assume that the first group is the aforementioned group A, and the second group is the aforementioned group B. Input the third prompt word and the first intermediate analysis result corresponding to group B to any agent in group A (assuming the first IO traffic agent), and the third prompt word may be "The first intermediate analysis result corresponding to your group is different from the first intermediate analysis results corresponding to other groups. Please point out the errors in the first intermediate analysis results corresponding to other groups, and generate your corresponding second intermediate analysis result based on the errors."

[0066] By inputting the third prompt word and the first intermediate analysis result corresponding to the second group to any intelligent agent in the first group, it is possible to achieve a dispute between different groups, ensuring that more accurate intermediate analysis results can be obtained later, thereby ensuring the accuracy of the entire fault root cause location work.

[0067] S4. Determine the target analysis result corresponding to the target agent group according to the second intermediate analysis results corresponding to each of the multiple groups. It should be understood that since the target agent group includes multiple agents, and the multiple agents are divided into multiple groups, after the multiple groups generate the corresponding second intermediate analysis results, in order to determine the target analysis result corresponding to the final target agent group, it is necessary to select from the multiple second intermediate analysis results. The specific selection process is as follows:

[0068] The second intermediate analysis results, fourth prompt words and operation abnormality data corresponding to each of the multiple groups are input into the above-mentioned large language model to obtain the selection result output by the large language model, and the second intermediate analysis result corresponding to the selection result is determined to be the target analysis result corresponding to the target intelligent agent group. The fourth prompt word is used to prompt the large language model to select the second intermediate analysis results corresponding to each of the multiple groups based on the operation abnormality data.

[0069] In actual applications, assuming that the second intermediate analysis result corresponding to Group A is XXX and the second intermediate analysis result corresponding to Group B is YYY, then the fourth prompt word can be "You are an expert in the field of QPS fault location. The current problem phenomenon is {abnormal operation data}. Please select the correct option from the following options: A, {XXX}; B, {YYY}". In specific implementation, the second intermediate analysis results, fourth prompt words and abnormal operation data corresponding to Group A and Group B are respectively input into the large language model, and the selection result output by the large language model (such as selecting A) can be obtained.

[0070] By inputting the second intermediate analysis results, fourth prompt words and operation abnormality data corresponding to multiple groups into the large language model to obtain the selection results output by the large language model, it is possible to automatically make a selection among the second intermediate analysis results corresponding to multiple groups, thereby improving work efficiency while ensuring the accuracy of the selection.

[0071] Based on the above, the embodiment of the present invention can divide multiple agents into groups according to the similarity between the initial analysis results corresponding to each of the multiple agents in the target agent group to obtain multiple groups, and control the agents in the multiple groups to cooperate within the groups and debate between the groups, so as to ensure the comprehensiveness and accuracy of the target analysis results corresponding to the target agent group finally obtained.

[0072] For ease of understanding, the solution of the present invention is described below by taking the target application system as a cloud computing system as an example:

[0073] Assume that there are three agent groups, namely: CPU agent group, memory agent group and IO traffic agent group, and the abnormal operation data of the cloud computing system is abnormal memory usage.

[0074] During specific implementation, a query question is generated based on the abnormal operation data (abnormal memory usage) and the first prompt word corresponding to the CPU agent group (you are an expert in the field of CPU fault location, please analyze the root cause of the fault from the perspective of CPU fault location).

[0075] Based on the query question, a search is performed in the set document set to obtain a reference document set corresponding to the CPU agent group, which includes reference document 1, reference document 2, reference document 3, ... reference document 10. Reference document 1, reference document 2, reference document 3, ... reference document 10 are randomly sent to the four CPU agents (including the first CPU agent, the second CPU agent, the third CPU agent, and the fourth CPU agent) included in the CPU agent group. Specifically:

[0076] Reference document 1, reference document 3, and reference document 4 are sent to the first CPU agent, and the first CPU agent inputs the query question and reference document 1, reference document 2, and reference document 3 into the large language model to obtain a first initial analysis result.

[0077] Reference document 2, reference document 7, and reference document 9 are sent to the second CPU agent, and the second CPU agent inputs the query question and reference document 2, reference document 7, and reference document 9 into the large language model to obtain a second initial analysis result.

[0078] The reference documents 5 and 6 are sent to the third CPU agent, and the third CPU agent inputs the query question and the reference documents 5 and 6 into the large language model to obtain a third initial analysis result.

