CN118349661A - A data processing method, device and equipment based on large language model - Google Patents

Abstract

The present application provides a data processing method, device and equipment based on a large language model, the method comprising: inputting an original input question and a prompt word into a large language model, processing the original input question based on the prompt word by the large language model to obtain a target question parameter and a target question description; wherein the prompt word includes a question sample, a question parameter sample and a question description sample; querying a target access parameter corresponding to the target question description from a feature database; wherein the feature database includes a correspondence between an access description and an access parameter; generating operation and maintenance system operation data based on the target access parameter and the target question parameter; operating the operation and maintenance system through the operation and maintenance system operation data to obtain the target operation and maintenance data corresponding to the original input question, and returning the target operation and maintenance data to the user. The technical solution of the present application can save computing resources and time overhead.

Description

A data processing method, device and equipment based on large language model

Technical Field

The present application relates to the field of artificial intelligence technology, and in particular to a data processing method, device and equipment based on a large language model (LLM).

Background technique

With the rapid development of network technology, there is more and more network data. In order to retrieve the data required by users from a large amount of network data, users are required to provide original input questions. The server retrieves the data required by users from a large amount of network data based on the original input questions and returns the data to the user.

However, due to the large amount of network data, the server needs to consume a lot of computing resources to retrieve the data required by the user from a large amount of network data, and the server needs to spend a lot of time, resulting in a large amount of server resource consumption, and the user can only obtain the required data for a long time.

Summary of the invention

The present application provides a data processing method based on a large language model, the method comprising:

Inputting the original input question and the prompt word into the large language model, and processing the original input question based on the prompt word by the large language model to obtain target question parameters and target question description; wherein the prompt word includes question examples, question parameter examples and question description examples;

Querying the target access parameter corresponding to the target problem description from a feature database; wherein the feature database includes a correspondence between the access description and the access parameter;

Generate operation data of the operation and maintenance system based on the target access parameter and the target problem parameter;

The operation and maintenance system is operated through the operation and maintenance system operation data to obtain target operation and maintenance data corresponding to the original input question, and the target operation and maintenance data is returned to the user.

The present application provides a data processing device based on a large language model, the device comprising:

A processing module, used to input the original input question and the prompt word into the large language model, and process the original input question based on the prompt word through the large language model to obtain target question parameters and target question description; wherein the prompt word includes question examples, question parameter examples and question description examples;

A query module, used to query the target access parameter corresponding to the target problem description from a feature database; wherein the feature database includes a correspondence between the access description and the access parameter;

An operation module is used to generate operation and maintenance system operation data based on the target access parameters and the target question parameters, and operate the operation and maintenance system through the operation and maintenance system operation data to obtain target operation and maintenance data corresponding to the original input question, and return the target operation and maintenance data to the user.

The present application provides an electronic device, comprising: a processor and a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions that can be executed by the processor; the processor is used to execute the machine-executable instructions to implement the above-mentioned data processing method based on a large language model.

It can be seen from the above technical solutions that in the embodiment of the present application, after obtaining the original input question, the target operation and maintenance data corresponding to the original input question is obtained with the help of a large language model, and the target operation and maintenance data is returned to the user. Only a small amount of computing resources is needed to obtain the target operation and maintenance data, and the time to obtain the target operation and maintenance data is shorter, thereby saving the computing resources of the server and saving time. It is possible to quickly determine the target operation and maintenance data corresponding to the original input question, provide logical analysis and reasoning capabilities, and use the analysis capabilities of the large language model to provide accurate and reliable target operation and maintenance data. Through a Copilot (assistant) agent as an intermediate layer, users can interact with the Copilot agent in natural language to complete the interaction of operation and maintenance requirements. The Copilot agent converts natural language into instructions and passes them to the operation and maintenance system to realize a conversational intelligent operation and maintenance system, improve the operation and maintenance experience, and realize barrier-free human-computer interaction experience between users and the operation and maintenance system, thereby improving user experience. Based on the capabilities of the large language model, the obstacles to natural language understanding are solved, and the user's input is changed from page clicks to human-computer language interaction. Based on the capabilities of the large language model, the automatic implementation capability based on language interaction is realized for the functions of the operation and maintenance system, such as overview, diagnosis, self-healing closed loop, and reporting.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a data processing method based on a large language model in one embodiment of the present application;

FIG2 is a schematic diagram of the structure of a Copilot agent in one embodiment of the present application;

FIG3 is a flow chart of a data processing method based on a large language model in one embodiment of the present application;

FIG4 is a schematic diagram of a data processing method based on a large language model in one embodiment of the present application;

FIG5 is a schematic diagram of four capabilities supported by the Copilot agent in one embodiment of the present application;

FIG6 is a schematic diagram of a single task process in one embodiment of the present application;

FIG7 is a schematic diagram of a multi-task process in one embodiment of the present application;

FIG8 is a structural diagram of a data processing device based on a large language model in one embodiment of the present application;

FIG. 9 is a hardware structure diagram of an electronic device in one embodiment of the present application.

Detailed ways

The present application embodiment proposes a data processing method based on a large language model, which can be applied to an electronic device, such as a server. Referring to FIG1 , which is a flow chart of the method, the method may include:

Step 101: input an original input question and a prompt word into a large language model, and process the original input question based on the prompt word through the large language model to obtain target question parameters and a target question description; wherein the prompt word may include a question sample, a question parameter sample, and a question description sample.

Step 102: query the target access parameter corresponding to the target problem description from the feature database; wherein the feature database may include a correspondence between the access description and the access parameter.

Step 103: Generate operation data of the operation and maintenance system based on the target access parameter and the target problem parameter.

Step 104: operate the operation and maintenance system through the operation data of the operation and maintenance system to obtain target operation and maintenance data corresponding to the original input question, and return the target operation and maintenance data to the user.

Exemplarily, querying the target access parameters corresponding to the target problem description from the feature database may include, but is not limited to: if there is a first access description matching the target problem description in the feature database, then a first data structure corresponding to the first access description may be obtained from the feature database, and the first data structure may include the first access description and access parameters. On this basis, the access parameters in the first data structure may be determined as the target access parameters corresponding to the target problem description.

Exemplarily, before querying the target access parameters corresponding to the target problem description from the feature database, private domain data of the operation and maintenance system can also be obtained, and the private domain data can include access description information and access parameter information; extract the first description keyword from the access description information, and generate an access description based on the first description keyword; extract the first parameter keyword from the access parameter information, and generate an access parameter based on the first parameter keyword; on this basis, the correspondence between the access description and the access parameter can be recorded in the feature database.

Exemplarily, querying the target access parameters corresponding to the target problem description from the feature database may include but is not limited to: if there is a second access description matching the target problem description in the feature database, obtaining a second data structure corresponding to the second access description from the feature database, the second data structure may include a second access description and multiple operation step descriptions; for each operation step description, if there is a third access description matching the operation step description in the feature database, obtaining a third data structure corresponding to the third access description from the feature database, the third data structure may include a third access description and access parameters, and determining the access parameters in the third data structure as the access parameters corresponding to the operation step description; and determining the access parameters corresponding to each operation step description as the target access parameters.

Exemplarily, before querying the target access parameters corresponding to the target problem description from the feature database, the private domain data of the operation and maintenance system can also be obtained, and the private domain data includes access description information and multiple operation step description information; wherein, for each operation step description information, the private domain data also includes access description information and access parameter information corresponding to the operation step description information. Extract the second description keyword from the access description information, and generate an access description based on the second description keyword; for each operation step description information, extract the operation step keyword from the operation step description information, and generate an operation step description based on the operation step keyword; record the correspondence between the access description and the operation step description in the feature database. For each operation step description information, extract the third description keyword from the access description information corresponding to the operation step description information, and generate an access description based on the third description keyword; extract the second parameter keyword from the access parameter information corresponding to the operation step description information, and generate an access parameter based on the second parameter keyword; record the correspondence between the access description and the access parameter in the feature database.

In a possible implementation, the target question parameter includes a target device identifier, the target question description includes a target URL description, the target access parameter includes a target URL address, and the target URL address may include a configured fixed device identifier. Based on this, the operation and maintenance system operation data is generated based on the target access parameter and the target question parameter, which may include but is not limited to: replacing the fixed device identifier in the target URL address with the target device identifier to obtain a modified URL address as the operation and maintenance system operation data. The operation and maintenance system is operated through the operation and maintenance system operation data to obtain the target operation and maintenance data corresponding to the original input question, which may include but is not limited to: accessing the operation and maintenance management page of the operation and maintenance system corresponding to the modified URL address, and the operation and maintenance management page may include the target operation and maintenance data.

