Configuration item attribute alignment method based on large language model

Abstract

The application provides a configuration item attribute alignment method based on a large language model, and relates to the technical field of data processing.The method comprises the following steps: identifying configuration item attributes based on a large language model, and obtaining an identification result; in the case where the attribute values of two configuration item attributes are the same and the attribute names are different, determining an alignment result of the two configuration item attributes according to the identification results corresponding to the two configuration item attributes respectively.The application makes the configuration item attribute alignment process intelligent, and improves the configuration item attribute alignment efficiency.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for aligning configuration item attributes based on a large language model. Background Technology

[0002] With the advancement of Internet technology (IT), enterprises are focusing on IT-based operations and maintenance. In situations with numerous IT systems, building a Configuration Management Database (CMDB) can better facilitate enterprise operations and maintenance, and a well-developed CMDB can also provide data services to other systems.

[0003] Because various participants have different needs for CMDB, the data quality varies greatly. Organizing CMDB data into a knowledge graph can help improve data quality. The specific process can be as follows: Data automatically collected from different data sources is processed to obtain configuration item data; configuration item attributes are aligned; and relationships between configuration items are established based on the alignment results, thereby constructing an operations and maintenance knowledge graph. However, in related technologies, the alignment of configuration item attributes is not very efficient. Summary of the Invention

[0004] This invention provides a configuration item attribute alignment method based on a large language model to address the shortcomings of low configuration item attribute alignment efficiency in related technologies, thereby improving the efficiency of configuration item attribute alignment.

[0005] This invention provides a configuration item attribute alignment method based on a large language model, comprising:

[0006] The configuration item attributes are identified based on a large language model, and the identification results are obtained.

[0007] When two configuration item attributes have the same attribute value but different attribute names, the alignment result of the two configuration item attributes is determined based on the recognition results corresponding to the two configuration item attributes respectively.

[0008] In some embodiments, the step of identifying configuration item attributes based on a large language model and obtaining the identification result includes:

[0009] Select a prompt word template from the prompt word template library that is related to the configuration item information corresponding to the configuration item attribute;

[0010] Write the configuration item information into the prompt word template to generate a prompt word instance;

[0011] The prompt word instance is identified based on a large language model, and the identification result is obtained.

[0012] In some embodiments, selecting a prompt word template from the prompt word template library that is related to the configuration item information corresponding to the configuration item attribute includes:

[0013] Based on the configuration information and the prompt information, a template type code is generated; the prompt information is used to provide prompts for the large language model.

[0014] Select the prompt word template corresponding to the template type code from the prompt word template library; one template type code corresponds to one prompt word template.

[0015] In some embodiments, after obtaining the recognition result, the method further includes:

[0016] Detect whether the recognition result meets the preset rules;

[0017] If the recognition result does not meet the preset rule, check whether the number of recognitions of the large language model exceeds the threshold.

[0018] If the number of recognition attempts by the large language model does not exceed the threshold, the configuration item attribute is re-recognized based on the large language model, and the recognition result is obtained again.

[0019] In some embodiments, the step of re-identifying the configuration item attributes based on a large language model and re-obtaining the identification result includes:

[0020] Update the prompt information; the prompt information is used to provide prompts to the large language model;

[0021] Based on the configuration item information and the updated prompt information, the template type code is regenerated;

[0022] The prompt word template corresponding to the newly generated template type code is selected again from the prompt word template library; one template type code corresponds to one prompt word template;

[0023] Write the configuration information into the newly selected prompt word template and regenerate the prompt word instance;

[0024] The regenerated prompt word instances are identified based on the large language model, and the identification results are obtained again.

[0025] In some embodiments, the step of identifying configuration item attributes based on a large language model and obtaining the identification result includes:

[0026] Tag the configuration item attributes based on the large language model to obtain a tag set;

[0027] Correspondingly, determining the alignment result of the two configuration item attributes based on the recognition results corresponding to the two configuration item attributes respectively includes:

[0028] Find the intersection of the tag sets corresponding to the two configuration item attributes respectively, and obtain the intersection result;

[0029] If the intersection result is not empty, it is determined that the two configuration item attributes are aligned.

[0030] In some embodiments, prior to identifying configuration item attributes based on a large language model, the method further includes:

[0031] The model parameters in the large language model are initialized; the model parameters include parameters related to random responses and parameters related to context.

[0032] In some embodiments, the prompt word template includes at least one of: background information, prior information, task description information, and adaptation information.

[0033] The configuration item attribute alignment method based on a large language model provided by this invention applies a large language model to the configuration item attribute alignment process. The large language model is used to identify the configuration item attributes. When two configuration item attributes have the same attribute value but different attribute names, the alignment result of the two configuration item attributes is determined according to the identification results corresponding to the two configuration item attributes respectively. This makes the configuration item attribute alignment process intelligent and improves the efficiency of configuration item attribute alignment. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0035] Figure 1 This is one of the flowcharts illustrating the configuration item attribute alignment method based on a large language model provided in this embodiment of the invention;

[0036] Figure 2 This is a schematic diagram illustrating the alignment of configuration item attributes provided by the present invention;

[0037] Figure 3 This is the second flowchart illustrating the configuration item attribute alignment method based on a large language model provided in this embodiment of the invention.

