Task execution method, task execution apparatus, electronic device and storage medium

Abstract

Disclosed in the embodiments of the present application are a task execution method, an electronic device and a storage medium, which are used for improving the efficiency of executing a task corresponding to a standard procedure. The task execution method comprises: receiving a task request; on the basis of the task request, acquiring a first standard procedure from a standard procedure information base; on the basis of the task request, acquiring a first sub-standard procedure from the first standard procedure, the first sub-standard procedure being a sub-standard procedure of the first standard procedure, the first sub-standard procedure being used for instructing a plurality of first sub-tasks to be successively executed on the basis of a jump relationship, the first sub-standard procedure comprising a first root node, and the first root node being used for instructing executing a start sub-task among the plurality of first sub-tasks; and, on the basis of the first sub-standard procedure, executing the plurality of first sub-tasks.

Description

A task execution method, a task execution device, an electronic device, and a storage medium.

[0001] This application claims priority to Chinese Patent Application No. 202411333787.3, filed with the State Intellectual Property Office of China on September 23, 2024, entitled "A Task Execution Method and Task Execution Device, Electronic Device and Storage Medium", the entire contents of which are incorporated herein by reference. Technical Field

[0002] This application relates to the task domain in computer technology, specifically to a task execution method and device, electronic device, and storage medium. Background Technology

[0003] The application scenarios for large models are constantly expanding. Large models can be used to solve complex problems, and agent frameworks integrate multiple general-purpose large models. Within the agent framework, agents are endowed with the general capabilities of large models to solve one or more specific types of problems, and can handle even more complex issues. Therefore, large models play a crucial role in task planning and execution.

[0004] In providing solutions to tasks, introducing standard procedures can handle a relatively fixed type of problem and achieve the completion of the entire task. By introducing standard procedures into the agent framework, the large model can focus more on task selection and execution.

[0005] Currently, the aforementioned standard procedures are obtained by manually writing them, which results in low task execution efficiency and low success rate. Summary of the Invention

[0006] This application provides a task execution method, a task execution device, an electronic device, and a storage medium to improve the execution efficiency and success rate of tasks corresponding to standard processes.

[0007] To address the aforementioned technical problems, this application provides the following technical solutions:

[0008] In a first aspect, embodiments of this application provide a task execution method, comprising: firstly receiving a task request; then obtaining a first standard process from a standard process information database according to the task request; nextly obtaining a first sub-standard process from the first standard process according to the task request, wherein the first sub-standard process is a sub-standard process of the first standard process, the first sub-standard process is used to instruct multiple first sub-tasks to be executed sequentially based on a jump relationship, the first sub-standard process includes a first root node, the first root node is used to instruct the execution of the starting sub-task among the multiple first sub-tasks; and finally executing the multiple first sub-tasks according to the first sub-standard process.

[0009] In the above scheme, the standard process information database is configured with standard processes. This database provides the first standard process corresponding to a task request, eliminating the need for manual configuration of standard processes for each request and simplifying the process. Furthermore, a first sub-standard process can be obtained from the first standard process based on the task request. This first sub-standard process instructs multiple first subtasks to be executed sequentially, and the first root node within the first sub-standard process indicates the starting subtask among these. This allows for the execution of multiple first subtasks based on the first sub-standard process. Since the first sub-standard process can be obtained from the first standard process based on the task request, different first sub-standard processes can be obtained for different task requests. This approach is applicable to complex and dynamic task requests, improving the flexibility of executing multiple first subtasks and increasing their execution efficiency and success rate.

[0010] In one possible implementation of the first aspect, obtaining the first standard process from the standard process information base according to the task request includes:

[0011] The task request is semantically parsed to obtain a target semantic representation;

[0012] Semantic matching is performed between the target semantic representation and the standard process information database to obtain a first semantic representation that matches the target semantic representation. The standard process information database includes: multiple standard processes and first semantic representations corresponding to the multiple standard processes respectively.

[0013] The first standard process is obtained from the standard process information database based on the first semantic representation that matches the target semantic representation.

