Resource arrangement method and device based on generative AI technology

Abstract

The invention relates to a resource arrangement method and device based on a generative AI technology. The method comprises the steps that cue words are obtained, a trained generative AI model is used for verifying and analyzing the cue words, if the cue words are wrong, a correction prompt is output, and if the cue words are correct, an arrangement view is generated according to the cue words; according to the arrangement view, matching underlying resources, and performing resource arrangement; and combining the arranged resources according to the arrangement view, performing function test on the combined resources, and outputting a final resource combination after the function test is correct. Compared with the prior art, the flexibility and the automation degree of resource arrangement and the reliability of an arrangement system are improved, and the labor cost, the time cost and the operation threshold required by arrangement are reduced.

Description

Technical Field

[0001] The present invention relates to the field of resource orchestration, and in particular to a resource orchestration method and device based on generative AI technology. Background Art

[0002] As an important technical means, resource orchestration is widely used in various cloud computing platforms, big data processing systems and distributed application environments. Its main purpose is to achieve efficient utilization and management of resources to meet the needs of different business scenarios. By automatically deploying, orchestrating and optimizing resources, resource orchestration can improve the flexibility, scalability and stability of the system and provide users with convenient and efficient resource management and orchestration solutions.

[0003] The traditional resource orchestration architecture is shown in the figure below: Figure 1 As shown, at present, resource combination orchestration is mostly implemented through hard coding or configuration according to the orchestration requirements, which cannot achieve free, fast and flexible orchestration; the user learning threshold is high, and the user determines the timing of resource orchestration, and the timing accuracy cannot be guaranteed; parameter mapping and transmission between resources requires a lot of combination development work, high labor cost investment, and a long delivery cycle.

[0004] Traditional resource orchestration process Figure 2 As shown in the figure, in the traditional solution, users submit orders through the computer page to activate the combination product. Then, the order data is verified by the hard-coded or configuration file arrangement module to ensure that the order meets the preset combination arrangement template requirements. Then, the manual arrangement development module intervenes to split, combine and develop the products, arrange from product 1 to product n, and complete the release and launch process after functional testing. In the whole process, there is a lot of manual intervention, from order verification to product arrangement to functional testing, all of which require subjective operations of personnel, making the process relatively time-consuming and complicated.

[0005] The disadvantages of the prior art are as follows:

[0006] The orchestration process lacks flexibility. Resource orchestration is implemented through hard coding or configuration files, resulting in a lack of flexibility and configurability in the orchestration process. This rigid implementation method makes it difficult for the orchestration solution to adapt to different demand changes, and it is impossible to quickly and flexibly perform free combination orchestration, which increases maintenance costs and the complexity of system updates.

[0007] Orchestration lacks automation and has a high threshold for operation. The existing resource orchestration technology has a high learning threshold, and users need to have in-depth professional knowledge and experience to perform orchestration operations. In addition, the timing of resource orchestration is completely dependent on the user, lacking automation and intelligence, and unable to guarantee the timing accuracy of resource combination orchestration. This dependence increases the user's workload and reduces operational efficiency and system reliability.

[0008] Orchestration requires a lot of combined development work. In order to ensure parameter mapping and transmission between resources, the existing resource orchestration combination requires a lot of combined development work. This means that a lot of human resources and time are needed for development and testing, which increases the cost and delivery cycle of the project. At the same time, due to human factors, there may be errors and vulnerabilities, which further increases the difficulty of development and maintenance. Summary of the invention

[0009] The purpose of the present invention is to provide a resource orchestration method and device based on generative AI technology. By using a generative AI model to perform resource orchestration and output correction prompts for incorrectly input prompt words, the flexibility and automation of resource orchestration and the reliability of the orchestration system are improved, and the manpower cost, time cost and operation threshold required for orchestration are reduced.

[0010] The purpose of the present invention can be achieved by the following technical solutions:

[0011] A resource orchestration method based on generative AI technology, the method comprising:

[0012] Obtain prompt words, and use the trained generative AI model to verify and parse the prompt words. If the prompt words are wrong, a correction prompt is output; if the prompt words are correct, an arrangement view is generated based on the prompt words.

[0013] According to the orchestration view, match the underlying resources and perform resource orchestration;

[0014] Combine the arranged resources according to the arrangement view, perform functional tests on the combined resources, and output the final resource combination after the functional tests are correct.

[0015] Furthermore, the prompt words include resource type, resource name, resource specification, operation type and belonging project, and the arrangement view includes resource combination and arrangement order.