[0038] Figure 4 This is a schematic diagram illustrating the principle of configuration item attribute alignment provided by the present invention;

[0039] Figure 5 This is the third flowchart illustrating the configuration item attribute alignment method based on a large language model provided in this embodiment of the invention. Detailed Implementation

[0040] It should be noted that, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0041] It should be further noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0042] In this invention, "at least one" means one or more, and "more than one" means two or more. The terms "first," "second," "third," "fourth," etc. (if present) in this invention are used to distinguish similar objects, rather than to describe a specific order or sequence.

[0043] In embodiments of the present invention, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" or "for example" in embodiments of the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0044] In related technologies, common methods for aligning configuration item attributes include the following:

[0045] (1) Rule-based configuration item attribute alignment method: This approach relies heavily on expert experience. It depends on manually defined rules, primarily including attribute annotations and attribute correspondences. Its advantage lies in the absence of data labeling, a lower barrier to entry, and suitability for the initial stage. However, the main limitation is the poor scalability of the rules. As configuration items are added and updated, existing attribute alignment rules may become inapplicable, leading to missing connections. Furthermore, as the number of rules increases, the workload for rule maintenance gradually increases, raising the risk of errors.

[0046] (2) Quantitative Alignment Method for Configuration Items: This method treats configuration item attribute names as text sequences and measures similarity through sequence matching to achieve alignment and association. The advantage of this method is that it requires no data labeling, has a low application threshold, and is particularly suitable for the initial stage. However, its main drawback is its limited applicability; it is only suitable for specific objects, i.e., attribute names consisting of the same words, abbreviations, affixes, etc. If multiple naming methods exist (such as synonyms or different abbreviations), the similarity calculation results may have significant deviations.

[0047] (3) Machine Learning-Based Configuration Attribute Alignment Method: This method involves training configuration items and related graph data using machine learning techniques. The data includes various information about the configuration items (e.g., configuration item name, attribute name, attribute value, and type), graph information (e.g., neighbor node set, topology), and text information. The model can be applied to various problems, such as binary classification, clustering, embedding representation, and link prediction. The greatest advantage of this method lies in its excellent generalization ability and high flexibility. Its disadvantage is the high data requirements; a significant amount of time needs to be invested in data preparation in the early stages, and the modeling is quite difficult, making it unsuitable for early-stage implementation.

[0048] To address the shortcomings of low alignment efficiency of configuration item attributes in related technologies, this invention provides a configuration item attribute alignment method based on a large language model, thereby improving the alignment efficiency of configuration item attributes.

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. It should be noted that, unless otherwise specified, the embodiments and features of the embodiments of this invention can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0050] Figure 1 This is one of the flowcharts illustrating the configuration item attribute alignment method based on a large language model provided in this embodiment of the invention, such as... Figure 1 As shown, this invention provides a configuration item attribute alignment method based on a large language model, comprising the following steps:

[0051] Step 110: Identify the configuration item attributes based on the large language model and obtain the identification results.

[0052] Specifically, large language models typically refer to models in the fields of machine learning and artificial intelligence that have a large number of parameters and high computational complexity. These models excel at handling complex tasks, capturing features from large-scale data, and performing advanced semantic understanding.

[0053] Large language models can extract key knowledge points and concepts from a large amount of text data, understand the contextual relationships of text, and thus identify and mine the connections between knowledge. This helps to integrate fragmented information and better organize and utilize domain knowledge.

[0054] A configuration item can have many attributes. The attributes of the configuration item are identified based on a large language model to distinguish each attribute and obtain the identification result.

[0055] Alternatively, the recognition method can be labeling, and the corresponding recognition result can be a set of labels; the recognition method can also be classification, and the corresponding recognition result can be a classification result.

[0056] Step 120: When the attribute values ​​of the two configuration items are the same but the attribute names are different, determine the alignment result of the two configuration item attributes based on the recognition results corresponding to the two configuration item attributes respectively.

[0057] Specifically, a configuration item attribute can have many characteristics, such as attribute value, attribute name, and attribute type. Each characteristic can have multiple candidate instances. For example, the attribute name of configuration item attribute 1 can be selected from attribute name 1, attribute name 2, and attribute name 3.

[0058] Configuration item names and attribute names are mainly composed of proprietary terms, affixes, and abbreviations related to the operations and maintenance field. Operations and maintenance experts can roughly understand their meaning through the names, that is, configuration item names and attribute names have the literal meaning of the operations and maintenance field.