[0079] The reference documents 8 and 10 are sent to the fourth CPU agent, and the fourth CPU agent inputs the query question and the reference documents 8 and 10 into the large language model to obtain a fourth initial analysis result.

[0080] Afterwards, the similarities between the first initial analysis result, the second initial analysis result, the third initial analysis result and the fourth initial analysis result are calculated, and it is obtained that: the similarity between the first initial analysis result and the second initial analysis result is greater than the set threshold, the similarity between the third initial analysis result and the fourth initial analysis result is also greater than 80%, and the similarity between the first initial analysis result and the third initial analysis result and the fourth initial analysis result is less than 80%, and the similarity between the second initial analysis result and the third initial analysis result and the fourth initial analysis result is also less than 80%. At this time, the first CPU agent and the second CPU agent are divided into the same group, which is referred to as Group A, and the third CPU agent and the fourth CPU agent are divided into the same group, which is referred to as Group B.

[0081] Input the second prompt word to the first CPU agent in group A (your initial analysis result is similar to the initial analysis results of other agents in this group. Please combine your initial analysis result to summarize the initial analysis results of each agent in this group to generate the first intermediate analysis result of this group). The first CPU agent inputs the second prompt word and the second initial analysis result into the large language model to obtain the first intermediate analysis result of group A. In the same way, the first intermediate analysis result of group B is obtained.

[0082] Input the third prompt word (the first intermediate analysis result corresponding to your group is different from the first intermediate analysis results corresponding to other groups. Please point out the errors in the first intermediate analysis results corresponding to other groups and generate your second intermediate analysis result based on the errors) and the first intermediate analysis result corresponding to group B to the first CPU agent in group A. The first CPU agent inputs the third prompt word and the first intermediate analysis result corresponding to group B into the large language model to obtain the second intermediate analysis result corresponding to group A (assuming XXX). In the same way, the second intermediate analysis result corresponding to group B (assuming YYY) is obtained.

[0083] Input the second intermediate analysis results, the fourth prompt word (You are an expert in the field of memory usage fault location, the current problem phenomenon is {operation abnormal data}, please select the correct option from the following options: A, {XXX}; B, {YYY}) and the operation abnormal data corresponding to Group A and Group B into the large language model, and obtain the selection result output by the large language model (such as A). At this point, it can be determined that the target analysis result corresponding to the CPU agent group is XXX.

[0084] Based on the same method as above, the target analysis results corresponding to the memory agent group (assuming ZZZ) and the target analysis results corresponding to the IO traffic agent group (assuming CCC) are obtained respectively. Afterwards, the summary agent summarizes "XXX, ZZZ, CCC" to obtain the final abnormal root cause information, such as "system parameter configuration error".

[0085] Figure 5 A flowchart of retrieving reference document sets corresponding to multiple agent groups provided by an embodiment of the present invention, such as Figure 5 As shown, according to multiple query questions, reference document sets corresponding to multiple agent groups are retrieved from the set document set respectively, which specifically includes the following steps:

[0086] 501. Input the target query problem corresponding to the target agent group into the embedding model to obtain a first embedding vector corresponding to the target query problem.

[0087] 502. Obtain a second embedding vector corresponding to each document in a set document set through an embedding model.

[0088] 503. Determine a reference document set corresponding to the target agent group based on the similarity between the first embedding vector and the second embedding vector corresponding to each document in the set document set.

[0089] In specific implementation, the target query corresponding to the target agent group is input into the trained embedding model so that the embedding model outputs the first embedding vector corresponding to the target query, and then each document in the set document set is input into the trained embedding model so that the embedding model outputs the second embedding vector corresponding to each document. After that, the similarity between the first embedding vector and the second embedding vector corresponding to each document in the set document set is determined, and the documents with similarity greater than the set similarity threshold (such as 80%) are taken as the documents corresponding to the target agent group to form a reference document set.

[0090] By setting an embedding model, and inputting the target query questions corresponding to the target agent group and each document in the set document set into the trained embedding model, a first embedding vector and a second embedding vector are obtained, and then the similarity between the first embedding vector and the second embedding vector corresponding to each document in the set document set is compared to obtain documents more relevant to the target agent group, thereby ensuring the target agent's ability to answer questions and laying the foundation for the subsequent use of the target agent to locate the root cause more accurately. The embedding model may be a BERT model, etc., but is not limited thereto.