In a possible implementation, the target problem parameter includes a target device identifier, the target problem description includes a target API description, the target access parameter includes a target API, and the target API may include a configured fixed device identifier. Based on this, the operation data of the operation and maintenance system is generated based on the target access parameter and the target problem parameter, which may include but is not limited to: replacing the fixed device identifier in the target API with the target device identifier to obtain a modified API as the operation data of the operation and maintenance system.

The operation and maintenance system is operated through the operation and maintenance system operation data to obtain target operation and maintenance data corresponding to the original input problem, which may include but is not limited to: calling the modified API to operate the operation and maintenance system to obtain operation result data, and determining the target operation and maintenance data based on the operation result data.

In a possible implementation, the target problem parameter includes a target data table identifier, the target problem description includes a target SQL description, and the target access parameter includes a target SQL definition. Based on this, the operation and maintenance system operation data is generated based on the target access parameter and the target problem parameter, including but not limited to: generating a target SQL statement based on the target SQL definition and the target data table identifier as the operation and maintenance system operation data, the target SQL statement includes the target data table identifier. The operation and maintenance system is operated through the operation and maintenance system operation data to obtain the target operation and maintenance data corresponding to the original input problem, including but not limited to: operating the data table corresponding to the target data table identifier of the operation and maintenance system by executing the target SQL statement to obtain the operation result data of the data table; determining the target operation and maintenance data based on the operation result data of the data table.

In one example, the above execution order is only for the convenience of describing the examples given. In practical applications, the execution order between the steps can also be changed, and there is no limitation on this execution order. Moreover, in other embodiments, the steps of the corresponding method are not necessarily executed in the order shown and described in this specification, and the steps included in the method may be more or less than those described in this specification. In addition, the single step described in this specification may be decomposed into multiple steps for description in other embodiments. The multiple steps described in this specification may also be combined into a single step for description in other embodiments.

It can be seen from the above technical solutions that in the embodiment of the present application, after obtaining the original input question, the target operation and maintenance data corresponding to the original input question is obtained with the help of a large language model, and the target operation and maintenance data is returned to the user. Only a small amount of computing resources is needed to obtain the target operation and maintenance data, and the time to obtain the target operation and maintenance data is shorter, thereby saving the computing resources of the server and saving time. It is possible to quickly determine the target operation and maintenance data corresponding to the original input question, provide logical analysis and reasoning capabilities, and use the analysis capabilities of the large language model to provide accurate and reliable target operation and maintenance data. Through a Copilot (assistant) agent as an intermediate layer, users can interact with the Copilot agent in natural language to complete the interaction of operation and maintenance requirements. The Copilot agent converts natural language into instructions and passes them to the operation and maintenance system to realize a conversational intelligent operation and maintenance system, improve the operation and maintenance experience, and realize barrier-free human-computer interaction experience between users and the operation and maintenance system, thereby improving user experience. Based on the capabilities of the large language model, the obstacles to natural language understanding are solved, and the user's input is changed from page clicks to human-computer language interaction. Based on the capabilities of the large language model, the automatic implementation capability based on language interaction is realized for the functions of the operation and maintenance system, such as overview, diagnosis, self-healing closed loop, and reporting.

The above technical solutions of the embodiments of the present application are described below in combination with specific application scenarios.

With the evolution of ICT (Information and Communications Technology) infrastructure and the development of enterprise digital transformation, users' O&M needs have increased, O&M functions have become increasingly complex, and the ease of use of O&M has been challenged. Automation and intelligence of O&M have become the theme. O&M refers to the operation and maintenance of networks, servers, and services at all stages of their life cycle.

In order to retrieve the data required by the user from the operation and maintenance system, the user needs to provide the original input question. The operation and maintenance system retrieves the data required by the user based on the original input question and returns the data to the user. Since the amount of data in the operation and maintenance system is large, a large amount of computing resources is required to retrieve the data required by the user, which takes a lot of time, resulting in a large amount of resource consumption, and the user takes a long time to obtain the required data.

In view of the above findings, the data can be analyzed with the help of a large language model (LLM). For example, with the help of a large language model, the target operation and maintenance data corresponding to the original input question can be obtained, and the target operation and maintenance data can be returned to the user. Only a small amount of computing resources are needed to obtain the target operation and maintenance data, and the time to obtain the target operation and maintenance data is shorter, thereby saving server computing resources and time costs. It can quickly determine the target operation and maintenance data corresponding to the original input question, provide logical analysis and reasoning capabilities, and use the analysis capabilities of the large language model to provide accurate and reliable target operation and maintenance data.

For example, ChatGPT is a large language model based on natural language processing and machine learning. It can simulate human language communication and realize intelligent dialogue with users. It is widely used in intelligent customer service, chatbots, voice assistants and other fields to provide people with more convenient, fast and intelligent services. Based on this, ChatGPT can be used to analyze the data and obtain the target operation and maintenance data corresponding to the original input question.

In the above application scenario, a data processing method based on a large language model is proposed in an embodiment of the present application, which can be applied to electronic devices, such as servers of an operation and maintenance system. By deploying a Copilot agent (assistant agent) on the server and using the Copilot agent as an intermediate layer, users (such as operation and maintenance personnel) can interact with the Copilot agent in natural language to convey business needs. The Copilot agent converts language into instructions and passes them to the operation and maintenance system, realizing a barrier-free human-computer interaction experience between users and the operation and maintenance system.

As shown in FIG. 2 , which is a schematic diagram of the structure of the Copilot agent, the user provides the Copilot agent with an original input question, which can be an original input question in voice format (Voice), an original input question in text format (Text), or an original input question in other formats, such as video format, etc., without limitation. Obviously, the user can interact with the Copilot agent through voice or text.

Copilot agent supports functions such as large language model (LLM), RAG (Retrieval Augmented Generation) and Agent. Based on the functions of the large language model, Copilot agent can recognize and understand natural language (such as original input questions). Through fine-tuning training, the large language model can have the operation and maintenance knowledge of the operation and maintenance system, so as to better recognize and understand the original input questions. Based on the capabilities of the large language model, the obstacles to natural language understanding are solved, and the user's input is changed from page clicking to human-computer language interaction. For the functions of the operation and maintenance system, such as overview function, diagnosis function, self-healing closed loop function, reporting function, etc., the automatic realization capability based on language interaction is realized.

Based on the functions of RAG, the Copilot agent can combine RAG and large language models to solve problems such as hallucinations and knowledge limitations of large language models, thereby improving the accuracy of large language models.

Based on the functions of the Agent, the Copilot agent can realize the call docking with the operation and maintenance system, thereby completing the tasks related to the original input problem and returning the target operation and maintenance data corresponding to the original input problem to the user. As shown in Figure 2, the Copilot agent can be docked with different operation and maintenance systems.

In the embodiment of the present application, a data processing method based on a large language model is proposed. In the operation and maintenance system, a conversational human-computer interaction based on a large language model (private domain large model) can be realized. Referring to FIG3 , a flow chart of the data processing method is shown. The data processing method may include:

Step 301: The Copilot agent obtains private domain data of the operation and maintenance system. For example, the private domain data may include access description information and access parameter information, and/or the private domain data may include access description information and multiple operation step description information; wherein, for each operation step description information, the private domain data may also include access description information and access parameter information corresponding to the operation step description information.

Exemplarily, during the data preparation phase, the Copilot agent can obtain private domain data related to the operation and maintenance system, and the private domain data can include access description information and access parameter information, access description information and operation step description information, access description information corresponding to the operation step description information, and access parameter information corresponding to the operation step description information. In addition, private domain data can also include operation and maintenance knowledge, etc., which is not limited.

Exemplarily, the access parameter information may be information related to access parameters, and the access parameters are parameters used to access the operation and maintenance system, such as URL addresses, APIs, SQL, etc., and there is no restriction on the access parameters. The access description information may be information related to access descriptions, and the access descriptions are information used to describe the problem, such as information describing the keywords of the problem, and there is no restriction on the access descriptions.

Exemplarily, when a task corresponds to multiple operation steps, the private domain data may include access description information and multiple operation step description information, where the access description information may be information related to the task and is description information for the task. For each operation step description information, the operation step description information may be description information for the operation step and may be a problem description for the operation step.

For each operation step description information, the private domain data may also include access description information and access parameter information corresponding to the operation step description information. The access parameter information may be information related to access parameters, and access parameters are parameters used to access the operation and maintenance system, such as URL address, API, SQL, etc. The access description information may be information related to the access description, and the access description is information used to describe the problem (i.e., the problem corresponding to the operation step description information), and there is no restriction on this access description.