[0059] Table 1. Example table of configuration items, attribute names, and attribute values.

[0060] Configuration items Attribute Name Attribute value WebLogic IP 10.201.82.30 Tomcat IP 10.201.82.30 OracleInstance monitoring_name 10.232.19.11_orapmt1 OracleService orasevice_ip_first 10.232.19.11_orapmt1 ApplicationProcess sys_prs_ip_port 10.246.214.154_10079 ZookeeperInstance monitoring_name 10.246.214.154_10079 appnodes Nodes_IP 10.240.86.231 vm VM_IPAddress 10.240.86.231 commmonCabinet deployarea 1 firewall device height 1

[0061] For example, Table 1 is an example table of configuration items, attribute names, and attribute values. As shown in Table 1, the attribute names and values ​​of the configuration items "WebLogic" and "Tomcat" are the same; the attribute values ​​of the configuration items "OracleInstance" and "OracleService" are the same, but the attribute names are different; the attribute values ​​of the configuration items "ApplicationProcess" and "ZookeeperInstance" are the same, but the attribute names are different; the attribute values ​​of the configuration items "appnodes" and "vm" are the same, but the attribute names are different; the attribute values ​​of the configuration items "commmonCabinet" and "firewall" are the same, but the attribute names are different.

[0062] If the attribute names and values ​​of two configuration items are the same, it means that the two configuration item attributes can be aligned; if the attribute names and values ​​of two configuration item attributes are different, it means that the two configuration item attributes are not aligned.

[0063] When two configuration item attributes have the same attribute value but different attribute names, the recognition results corresponding to the two configuration item attributes are determined. Based on the intersection type between the two recognition results, the alignment result of the two configuration item attributes is determined. If the intersection type is an empty set, it indicates that the two configuration item attributes are not aligned; if the intersection type is not empty, it indicates that the two configuration item attributes can be aligned.

[0064] Figure 2 This is a schematic diagram illustrating the alignment of configuration item attributes provided by the present invention, such as... Figure 2 As shown, attribute A of configuration item 1 and attribute B of configuration item 2 can be aligned. An association edge is established between configuration item 1 and configuration item 2, which represents the alignment relationship between attribute A and attribute B.

[0065] The configuration item attribute alignment method based on a large language model provided in this invention applies a large language model to the configuration item attribute alignment process. The large language model is used to identify the configuration item attributes. When two configuration item attributes have the same attribute value but different attribute names, the alignment result of the two configuration item attributes is determined according to the identification results corresponding to the two configuration item attributes respectively. This makes the configuration item attribute alignment process intelligent and improves the efficiency of configuration item attribute alignment.

[0066] Figure 3 This is the second flowchart illustrating the configuration item attribute alignment method based on a large language model provided in this embodiment of the invention. Figure 3 As shown in the figure, the configuration item attribute alignment method based on a large language model provided in this embodiment of the invention includes the following steps:

[0067] Step 310: Select a configuration item attribute pair.

[0068] Specifically, two configuration item attributes are randomly selected to form a configuration item attribute pair. These two configuration item attributes can be attributes of different configuration items, or different attributes of the same configuration item.

[0069] Step 320, text preprocessing.

[0070] Specifically, text preprocessing is performed on both configuration item attributes. Text preprocessing may include converting uppercase letters to lowercase, removing spaces, removing special characters, and standardizing connectors.

[0071] Step 330: Determine if the attribute values ​​are the same.

[0072] Specifically, determine whether the attribute values ​​of the two configuration items are the same. If the attribute values ​​are the same, proceed to step 340; if the attribute values ​​are different, proceed to step 370.

[0073] Step 340: Determine if the attribute names are the same.

[0074] Specifically, determine whether the attribute names of the two configuration items are the same. If the attribute names are the same, proceed to step 380; if the attribute names are different, proceed to step 350.

[0075] Step 350: Identify configuration item attribute pairs based on the large language model.

[0076] Specifically, the two configuration item attributes are identified based on a large language model, and two identification results are obtained.

[0077] Step 360: Determine if there is any overlap in the recognition results.

[0078] Specifically, determine whether the two recognition results have any overlap. If the two recognition results have overlap, proceed to step 380; if the two recognition results do not have overlap, proceed to step 370.

[0079] Step 370: Do not perform attribute alignment.

[0080] Specifically, attribute alignment is not performed, meaning that no association edge representing the alignment relationship between the attributes of the two configuration items is established between them.

[0081] Step 380: Perform attribute alignment.

[0082] Specifically, attribute alignment is performed, which means establishing an association edge between two configuration items to represent the alignment relationship between the attributes of the two configuration items.

[0083] In some embodiments, configuration item attributes are identified based on a large language model to obtain the identification results, including:

[0084] The configuration item attributes are tagged based on a large language model to obtain a tag set.