[0016] Furthermore, the training steps of the generative AI model include:

[0017] Collect training data, including resource type, resource name, resource specification, operation type, and project to which it belongs;

[0018] Preprocess the training data to obtain training samples;

[0019] Select the generative AI model to be trained and determine the loss function required for training;

[0020] Use training samples to train generative AI models;

[0021] Carry out training for more than 2 times, and select the generative AI model with the smallest loss function after training as the final generative AI model.

[0022] Furthermore, the preprocessing includes:

[0023] Randomly combine training data and construct expected prompt words as correct input;

[0024] Generate an orchestration view based on the correct input;

[0025] Randomly combine training data and construct unexpected prompt words as error input;

[0026] Generate correction prompts based on incorrect input;

[0027] Integrate constructed data and generated data as training samples.

[0028] Furthermore, the loss function is:

[0029]

[0030] Among them, λ1 is the completeness matching loss weight, λ2 is the confidence loss weight, and N c is the number of correct inputs, N i is the number of incorrect inputs, Used to evaluate the completeness of generated orchestration views, Used to evaluate the completeness of generated correction tips, represents the confidence loss of the kth training sample.

[0031] Furthermore, the L ci The specific solution formula is:

[0032]

[0033] Among them, α is the adjustment parameter, C i Indicates the completeness of the i-th orchestration view, represents the arrangement view word vector generated by training sample i, u i Represents the correct arrangement view word vector of training sample i;

[0034] Said The specific solution formula is:

[0035]

[0036] Among them, β is the adjustment parameter, C j Indicates the completeness of the jth revision prompt, represents the modified prompt word vector generated by training sample j, v jRepresents the correct modified hint word vector of training sample j;

[0037] Said The specific solution formula is:

[0038]

[0039] Among them, y k is a parameter used to adjust the weights of correct and incorrect judgments, p k Represents the judgment result of the kth training sample.

[0040] Furthermore, the verification and analysis of the prompt word also includes:

[0041] When special characters and noise are detected in the prompt word, the special characters and noise are removed from the prompt word;

[0042] When the length of the prompt word exceeds the maximum input length limit of the model, the prompt word is staged and padded and input into the model in segments.

[0043] Furthermore, the underlying resources include a cloud host, a cloud hard disk and a cloud database.

[0044] Furthermore, the resource orchestration method includes performing resource orchestration tasks sequentially and performing multiple resource orchestration tasks asynchronously.

[0045] A resource orchestration device based on generative AI technology is designed using any of the resource orchestration methods based on generative AI technology as described above, comprising:

[0046] ROE (Resource Orchestration Engine) module, used to obtain prompt words, train generative AI models, and use the trained generative AI models to verify and parse prompt words. If the prompt words are wrong, a correction prompt is output. If the prompt words are correct, an orchestration view is generated according to the prompt words.

[0047] Resource registration module, used to manage resource list, resource operation, resource registration, resource record and resource calling mode, including synchronous calling and asynchronous calling;

[0048] Resource management module, used to manage underlying resource corresponding plug-ins and develop underlying resource plug-ins;

[0049] The task-scheduled (resource orchestration) module is used to parse the orchestration view, query, match and call the underlying resources through the resource registration module and resource management module, and perform resource orchestration.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] 1. The present invention uses a generative AI model to intelligently verify and parse the input prompt words, generate an orchestration view to orchestrate cloud resources, and resources can be randomly combined and orchestrated, which improves the flexibility and automation of resource orchestration and reduces the manpower and time costs required for orchestration. At the same time, the generative AI model can output correction prompts for input erroneous prompt words, guide users to input standardized prompt words, lower the threshold of orchestration operations, and improve system reliability;

[0052] 2. During the orchestration process, resource orchestration tasks can be performed sequentially or multiple resource orchestration tasks can be performed asynchronously, which improves the efficiency and accuracy of resource orchestration. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 Orchestrate architecture diagrams for traditional resources;

[0054] Figure 2 Compile flow charts for traditional resources;

[0055] Figure 3 is a flow chart of the method of the present invention;

[0056] Figure 4 This is an example diagram of the device of the present invention. DETAILED DESCRIPTION

[0057] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0058] Example 1

[0059] This embodiment discloses a resource orchestration method based on generative AI technology. Figure 3 As shown, including:

[0060] Obtain prompt words, and use the trained generative AI model to verify and parse the prompt words. If the prompt words are wrong, a correction prompt is output; if the prompt words are correct, an arrangement view is generated based on the prompt words.