[0091] The following is a detailed description of the training process of the embedding model from three aspects based on three losses:

[0092] First, the training process of the embedding model includes:

[0093] Obtaining a first question sample, and performing keyword extraction on the first question sample to obtain corresponding keywords;

[0094] Inputting the first question sample, the keyword, and a plurality of documents in the set document set into the embedding model respectively, obtaining a third embedding vector corresponding to the first question sample, a fourth embedding vector corresponding to the keyword, and a fifth embedding vector corresponding to each of the plurality of documents;

[0095] Determine a first document retrieval probability distribution corresponding to the first question sample according to the third embedding vector and fifth embedding vectors corresponding to each of the plurality of documents, and determine a second document retrieval probability distribution corresponding to the keyword according to the fourth embedding vector and fifth embedding vectors corresponding to each of the plurality of documents;

[0096] Determining a first loss with the goal of making the first document retrieval probability distribution equal to the second document retrieval probability distribution;

[0097] The embedding model is trained according to the first loss.

[0098] In practical applications, when retrieving documents from a set of documents according to questions, not all words in the question are equally important. In most cases, the keywords in the question can directly convey the semantic information of the question. At the same time, there are many words in the question that have no practical meaning. If these words with no practical meaning are used for document retrieval, it will burden the document retrieval work. Therefore, in the process of document retrieval, in order to improve the retrieval efficiency, it is very important to train the embedding model based on keywords. In the specific embedding model training process, if keywords are manually annotated in multiple questions, it will increase human resource consumption and labor costs. In this regard, a large language model can be used to annotate keywords to reduce labor costs and improve work efficiency.

[0099] For ease of understanding, the process of determining the first loss is explained below with the formula:

[0100] First of all, it should be noted that the first question sample is any one of the preset question sample sets, which is different from the query question formed based on the actually collected operation abnormality data in the above text.

[0101] In a specific implementation, the first question sample, keywords and multiple documents in a set document set are respectively input into the embedding model to obtain a third embedding vector corresponding to the first question sample, a fourth embedding vector corresponding to the keyword and fifth embedding vectors corresponding to each of the multiple documents.

[0102] Afterwards, the third embedding vector is calculated for similarity with the fifth embedding vector corresponding to each of the multiple documents to obtain the similarity calculation results between the first question sample and each document. These similarity calculation results form a document retrieval probability distribution, namely, the first document retrieval probability distribution Q(x). Similarly, the fourth embedding vector is calculated for similarity with the fifth embedding vector corresponding to each of the multiple documents to obtain the similarity calculation results between the keyword and each document. These similarity calculation results form a document retrieval probability distribution, namely, the second document retrieval probability distribution K(x). It should be understood that in order to make the first question sample and the keyword have the same document query effect, the first document retrieval probability distribution Q(x) and the second document retrieval probability distribution K(x) can be made equal as the goal to determine the first loss L distill :

[0103]

[0104] in, Refers to the probability distribution of document retrieval based on keywords. It refers to the document retrieval probability distribution based on the first question sample, and x is each document in the set document set.

[0105] After determining the first loss, the embedding model can be trained according to the first loss. By training the embedding model according to the first loss, the ability of the embedding model to retrieve documents based on keywords is improved, so that the retrieval efficiency is improved in the process of document retrieval based on keywords. It should be noted that in the inference stage of using the embedding model, it is not necessary to use the large language model to perform keyword recognition on the input query question before using the embedding model, but the query question can be directly input into the embedding model. Through the above training, the embedding model has the ability to recognize keywords.

[0106] Secondly, the training process of the embedding model also includes:

[0107] Obtain a second question sample, a first document, and a second document, wherein the first document is a document related to the second question sample, and the second document is a document unrelated to the second question sample;

[0108] Performing contrastive learning based on the second question sample, the first document, and the second document to obtain a first contrastive learning loss term;

[0109] Obtain a third document, a third question sample, and a fourth question sample, wherein the third question sample is a question related to the third document, and the fourth question sample is a question unrelated to the third document;

[0110] Perform contrastive learning based on the third document, the third question sample, and the fourth question sample to obtain a second contrastive learning loss item;

[0111] Determining a second loss based on the first contrastive learning loss term and the second contrastive learning loss term;

[0112] The embedding model is trained according to the second loss.

[0113] It should be understood that in order to accurately complete document retrieval, it is necessary to construct the relevance between questions and documents. By constructing the relevance between questions and documents, it can be ensured that the retrieved documents can answer the corresponding questions. In view of this, in a second aspect, an embodiment of the present invention trains an embedding model based on the relevance between questions and documents.