[0085] Correspondingly, based on the recognition results corresponding to the two configuration item attributes, the alignment result of the two configuration item attributes is determined, including:

[0086] Find the intersection of the tag sets corresponding to the two configuration item attributes to obtain the intersection result;

[0087] If the intersection result is not empty, it is determined that the two configuration item attributes are aligned.

[0088] Specifically, when the identification method is labeling, the configuration item attributes are labeled based on a large language model to obtain a label set.

[0089] When two configuration item attributes have the same attribute value but different attribute names, obtain the tag sets corresponding to the two configuration item attributes respectively, find the intersection of the two tag sets, and obtain the intersection result.

[0090] If the intersection result is not empty, it is determined that the two configuration item attributes are aligned, that is, an association edge representing the alignment relationship between the two configuration item attributes is established between the two configuration items.

[0091] Figure 4 This is a schematic diagram illustrating the principle of configuration item attribute alignment provided by the present invention, such as... Figure 4 As shown, the large language model labels configuration item attributes based on configuration item name, attribute name, and attribute value, resulting in a tag set. The intersection of two tag sets is calculated. If the intersection is not empty, an association edge representing the alignment relationship between the two configuration item attributes is established.

[0092] For example, the configuration items "appnodes" and "vm" have the same attribute value of 10.240.86.231, but the attribute name of the configuration item "appnodes" is "Nodes_IP" and the attribute name of the configuration item "vm" is "VM_IPAddress", which are different attribute names.

[0093] Tagging the attribute "Nodes_IP" of the configuration item "appnodes" using the large language model yields tag result A; tagging the attribute "VM_IPAddress" of the configuration item "vm" using the large language model yields tag result B.

[0094] Suppose that label result A includes label 1 and label 2, and label result B includes label 1, label 2 and label 3.

[0095] Find the intersection of label result A and label result B. The intersection result is label 1 and label 2. The intersection result is not an empty set. Establish an association edge between configuration item "appnodes" and configuration item "vm" to represent the alignment relationship between configuration item attributes "Nodes_IP" and "VM_IPAddress".

[0096] The configuration item attribute alignment method based on a large language model provided in this invention identifies configuration item attributes by tagging them, which further facilitates the intelligentization of the configuration item attribute alignment process.

[0097] In some embodiments, prior to identifying configuration item attributes based on a large language model, the method further includes:

[0098] Initialize the model parameters in the large language model; the model parameters include parameters related to random responses and parameters related to context.

[0099] Specifically, to limit the randomness of the large language model's responses, the parameters related to random responses in the large language model are initialized before identifying configuration item attributes based on the large language model. Parameters related to random responses include: sampling temperature, random sampling cumulative probability limit (top_p), etc.

[0100] Optionally, the sampling temperature can be set to a small value (default 0-0.2, keep it as small as possible). A lower sampling temperature can make the output of the large language model more deterministic.

[0101] Optionally, the cumulative probability limit (top_p) of random sampling can be set to a small value (default 0.1, the smaller the better). Tokens with a cumulative probability exceeding top_p are randomly sampled as candidate tokens. The fewer tokens sampled, the better.

[0102] To limit the large language model's contextual association and prevent it from considering contextual information when responding, the parameters related to contextual association in the large language model are initialized before identifying configuration item attributes based on the large language model. These parameters include: the number of accompanying historical messages (history_m), the historical summary (historicalSummary), and the dialogue role (role).

[0103] Optionally, the number of attached historical messages can be set to 0, meaning that the number of historical messages carried in each request is 0.

[0104] Optionally, setting the history summary to invalid (false) skips the automatic chat history compression step and does not send context.

[0105] Optionally, the dialogue roles can be set to empty, with no role restrictions and only the current problem scenario considered.

[0106] The configuration item attribute alignment method based on a large language model provided in this invention initializes the parameters related to random answers and the parameters related to context association in the large language model before the large language model identifies the configuration item attributes, thereby limiting the randomness of the large language model's answers and limiting the large language model's context association.

[0107] In some embodiments, configuration item attributes are identified based on a large language model to obtain the identification results, including:

[0108] Select a prompt word template from the prompt word template library that is related to the configuration item information corresponding to the configuration item attribute;

[0109] Write the configuration information into the prompt word template to generate a prompt word instance;

[0110] The system identifies prompt word instances based on a large language model and obtains the identification results.

[0111] Specifically, the prompt word template library is a template library composed of prompt word templates. Prompt word templates are used to enable a large language model to understand that this is an IT operations domain, and the problem it addresses is CMDB configuration item attribute alignment.

[0112] In some embodiments, the prompt template may include at least one of the following: background information, prior information, task description information, and adaptation information. The content of the prompt template can be added or adjusted as needed.

[0113] Background information is used to provide domain-specific context. For example, background information includes identity information, which allows the large language model to role-play and define its own identity. Identity information could be that of an IT operations expert or a senior CMDB operations expert, etc. Background information can also provide relevant task context, such as general solution technology information.