[0061] According to the orchestration view, match the underlying resources and perform resource orchestration;

[0062] Combine the arranged resources according to the arrangement view, perform functional tests on the combined resources, and output the final resource combination after the functional tests are correct.

[0063] The user enters the prompt word through the interactive interface or command line, and the prompt word content needs to include the type of required resources, operation and other necessary parameter information. For example, the user gives the prompt word: "Create a virtual machine with 2 cores and 4GB memory in project A".

[0064] To use the generative AI model to verify and parse the prompt words, you first need to train the AI ​​model. The specific training process is as follows:

[0065] Step 1: Collect training data and randomly combine them to construct prompt words that meet expectations as correct inputs. Generate an arrangement view based on the correct inputs. Randomly combine training data to construct prompt words that do not meet expectations as incorrect inputs. Generate correction prompts based on incorrect inputs. Integrate constructed data and generated data as training samples.

[0066] Step 2: Model selection. We chose GPT (Generative Pre-trained Transformer) as the model. GPT is a pre-trained language model based on the Transformer architecture with strong generation and generalization capabilities.

[0067] Step 3: Determine the loss function. For the correct prompt word, the arrangement view generated for each prompt word matches the correct arrangement view, and the completeness of the generated arrangement view is evaluated. The formula is as follows:

[0068]

[0069] Among them, α is the adjustment parameter, C i Indicates the completeness of the i-th orchestration view, represents the arrangement view word vector generated by training sample i, u i Represents the correct arrangement view word vector of training sample i;

[0070] For incorrect prompt words, the correction words generated for each prompt word are matched with the correct correction words to evaluate the matching completeness. The formula is as follows:

[0071]

[0072] Among them, β is the adjustment parameter, C j Indicates the completeness of the jth revision prompt, represents the modified prompt word vector generated by training sample j, v j Represents the correct modified hint word vector of training sample j;

[0073] In addition, the correct prompt word generates an arrangement view and is recorded as true, and the generated correction word is recorded as false. The incorrect prompt word generates an arrangement view and is recorded as false, and the generated correction word is recorded as true. This scheme uses probability p k It represents the judgment result of the kth sample, which indicates the confidence level of the model in the completeness of the prompt word. The confidence loss of the kth sample is expressed by cross entropy. The specific solution formula is:

[0074]

[0075] Among them, y k is a parameter used to adjust the weights of correct and incorrect judgments, p k Represents the judgment result of the kth training sample.

[0076] The overall loss function is as follows:

[0077]

[0078] Among them, λ1 is the completeness matching loss weight, λ2 is the confidence loss weight, and N c is the number of correct inputs, N i is the number of incorrect inputs, Used to evaluate the completeness of generated orchestration views, Used to evaluate the completeness of generated correction tips, represents the confidence loss of the kth training sample.

[0079] Step 4: Model training. Use the preprocessed data in step 1 as training samples to fine-tune the GPT model.

[0080] Step 5: Model selection. After multiple trainings, the training model with the smallest loss is selected as the final model for generating views.

[0081] When collecting training data, common resource management requirements include basic resource information, status, operation type, and dependencies with other resources. This data will provide important reference and support for subsequent resource orchestration and management.

[0082] The specific resources are shown in Table 1 below.

[0083] Table 1 Resource management specific data table

[0084]

[0085] The model will generate an arrangement view for correct prompt word input, and will generate corresponding correction prompts for incorrect prompt word input to help users correct the incorrect input until the user inputs correctly.

[0086] For example, the correct prompt may be: "Create a virtual machine with 4 cores and 8GB of memory in Project A". The wrong situation is that when the resource type is "virtual machine", the resource name is set to "Network001" or the resource specification is set to "8GB RAM".

[0087] When the user arranges input: "Create a virtual machine with 4 cores and 8GB of memory in Network001" or "Create a virtual machine with 8GB of RAM in ProjectA", AI detects the input error and can give a correction prompt: "Please provide the resource specifications of the virtual machine, for example: a virtual machine with 4 cores and 8GB of memory".

[0088] In this way, validation can not only help users describe their needs more accurately, but also improve the system's ability to understand and process user input, thereby improving the overall user experience and system reliability.

[0089] Based on the instructions entered by the user and the knowledge trained by the model, the resource list submitted by the user is dynamically arranged and an arrangement view is generated. This view shows the combination and arrangement order of resources, providing guidance and basis for subsequent resource arrangement. The arrangement view is shown below.