Step 302: The Copilot agent uses the private domain data to perform incremental training on the large language model, that is, fine-tunes the large language model to obtain a trained large language model, so that the large language model learns the knowledge of the private domain data. The incremental training process of the large language model is not limited in this embodiment.

Step 303: The Copilot agent uses private domain data to maintain the feature database of the RAG function.

For example, in order to solve the hallucination and inaccuracy problems of the large language model, the private domain data can be extracted, sorted, vectorized (embedded), and injected into the feature database of the RAG function, thereby improving the accuracy of the large language model with the help of the feature database of the RAG function.

For example, if the private domain data includes access description information and access parameter information, the access description information can be processed to obtain an access description, which can be information used to describe the problem. The access parameter information can be processed to obtain access parameters, which can be parameters used to access the operation and maintenance system, such as URL addresses, APIs, SQL, etc.

Processing the access description information (the processing process of the access parameter information is similar) to obtain the access description means: extracting the first description keyword from the access description information, and generating the access description based on the first description keyword. For example, invalid information in the access description information can be filtered, and the access description information can be processed by keyword extraction, sorting, and word segmentation to obtain the first description keyword in the access description information (such as high AP channel utilization), and the first description keyword is used as the access description, and the above processing process is not limited. Alternatively, after obtaining the first description keyword, the first description keyword can also be vectorized and the vectorized keyword can be used as the access description.

Vectorizing the first description keyword of the access description information refers to the process of mapping high-dimensional data (such as text, pictures, audio, etc.) into a low-dimensional space. The embedding vector is a vector composed of real numbers, which represents the input data as a point in a continuous numerical space.

After obtaining the access description corresponding to the access description information and the access parameters corresponding to the access parameter information, the corresponding relationship between the access description and the access parameters may be recorded in the feature database.

Exemplarily, if the private domain data includes access description information and multiple operation step description information, the access description information can be processed to obtain an access description (such as extracting a second description keyword from the access description information and generating an access description based on the second description keyword), and the access description can be description information for the task. For each operation step description information, the operation step description information can be processed to obtain an operation step description (such as extracting an operation step keyword from the operation step description information and generating an operation step description based on the operation step keyword), and the operation step description is problem description information for the operation step.

After obtaining the access description corresponding to the access description information and the operation step description corresponding to the operation step description information, the corresponding relationship between the access description and the operation step description is recorded in the feature database.

For each operation step description information, the private domain data may include the access description information and access parameter information corresponding to the operation step description information. Based on this, for each operation step description information, the access description information corresponding to the operation step description information may also be processed to obtain an access description (such as extracting a third description keyword from the access description information corresponding to the operation step description information, and generating an access description based on the third description keyword), and the access parameter information corresponding to the operation step description information may be processed to obtain an access parameter (such as extracting a second parameter keyword from the access parameter information corresponding to the operation step description information, and generating an access parameter based on the second parameter keyword). After obtaining the access description corresponding to the access description information and the access parameter corresponding to the access parameter information, the corresponding relationship between the access description and the access parameter is recorded in the feature database.

For example, see Table 1 and Table 2, which are examples of the feature database.

Table 1

Table 2

Step 304: The Copilot agent obtains the original input question, which may be an original input question in text format input by the user, or an original input question in voice format, or an original input question in another format. There is no limitation to this, and the original input question in text format is used as an example below.

Step 305: The Copilot agent inputs the original input question and the prompt word into the large language model. The large language model processes the original input question based on the prompt word to obtain target question parameters and target question description. The prompt word may include question examples, question parameter examples, and question description examples.

For example, after obtaining the original input question, the Copilot agent can input the original input question and the prompt word together into the large language model. The large language model can perform structural decomposition on the original input question based on the semantic understanding of the prompt word to obtain the target question parameters and the target question description.

For example, prompt words are used to assist the large language model in performing structural splitting of the original input question. That is, the large language model can understand the structural splitting method of the original input question based on the prompt words, and then perform structural splitting on the original input question to obtain the target question parameters and target question description.

For example, the prompt words may include question examples, question parameter examples, and question description examples, and the original input question corresponds to the question example. Based on the representation of the question parameter examples in the question example, the large language model can separate the target question parameters from the original input question, and the target question parameters correspond to the question parameter examples. Based on the representation of the question description examples in the question example, the large language model can separate the target question description from the original input question, and the target question description corresponds to the question description example.

Step 306: The Copilot agent queries the target access parameters corresponding to the target question description from the feature database. For example, the Copilot agent queries the target access parameters from the feature database using a large language model.

Exemplarily, if there is a first access description matching the target problem description in the feature database, a first data structure corresponding to the first access description can be obtained from the feature database, and the first data structure can include the first access description and access parameters. The access parameters in the first data structure are determined as target access parameters corresponding to the target problem description, and the large language model can obtain the target access parameters.

For example, if the feature database stores access descriptions in vector form, after vectorizing the target problem description, the similarity between each access description in the feature database and the target problem description is compared. If the feature database stores access descriptions in keyword form, the similarity between each access description in the feature database and the target problem description (not vectorized) is directly compared. If the similarity is greater than a threshold, the access description is the first access description that matches the target problem description. If the similarity is not greater than the threshold, the access description is not the first access description that matches the target problem description. Based on this, the first access description can be found.

After obtaining the first access description, the access parameter corresponding to the first access description can be used as the target access parameter, and the target access parameter can be returned to the large language model. After obtaining the target access parameter, the large language model can perform a semantic check on the target access parameter. If it is found that the target access parameter is a single task result, such as a URL address or API, etc., it is confirmed that the target access parameter is obtained and the subsequent steps are executed.

Exemplarily, if there is a second access description matching the target problem description in the feature database, a second data structure corresponding to the second access description is obtained from the feature database, and the second data structure includes the second access description and multiple operation step descriptions. On this basis, for each operation step description: if there is a third access description matching the operation step description in the feature database, a third data structure corresponding to the third access description is obtained from the feature database, and the third data structure may include the third access description and access parameters, and the access parameters in the third data structure are determined as the access parameters corresponding to the operation step description; the access parameters corresponding to each operation step description are determined as the target access parameters corresponding to the target problem description, and the large language model can obtain the target access parameters (multiple access parameters).

For example, the similarity between each access description in the feature database and the target problem description is compared. If the similarity is greater than a threshold, the access description is a second access description that matches the target problem description. If the similarity is not greater than a threshold, the access description is not a second access description that matches the target problem description. After obtaining the second access description, as shown in Table 1, if the second access description corresponds to multiple operation step descriptions rather than access parameters, the multiple operation step descriptions can be returned to the large language model.

For each operation step description, the large language model uses the operation step description as an input question (similar to the original input question) and processes the operation step description to obtain a problem description. For example, the large language model uses prompt words to decompose the thought chain into multiple steps, and after structuring the problem of each operation step description, the problem description corresponding to the operation step description is obtained. For example, the operation step description can be used as a problem description, or keywords can be extracted from the operation step description as a problem description. There is no restriction on this, and the operation step description is used as a problem description as an example in the following.

Compare the similarity between each access description in the feature database and the operation step description. If the similarity is greater than a threshold, the access description is a third access description that matches the operation step description; if the similarity is not greater than the threshold, the access description is not a third access description that matches the operation step description.

After obtaining the third access description, as shown in Table 2, if the third access description corresponds to an access parameter, the access parameter corresponding to the third access description can be used as the target access parameter, and the target access parameter is returned to the large language model, so that the large language model can obtain the target access parameter.

After obtaining multiple target access parameters corresponding to multiple operation step descriptions, the large language model can perform semantic checks on these target access parameters. If it is found that the target access parameters are multiple task results, such as URL addresses or APIs, the target access parameters are confirmed to be obtained and subsequent steps are executed.

Step 307: The Copilot agent generates operation data of the operation system based on the target access parameter and the target problem parameter. For example, the target problem parameter can be added to the target access parameter to obtain the modified access parameter, and the modified access parameter can be used as the operation data of the operation system.

For example, the target access parameter may include a fixed device identifier, the target problem parameter may be a target device identifier, the fixed device identifier in the target access parameter may be replaced with the target device identifier to obtain a modified access parameter, and the operation data of the operation and maintenance system is the modified access parameter.

Step 308: The Copilot agent operates the operation and maintenance system through the operation data of the operation and maintenance system to obtain target operation and maintenance data corresponding to the original input question, and returns the target operation and maintenance data to the user.

Exemplarily, the Copilot agent can support the functions of the Agent. Based on the functions of the Agent, the Copilot agent can realize the call docking with the operation and maintenance system. Therefore, the operation data of the operation and maintenance system can be provided to the Agent. The Agent calls the operation data of the operation and maintenance system to operate the operation and maintenance system, obtains the target operation and maintenance data corresponding to the original input problem, and returns the target operation and maintenance data to the user.