[0114] For ease of understanding, the following example illustrates the training process of the embedding model:

[0115] First, considering the semantic differences of words in different sentences, the similarity between the question and the document is calculated at the word level to obtain the relevance score of the question and the document. The specific formula is as follows:

[0116]

[0117] in, represents the jth word in question Q, Represents the kth word in document i.

[0118] Furthermore, in order to make the similarity between the question and the relevant documents much greater than the similarity between the question and the irrelevant documents, contrastive learning can be used to distinguish the relevant documents from the irrelevant documents. In the specific implementation, two triples are constructed, namely: {Second question sample Q, first document D + , the second document D -}, and {third document D, third question sample Q + , the fourth question sample Q -}, wherein the first document is a document related to the second question sample, the second document is a document unrelated to the second question sample, the third question sample is a question related to the third document, and the fourth question sample is a question unrelated to the third document.

[0119] Based on the second question sample, the first document and the second document, comparative learning is performed to obtain a first comparative learning loss term. Based on the third document, the third question sample and the fourth question sample, comparative learning is performed to obtain a second comparative learning loss term. The above two triples are optimized according to the first comparative learning loss term and the second comparative learning loss term, so that the score of the question and the relevant documents is high, and the score of the question and the irrelevant documents is low, so that the document containing the answer to the question is retrieved according to the relevance score. The second loss term The specific calculation formula is as follows:

[0120]

[0121] in, is the second loss, which can be a cross entropy loss, is the first contrastive learning loss term, is the second contrastive learning loss term, represents the relevance score between the third document D and the third question sample Q+, represents the relevance score between the third document D and the fourth question sample Q-, represents the relevance score between the second question sample Q and the first document D+, It represents the relevance score between the second question sample Q and the first document D-. ‌exp refers to the exponential function with the mathematical constant e (‌approximately equal to 2.718)‌ as the base.

[0122] After determining the second loss, the embedding model can be trained according to the second loss. By training the embedding model according to the second loss, it is ensured that there is sufficient relevance between the question and the document, so that the answer corresponding to the question can be accurately found in the document later, further improving the ability of the embedding model to perform document retrieval, while ensuring the retrieval accuracy and improving the retrieval efficiency.

[0123] Thirdly, the training process of the embedding model also includes:

[0124] Obtain embedding vectors of multiple documents in a set of documents through an embedding model;

[0125] Initializing a plurality of cluster center vectors, wherein some of the cluster center vectors are initialized based on clustering results of embedding vectors of a set number of documents in a set document set that respectively correspond to a plurality of fault location fields;

[0126] For any document among the multiple documents, according to the distances between the embedding vector of any document and the multiple cluster center vectors, determine the target cluster center vector to which any document belongs, and update the target cluster center vector with the embedding vector of any document;

[0127] Determine a third loss with the goal of minimizing the distance between the embedding vectors of each document belonging to the same updated cluster center vector and the same updated cluster center vector;

[0128] The embedding model is trained according to the third loss.

[0129] In actual applications, when performing document retrieval, the relationship between documents is often ignored. For example, some documents may be obtained by segmenting a large document, and different large documents may have some similar parts. In order to more accurately select the required documents in the set document set later, the embodiment of the present invention trains the embedding model based on the relationship between documents in a prototype learning manner.

[0130] For ease of understanding, the following example illustrates the training process of the embedding model:

[0131] Assuming that 10 cluster center vectors (also called prototypes) are preset, these 10 cluster center vectors need to be initialized first. Optionally, the specific steps for initializing these 10 cluster center vectors are as follows:

[0132] Assuming that the multiple fault location fields set above include: CPU fault location field, memory fault location field and network connection fault location field, multiple documents corresponding to each fault location field can be selected from the set document set, for example, 5 documents, that is, 5 corresponding documents are selected for CPU usage; 5 corresponding documents are selected for memory usage; and 5 corresponding documents are selected for the number of network connections. Afterwards, the cluster center vector corresponding to each fault location field is initialized with the embedding vectors corresponding to the 5 documents corresponding to each fault location field.