[0014] In the above scheme, the standard process information database stores multiple standard processes and their corresponding first semantic representations. Therefore, the task execution device performs semantic matching between the target semantic representation and the standard process information database to obtain a first semantic representation that matches the target semantic representation. Finally, the task execution device retrieves the first standard process from the standard process information database based on the first semantic representation that matches the target semantic representation. In this embodiment, the first standard process can be quickly retrieved from the standard process information database by matching the target semantic representation with the first semantic representation.

[0015] In one possible implementation of the first aspect, obtaining the first sub-standard process from the first standard process according to the task request includes:

[0016] A sub-standard process information database is obtained according to the first standard process. The sub-standard process information database includes multiple sub-standard processes, and the multiple sub-standard processes belong to the first standard process.

[0017] The first substandard process is retrieved from the substandard process information database according to the task request.

[0018] In the above scheme, the task execution device establishes a sub-standard process information database. The multiple sub-standard processes included in the sub-standard process information database belong to the first standard process. The task execution device can also obtain the first sub-standard process by searching the sub-standard process information database. The retrieval of the sub-standard process information database improves the task execution efficiency. The execution of the standard process starts from the starting subtask indicated by the first root node included in the first sub-standard process, reducing unnecessary calculations.

[0019] In one possible implementation of the first aspect, obtaining the first sub-standard process from the sub-standard process information database according to the task request includes:

[0020] The task request is semantically parsed to obtain a target semantic representation;

[0021] Semantic matching is performed between the target semantic representation and the sub-standard process information base to obtain a second semantic representation that matches the target semantic representation. The sub-standard process information base includes: multiple sub-standard processes of the first standard process and second semantic representations corresponding to the multiple sub-standard processes of the first standard process, respectively.

[0022] The first substandard process is obtained from the substandard process information database based on the second semantic representation that matches the target semantic representation.

[0023] In the above scheme, the sub-standard process information database stores multiple sub-standard processes and their corresponding second semantic representations. Therefore, the task execution device performs semantic matching between the target semantic representation and the sub-standard process information database to obtain a second semantic representation that matches the target semantic representation. Finally, the task execution device retrieves the first sub-standard process from the sub-standard process information database based on the second semantic representation that matches the target semantic representation. In this embodiment, the first sub-standard process can be quickly retrieved from the sub-standard process information database by matching the target semantic representation with the second semantic representation.

[0024] In one possible implementation of the first aspect, the method further includes: when the first standard process is not obtained from the standard process information library, obtaining a second standard process according to the task request, wherein the second standard process is standard process information generated based on the task request and the experience database; and storing the second standard process in the standard process information library and the sub-standard process information library.

[0025] In the above scheme, the task execution device and the terminal device interact. The task execution device receives a task request from the terminal device, parses the task request, and obtains the task request to be executed. The task execution device can access an experience database, which provides expert experience information for generating standard processes. Using the experience database and the task request, a second standard process can be generated. The second standard process is stored in a standard process information database and a sub-standard process information database to achieve dynamic updates to these databases. In some optional implementations, the second standard process can also be stored in a sub-standard process information database to achieve dynamic updates to the sub-standard process information database.

[0026] In one possible implementation of the first aspect, the method further includes: when the first standard process is not obtained from the standard process information library, obtaining first multimodal information according to the task request, the first multimodal information including multiple types of information required to generate standard process information; inputting the first multimodal information into a large model, outputting a third standard process through the large model, the large model being used to generate standard process information corresponding to multiple tasks respectively; and storing the third standard process in the standard process information library and the sub-standard process information library.

[0027] In the above scheme, the task execution device can train a large model and use the trained large model to predict the first multimodal information to output a third standard procedure. The large model is used to generate standard procedure information corresponding to various tasks. When the first standard procedure is not found in the standard procedure information database, the third standard procedure is stored in the standard procedure information database to achieve dynamic updates to the standard procedure information database. This ensures that the standard procedure information database and sub-standard procedure information databases can retrieve the standard procedure and sub-standard procedure matching the task request during the next retrieval. Alternatively, the third standard procedure can be stored in the standard procedure information database and sub-standard procedure information database to achieve dynamic updates to the standard procedure information database and sub-standard procedure information databases, ensuring that the standard procedure information database and sub-standard procedure information databases can retrieve the standard procedure and sub-standard procedure matching the task request during the next retrieval.