In one example, see Figure 4, which is a schematic diagram of a data processing method based on a large language model. The red arrows 1, 2, and 3 correspond to steps 301-303, 1 represents obtaining the private domain data (private domain knowledge corpus) of the operation and maintenance system, and 2 and 3 represent maintaining the feature database (vector database). Embedding model (vectorization model) represents vectorization processing of private domain data, and Semantic Vector (semantic vector) represents semantic vector processing of private domain data. After the above processing, the feature database can be maintained.

The blue arrows 1-11 correspond to steps 304-308, 1 represents obtaining the original input question (user question), and 2 represents inputting the original input question and prompt words into the large language model (LLM large model).

3-8 represent querying the target access parameters, 3 represents inputting the target problem description into the Embedding model for vectorization, 4 and 5 represent comparing the similarity between each access description in the feature database and the target problem description to determine the first access description. If the first access description corresponds to the access parameter, the access parameter corresponding to the first access description can be used as the target access parameter.

6 indicates that the second access description corresponds to multiple operation step descriptions (knowledge), 7 indicates that multiple operation step descriptions and prompt words are input into the large language model, and 8 indicates that the large language model determines the target access parameters corresponding to each operation step description (i.e., decomposing the complex task to obtain multiple target access parameters).

9 indicates that the operation data of the operation system is generated based on the target access parameter and the target problem parameter, and the operation data of the operation system is input to the Agents. 10 indicates that the Agents operate the operation system through the operation data of the operation system to obtain the target operation data, and 11 indicates that the target operation data is returned to the user.

In an example, the conversational operation and maintenance system can provide four capabilities, as shown in Figure 5, which is a schematic diagram of the four capabilities supported by the Copilot agent. 1. The operation and maintenance system supports the webpage access function, the Copilot agent supports the Text2URL function, and the user interface supports the navigator function. Based on the Text2URL function, the Copilot agent can convert the original input question into a URL address, access the webpage of the operation and maintenance system through the URL address, and display the access result of the webpage in the user interface.

With this feature, the Copilot agent can convert the user's natural language needs into any operation and maintenance management page. The user can directly access the operation and maintenance management page through language communication without the user having to search and click on the link themselves. In this scenario, the user can view a direct overview and detailed data.

2. The operation and maintenance system supports the RestAPI call function, the Copilot agent supports the Text2API function, and the user interface supports the troubleshooting (diagnosis) function. Based on the Text2API function, the Copilot agent can convert the original input problem into an API, call the API to operate the operation and maintenance system, obtain the operation result, and return the operation result to the user. The user can obtain the diagnosis result of the operation and maintenance system based on the operation result.

With this function, the Copilot agent can convert the user's natural language requirements into API calls. For requirements for non-original pages or action delivery requirements, it can complete the REST API call after understanding the semantics to display and deliver related functions. In this scenario, it can provide a closed-loop solution to the problem that the user wants to diagnose.

3. The operation and maintenance system supports the SQLDB (database query) function, the Copilot agent supports the Text2SQL function, and the user interface supports the Report function. Based on the Text2SQL function, the Copilot agent can convert the original input question into an SQL statement, query the data table of the operation and maintenance system through the SQL statement, obtain the data table query result, and return the data table query result to the user.

With this function, the Copilot agent can convert the user's natural language requirements into SQL, support some of the user's chart summaries and queries, and provide users with personalized statistical reports in this scenario.

4. The operation and maintenance system supports the Knowledge function, the Copilot agent supports the QA function, and the user interface supports the Help function. Based on the QA function, the Copilot agent can convert the original input question into a question and answer sentence, and query the question and answer results from the operation and maintenance system through the question and answer sentence.

With this feature, Copilot agents can convert users' natural language questions into answers, and can support users in asking professional technical questions and asking questions about the software itself.

In view of the fact that the Copilot agent supports the Text2URL function, an embodiment of the present application proposes a data processing method based on a large language model, which may include:

Step S11, the Copilot agent obtains private domain data of the operation and maintenance system, and the private domain data may include URL description information and URL address information. The URL address information may be information related to the URL address, and the URL address is a parameter for accessing the operation and maintenance system. The URL description information may be information related to the URL description, and the URL description is information for describing the problem.

Step S12: The Copilot agent uses private domain data to perform incremental training on the large language model.

Step S13: The Copilot agent uses private domain data to maintain the feature database of the RAG function.

If the private domain data includes URL description information and URL address information, the URL description information can be processed to obtain a URL description, which can be information used to describe the problem, such as description information of the page corresponding to the URL address, description information of the problem for accessing the page, etc., without limitation. The URL address information can be processed to obtain a URL address, which can be a parameter for accessing the operation and maintenance system. After obtaining the URL description and URL address, the corresponding relationship between the URL description and the URL address can be recorded in the feature database.

Step S14: The Copilot agent obtains the original input question.

Step S15, the Copilot agent inputs the original input question and prompt words into the large language model, and the large language model processes the original input question based on the prompt words to obtain the target device identification and target URL description. The prompt words may include question samples, question parameter samples, and question description samples.

Step S16: The Copilot agent queries the target URL address corresponding to the target URL description from the feature database, and the target URL address may include a configured fixed device identifier.

Exemplarily, if the feature database contains a first URL description that matches the target URL description, and the first URL description corresponds to a URL address, the URL address corresponding to the first URL description is determined as the target URL address corresponding to the target URL description, and the large language model can obtain the target URL address.

Step S17: The Copilot agent replaces the fixed device identifier in the target URL address (ie, the device identifier pre-configured in the URL address) with the target device identifier to obtain a modified URL address.

Step S18: The Copilot agent accesses the operation and maintenance management page of the operation and maintenance system corresponding to the modified URL address, and returns the target operation and maintenance data of the operation and maintenance management page to the user.

For example, in the display interface of the conversational operation and maintenance system, the Copilot agent can receive natural language instructions on the display interface, directly answer simple text questions or generate a chart on the display interface, and jump to the maintenance management page on the display interface to display the target operation and maintenance data of the maintenance management page.

The following describes the processing of step S11 to step S18 in conjunction with a single task flow example.

See Figure 6, which is a schematic diagram of a single task process. Taking viewing AP information as an example, the process of single task calling is described. The specific questions and format descriptions in Figure 6 are all examples and there is no fixed restriction on this.

First, the user needs to view the basic information of a certain AP. The user sends the original input question "View the information of the AP named A1208-11" to the Copilot agent (Copilot assistant).

Then, in order to improve the vector matching degree of the knowledge base, it is necessary to accurately extract the question parameters in the original input question. For this purpose, the Copilot agent also needs to obtain the prompt word, and the Copilot agent sends the original input question and the prompt word to the large language model (LLM large model). By using the one-shot prompt word technique of the large language model, by constructing a question sample, a question parameter sample, and a question description sample in the prompt word, the generalization ability of the large language model can be stimulated, so that the large language model can complete the subsequent task requirements according to the question sample. In Figure 6, "Parse the question according to the parameter part and the intent part, and return it in json; example: ...}" represents the prompt word, and "The question is: View the basic information of the AP named A1208-11" represents the original input question.

Then, the large language model performs structured processing on the original input question based on the prompt word, and obtains a structured question of {apname: A1208-11; desc: AP information}. In this structured question, A1208-11 after apname represents the target device identifier (i.e., the device name), indicating that the device corresponding to the target device identifier needs to be operated, and AP information represents the target URL description, which is the question description for the original input question, indicating that the page corresponding to the device's "AP information" needs to be operated.

Then, the large language model queries the target URL address corresponding to the AP information from the feature database. For example, since the similarity between "AP basic information" and AP information is relatively large, "AP basic information" is the first URL description, and the URL address "http://url?apname=$apname" corresponding to "AP basic information" is determined as the target URL address, so that the large language model can obtain the target URL address.

In Figure 6, AP basic information is the main description information of the URL description, AP radio frequency and AP channel are auxiliary description information of the URL description. If the similarity between the AP basic information and the target URL description does not meet the conditions, but the similarity between the AP radio frequency and the target URL description meets the conditions, and/or the similarity between the AP channel and the target URL description meets the conditions, the AP basic information can also be used as the first URL description.

Then, the large language model performs replacement based on the target URL address and structured parameters, such as replacing $apname in the target URL address "http://url?apname=$apname" with A1208-11 to obtain the modified URL address "http://url?apname=A1208-11", where $apname represents a fixed device identifier.

Then, the agent calls http://url?apname=A1208-11, and the operation and maintenance system automatically navigates to the corresponding operation and maintenance management page. The user views the data on the operation and maintenance management page, realizing page access in a human-computer interactive manner.