[0133] For ease of understanding, the following is a specific explanation of the initialization of the cluster center vector corresponding to the CPU usage: the five documents corresponding to the CPU usage are input into the embedding model respectively, and the five embedding vectors output by the embedding model are obtained. These five embedding vectors are clustered using a certain set clustering algorithm to obtain the clustering results as the initialization results of the corresponding cluster center vectors. Similarly, the cluster center vectors corresponding to the other two fault location fields can also be initialized in the above manner. In this way, the initialization of three cluster center vectors is completed, and the remaining seven cluster center vectors of the ten cluster center vectors can be directly randomly initialized.

[0134] After obtaining the embedding vectors of multiple documents in the set document set through the embedding model and initializing multiple cluster center vectors, the target cluster center vector to which any document belongs is determined according to the distance (i.e., similarity) between the embedding vector of any document in the multiple documents and the multiple cluster center vectors, and the target cluster center vector is updated with the embedding vector of the any document. For example, assuming that the target cluster center vector previously corresponds to 5 documents, and after obtaining the distance between the embedding vector of any document and the multiple cluster center vectors, it is found that the distance between the embedding vector of the any document and the target cluster center vector is smaller, i.e., the similarity is greater. At this time, it can be determined that the any document belongs to the target cluster center vector among the multiple cluster center vectors, and the target cluster center vector can be updated with the embedding vector of the any document. At this time, the target cluster center vector corresponds to 6 documents. Among them, the updating process is to use the above-mentioned clustering algorithm to re-cluster the embedding vectors of the 6 documents belonging to the same target distance center vector at this time, and the clustering result is used as the new target cluster center vector.

[0135] Based on the above process, each document is traversed one by one, the cluster center vector to which it belongs is determined, and the corresponding cluster center vector is updated. After traversing all documents, the third loss L can be determined based on the goal of "minimizing the distance between the embedding vectors of each document belonging to the final updated target cluster center vector and the final updated target cluster center vector". cluster :

[0136]

[0137] Where I represents the exponential function, N represents the number of documents, represents the i-th document, Represents the j-th cluster center vector. Indicates the distance between the two. The meaning of the part is: when the distance between the i-th document and the j-th cluster center vector is the smallest, the i-th document belongs to the j-th cluster center vector, and the indicator function value is 1.

[0138] Among them, it is assumed that m documents are finally determined to belong to Yes, then at this time is the cluster center vector updated after the mth document is added. The above goal is to make the embedding vectors of these m documents consistent with the final updated one The distance is the smallest.

[0139] After determining the third loss, the embedding model can be trained according to the third loss. By training the embedding model according to the third loss, aggregation between similar documents is achieved, so that the document corresponding to the question can be found quickly and accurately in the set document set in the future, further improving the ability of the embedding model to perform document retrieval.

[0140] It is understandable that when actually training the embedding model, the total loss can be the cumulative sum of the above three losses.

[0141] The following will describe in detail the fault root cause location device of one or more embodiments of the present invention. Those skilled in the art will appreciate that these devices can be configured using commercially available hardware components through the steps taught in this solution.

[0142] Figure 6 A schematic diagram of a fault root cause location device provided by an embodiment of the present invention is shown in FIG. Figure 6 As shown, the device includes: a generating module 11, a retrieving module 12, a distributing module 13, an acquiring module 14 and a summarizing module 15.

[0143] The generation module 11 is used to generate a plurality of query questions corresponding to the plurality of intelligent agent groups according to the operation abnormality data corresponding to the target application system and the first prompt words corresponding to the plurality of intelligent agent groups respectively, wherein the plurality of intelligent agent groups correspond one-to-one to the plurality of fault location fields of the target application system and the plurality of query questions, and the first prompt words are used to prompt the corresponding intelligent agent group to perform fault analysis in the corresponding fault location field.

[0144] The retrieval module 12 is used to retrieve reference document sets corresponding to the multiple agent groups respectively in the set document set according to the multiple query questions.

[0145] The distribution module 13 is used to distribute the multiple reference documents contained in the reference document set corresponding to each agent group to the multiple agents in the corresponding agent group, so that the multiple agents input the corresponding query questions and the respective distributed reference documents into the large language model to obtain the initial analysis results corresponding to the multiple agents, wherein the reference documents distributed to the multiple agents are different.

[0146] The acquisition module 14 is used to obtain the target analysis results corresponding to the multiple agent groups respectively, wherein the target analysis results corresponding to the target agent group are obtained from the initial analysis results corresponding to the multiple agents in the target agent group.

[0147] The summary module 15 is used to obtain abnormal root cause information based on the target analysis results corresponding to the multiple agent groups.