[0028] In one possible implementation of the first aspect, the method further includes: acquiring a fourth standard process and storing the fourth standard process in the standard process information database, wherein the fourth standard process is standard process information generated based on an experience database; or, acquiring second multimodal information, wherein the second multimodal information includes multiple types of information required to generate standard process information; inputting the second multimodal information into a large model, and outputting a fifth standard process through the large model, wherein the large model is used to generate standard process information corresponding to multiple tasks respectively; and storing the fifth standard process in the standard process information database and the sub-standard process information database.

[0029] In the above scheme, the experience database can provide expert experience information for generating standard processes. When no task request is required, a fourth standard process can be generated from the experience database and stored in a standard process information repository. This allows for dynamic updates to the standard process information repository, ensuring that the repository retrieves the standard process and sub-standard processes matching the task request during subsequent searches. Similarly, the fourth standard process can be stored in both the standard process information repository and the sub-standard process information repository, enabling dynamic updates to both repositories and ensuring that the repositories retrieve the standard process and sub-standard processes matching the task request during subsequent searches.

[0030] In the above scheme, the task execution device can train a large model and use the trained large model to predict the second multimodal information to output a fifth standard procedure. The large model is used to generate standard procedure information corresponding to various tasks. The fifth standard procedure is stored in a standard procedure information database and a sub-standard procedure information database to achieve dynamic updates to the standard procedure information database, so that the database can retrieve the standard procedure and sub-standard procedure matching the task request during the next retrieval. Alternatively, the fifth standard procedure can be stored in a standard procedure information database and a sub-standard procedure information database to achieve dynamic updates to the database, so that the database can retrieve the standard procedure and sub-standard procedure matching the task request during the next retrieval.

[0031] In one possible implementation of the first aspect, the method further includes: during the execution of multiple first sub-tasks through an action model, obtaining an action sequence corresponding to the first sub-standard process, wherein multiple actions in the action sequence are used to execute multiple first sub-tasks or sub-task jumps, and the multiple actions in the action sequence are implemented based on sub-task information or sub-task jump logic information of multiple sub-tasks; performing anomaly evaluation on the action sequence through a first evaluation model based on the execution results of the multiple actions in the action sequence, the sub-task information of the multiple sub-tasks, and the sub-task jump logic information, to obtain suspicious actions from the action sequence; and performing anomaly detection on the suspicious actions through a second evaluation model based on the execution results of the suspicious actions and the sub-task information of the first sub-tasks corresponding to the suspicious actions, or based on the execution results of the suspicious actions and the sub-task jump logic information of the first sub-tasks corresponding to the suspicious actions, to determine that the suspicious actions are abnormal actions.

[0032] In the above scheme, the task execution device can train a second evaluation model. The task execution device obtains the execution result of the suspicious action. Based on the execution result of the suspicious action and the subtask information or subtask jump logic information of the first subtask corresponding to the suspicious action, the trained second evaluation model is used to perform anomaly detection on the suspicious action. Here, anomaly detection means that the second evaluation model is used to locally evaluate whether the suspicious action is abnormal. When the suspicious action is abnormal, it is determined to be an abnormal action.

[0033] In one possible implementation of the first aspect, the method further includes: optimizing the abnormal action using an optimization model based on the execution result of the abnormal action and the subtask information of the subtask corresponding to the abnormal action, or based on the execution result of the suspicious action and the subtask jump logic information of the first subtask corresponding to the suspicious action, to obtain the execution result of the optimized action; and evaluating the execution result of the optimized action using a second evaluation model to obtain a first evaluation result.

[0034] In the above scheme, the task execution device calls the second evaluation model again for local evaluation. The second evaluation model is used to evaluate the execution result of the optimized action to obtain the first evaluation result. In this embodiment of the application, by optimizing the abnormal action through the optimization model and evaluating the execution result of the optimized action through the second evaluation model, it is possible to detect whether the optimized action still has problems, so as to determine whether the optimization of the abnormal action is successful.