In an example, when the large language model queries the target URL address corresponding to the AP information from the feature database, if multiple target URL addresses are obtained, such as the URL address http://url?apname=$apname and the URL address http://url2?apname=$apname, two modified URL addresses can be obtained, namely "http://url?apname=A1208-11" and "http://url2?apname=A1208-11".

Then, the two modified URL addresses are returned to the user, and the user selects one of the modified URL addresses, that is, clicks the link of the selected URL address, so that the user can automatically navigate to the operation and maintenance management page corresponding to the link of this URL address, and the user can view the data on the operation and maintenance management page.

In view of the fact that the Copilot agent supports the Text2URL function, an embodiment of the present application proposes a data processing method based on a large language model, which may include:

Step S21, the Copilot agent obtains private domain data of the operation and maintenance system, and the private domain data may include URL description information and multiple operation step description information. The URL description information may be information related to the task, which is description information for the task. For each operation step description information, the operation step description information may be description information for the operation step, which may be a description of the problem for the operation step.

For each operation step description information, the private domain data may also include URL description information and URL address information corresponding to the operation step description information. URL address information is information related to the URL address, and the URL address is used to access the operation and maintenance system. URL description information is information related to the URL description, and the URL description is information used to describe the problem (i.e., the problem corresponding to the operation step description information).

Step S22: The Copilot agent uses private domain data to perform incremental training on the large language model.

Step S23: The Copilot agent uses private domain data to maintain the feature database of the RAG function.

Exemplarily, if the private domain data includes URL description information and multiple operation step description information, the URL description information can be processed to obtain a URL description, which can be description information for the task. For each operation step description information, the operation step description information can be processed to obtain an operation step description, which is a problem description information for the operation step. Then, the correspondence between the URL description and each operation step description can be recorded in the feature database.

For each operation step description information, if the private domain data may include URL description information and URL address information corresponding to the operation step description information, the URL description information may be processed to obtain the URL description, and the URL address information may be processed to obtain the URL address. After obtaining the URL description and the URL address, the corresponding relationship between the URL description and the URL address is recorded in the feature database.

Step S24: The Copilot agent obtains the original input question.

Step S25, the Copilot agent inputs the original input question and prompt words into the large language model, and the large language model processes the original input question based on the prompt words to obtain the target device identification and target URL description. The prompt words may include question samples, question parameter samples, and question description samples.

Step S26: The Copilot agent queries the target URL address corresponding to the target URL description from the feature database, and the target URL address may include a configured fixed device identifier.

Exemplarily, if there is a first URL description matching the target URL description in the feature database, and the first URL description corresponds to multiple operation step descriptions, then for each operation step description: if there is a second URL description matching the operation step description in the feature database, and the second URL description corresponds to a URL address, then the URL address corresponding to the second URL description is determined as the target URL address corresponding to the target URL description, and the large language model can obtain multiple target URL addresses.

Step S27: The Copilot agent replaces the fixed device identifier in the target URL address (ie, the device identifier pre-configured in the URL address) with the target device identifier to obtain a modified URL address.

Step S28, the Copilot agent accesses the operation and maintenance management page of the operation and maintenance system corresponding to the modified URL address, and returns the target operation and maintenance data of the operation and maintenance management page to the user.

For example, since the large language model can obtain multiple target URL addresses, multiple modified URL addresses can be obtained, and the Copilot agent can sequentially access each operation and maintenance management page corresponding to the modified URL address, and sequentially display each operation and maintenance management page to the user. Alternatively, the operation and maintenance management pages corresponding to multiple modified URL addresses can be accessed simultaneously, and multiple operation and maintenance management pages can be displayed to the user at the same time, that is, multiple operation and maintenance management pages can be displayed on one display interface.

In view of the fact that the Copilot agent supports the Text2API function, a data processing method based on a large language model is proposed in an embodiment of the present application. The data processing method may include:

Step S31, the Copilot agent obtains private domain data of the operation and maintenance system, and the private domain data may include API description information and API information. The API information may be information related to an API (API represents an operation method), and an API is a parameter used to access the operation and maintenance system. The API description information may be information related to an API description, and the API description is information used to describe the problem.

Step S32: The Copilot agent uses private domain data to perform incremental training on the large language model.

Step S33: The Copilot agent uses private domain data to maintain the feature database of the RAG function.

If the private domain data includes API description information and API information, the API description information can be processed to obtain the API description, which can be information used to describe the problem, such as description information of the API corresponding function/operation, description information of the problem for calling the API function/operation, etc., without limitation. The API information can be processed to obtain the API, which can be a parameter for accessing the operation and maintenance system. The correspondence between the API description and the API is recorded in the feature database.

Step S34: The Copilot agent obtains the original input question.

Step S35, the Copilot agent inputs the original input question and prompt words into the large language model, and the large language model processes the original input question based on the prompt words to obtain the target device identification and target API description. The prompt words may include question samples, question parameter samples, and question description samples.

Step S36: The Copilot agent queries the target API corresponding to the target API description from the feature database, and the target API may include a configured fixed device identifier.

Exemplarily, if the feature database has a first API description that matches the target API description, and the first API description corresponds to an API, the API corresponding to the first API description is determined as the target API corresponding to the target API description, and the large language model can obtain this target API.

Step S37: The Copilot agent replaces the fixed device identifier in the target API (ie, the device identifier pre-configured in the API) with the target device identifier to obtain a modified API.

Step S38, the Copilot agent calls the modified API to operate the operation and maintenance system to obtain operation result data, and determines target operation and maintenance data based on the operation result data (such as using the operation result data as target operation and maintenance data), and returns the target operation and maintenance data to the user.

In view of the fact that the Copilot agent supports the Text2API function, a data processing method based on a large language model is proposed in an embodiment of the present application. The data processing method may include:

Step S41, the Copilot agent obtains private domain data of the operation and maintenance system, and the private domain data may include API description information and multiple operation step description information. Among them, the API description information may be information related to the task, which is the description information for the task. For each operation step description information, the operation step description information may be the description information for the operation step, which may be the problem description for the operation step.

For each operation step description information, the private domain data may also include API description information and API information corresponding to the operation step description information. API information is information related to the API, and the API is used to access the operation and maintenance system. API description information is information related to the API description, and the API description is information used to describe the problem (i.e., the problem corresponding to the operation step description information).

Step S42: The Copilot agent uses private domain data to perform incremental training on the large language model.

Step S43: The Copilot agent uses private domain data to maintain the feature database of the RAG function.

Exemplarily, if the private domain data includes API description information and multiple operation step description information, the API description information can be processed to obtain the API description. For each operation step description information, the operation step description information can be processed to obtain the operation step description. Then, the correspondence between the API description and each operation step description can be recorded in the feature database. For each operation step description information, if the private domain data can include the API description information and API information corresponding to the operation step description information, the API description information can also be processed to obtain the API description, the API information can be processed to obtain the API, and the correspondence between the API description and the API can be recorded in the feature database.

Step S44: The Copilot agent obtains the original input question.

Step S45, the Copilot agent inputs the original input question and prompt words into the large language model, and the large language model processes the original input question based on the prompt words to obtain the target device identification and target API description. The prompt words may include question samples, question parameter samples, and question description samples.

Step S46: The Copilot agent queries the target API corresponding to the target API description from the feature database, and the target API may include a configured fixed device identifier.

Exemplarily, if there is a first API description in the feature database that matches the target API description, and the first API description corresponds to multiple operation step descriptions, then for each operation step description: if there is a second API description in the feature database that matches the operation step description, and the second API description corresponds to an API, then the API corresponding to the second API description is determined as the target API corresponding to the target API description.

Step S47: The Copilot agent replaces the fixed device identifier in the target API (ie, the device identifier pre-configured in the API) with the target device identifier to obtain a modified API.

Step S48, the Copilot agent calls the modified API to operate the operation and maintenance system to obtain operation result data, and determines target operation and maintenance data based on the operation result data (such as using the operation result data as target operation and maintenance data), and returns the target operation and maintenance data to the user.

For example, since the large language model can obtain multiple target APIs, multiple modified APIs can be obtained, and the Copilot agent can call each modified API in turn to operate the operation and maintenance system.

The processing of step S41 to step S48 is described below with reference to a multi-task process example.

See FIG. 7 , which is a schematic diagram of a multi-task process. Taking the diagnosis of the cause of high AP channel utilization as an example, the multi-task calling process is described. The specific questions and format descriptions are examples and are not fixed.

First, the user needs to check why the channel utilization of a certain AP is high. The user sends the original input question "Why is the channel utilization of the AP named A1208-11 high?" to the Copilot agent.