[0148] Optionally, the multiple agent groups share the large language model.

[0149] Among them, optionally, the retrieval module 12 is specifically used to: input the target query question corresponding to the target intelligent agent group into the embedding model to obtain a first embedding vector corresponding to the target query question; obtain the second embedding vector corresponding to each document in the set document set through the embedding model; based on the similarity between the first embedding vector and the second embedding vector corresponding to each document in the set document set, determine the reference document set corresponding to the target intelligent agent group.

[0150] Among them, optionally, the acquisition module 14 is specifically used to: divide the multiple agents in the target agent group into groups to obtain multiple groups according to the similarity between the initial analysis results corresponding to each of the agents in the target agent group, wherein the similarity between the initial analysis results corresponding to each of the agents in the same group is greater than a set threshold; control the agents in the same group to summarize the initial analysis results corresponding to each of the agents in the same group through a cooperative integration method to obtain a first intermediate analysis result corresponding to each group; control the agents in different groups to identify error information for the first intermediate analysis results corresponding to different groups through a debate-style integration method to obtain a second intermediate analysis result corresponding to each group; determine the target analysis result corresponding to the target agent group according to the second intermediate analysis results corresponding to each of the multiple groups.

[0151] Among them, optionally, the acquisition module 14 is also specifically used to: input a second prompt word and the initial analysis results corresponding to other intelligent agents in the first group to any intelligent agent in the first group, so that any intelligent agent outputs a first intermediate analysis result corresponding to the first group through the large language model, the first group is any one of the multiple groups, and the second prompt word is used to prompt any intelligent agent to summarize the initial analysis results corresponding to each intelligent agent in the first group; and input a third prompt word and the first intermediate analysis result corresponding to the second group to any intelligent agent in the first group, so that any intelligent agent outputs a second intermediate analysis result corresponding to the first group through the large language model, and the third prompt word is used to prompt any intelligent agent The intelligent agent determines the error information in the first intermediate analysis result of the second group based on the first intermediate analysis result of the first group, and generates the second intermediate analysis result corresponding to the first group based on the error information, wherein the first group and the second group are any two of the multiple groups; and the second intermediate analysis results corresponding to each of the multiple groups, a fourth prompt word and the operation abnormality data are input into the large language model; to obtain the selection result output by the large language model, wherein the fourth prompt word is used to prompt the large language model to select the second intermediate analysis results corresponding to each of the multiple groups based on the operation abnormality data; and determines that the second intermediate analysis result corresponding to the selection result is the target analysis result corresponding to the target intelligent agent group.

[0152] Optionally, the device further includes: a training module, wherein the training module is used to:

[0153] Obtain a first question sample, and perform keyword extraction on the first question sample to obtain corresponding keywords; input the first question sample, the keywords, and a plurality of documents in a set document set into an embedding model respectively to obtain a third embedding vector corresponding to the first question sample, a fourth embedding vector corresponding to the keywords, and a fifth embedding vector corresponding to each of the plurality of documents; determine a first document retrieval probability distribution corresponding to the first question sample based on the third embedding vector and the fifth embedding vectors corresponding to each of the plurality of documents, and determine a second document retrieval probability distribution corresponding to the keywords based on the fourth embedding vector and the fifth embedding vectors corresponding to each of the plurality of documents; determine a first loss with the goal of making the first document retrieval probability distribution equal to the second document retrieval probability distribution; and train the embedding model based on the first loss.

[0154] Optionally, the training module is also used to: obtain a second question sample, a first document, and a second document, wherein the first document is a document related to the second question sample, and the second document is a document unrelated to the second question sample; perform comparative learning based on the second question sample, the first document, and the second document to obtain a first comparative learning loss term; obtain a third document, a third question sample, and a fourth question sample, wherein the third question sample is a question related to the third document, and the fourth question sample is a question unrelated to the third document; perform comparative learning based on the third document, the third question sample, and the fourth question sample to obtain a second comparative learning loss term; determine a second loss based on the first comparative learning loss term and the second comparative learning loss term; and train the embedding model according to the second loss.