[0035] In one possible implementation of the first aspect, the method further includes: when the first evaluation result indicates that there is a problem with the execution result of the optimized action, backtracking is performed using a backtracking model based on the execution result of the optimized action, subtask information of multiple first subtasks, and subtask jump logic information to obtain the first associated action that caused the problem from the action sequence and obtain historical backtracking information; optimization is performed using the optimization model starting from the first associated action that caused the problem until the abnormal action is reached, to obtain the execution result of the optimized first associated action; and the execution result of the optimized first associated action is evaluated using a second evaluation model to obtain a second evaluation result.

[0036] In the above scheme, the task execution device uses a backtracking model to determine the first associated action that caused the problem from the action sequence. It then calls the optimization model again to optimize from this first associated action until an abnormal action is reached, obtaining the execution result of the optimized first associated action. Next, a second evaluation model is called to evaluate the execution result of the optimized first associated action, obtaining a second evaluation result. In this embodiment, by optimizing from the first associated action to the abnormal action using the optimization model, and by evaluating the execution result of the optimized first associated action using the second evaluation model, it is possible to detect whether the action still has problems after further optimization, thus determining whether the optimization of the first associated action up to the abnormal action was successful.

[0037] In one possible implementation of the first aspect, the method further includes: when the second evaluation result indicates that there is a problem with the execution result of the optimized first associated action, backtracking is performed using the backtracking model based on the historical backtracking information, the execution result of the optimized first associated action, and the subtask information and subtask jump logic information of multiple first subtasks, to obtain the second associated action that caused the problem from the action sequence; optimization is performed using the optimization model starting from the second associated action until the abnormal action is reached, to obtain the execution result of the optimized second associated action; and the execution result of the optimized second associated action is evaluated using the second evaluation model to obtain a third evaluation result.

[0038] In the above scheme, the second associated action is optimized by an optimization model until an abnormal action is reached. The execution result of the second associated action after optimization by a second evaluation model is evaluated to detect whether there are still problems with the action after further optimization, so as to determine whether the optimization of the second associated action up to the abnormal action is successful.

[0039] In one possible implementation of the first aspect, when the first evaluation result indicates that there is no problem with the execution result of the optimized action, the method further includes: updating the standard process information base using the optimized action; and / or, training at least one of the following models using the optimized action: an action model, a first evaluation model, and a second evaluation model.

[0040] In the above scheme, when the first evaluation result indicates that there are no problems with the execution of the optimized action, the task execution device can use the optimized action to update the information base and / or optimize the model. This information base may include a standard process information base, or the updated database may include a standard process information base and a sub-standard process information base. The model is then trained: an action model, a first evaluation model, and a second evaluation model. By updating the information base and / or optimizing the model, the information base and model can be used for subsequent task execution, improving the efficiency and success rate of task execution.

[0041] In one possible implementation of the first aspect, when the second evaluation result indicates that there is no problem with the execution result of the optimized first associated action, the method further includes: updating the standard process information base using the optimized first associated action; and / or, training at least one of the following models using the optimized first associated action: an action model, a first evaluation model, a second evaluation model, and an optimization model.

[0042] In the above scheme, when the second evaluation result indicates that there are no problems with the execution result of the optimized first associated action, the task execution device can use the optimized action to update the information base and / or optimize the model. This information base may include a standard process information base, or the updated database may include a standard process information base and a sub-standard process information base. The model is then trained: an action model, a first evaluation model, a second evaluation model, and an optimization model. By updating the information base and / or optimizing the model, the information base and model can be used for subsequent task execution, improving the efficiency and success rate of task execution.

[0043] In one possible implementation of the first aspect, when the third evaluation result indicates that there is no problem with the execution result of the optimized second associated action, the method further includes: updating the standard process information base using the optimized second associated action; and / or, training at least one of the following models using the optimized second associated action: an action model, a first evaluation model, a second evaluation model, a backtracking model, and an optimization model.