[0114] Prior information is used to provide prior knowledge, enabling large language models to understand the problem domain (providing hints, specifications, and constraints to the large language model). For example, prior information may include label definition information and label examples.

[0115] For example, taking "db" as an example, "represents database" is the label definition, and "db2, sequoiadb, redis, oracle, tidb" are database-related terms (analogous terms).

[0116] For example, taking datacenter as an example, "represents a data center" is the label definition, and "deploydb represents a deployed data center" is the specification.

[0117] The task description is used to organize and describe the specific task. The task description can be set to identify the attributes of the configuration items based on the configuration item information.

[0118] Adaptation information is used to improve the adaptability of large language models, such as their ability to connect to downstream tasks and optimize their understanding capabilities. Other information may include output format and model parameters of the large language model.

[0119] Different prompt templates can include different information types. For example, prompt template 1 includes background information, task description information, and adaptation information; prompt template 2 includes background information, prior information, task description information, and adaptation information.

[0120] Different prompt templates can include the same information type, but the specific information corresponding to the same information type may differ. For example, prompt template 1 and prompt template 2 both include background information, prior information, task description information, and adaptation information, but the configuration items in the task description information are different.

[0121] The configuration item information corresponding to a configuration item attribute may include at least one of the following: configuration item name, attribute name, and attribute value. For example, configuration item information may include configuration item name and attribute name; or configuration item information may include configuration item name, attribute name, and attribute value.

[0122] Select a prompt word template from the prompt word template library that is related to the configuration item information corresponding to the configuration item attribute. Specifically, this process can be as follows: based on the configuration item information, select a prompt word template from the prompt word template library that contains the information type of that configuration item information. If multiple prompt word templates are selected, one can be randomly chosen.

[0123] The configuration information is written to the selected prompt word template to generate a prompt word instance. The prompt word instance is submitted to the large language model by calling the application programming interface (API). The large language model recognizes the prompt word instance (that is, it answers the prompt word instance as a question) and obtains the recognition result (i.e., the question-and-answer result).

[0124] The configuration item attribute alignment method based on a large language model provided in this invention improves the recognition efficiency of configuration item information by converting configuration item information into prompt word instances and recognizing the prompt word instances based on a large language model, which further helps to improve the configuration item attribute alignment efficiency.

[0125] In some embodiments, a prompt word template related to the configuration item information corresponding to the configuration item attribute is selected from the prompt word template library, including:

[0126] Based on configuration information and prompts, a template type code is generated; the prompts are used to provide suggestions for the large language model.

[0127] Select the prompt word template that corresponds to the template type code from the prompt word template library; one template type code corresponds to one prompt word template.

[0128] Specifically, the information covered in the prompt word template library is mapped to codes, including configuration item information and prompt information. The prompt information can be background information, prior information, task description information, and adaptation information that can provide hints to the large language model to help it understand and answer questions.

[0129] For example, different configuration item information can be mapped to codes. For example, "0" indicates that the configuration item information includes the configuration item name, attribute name, and attribute value; "1" indicates that the configuration item information includes the configuration item name and attribute name.

[0130] For example, different prior information can be mapped to codes. For instance, "000" indicates no prior information, "001" indicates existing information, and other codes represent different types of prior information. For example, "002" indicates prior information containing constraints, and "003" indicates prior information containing specifications.

[0131] For example, different question formats can be mapped to codes. "000" represents the default question format, and "001" represents a concise question format. Other codes represent more different types of question formats. For example, "002" represents a question format where output constraints are stated first, and "003" represents a question format where prior information is stated later.

[0132] For example, the output specifications are mapped to the encoding. For instance, "0" represents the default output, and "1" represents enhanced output constraints. Enhanced output constraints limit the recognition results output by the large language model. For example, if the default output is 5 recognition results, after applying enhanced output constraints, only 2 recognition results will be output.

[0133] After mapping information to codes, the codes corresponding to the information are combined in a certain order to form template type codes. For example, template type codes can be formed in the order of configuration item information - prior information - problem description - output specifications.

[0134] For example, <configuration item name, attribute name> <no prior information> <other problem description> <default output specification>, the template type code is 1-000-001-0.

[0135] For example, <configuration item name, attribute name> <preliminary information> <default problem description> <default output specification>, the template type code is 1-001-000-0.

[0136] For example, <configuration item name, attribute name, attribute value> <preliminary information> <default problem description> <default output specification>, template type code is 0-001-000-0.

[0137] For example, <Configuration item name, attribute name, attribute value><Prerequisite information><Other problem descriptions><Default output specification>, the template type code is 0-001-001-0.

[0138] For example, <configuration item name, attribute name, attribute value> <with prior information> <default problem description> <enhanced output restrictions>, the template type code is 0-001-000-1.