[0090]

[0091] Then, according to the orchestration view, the underlying resources are matched and resource orchestration is performed. These resources can be cloud hosts, cloud storage, etc. The execution of resource orchestration tasks can be sequential. After the first resource orchestration task is completed, the result is passed through and the next resource orchestration task is executed. At the same time, the resource orchestration task can also be executed asynchronously, and multiple resource orchestration tasks are executed at the same time until the last task of the orchestration view is completed.

[0092] Finally, the orchestrated resources are combined according to the orchestration view, and functional tests are performed on the combined resources. After the functional tests are correct, the final resource combination is output to achieve rapid development and deployment of the entire product.

[0093] Example 2

[0094] This embodiment provides a resource orchestration device based on generative AI technology based on the specific implementation steps of a resource orchestration method based on generative AI technology disclosed in the above embodiment 1, such as Figure 4 As shown, including:

[0095] ROE (Resource Orchestration Engine) module, used to obtain prompt words, train generative AI models, and use the trained generative AI models to verify and parse prompt words. If the prompt words are wrong, a correction prompt is output. If the prompt words are correct, an orchestration view is generated according to the prompt words.

[0096] Resource registration module, used to manage resource list, resource operation, resource registration, resource record and resource calling mode, including synchronous calling and asynchronous calling;

[0097] Resource management module, used to manage underlying resource corresponding plug-ins and develop underlying resource plug-ins;

[0098] The task-scheduled (resource orchestration) module is used to parse the orchestration view and use the resource registration module and resource management module to query, match and call the underlying resources for resource orchestration.

[0099] The user inputs instructions through the interactive interface or command line, and the ROE module is responsible for receiving and parsing the user's instructions, identifying the resource type and corresponding operation contained therein as the basis for subsequent arrangement. The prompt word content needs to include the type of required resources, operation and other necessary parameter information.

[0100] The ROE module has a built-in self-trained CMP (Cloud Management Platform)-ROE resource orchestration model. The model training takes into account the pre-dependency order relationship between resources. Through the pre-training of the large model, the dependency order between resources is learned, thereby realizing intelligent orchestration and combination, making the resource orchestration timing more flexible and accurate.

[0101] The ROE module verifies the instructions and parameters entered by the user to ensure that they meet the expected format and specifications. When the prompt words entered by the user are unclear or incomplete, the ROE module can ask the user corresponding questions to obtain more information, thereby improving the user's input and avoiding problems in subsequent editing caused by input errors.

[0102] The resource registration module is used to manage different types of resource lists, such as ECS (Elastic Compute Service), EVS (Elastic Volume Service), etc. The resource list records the basic information and properties of each resource, including the resource type, name, configuration specifications, etc.

[0103] The resource registration module is used to manage various operations of resources, such as starting, stopping, and configuring resources. For each operation, the system records the parameters and return results of the operation for subsequent calls and processing.

[0104] The resource registration module is used to uniformly register the underlying resources to the registration center and record detailed information of the resources. This information includes the resource code, name, operation type, interface address, etc. Through the resource registration module, the system can uniformly manage and access various resources, improving resource availability and management efficiency.

[0105] The resource registration module can also manage the resource calling method synchronously or asynchronously, and manage and record the calling method. In this way, the appropriate calling method can be selected according to actual needs and the smooth execution of the call can be ensured.

[0106] The resource management module is responsible for managing the plug-ins corresponding to the underlying cloud resources. This includes operations such as plug-in registration, loading, and unloading. During the plug-in management process, the system will define the plug-in interface specifications, including the input parameters, output results, and supported operation types required by the plug-in. This ensures that the plug-in can be correctly called and used in the system.

[0107] The resource management module is used to develop plug-ins for various underlying cloud resources. Developers develop corresponding plug-ins based on specific needs to implement corresponding functions, such as cloud hosts, cloud hard disks, cloud databases, etc. Through plug-in development, the system can support various types of cloud resources and provide users with rich functions and services.

[0108] The pluggable resource support capability of resource plug-ins is an important feature, which enables the system to dynamically load and use different resource plug-ins as needed, thereby achieving flexible resource support and management. Users can choose the appropriate resource plug-in according to actual needs without being restricted by fixed resource types and implementation methods.

[0109] When the task-scheduled module runs, it first parses the orchestration view output by the ROE module, which describes the combination and scheduling order of resources. The parsed orchestration view is converted into a data structure that the system can process for subsequent resource matching and scheduling. The parsing process involves extracting and parsing information such as resource type, quantity, and operation type in the orchestration view to ensure that the system can accurately understand user needs and resource scheduling planning.