[0044] In the above scheme, when the third evaluation result indicates that the execution result of the optimized second associated action is problem-free, the task execution device can use the optimized action to update the information base and / or optimize the model. This information base may include a standard process information base, or the updated database may include a standard process information base and a sub-standard process information base. The model is then trained as follows: an action model, a first evaluation model, a second evaluation model, a backtracking model, and an optimization model. By updating the information base and / or optimizing the model, the information base and model can be used for subsequent task execution, improving the efficiency and success rate of task execution.

[0045] In one possible implementation of the first aspect, the task execution method provided in the first aspect can be implemented by a task execution device, which may specifically be a terminal device or a server. For example, the terminal device or server includes a processor, which is used to execute the task execution method in the first aspect.

[0046] For example, a standard process may specifically include computer program code, and a task execution device may specifically be a code executor.

[0047] For example, the task execution device can specifically be an agent, which incorporates a large model. The agent can be used to execute the task execution method provided in the first aspect above.

[0048] Within the framework of intelligent agents, agents are endowed with the general capabilities of large models to solve one or more specific problems. Different types of intelligent agents can be automatically linked together to solve more complex problems.

[0049] In one possible implementation of the first aspect, the task request in the task execution method provided in the first aspect can be a task from various business scenarios, which is not limited here. Specifically, the task request can be from the fields of smart home, autonomous driving, gaming, intelligent question answering, and communication. For example, in the smart home field, the task request can be a user initiating fault detection of smart home appliances, or a user controlling a robot vacuum cleaner to customize a route throughout the house. In the autonomous driving field, the task request can be a user initiating automatic obstacle avoidance of a smart vehicle, or controlling a smart vehicle to update the map. In the gaming field, the task request can be a game player controlling a game character to automatically find a path, or controlling a game character to perform multi-task allocation and execution with other game characters. In the communication field, the task request can specifically be evaluating the signal quality of attached terminals in a base station cell. According to the task execution method provided in the embodiments of this application, the task execution device is configured with a standard process in a standard process information library, so as to provide a first standard process corresponding to the task request in the task request, thereby eliminating the need for manual configuration of standard processes for tasks and simplifying the method of providing standard processes for tasks. In addition, the task execution device is configured with sub-standard processes in the sub-standard process information database, so as to provide the ability to determine the starting subtask of the task request in the first standard process corresponding to the first sub-standard process. This allows for flexible selection of the starting subtask in the first standard process according to the task request, thus making it applicable to different complex and dynamic tasks, improving the flexibility of executing task requests, and improving the execution efficiency of tasks corresponding to the standard process.

[0050] In one possible implementation of the first aspect, the standard process information base is constructed in the following manner:

[0051] Obtain subtask information of multiple second subtasks and subtask jump logic information between the multiple second subtasks. The subtask jump logic information between the multiple second subtasks is used to indicate the subtask jump method between the multiple second subtasks. The multiple second subtasks are used to execute the jump of the second subtasks sequentially according to the jump relationship indicated by the sixth standard process.

[0052] The second subtask is jumped according to the reverse jump method between the multiple second subtasks, and the subtask information of the second subtask after the jump is added to the node of the second subtask after the jump, so as to obtain the description information of the sixth standard process. The reverse jump method of the subtask is opposite to the jump method of the subtask indicated by the subtask jump logic information.

[0053] The standard process information database is constructed based on the description information of the sixth standard process, wherein the sixth standard process is any one of the standard processes in the standard process information database.

[0054] In the above scheme, the description information of the sixth standard process is obtained based on the subtask jump logic information of multiple second subtasks. Similarly, the description information of multiple standard processes in the standard process information library can be obtained, thus forming a standard process information library. This application embodiment constructs a standard process information library through the description information of multiple standard processes, thereby providing a standard process information library for various different tasks and simplifying the method of providing standard processes.