[0139] Based on the configuration item information, a template type code containing the corresponding encoding of the configuration item information is generated. For example, if the configuration item information is configuration item name, attribute name, and attribute value, the generated template type code can be 0-001-000-0, 0-001-001-0, and 0-001-000-1, etc.

[0140] To narrow down the range of template type codes, a unique template type code can be determined by combining it with the prompt information (e.g., prior information, problem description, and output specifications).

[0141] Select the prompt word template that corresponds to the template type code from the prompt word template library. One template type code corresponds to one prompt word template.

[0142] The configuration item attribute alignment method based on a large language model provided in this invention generates a template type code based on configuration item information and prompt information, and selects a prompt word template corresponding to the template type code from the prompt word template library, thereby achieving rapid selection of prompt word templates and further improving the recognition efficiency of prompt word instances by the large language model.

[0143] In some embodiments, after obtaining the recognition result, the method further includes:

[0144] Check whether the detection and recognition results meet the preset rules;

[0145] If the recognition result does not meet the preset rules, check whether the number of recognition attempts of the large language model exceeds the threshold.

[0146] If the number of recognition attempts by the large language model does not exceed the threshold, the configuration item attributes are re-recognized based on the large language model, and the recognition results are obtained again.

[0147] Specifically, after obtaining the recognition results, it is checked whether the recognition results meet the preset rules.

[0148] Optionally, preset rules may include: the recognition result does not contain content unrelated to the recognition result, and the recognition result meets the requirements. Content unrelated to the recognition result may include duplicate questions and result descriptions. The recognition result meeting the requirements may include: the number of recognition results meets the specified number, and the recognition result is in a given result list.

[0149] If the recognition result does not meet the preset rules, an abnormal situation has occurred:

[0150] For example, if the recognition results contain content unrelated to the recognition results, it indicates that the large language model does not understand the problem, and it is necessary to strengthen the output constraints and update the encoding of the output rules.

[0151] For example, if the number of recognition results is not the specified number, it indicates that the understanding ability of the large language model is insufficient, and the template type code needs to be switched.

[0152] For example, if the recognition result cannot be returned, it indicates that the large language model cannot understand the question and therefore cannot return a structure relevant to the question. A change in template type code is required.

[0153] If the recognition result does not meet the preset rules, check whether the number of recognition attempts by the large language model exceeds a threshold. The threshold can be set according to the recognition capability of the large language model.

[0154] If the number of recognition attempts by the large language model does not exceed the threshold, the configuration item attributes will be re-recognized based on the large language model, and the recognition results will be obtained again until the recognition results meet the preset rules, or the number of recognition attempts by the large language model exceeds the threshold. If the number of recognition attempts by the large language model exceeds the threshold, an error will be reported.

[0155] The configuration item attribute alignment method based on a large language model provided in this invention performs a post-hoc analysis on the recognition results. If the post-hoc analysis does not meet the requirements, the configuration item attributes are re-identified based on the large language model, and the recognition results are re-obtained, thereby achieving automatic updating of the recognition results and further improving the intelligence of the configuration item attribute alignment process.

[0156] In some embodiments, the configuration item attributes are re-identified based on a large language model, and the identification results are re-obtained, including:

[0157] Update the prompt message; the prompt message is used to provide suggestions to the large language model;

[0158] Based on the configuration information and the updated prompts, regenerate the template type code;

[0159] Select a new prompt word template from the prompt word template library that corresponds to the newly generated template type code; one prompt word template corresponds to one template type code.

[0160] Write the configuration information into the newly selected prompt word template and regenerate the prompt word instance;

[0161] The regenerated prompt word instances are identified based on the large language model, and the identification results are obtained again.

[0162] Specifically, if the recognition results output by the large language model do not meet the preset rules, it indicates that the template type code needs to be changed. Therefore, the prompt information needs to be updated. For example, if the original prompt information did not contain prior information, the updated prompt information will include prior information; for example, if the original prompt information contained the default output specifications, the updated prompt information could contain stronger output restrictions.

[0163] Based on the configuration information and the updated prompts, the template type code is regenerated. Because the prompts have changed, the regenerated template type code is different from the original template type code. For example, if the original template type code is 0-001-000-0, the regenerated template type code will be 0-001-001-1.

[0164] Select a new prompt word template from the prompt word template library that corresponds to the newly generated template type code; write the configuration information into the newly selected prompt word template and regenerate the prompt word instance; recognize the newly generated prompt word instance based on the large language model and obtain the recognition result again.

[0165] The configuration item attribute alignment method based on a large language model provided in this embodiment of the invention can automatically update the recognition results and improve the probability that the recognition results meet the preset rules when the recognition results output by the large language model do not meet the preset rules. This is achieved by updating the prompt information, then regenerating the template type code, reselecting the prompt word template, regenerating the prompt word instance, and re-recognizing the prompt word instance.