[0155] Optionally, the training module is also used to: obtain embedding vectors of multiple documents in the set document set through the embedding model; initialize multiple cluster center vectors, wherein some cluster center vectors are initialized based on clustering results of embedding vectors of a set number of documents in the set document set corresponding to the multiple fault location fields; for any document in the multiple documents, determine a target cluster center vector to which the any document belongs based on the distances between the embedding vector of the any document and the multiple cluster center vectors, and update the target cluster center vector with the embedding vector of the any document; determine a third loss with the goal of minimizing the distance between the embedding vectors of each document belonging to the same updated cluster center vector and the same updated cluster center vector; and train the embedding model based on the third loss.

[0156] Figure 6 The device shown can execute the steps provided in the aforementioned embodiments. For detailed execution process and technical effects, please refer to the description in the aforementioned embodiments, which will not be repeated here.

[0157] In one possible design, the above Figure 6 The structure of the fault root cause location device shown can be implemented as an electronic device. Figure 7 As shown, the electronic device may include: a processor 21, a memory 22, and a communication interface 23. The memory 22 stores executable codes, and when the executable codes are executed by the processor 21, the processor 21 can at least implement the fault root cause location method provided in the above embodiments.

[0158] In addition, an embodiment of the present invention provides a non-temporary machine-readable storage medium, on which executable code is stored. When the executable code is executed by a processor of an electronic device, the processor can at least implement the fault root cause locating method provided in the aforementioned embodiment.

[0159] In addition, an embodiment of the present invention provides a computer program product, which includes: a computer program, when the computer program is executed by a processor of an electronic device, the processor can at least implement the fault root cause locating method provided in the above embodiment.

[0160] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. Those of ordinary skill in the art may understand and implement the present invention without creative effort.

[0161] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by adding a necessary general hardware platform, and of course can also be implemented by combining hardware and application programs. Based on such an understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a computer product, and the present invention can be in the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0162] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for locating the root cause of a fault, characterized in that: include: Generate multiple query questions corresponding to the multiple agent groups according to the operation abnormality data corresponding to the target application system and the first prompt words respectively corresponding to the multiple agent groups, wherein the multiple agent groups correspond to the multiple fault location fields of the target application system and the multiple query questions one by one, and the first prompt words are used to prompt the corresponding agent group to perform fault analysis in the corresponding fault location field; Retrieving reference document sets corresponding to the multiple agent groups respectively from the set document set according to the multiple query questions; Distributing a plurality of reference documents contained in a reference document set corresponding to each agent group to a plurality of agents in the corresponding agent group, so that the plurality of agents input the corresponding query question and the reference documents respectively distributed into a large language model, and obtain initial analysis results corresponding to the plurality of agents respectively, wherein the reference documents distributed to the plurality of agents are different; Obtaining target analysis results corresponding to the plurality of agent groups, respectively, wherein the target analysis results corresponding to the target agent group are obtained from initial analysis results corresponding to the plurality of agents in the target agent group; The abnormal root cause information is obtained according to the target analysis results corresponding to the multiple agent groups.

2. The method according to claim 1, characterized in that The plurality of agent groups share the large language model.

3. The method according to claim 1, characterized in that The retrieving reference document sets corresponding to the plurality of agent groups respectively from the set document set according to the plurality of query questions comprises: Inputting the target query problem corresponding to the target agent group into the embedding model to obtain a first embedding vector corresponding to the target query problem; Obtaining a second embedding vector corresponding to each document in the set document set through the embedding model; Based on the similarity between the first embedding vector and the second embedding vector corresponding to each document in the set document set, a reference document set corresponding to the target agent group is determined.

4. The method according to claim 1, characterized in that: The obtaining of target analysis results respectively corresponding to the plurality of agent groups comprises: According to the similarity between the initial analysis results corresponding to the multiple agents in the target agent group, the multiple agents are divided into groups to obtain multiple groups, wherein the similarity between the initial analysis results corresponding to the agents in the same group is greater than a set threshold; Controlling the agents in the same group to aggregate the initial analysis results corresponding to the agents in the same group through a cooperative integration method, so as to obtain the first intermediate analysis result corresponding to each group; Control the agents in different groups to identify the error information of the first intermediate analysis results corresponding to the different groups through a debate-style integration method, so as to obtain the second intermediate analysis results corresponding to each group; The target analysis result corresponding to the target agent group is determined according to the second intermediate analysis results corresponding to each of the multiple groups.

5. The method according to claim 4, characterized in that The controlling the agents in the same group to aggregate the initial analysis results corresponding to the agents in the same group in a cooperative integration manner to obtain the first intermediate analysis result corresponding to each group includes: A second prompt word and initial analysis results corresponding to other agents in the first group are input to any agent in the first group, so that the any agent outputs a first intermediate analysis result corresponding to the first group through the large language model, the first group is any one of the multiple groups, and the second prompt word is used to prompt the any agent to summarize the initial analysis results corresponding to each agent in the first group.