[0055] In one possible implementation of the first aspect, the step of jumping the second subtasks according to the reverse jumping method between the plurality of second subtasks, and adding the subtask information of the second subtask after the jump to the node of the second subtask after the jump, to obtain the description information of the sixth standard process, includes:

[0056] Jump according to the reverse jump method between the multiple second subtasks, and add the subtask information of the second subtask after the jump to the node of the second subtask after the jump, to obtain the description information of multiple sub-standard processes of the sixth standard process.

[0057] The description information of the sixth standard process is generated based on the description information of multiple sub-standard processes of the sixth standard process.

[0058] In the above scheme, the jump of the second subtask is compacted by using the subtask reverse jump method, and the jump information of the second subtask is added to the node of the second subtask after the jump, thus obtaining the description information of multiple sub-standard processes. Therefore, the description information of the sixth standard process can be generated based on the description information of multiple sub-standard processes, simplifying the way to generate the description information of the sixth standard process.

[0059] In one possible implementation of the first aspect, constructing the standard process information database based on the description information of the six standard processes includes:

[0060] Generate a first semantic representation corresponding to the sixth standard process based on the description information of the sixth standard process;

[0061] The standard process information database is constructed based on the first semantic representation corresponding to the sixth standard process.

[0062] In the above scheme, the first semantic representation corresponding to the sixth standard process is one or more vector representations corresponding to the standard process. This first semantic representation can be sparse or dense. The task execution device generates a standard process information database after obtaining the first semantic representations corresponding to multiple standard processes, based on the first semantic representation generated for the sixth standard process. In this embodiment, the first semantic representation corresponding to each standard process can be used for retrieval to achieve matching between tasks and standard processes.

[0063] In one possible implementation of the first aspect, the sub-standard process information base is constructed in the following manner:

[0064] Obtain subtask information of multiple first subtasks and subtask jump logic information between the multiple first subtasks. The subtask jump logic information between the multiple first subtasks is used to indicate the subtask jump method between the multiple first subtasks.

[0065] The first subtask is jumped according to the reverse jump method between the multiple first subtasks, and the subtask information of the first subtask after the jump is added to the position node of the first subtask after the jump, so as to obtain the description information of multiple sub-standard processes of the first standard process.

[0066] The sub-standard process information database is constructed based on the description information of multiple sub-standard processes of the first standard process.

[0067] In the above scheme, a subtask reverse jump method is used to compact the jump of the second subtask, and the jump information of the second subtask is added to the node of the second subtask after the jump, thereby obtaining the description information of multiple sub-standard processes of the first standard process. Based on the obtained description information of multiple sub-standard processes of the first standard process, a sub-standard process information library is constructed. The embodiments of this application construct a sub-standard process information library by using the description information of multiple sub-standard processes, thereby providing a sub-standard process information library for a variety of different tasks and simplifying the way sub-standard processes are provided.

[0068] In one possible implementation of the first aspect, constructing the sub-standard process information database based on the description information of multiple sub-standard processes of the first standard process includes:

[0069] Based on the description information of the multiple sub-standard processes of the first standard process, generate second semantic representations corresponding to the multiple sub-standard processes of the first standard process respectively.

[0070] The sub-standard process information database is constructed based on the second semantic representations corresponding to the multiple sub-standard processes of the first standard process.

[0071] In the above scheme, the second semantic representation corresponding to each of the multiple sub-standard processes of the first standard process is one or more vector representations corresponding to the multiple sub-standard processes. After obtaining the second semantic representations corresponding to the multiple standard processes according to the generation of the first standard process, a sub-standard process information database is generated. In this embodiment, the second semantic representation corresponding to each sub-standard process can be used for retrieval to achieve task matching with the sub-standard process.

[0072] Secondly, embodiments of this application also provide a task execution device, comprising:

[0073] The receiving module is used to receive task requests;

[0074] The standard process acquisition module is used to acquire a first standard process from the standard process information database according to the task request;

[0075] The sub-standard process acquisition module is used to acquire a first sub-standard process from the first standard process according to the task request. The first sub-standard process is a sub-standard process of the first standard process. The first sub-standard process is used to instruct multiple first subtasks to be executed sequentially based on a jump relationship. The first sub-standard process includes a first root node, which is used to instruct the execution of the starting subtask among the multiple first subtasks.