[0166] Figure 5 This is the third flowchart illustrating the configuration item attribute alignment method based on a large language model provided in this embodiment of the invention. Figure 5 As shown in the figure, the configuration item attribute alignment method based on a large language model provided in this embodiment of the invention includes the following steps:

[0167] Step 510: Initialize the large language model.

[0168] Specifically, the model parameters in the large language model are initialized; the model parameters include parameters related to random responses and parameters related to context.

[0169] Step 520: Generate template type code based on configuration item information and prompt information.

[0170] Step 530: Select the prompt word template that corresponds to the template type code from the prompt word template library.

[0171] Step 540: Write the configuration item information into the prompt word template to generate a prompt word instance.

[0172] Step 550: Identify prompt word instances based on a large language model and obtain the identification results.

[0173] Step 560: Determine if the recognition result is abnormal.

[0174] Specifically, if the recognition result is abnormal, that is, if the recognition result does not meet the preset rules, step 580 is executed; if the recognition result is normal, that is, if the recognition result meets the preset rules, step 570 is executed.

[0175] Step 570: Analyze the recognition results to obtain the tag set.

[0176] Step 580: Determine whether the number of recognition attempts of the large language model exceeds the threshold.

[0177] Specifically, if the number of recognitions by the large language model does not exceed the threshold, proceed to step 590; if the number of recognitions by the large language model exceeds the threshold, proceed to step 5100.

[0178] Step 590: Update the prompt message.

[0179] Specifically, after updating the prompt message, step 520 is executed, which is to regenerate the template type based on the configuration item information and the updated prompt message.

[0180] Step 5100: Report an error.

[0181] To facilitate understanding of this invention, specific examples are provided below:

[0182] Example 1: Define 7 labels: port, app, db, ip, cabinet, datacenter, and device. These labels and their definitions are: [port represents the port number; app represents the application service; db represents the database; ip represents the IP address; cabinet represents the server rack; datacenter represents the data center; device represents the identification device].

[0183] Configuration item information (configuration item name: flowanalysis, attribute name: netdevice_currentuse, attribute value is "Mars Network Traffic Monitoring"). Here, we take the large language model llama2-7b-v2 as an example.

[0184] Step 1: Initialize configuration.

[0185] Specifically, the randomness of the large language model's responses is limited. First, the sampling temperature is set to 0; second, the cumulative probability limit for random sampling, top_p, is set to 0.1.

[0186] To eliminate the influence of the large language model on contextual association, the following measures are implemented: First, the number of accompanying historical messages (history_m) is set to 0; second, the historical summary (historicalSummary) is set to false; and third, the dialogue role is set to empty.

[0187] Step 2: Generate template type code.

[0188] Specifically, a template type code is generated based on the configuration item information (configuration item name, attribute name, attribute value) and prompt information (preliminary information, use default expression, default output specification).

[0189] Step 3: Select a prompt word template.

[0190] Specifically, select the prompt word template that corresponds to the template type code from the prompt word template library.

[0191] Step 4: Generate prompt word instances.

[0192] Specifically, configuration information is added to the prompt word template to generate prompt word instances.

[0193] Step 5: The large language model identifies the prompt word instances and obtains the identification results.

[0194] Specifically, the prompt word instance is submitted to the large language model by calling the API, and the large language model treats the prompt word instance as a question and answers it.

[0195] Let's take a large language model as an example: I am an IT operations expert, and I'm dealing with the relationships between CMDB configuration item attributes. I need to tag the configuration item attributes; there are a total of 7 tags. These tags and their definitions are as follows:

[0196] port: Indicates the port number;

[0197] app: refers to application services;

[0198] db: refers to a database; for example, db2, sequeoiadb, redis, oracle, and tidb, etc.

[0199] ip: Represents an IP address;

[0200] cabinet: refers to server rack;

[0201] datacenter: refers to a data center;

[0202] device: indicates the identification of a device, such as device number, device serial number, etc.

[0203] Tag the "netdevice_currentuse" attribute (value: "Mars Network Traffic Monitoring") of the configuration item "flowanalysis". Select the two most likely tags from the seven tags listed above and output the tags in descending order of likelihood. The output tags are: <device> <app>.

[0204] Step 6: Detection of recognition results.

[0205] The output consists of two tags. <device> <app>According to preset rules, a full anomaly detection is performed on the content of the recognition results. If no anomalies are found, a label is output and the dialogue ends.

[0206] Example 2: Seven labels are defined: ['port', 'app', 'dbname', 'ip', 'cabinet', 'datacenter', 'device', 'db']. Question: Which three standardized attribute names in the above list are most relevant to the "wanrouter" attribute of the configuration item "networkline"? This example uses the large language model llama2-7b-v2.

[0207] Step 1: Initialize configuration.

[0208] Specifically, the randomness of the large language model's responses is limited. First, the sampling temperature is set to 0; second, the cumulative probability limit for random sampling, top_p, is set to 0.1.