6. The method according to claim 4, characterized in that The controlling of the intelligent agents in different groups to identify the difference information of the first intermediate analysis results corresponding to the different groups through a debate-style integration method to obtain the second intermediate analysis results corresponding to each group includes: A third prompt word and a first intermediate analysis result corresponding to the second group are input to any agent in the first group, so that the any agent outputs a second intermediate analysis result corresponding to the first group through the large language model, the third prompt word is used to prompt the any agent to determine error information in the first intermediate analysis result of the second group based on the first intermediate analysis result of the first group, and generate a second intermediate analysis result corresponding to the first group based on the error information, the first group and the second group are any two of the multiple groups.

7. The method according to claim 4, characterized in that Determining the target analysis result corresponding to the target agent group according to the second intermediate analysis results corresponding to each of the plurality of groups includes: Inputting the second intermediate analysis results, the fourth prompt word and the abnormal operation data corresponding to each of the multiple groups into the large language model to obtain a selection result output by the large language model, wherein the fourth prompt word is used to prompt the large language model to select the second intermediate analysis results corresponding to each of the multiple groups based on the abnormal operation data; Determine that the second intermediate analysis result corresponding to the selection result is the target analysis result corresponding to the target agent group.

8. The method according to claim 3, characterized in that The training process of the embedding model includes: Obtaining a first question sample, and performing keyword extraction on the first question sample to obtain corresponding keywords; Inputting the first question sample, the keyword, and a plurality of documents in a set document set into an embedding model respectively, obtaining a third embedding vector corresponding to the first question sample, a fourth embedding vector corresponding to the keyword, and a fifth embedding vector corresponding to each of the plurality of documents; Determine a first document retrieval probability distribution corresponding to the first question sample according to the third embedding vector and fifth embedding vectors corresponding to each of the plurality of documents, and determine a second document retrieval probability distribution corresponding to the keyword according to the fourth embedding vector and fifth embedding vectors corresponding to each of the plurality of documents; Determining a first loss with the goal of making the first document retrieval probability distribution equal to the second document retrieval probability distribution; The embedding model is trained according to the first loss.

9. The method according to claim 8, characterized in that The training process of the embedding model also includes: Acquire a second question sample, a first document, and a second document, wherein the first document is a document related to the second question sample, and the second document is a document unrelated to the second question sample; Performing contrastive learning based on the second question sample, the first document, and the second document to obtain a first contrastive learning loss term; Acquire a third document, a third question sample, and a fourth question sample, wherein the third question sample is a question related to the third document, and the fourth question sample is a question unrelated to the third document; Performing contrastive learning based on the third document, the third question sample, and the fourth question sample to obtain a second contrastive learning loss term; determining a second loss based on the first contrastive learning loss term and the second contrastive learning loss term; The embedding model is trained according to the second loss.

10. The method according to claim 9, characterized in that The training process of the embedding model also includes: Obtaining embedding vectors of multiple documents in the set document set through the embedding model; Initializing a plurality of cluster center vectors, wherein some of the cluster center vectors are initialized based on clustering results of embedding vectors of a set number of documents in the set document set that respectively correspond to the plurality of fault location fields; For any document among the multiple documents, determine a target cluster center vector to which the document belongs according to distances between the embedding vector of the document and the multiple cluster center vectors, and update the target cluster center vector with the embedding vector of the document; Determining a third loss with the goal of minimizing the distance between the embedding vectors of each document belonging to the same updated cluster center vector and the same updated cluster center vector; The embedding model is trained according to the third loss.

11. An electronic device, characterized in that: include: A memory, a processor, and a communication interface; wherein the memory stores executable code, and when the executable code is executed by the processor, the processor executes the fault root cause locating method as described in any one of claims 1 to 10.

12. A non-transitory machine-readable storage medium, characterized in that: The non-transitory machine-readable storage medium stores executable code, and when the executable code is executed by a processor of an electronic device, the processor is caused to execute the fault root cause locating method according to any one of claims 1 to 10.

13. A computer program product, characterized in that include: A computer program, when executed by a processor of an electronic device, causes the processor to execute the fault root cause locating method according to any one of claims 1 to 10.

Images