[0076] The task execution module is used to execute the plurality of first sub-tasks according to the first sub-standard process.

[0077] In the above scheme, the standard process information database is configured with standard processes to provide the first standard process corresponding to the task request in the execution task request. This eliminates the need for manual configuration of standard processes for each task, simplifying the method of providing standard processes for tasks. Furthermore, the sub-standard process information database is configured with sub-standard processes to determine the starting subtask of the task request within the first standard process corresponding to the first sub-standard process. This allows for flexible selection of the starting subtask within the first standard process based on the task request, making it applicable to complex and dynamic tasks, improving the flexibility of task request execution, and increasing the execution efficiency of tasks corresponding to the standard processes.

[0078] In the second aspect of this application, the constituent modules of the task execution device can also perform the steps described in the first aspect and various possible implementations, as detailed in the foregoing description of the first aspect and various possible implementations.

[0079] Thirdly, embodiments of this application provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described in the first aspect above.

[0080] Fourthly, embodiments of this application provide a computer program product containing instructions that, when run on a computer, cause the computer to perform the method described in the first aspect above.

[0081] Fifthly, embodiments of this application provide a communication device, which may include entities such as terminal devices or chips. The communication device includes: a processor and a memory; the memory is used to store instructions; the processor is used to execute the instructions in the memory, causing the communication device to perform the method as described in any one of the preceding first aspects.

[0082] Sixthly, this application provides a chip system including a processor for supporting a task execution device in performing the functions involved in the foregoing aspects, such as transmitting or processing data and / or information involved in the foregoing methods. In one possible design, the chip system further includes a memory for storing program instructions and data necessary for the task execution device. This chip system may be composed of chips or may include chips and other discrete devices.

[0083] In a seventh aspect, embodiments of this application provide a chip including one or more interface circuits and one or more processors; the interface circuits are used to receive signals from the memory of an electronic device and send signals to the processors, the signals including computer instructions stored in the memory; when the processor executes the computer instructions, it causes the electronic device to perform the task execution method in the first aspect or any possible implementation of the first aspect.

[0084] The seventh aspect and any implementation thereof correspond to the first aspect and any implementation thereof, respectively. The technical effects of the seventh aspect and any implementation thereof are similar to those of the first aspect and any implementation thereof, and will not be repeated here. Attached Figure Description

[0085] Figure 1 is a schematic diagram of a data center architecture provided in an embodiment of this application;

[0086] Figure 2a is a schematic diagram of an exemplary system provided in an embodiment of this application;

[0087] Figure 2b is a schematic diagram of another exemplary system provided in an embodiment of this application;

[0088] Figure 3 is a schematic diagram of an embodiment of a task execution method provided in this application;

[0089] Figure 4 is a schematic diagram of a standard process provided in an embodiment of this application;

[0090] Figure 5 is a schematic diagram of an application scenario of a task execution method provided in an embodiment of this application;

[0091] Figure 6 is a schematic diagram of a dynamic retrieval of SOP and task entry provided in an embodiment of this application;

[0092] Figure 7 is a schematic diagram of an improved SOP execution provided in an embodiment of this application;

[0093] Figure 8 is a schematic diagram of an embodiment of a task execution device provided in this application;

[0094] Figure 9 is a structural schematic diagram of a computing device provided in an embodiment of this application;

[0095] Figure 10 is a schematic diagram of a computing device cluster provided in an embodiment of this application;

[0096] Figure 11 is a schematic diagram of a computing device cluster provided in an embodiment of this application. Detailed Implementation

[0097] This application provides a task execution method, a task execution device, an electronic device, and a storage medium to improve the execution efficiency of standard process tasks.

[0098] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0099] As will be known to those skilled in the art, with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0100] In this application, "at least one" means one or more, and "more than one" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of singular or plural items. The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, system, product, or apparatus that comprises a series of units is not necessarily limited to those units, but may include other units not expressly listed or inherent to those processes, methods, products, or apparatus.

[0101] This application provides a task execution method that can be applied to a computing device cluster, which may include one or more computing devices.