Then, in order to improve the vector matching degree of the knowledge base, it is necessary to accurately extract the question parameters in the original input question. For this purpose, the Copilot agent needs to obtain the prompt word, and the Copilot agent can send the original input question and the prompt word together to the large language model. In this way, by using the one-shot prompt word technique of the large language model, by constructing a question sample, a question parameter sample, and a question description sample in the prompt word, the generalization ability of the large language model can be stimulated, so that the large language model can complete the subsequent task requirements according to the question sample. In Figure 7, "Parse the question according to the parameter part and the intent part, and return it in json; example: ...}" represents the prompt word, and "The question is: What is the reason for the high channel utilization of the AP named A1208-11" represents the original input question.

Then, the large language model performs structured processing on the original input question based on the prompt word, and obtains the structured question {apname: A1208-11; desc: AP channel utilization is high}. In this structured question, A1208-11 after apname represents the target device identifier (i.e., the device name), indicating that the device corresponding to the target device identifier needs to be operated. AP channel utilization is high, which represents the target API description, which is the problem description for the original input question, indicating that the reason for the high AP channel utilization needs to be queried.

Then, for the structured problem, the large language model performs vector matching of "AP channel utilization is high" with the feature database, and the corresponding result content is: {desc: AP channel utilization is high; alias: diagnosis reason; content: 1. Check the AP while using the terminal; 2. Check the AP's low-speed terminal; 3. Check the AP's high-traffic terminal}, and the large language model obtains the above result content.

In the above results, "AP channel utilization is high" represents the first API description corresponding to the target API description, and "View AP while using the terminal", "View AP's low-speed terminal" and "View AP's high-traffic terminal" represent multiple operation step descriptions corresponding to the first API description.

The large language model breaks down the thought chain based on the returned result content. For example, you can continue to use oneshot prompt words to stimulate the thinking ability of the large language model and break down the diagnostic steps into three steps.

In the first step, the large language model performs structural processing on "viewing AP while using terminal" based on the prompt word, that is, the operation step description of "viewing AP while using terminal" is used as the original input question, and the original input question is structured to obtain the target API description of "viewing AP while using terminal", and A1208-11 continues to be the target device identifier (i.e., device name). The large language model can query the target API corresponding to "viewing AP while using terminal" from the feature database, as shown in Table 2. The target API can be API-1, taking http://API?apname=$apname as an example.

In the second step, the large language model performs structural processing on "view the low-rate terminal of AP" based on the prompt word, that is, the operation step description of "view the low-rate terminal of AP" is used as the original input question, and the original input question is structured to obtain the target API description of "low-rate terminal of AP", and A1208-11 is used as the target device identifier. The large language model queries the target API corresponding to "low-rate terminal of AP" from the feature database, as shown in Table 2, and the target API can be API-2.

In the third step, the large language model performs structural processing on "view AP's high-traffic terminal" based on the prompt word, that is, the operation step description of "view AP's high-traffic terminal" is used as the original input question, and the original input question is structured to obtain the target API description of "AP's high-traffic terminal", and A1208-11 is used as the target device identifier. The large language model queries the target API corresponding to "AP's high-traffic terminal" from the feature database. As shown in Table 2, the target API can be API-3.

Then, the large language model performs replacement based on the target API (i.e., API-1) and the structured parameters (i.e., the target device identifier "A1208-11"), such as replacing $apname in the target API "http://API?apname=$apname" with A1208-11 to obtain the modified API-1 "http://API?apname=A1208-11".

Similarly, $apname in API-2 can be replaced with A1208-11 to obtain the modified API-2, and $apname in API-3 can be replaced with A1208-11 to obtain the modified API-3.

After obtaining the modified API-1, the modified API-2, and the modified API-3, the modified API-1 can be called according to the single-task process to operate the operation and maintenance system, that is, to check the terminals used at the same time by the AP, so as to obtain the query results of whether there are too many terminals used at the same time, and return the query results of whether there are too many terminals used at the same time (target operation and maintenance data) to the user. In addition, the modified API-2 can be called according to the single-task process to operate the operation and maintenance system, that is, to check whether there are low-speed terminals on the AP, so as to obtain the query results of whether there are low-speed terminals, and return the query results of whether there are low-speed terminals to the user. In addition, the modified API-3 can be called according to the single-task process to operate the operation and maintenance system, that is, to check whether there are high-traffic terminals on the AP, so as to obtain the query results of whether there are high-traffic terminals, and return the query results of whether there are high-traffic terminals to the user.

After returning the result of each step to the user, the user is asked whether to execute the recovery strategy. The user can choose whether to complete the problem loop based on the operation recommended by the large language model.

In view of the fact that the Copilot agent supports the Text2SQL function, an embodiment of the present application proposes a data processing method based on a large language model, which may include:

Step S51, the Copilot agent obtains private domain data of the operation and maintenance system, and the private domain data may include SQL description information and SQL definition information. The SQL definition information may be information related to SQL statements, and SQL statements are parameters used to access the operation and maintenance system. In addition, the SQL description information may be information related to SQL descriptions, and SQL descriptions are information used to describe problems.

Step S52: The Copilot agent uses private domain data to perform incremental training on the large language model.

Step S53: The Copilot agent uses private domain data to maintain the feature database of the RAG function.

Exemplarily, if the private domain data includes SQL description information and SQL definition information, the SQL description information can be processed to obtain the SQL description, which can be information used to describe the problem. The SQL definition information can be processed to obtain the SQL definition, which can also be understood as the SQL function, such as the SQL definition can be select, insert, update, delete, etc., and there is no restriction on this SQL definition. Select can represent the query function of the data table, insert can represent the addition function of the data table, update can represent the update function of the data table, and delete can represent the deletion function of the data table.

Then, the correspondence between the SQL description and the SQL definition may be recorded in the feature database.

Step S54: The Copilot agent obtains the original input question.

Step S55, the Copilot agent inputs the original input question and prompt words into the large language model, and the large language model processes the original input question based on the prompt words to obtain the target data table identifier and target SQL description. The prompt words may include question examples, question parameter examples, and question description examples.

Step S56: The Copilot agent queries the target SQL definition corresponding to the target SQL description from the feature database. Exemplarily, if the feature database has a first SQL description that matches the target SQL description, and the first SQL description corresponds to a SQL definition, then the SQL definition corresponding to the first SQL description is determined as the target SQL definition corresponding to the target SQL description, and the large language model can obtain the target SQL definition.

Step S57: The Copilot agent generates a target SQL statement based on the target SQL definition and the target data table identifier, and the target SQL statement may include the target data table identifier.

For example, if the target SQL is defined as select, a target SQL statement for implementing select can be generated, and the target SQL statement includes a target data table identifier, and the target SQL statement is used to perform a select operation on the data table corresponding to the target data table identifier. If the target SQL is defined as insert, update, delete, etc., the generation method and function of the target SQL statement are similar and will not be repeated here.

Step S58, the Copilot agent operates on the data table corresponding to the target data table identifier of the operation and maintenance system by executing the target SQL statement, obtains the operation result data of the data table, determines the target operation and maintenance data based on the operation result data of the data table, and returns the target operation and maintenance data to the user.

It can be seen from the above technical solutions that in the embodiment of the present application, through a Copilot agent as an intermediate layer, the user can interact with the Copilot agent in natural language to complete the interaction of operation and maintenance needs. The Copilot agent converts natural language into instructions and passes them to the operation and maintenance system to realize a conversational intelligent operation and maintenance system, improve the operation and maintenance experience, and realize barrier-free human-computer interaction experience between users and the operation and maintenance system, thereby improving user experience. Based on the capabilities of the large language model, the obstacles to natural language understanding are solved, and the user's input is changed from page clicks to human-computer language interaction. Based on the capabilities of the large language model, the automatic implementation capability based on language interaction is realized for the functions of the operation and maintenance system, such as overview, diagnosis, self-healing closed loop, reports, etc.

The Copilot agent converts user intent into actions to interact with the operation and maintenance system on pages, APIs, data, and knowledge, thereby implementing a conversational intelligent operation and maintenance system and improving the operation and maintenance experience. Through incremental training and fine-tuning of the large language model, the large language model is endowed with industry knowledge attributes, including but not limited to the correspondence between the page description and page URL of the operation and maintenance system, the correspondence between the function description and API of the operation and maintenance system, the operation and maintenance experience of the operation and maintenance system, and basic industry operation and maintenance knowledge. In order to solve the hallucination problem of the large language model and improve accuracy, the RAG knowledge base is used to construct the above knowledge in the form of a knowledge base.

In order to improve the accurate recall rate of user questions and the knowledge base, the user's questions are understood and structured through a large language model, the user's questions are semantically decomposed into subject, predicate and object, the specific parameters are disassembled, and the core question part is used for vector matching with the knowledge base to improve the matching degree.