[0110] According to the parsed orchestration view information, the task-scheduled module uses the resource registration module and the resource management module to query, match, and map resources without manual intervention. This automated process includes querying, matching, and appropriate resource mapping for different types of resources to ensure that the system efficiently allocates and orchestrates resources based on user needs and system status, maximizing user needs and optimizing system performance.

[0111] Once the resource orchestration task is completed, the task-scheduled module is responsible for receiving the synchronization results returned by the resource plug-in set and passing the results to the next resource orchestration task. At the same time, it is also responsible for asynchronously receiving the orchestration results and executing the corresponding next task in the orchestration view until the last task in the orchestration view is completed.

[0112] The specific details of the above modules can be understood by referring to the relevant descriptions and effects in Example 1.

[0113] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A resource orchestration method based on generative AI technology, characterized in that: The method comprises: Obtain prompt words, use the trained generative AI model to verify and parse the prompt words, output correction prompts if the prompt words are wrong, and generate an arrangement view based on the prompt words if the prompt words are correct; According to the orchestration view, match the underlying resources and perform resource orchestration; Combine the arranged resources according to the arrangement view, perform functional tests on the combined resources, and output the final resource combination after the functional tests are correct.

2. According to claim 1, a resource orchestration method based on generative AI technology is characterized in that: The prompt words include resource type, resource name, resource specification, operation type and belonging project, and the arrangement view includes resource combination and arrangement sequence.

3. According to the resource arrangement method based on generative AI technology according to claim 1, it is characterized in that: The training steps of the generative AI model include: Collect training data, including resource type, resource name, resource specification, operation type, and project to which it belongs; Preprocess the training data to obtain training samples; Select the generative AI model to be trained and determine the loss function required for training; Use training samples to train generative AI models; Carry out training for more than 2 times, and select the generative AI model with the smallest loss function after training as the final generative AI model.

4. According to claim 3, a resource orchestration method based on generative AI technology is characterized in that: The pre-processing comprises: Randomly combine training data and construct expected prompt words as correct input; Generate an orchestration view based on the correct input; Randomly combine training data and construct unexpected prompt words as error inputs; Generate correction prompts based on incorrect input; Integrate constructed data and generated data as training samples.

5. According to the resource arrangement method based on generative AI technology in claim 3, it is characterized in that: The loss function is: Among them, λ1 is the completeness matching loss weight, λ2 is the confidence loss weight, and N c is the number of correct inputs, N i is the number of incorrect inputs, Used to evaluate the completeness of generated orchestration views, Used to evaluate the completeness of generated correction tips, represents the confidence loss of the kth training sample.

6. A resource orchestration method based on generative AI technology according to claim 5, characterized in that: Said The specific solution formula is: Among them, α is the adjustment parameter, C i Indicates the completeness of the i-th orchestration view, represents the arrangement view word vector generated by training sample i, u i Represents the correct arrangement view word vector of training sample i; Said The specific solution formula is: Among them, β is the adjustment parameter, C j Indicates the completeness of the jth revision tip, represents the modified prompt word vector generated by training sample j, v j Represents the correct modified hint word vector of training sample j; Said The specific solution formula is: Among them, y k is a parameter used to adjust the weights of correct and incorrect judgments, p k Represents the judgment result of the kth training sample.

7. The resource orchestration method based on generative AI technology according to claim 1, characterized in that: The verification and analysis of the prompt word also includes: When special characters and noise are detected in the prompt word, the special characters and noise are removed from the prompt word; When the length of the prompt word exceeds the maximum input length limit of the model, the prompt word is staged and padded and input into the model in segments.

8. The resource orchestration method based on generative AI technology according to claim 1, characterized in that: The underlying resources include cloud hosts, cloud hard disks, and cloud databases.

9. The resource orchestration method based on generative AI technology according to claim 1, characterized in that: The resource orchestration method includes performing resource orchestration tasks sequentially and performing multiple resource orchestration tasks asynchronously.

10. A resource orchestration device based on generative AI technology, characterized in that: include: ROE module, used to obtain prompt words, train the generative AI model, and use the trained generative AI model to verify and parse the prompt words. If the prompt words are wrong, a correction prompt is output. If the prompt words are correct, an arrangement view is generated according to the prompt words. Resource registration module, used to manage resource list, resource operation, resource registration, resource record and resource calling mode, including synchronous calling and asynchronous calling; Resource management module, used to manage and develop underlying resources; The task-scheduled module is used to parse the orchestration view, query, match and call the underlying resources through the resource registration module and resource management module, and perform resource orchestration.

Images