[0209] To eliminate the influence of the large language model on contextual association, the following measures are implemented: First, the number of accompanying historical messages (history_m) is set to 0; second, the historical summary (historicalSummary) is set to false; and third, the dialogue role is set to empty.

[0210] Step 2: Generate template type code.

[0211] Specifically, a template type code is generated based on the configuration item information (configuration item name and attribute name) and prompt information (no prior information, use other expressions, default output specification).

[0212] Step 3: Select a prompt word template.

[0213] Specifically, select the prompt word template that corresponds to the template type code from the prompt word template library.

[0214] Step 4: Generate prompt word instances.

[0215] Specifically, configuration information is added to the prompt word template to generate prompt word instances.

[0216] Step 5: The large language model identifies the prompt word instances and obtains the identification results.

[0217] Specifically, the prompt word instance is submitted to the large language model by calling the API, and the large language model treats the prompt word instance as a question and answers it.

[0218] Let's take a large language model as an example: I am a senior CMDB operations expert, and I know the following list of standardized attribute names: ['port', 'app', 'dbname', 'ip', 'cabinet', 'datacenter', 'device', 'db']. Question: Which three standardized attribute names in the above list are most closely related to the attribute "wanrouter" of the configuration item "networkline"? Output in the following format: Related attributes: <list>.

[0219] The output is "Sorry, I do not have any information about the configuration item 'networkline', therefore I cannot perform association attribute analysis on it."

[0220] Step 6: Detection of recognition results.

[0221] According to preset rules, a full anomaly detection is performed on the recognition results. If anomalies are found, the prompt information is updated, and the process returns to step 2 to regenerate the template type code. This continues until the recognition results meet the preset rules, at which point the label is output. <device> <ip>The dialogue ends. Alternatively, if the large language model reaches its maximum recognition threshold, an error message is displayed and the dialogue ends.

[0222] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.< / ip> < / device> < / app> < / device> < / app> < / device>

Claims

1. A method for aligning configuration item attributes based on a large language model, characterized in that, include: The configuration item attributes are identified based on a large language model, and the identification results are obtained. When two configuration item attributes have the same attribute value but different attribute names, the alignment result of the two configuration item attributes is determined based on the recognition results corresponding to the two configuration item attributes respectively. The process of identifying configuration item attributes based on a large language model and obtaining the identification results includes: Select a prompt word template from the prompt word template library that is related to the configuration item information corresponding to the configuration item attribute; Write the configuration item information into the prompt word template to generate a prompt word instance; The prompt word instance is identified based on a large language model, and the identification result is obtained; The process of identifying configuration item attributes based on a large language model and obtaining the identification results also includes: Tag the configuration item attributes based on the large language model to obtain a tag set; The step of determining the alignment result of the two configuration item attributes based on the recognition results corresponding to the two configuration item attributes respectively includes: Find the intersection of the tag sets corresponding to the two configuration item attributes respectively, and obtain the intersection result; If the intersection result is not empty, it is determined that the two configuration item attributes are aligned.

2. The configuration item attribute alignment method based on a large language model according to claim 1, characterized in that, The step of selecting a prompt word template from the prompt word template library that is related to the configuration item information corresponding to the configuration item attribute includes: Based on the configuration information and the prompt information, a template type code is generated; the prompt information is used to provide prompts for the large language model. Select the prompt word template corresponding to the template type code from the prompt word template library; one template type code corresponds to one prompt word template.

3. The configuration item attribute alignment method based on a large language model according to claim 1, characterized in that, After obtaining the recognition result, the process further includes: Detect whether the recognition result meets the preset rules; If the recognition result does not meet the preset rule, check whether the number of recognitions of the large language model exceeds the threshold. If the number of recognition attempts by the large language model does not exceed the threshold, the configuration item attribute is re-recognized based on the large language model, and the recognition result is obtained again.

4. The configuration item attribute alignment method based on a large language model according to claim 3, characterized in that, The process of re-identifying the configuration item attributes based on a large language model and re-obtaining the identification results includes: Update the prompt information; the prompt information is used to provide prompts to the large language model; Based on the configuration item information and the updated prompt information, the template type code is regenerated; The prompt word template corresponding to the newly generated template type code is selected again from the prompt word template library; one template type code corresponds to one prompt word template; Write the configuration information into the newly selected prompt word template and regenerate the prompt word instance; The regenerated prompt word instances are identified based on the large language model, and the identification results are obtained again.

5. The configuration item attribute alignment method based on a large language model according to claim 1, characterized in that, Before identifying configuration item attributes based on a large language model, the following steps are also included: The model parameters in the large language model are initialized; the model parameters include parameters related to random responses and parameters related to context.

6. The configuration item attribute alignment method based on a large language model according to any one of claims 1, 2, or 4, characterized in that, The prompt template includes at least one of the following: background information, prior information, task description information, and adaptation information.

Images