For multi-step operations, the large language model splits the thought chain, structures the operation description of each step, and matches it with the knowledge base to find the corresponding operation URL or API, replace the structured parameters with the function-call action, and execute the step-by-step operation after the user confirms.

Based on the same application concept as the above method, a data processing device based on a large language model is proposed in the embodiment of the present application. See FIG8 , which is a schematic diagram of the structure of the device. The device includes:

The processing module 81 is used to input the original input question and the prompt word into the large language model, and the large language model processes the original input question based on the prompt word to obtain the target question parameters and the target question description; wherein the prompt word includes a question sample, a question parameter sample and a question description sample;

A query module 82, configured to query a target access parameter corresponding to the target problem description from a feature database; wherein the feature database includes a correspondence between access descriptions and access parameters;

The operation module 83 is used to generate operation and maintenance system operation data based on the target access parameters and the target question parameters, and operate the operation and maintenance system through the operation and maintenance system operation data to obtain the target operation and maintenance data corresponding to the original input question, and return the target operation and maintenance data to the user.

Exemplarily, when the query module 82 queries the target access parameters corresponding to the target problem description from the feature database, it is specifically used to: if there is a first access description matching the target problem description in the feature database, obtain a first data structure corresponding to the first access description from the feature database, the first data structure including the first access description and access parameters, and determine the access parameters in the first data structure as the target access parameters corresponding to the target problem description.

Exemplarily, the processing module 81 is also used to obtain private domain data of the operation and maintenance system before querying the target access parameters corresponding to the target problem description from the feature database, the private domain data including access description information and access parameter information; extract the first description keyword from the access description information, and generate an access description based on the first description keyword; extract the first parameter keyword from the access parameter information, and generate an access parameter based on the first parameter keyword; and record the correspondence between the access description and the access parameter in the feature database.

Exemplarily, when the query module 82 queries the target access parameters corresponding to the target problem description from the feature database, it is specifically used to: if there is a second access description matching the target problem description in the feature database, obtain a second data structure corresponding to the second access description from the feature database, the second data structure including the second access description and multiple operation step descriptions; for each operation step description, if there is a third access description matching the operation step description in the feature database, obtain a third data structure corresponding to the third access description from the feature database, the third data structure including the third access description and access parameters, and determine the access parameters in the third data structure as the access parameters corresponding to the operation step description; determine the access parameters corresponding to each operation step description as the target access parameters corresponding to the target problem description.

Exemplarily, the processing module 81 is also used to obtain private domain data of the operation and maintenance system before querying the target access parameters corresponding to the target problem description from the feature database, the private domain data including access description information and multiple operation step description information; for each operation step description information, the private domain data also includes access description information and access parameter information corresponding to the operation step description information; extract the second description keyword from the access description information, and generate an access description based on the second description keyword; for each operation step description information, extract the operation step keyword from the operation step description information, and generate an operation step description based on the operation step keyword; record the correspondence between the access description and the operation step description in the feature database; for each operation step description information, extract the third description keyword from the access description information corresponding to the operation step description information, and generate an access description based on the third description keyword; extract the second parameter keyword from the access parameter information corresponding to the operation step description information, and generate an access parameter based on the second parameter keyword; and record the correspondence between the access description and the access parameter in the feature database.

Exemplarily, the target problem parameters include a target device identifier, the target problem description includes a target URL description, the target access parameters include a target URL address, and the target URL address includes a configured fixed device identifier; when the operation module 83 generates the operation and maintenance system operation data based on the target access parameters and the target problem parameters, it is specifically used to: replace the fixed device identifier in the target URL address with the target device identifier to obtain a modified URL address as the operation and maintenance system operation data; when the operation module 83 operates the operation and maintenance system through the operation and maintenance system operation data to obtain the target operation and maintenance data corresponding to the original input problem, it is specifically used to: access the operation and maintenance management page of the operation and maintenance system corresponding to the modified URL address, and the operation and maintenance management page includes the target operation and maintenance data.

Exemplarily, the target problem parameters include a target device identifier, the target problem description includes a target API description, the target access parameters include a target API, and the target API includes a configured fixed device identifier; when the operation module 83 generates the operation and maintenance system operation data based on the target access parameters and the target problem parameters, it is specifically used to: replace the fixed device identifier in the target API with the target device identifier to obtain a modified API as the operation and maintenance system operation data; when the operation module 83 operates the operation and maintenance system through the operation and maintenance system operation data to obtain the target operation and maintenance data corresponding to the original input problem, it is specifically used to: call the modified API to operate the operation and maintenance system to obtain operation result data, and determine the target operation and maintenance data based on the operation result data.

Exemplarily, the target problem parameters include a target data table identifier, the target problem description includes a target SQL description, and the target access parameters include a target SQL definition; when the operation module 83 generates the operation and maintenance system operation data based on the target access parameters and the target problem parameters, it is specifically used to: generate a target SQL statement based on the target SQL definition and the target data table identifier as the operation and maintenance system operation data, and the target SQL statement includes the target data table identifier; when the operation module 83 operates the operation and maintenance system through the operation and maintenance system operation data to obtain the target operation and maintenance data corresponding to the original input problem, it is specifically used to: operate the data table corresponding to the target data table identifier of the operation and maintenance system by executing the target SQL statement to obtain the operation result data of the data table; determine the target operation and maintenance data based on the operation result data of the data table.

Based on the same application concept as the above method, an electronic device is proposed in an embodiment of the present application, as shown in Figure 9, including: a processor 91 and a machine-readable storage medium 92, the machine-readable storage medium 92 stores machine-executable instructions that can be executed by the processor 91; the processor 91 is used to execute the machine-executable instructions to implement the data processing method based on the large language model disclosed in the above example of the present application.

Based on the same application concept as the above method, an embodiment of the present application also provides a machine-readable storage medium, on which a number of computer instructions are stored. When the computer instructions are executed by a processor, the data processing method based on a large language model disclosed in the above example of the present application can be implemented.

The above-mentioned machine-readable storage medium may be any electronic, magnetic, optical or other physical storage device, which may contain or store information, such as executable instructions, data, etc. For example, the machine-readable storage medium may be: RAM (Radom Access Memory), volatile memory, non-volatile memory, flash memory, storage drive (such as hard disk drive), solid state drive, any type of storage disk (such as CD, DVD, etc.), or similar storage medium, or a combination thereof.

Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the embodiments of the present application may adopt 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.

The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims (15)

1. A method for processing data based on a large language model, the method comprising:

Inputting an original input problem and a prompt word into a large language model, and processing the original input problem based on the prompt word through the large language model to obtain a target problem parameter and a target problem description; the prompt word comprises a problem sample, a problem parameter sample and a problem description sample;

inquiring target access parameters corresponding to the target problem description from a feature database; the feature database comprises a corresponding relation between access description and access parameters;

generating operation data of an operation and maintenance system based on the target access parameter and the target problem parameter;

And operating the operation and maintenance system through the operation and maintenance system operation data to obtain target operation and maintenance data corresponding to the original input problem, and returning the target operation and maintenance data to a user.

2. The method of claim 1, wherein the step of determining the position of the substrate comprises,

The querying the target access parameters corresponding to the target problem description from the feature database comprises the following steps:

If the first access description matched with the target problem description exists in the feature database, a first data structure corresponding to the first access description is obtained from the feature database, the first data structure comprises the first access description and access parameters, and the access parameters in the first data structure are determined to be target access parameters corresponding to the target problem description.

3. The method of claim 2, wherein prior to querying the target access parameters corresponding to the target problem description from the feature database, the method further comprises:

acquiring private domain data of an operation and maintenance system, wherein the private domain data comprises access description information and access parameter information; extracting a first description keyword from the access description information, and generating an access description based on the first description keyword; extracting a first parameter keyword from the access parameter information, and generating an access parameter based on the first parameter keyword;

And recording the corresponding relation between the access description and the access parameter in the characteristic database.

4. The method of claim 1, wherein the step of determining the position of the substrate comprises,

The querying the target access parameters corresponding to the target problem description from the feature database comprises the following steps:

If a second access description matched with the target problem description exists in the characteristic database, acquiring a second data structure corresponding to the second access description from the characteristic database, wherein the second data structure comprises the second access description and a plurality of operation step descriptions;

For each operation step description, if a third access description matched with the operation step description exists in the feature database, acquiring a third data structure corresponding to the third access description from the feature database, wherein the third data structure comprises the third access description and access parameters, and determining the access parameters in the third data structure as the access parameters corresponding to the operation step description;

and determining the access parameter corresponding to each operation step description as the target access parameter.

5. The method of claim 4, wherein prior to querying the target access parameters corresponding to the target problem description from the feature database, the method further comprises:

Acquiring private domain data of an operation and maintenance system, wherein the private domain data comprises access description information and a plurality of operation step description information; the private domain data further comprises access description information and access parameter information corresponding to the operation step description information aiming at each operation step description information;

Extracting a second description keyword from the access description information, and generating an access description based on the second description keyword; extracting operation step keywords from the operation step description information for each operation step description information, and generating operation step descriptions based on the operation step keywords; recording the corresponding relation between the access description and the operation step description in a characteristic database;

For each operation step description information, extracting a third description keyword from access description information corresponding to the operation step description information, and generating access description based on the third description keyword; extracting a second parameter keyword from the access parameter information corresponding to the operation step description information, and generating an access parameter based on the second parameter keyword; and recording the corresponding relation between the access description and the access parameter in the characteristic database.

6. The method of any of claims 1-5, wherein the target problem parameter comprises a target device identification, the target problem description comprises a target URL description, the target access parameter comprises a target URL address, and the target URL address comprises a configured fixed device identification;

the generating operation data of the operation and maintenance system based on the target access parameter and the target problem parameter comprises: replacing the fixed equipment identifier in the target URL address with the modified URL address obtained by the target equipment identifier, and taking the modified URL address as the operation data of the operation and maintenance system;

The operation of the operation and maintenance system through the operation and maintenance system operation data to obtain a target operation and maintenance data corresponding to the original input problem comprises the following steps: and accessing an operation and maintenance management page of the operation and maintenance system corresponding to the modified URL address, wherein the operation and maintenance management page comprises the target operation and maintenance data.

7. The method of any of claims 1-5, wherein the target problem parameter comprises a target device identification, the target problem description comprises a target API description, the target access parameter comprises a target API, and the target API comprises a configured fixed device identification;

The generating operation data of the operation and maintenance system based on the target access parameter and the target problem parameter comprises: replacing the fixed equipment identifier in the target API with the modified API obtained by the target equipment identifier, and taking the modified API as the operation data of the operation and maintenance system;

the operation of the operation and maintenance system through the operation and maintenance system operation data to obtain a target operation and maintenance data corresponding to the original input problem comprises the following steps: and calling the modified API to operate the operation and maintenance system to obtain operation result data, and determining the target operation and maintenance data based on the operation result data.

8. The method of any of claims 1-5, wherein the target problem parameter comprises a target data table identification, the target problem description comprises a target SQL description, and the target access parameter comprises a target SQL definition;

The generating operation data of the operation and maintenance system based on the target access parameter and the target problem parameter comprises: generating a target SQL statement based on the target SQL definition and the target data table identifier, wherein the target SQL statement comprises the target data table identifier as the operation data of the operation and maintenance system;

The operation of the operation and maintenance system through the operation and maintenance system operation data to obtain a target operation and maintenance data corresponding to the original input problem comprises the following steps: operating a data table corresponding to the target data table identifier of the operation and maintenance system by executing the target SQL statement to obtain operation result data of the data table; and determining the target operation data based on the operation result data of the data table.

9. A large language model based data processing apparatus, the apparatus comprising:

The processing module is used for inputting the original input problem and the prompt word into a large language model, and processing the original input problem based on the prompt word through the large language model to obtain target problem parameters and target problem description; the prompt word comprises a problem sample, a problem parameter sample and a problem description sample;

the query module is used for querying target access parameters corresponding to the target problem description from the feature database; the feature database comprises a corresponding relation between access description and access parameters;

and the operation module is used for generating operation data of the operation and maintenance system based on the target access parameters and the target problem parameters, operating the operation and maintenance system through the operation data of the operation and maintenance system, obtaining target operation and maintenance data corresponding to the original input problem, and returning the target operation and maintenance data to a user.

10. The apparatus of claim 9, wherein the device comprises a plurality of sensors,

The query module is specifically configured to, when querying the target access parameter corresponding to the target problem description from the feature database: if a first access description matched with the target problem description exists in the feature database, acquiring a first data structure corresponding to the first access description from the feature database, wherein the first data structure comprises the first access description and access parameters, and determining the access parameters in the first data structure as target access parameters corresponding to the target problem description;

The processing module is further used for acquiring private domain data of the operation and maintenance system before inquiring target access parameters corresponding to the target problem description from the characteristic database, wherein the private domain data comprises access description information and access parameter information; extracting a first description keyword from the access description information, and generating an access description based on the first description keyword; extracting a first parameter keyword from the access parameter information, and generating an access parameter based on the first parameter keyword; and recording the corresponding relation between the access description and the access parameter in the characteristic database.

11. The apparatus of claim 9, wherein the device comprises a plurality of sensors,

The query module is specifically configured to, when querying the target access parameter corresponding to the target problem description from the feature database: if a second access description matched with the target problem description exists in the characteristic database, acquiring a second data structure corresponding to the second access description from the characteristic database, wherein the second data structure comprises the second access description and a plurality of operation step descriptions; for each operation step description, if a third access description matched with the operation step description exists in the feature database, acquiring a third data structure corresponding to the third access description from the feature database, wherein the third data structure comprises the third access description and access parameters, and determining the access parameters in the third data structure as the access parameters corresponding to the operation step description; determining access parameters corresponding to each operation step description as target access parameters corresponding to the target problem description;

The processing module is further used for acquiring private domain data of the operation and maintenance system before inquiring target access parameters corresponding to the target problem description from the feature database, wherein the private domain data comprises access description information and a plurality of operation step description information; for each operation step description information, the private domain data also comprises access description information and access parameter information corresponding to the operation step description information; extracting a second description keyword from the access description information, and generating an access description based on the second description keyword; extracting operation step keywords from the operation step description information for each operation step description information, and generating operation step descriptions based on the operation step keywords; recording the corresponding relation between the access description and the operation step description in a characteristic database; for each operation step description information, extracting a third description keyword from access description information corresponding to the operation step description information, and generating access description based on the third description keyword; extracting a second parameter keyword from the access parameter information corresponding to the operation step description information, and generating an access parameter based on the second parameter keyword; and recording the corresponding relation between the access description and the access parameter in the characteristic database.

12. The apparatus of any of claims 9-11, wherein the target issue parameter comprises a target device identification, the target issue description comprises a target URL description, the target access parameter comprises a target URL address, and the target URL address comprises a configured fixed device identification;

The operation module is specifically configured to, when generating operation data of the operation and maintenance system based on the target access parameter and the target problem parameter: replacing the fixed equipment identifier in the target URL address with the modified URL address obtained by the target equipment identifier, and taking the modified URL address as the operation data of the operation and maintenance system;

The operation module operates the operation and maintenance system according to the operation and maintenance system operation data, and is specifically used for obtaining target operation and maintenance data corresponding to the original input problem: and accessing an operation and maintenance management page of the operation and maintenance system corresponding to the modified URL address, wherein the operation and maintenance management page comprises the target operation and maintenance data.

13. The apparatus of any of claims 9-11, wherein the target issue parameter comprises a target device identification, the target issue description comprises a target API description, the target access parameter comprises a target API, and the target API comprises a configured fixed device identification;

The operation module is specifically configured to, when generating operation data of the operation and maintenance system based on the target access parameter and the target problem parameter: replacing the fixed equipment identifier in the target API with the modified API obtained by the target equipment identifier, and taking the modified API as the operation data of the operation and maintenance system;

The operation module operates the operation and maintenance system according to the operation and maintenance system operation data, and is specifically used for obtaining target operation and maintenance data corresponding to the original input problem: and calling the modified API to operate the operation and maintenance system to obtain operation result data, and determining the target operation and maintenance data based on the operation result data.

14. The apparatus of any of claims 9-11, wherein the target problem parameter comprises a target data table identification, the target problem description comprises a target SQL description, and the target access parameter comprises a target SQL definition; the operation module is specifically configured to, when generating operation data of the operation and maintenance system based on the target access parameter and the target problem parameter: generating a target SQL statement based on the target SQL definition and the target data table identifier, wherein the target SQL statement comprises the target data table identifier as the operation data of the operation and maintenance system;

The operation module operates the operation and maintenance system according to the operation and maintenance system operation data, and is specifically used for obtaining target operation and maintenance data corresponding to the original input problem: operating a data table corresponding to the target data table identifier of the operation and maintenance system by executing the target SQL statement to obtain operation result data of the data table; and determining the target operation data based on the operation result data of the data table.

15. An electronic device, comprising: a processor and a machine-readable storage medium storing machine-executable instructions executable by the processor; the processor is configured to execute machine executable instructions to implement the method of any of claims 1-8.