[0102] The type of computing device is not limited here. For example, any computing device can be a terminal device, a cloud server, a container, or a virtual machine, etc.

[0103] The computing device cluster can provide a data processing platform to implement the task execution method of the embodiments of this application. Specifically, a data processing platform can be provided in the terminal device and in the cloud server. Based on the respective data processing platforms of the terminal device or the cloud server, consistency and computing accuracy between the terminal side and the cloud side can be achieved, while also making full use of the differences in computing power between different devices. The specific form of the computing device cluster and the corresponding data processing platform is not limited here.

[0104] In one example, the cluster of computing devices can be used to implement a cloud management platform; in other words, the data processing platform of this application embodiment can be implemented through a cloud management platform.

[0105] A cloud management platform is used to manage the infrastructure that provides cloud services. It can provide computing, networking, and storage capabilities based on hardware and software resources. For example, a cloud management platform may include one or more data centers to provide cloud resources through one or more data centers.

[0106] The following section introduces the data center with reference to an architecture diagram shown in Figure 1.

[0107] In Figure 1, the cloud management platform in this data center interacts with one or more servers (Server 1 and Server 2 in Figure 1) through the data center's internal network. A server comprises a hardware layer and a software layer. The hardware layer includes the server's hardware configuration, such as peripheral component interconnect (PCI) devices (e.g., network interface cards, graphics processing units, GPUs), offloading cards, etc.) that can be plugged into peripheral component interconnect (PCI) or peripheral component interconnect express (PCIe) slots. The software layer includes the operating system installed and running on the server (the operating system relative to a virtual machine can be called the host operating system). The host operating system contains a virtual machine manager (also known as a hypervisor), whose role is to implement compute virtualization, network virtualization, and storage virtualization of virtual machines, and to manage the virtual machines. A virtual machine (VM) refers to a complete computer system simulated by software, possessing full hardware system functionality, and running in a completely isolated environment. In the system architecture shown in Figure 1, a data center has multiple servers, which can be used to run virtual machines. The specifications of the virtual machines can be the same or different. Virtual machines can also be called Elastic Compute Service (ECS), Elastic Instances, etc., depending on the cloud service provider.

[0108] In one example of the embodiments of this application, the cloud management platform can be a public cloud platform. In this case, cloud service providers such as individuals or software developers with cloud resource development capabilities can provide cloud services to users. Users obtain cloud services through the Internet, but do not own cloud computing resources.

[0109] Specifically, in the example shown in Figure 1, the cloud management platform can provide an access interface (such as a user interface or application programming interface (API)). Users of the cloud management platform and cloud service providers can operate the client to remotely access the access interface to register a cloud account and password on the cloud management platform. After the cloud management platform successfully authenticates the cloud account and password, they can log in to the cloud management platform to create, manage, log in to and operate virtual machines in the cloud data center, thereby performing corresponding tasks through the cloud resources of the cloud management platform.

[0110] For example, as shown in the example in Figure 2a, the data processing platform that implements the task execution method in the embodiments of this application can be provided to users in the form of cloud services.

[0111] When data processing tasks require cross-device storage, some enterprises, organizations, or individuals can purchase cloud services that include the data processing platform. They can then use the cloud resources of the cloud management platform to perform the relevant data processing tasks and obtain the processing results of the input data from the cloud management platform.

[0112] This data processing platform can be implemented based on cloud resources such as one or more cloud data centers in the cloud management platform.

[0113] This data processing platform can be provided to users as an independent cloud service, and it is also compatible with computing frameworks such as TensorFlow and PyTorch. In other words, it can be used as an operator in computing frameworks such as TensorFlow and PyTorch, providing functions such as preprocessing, computation, feature extraction and / or other data processing for real-time and offline data.

[0114] Of course, the cloud management platform can also be other types of cloud management platforms, and this application embodiment does not limit this.

[0115] In another example, as shown in Figure 2b, the data processing platform can be provided to the user in the form of a software product or a functional service within a software product. The software product can be deployed on the user's terminal device or on the enterprise user's server cluster. The user can purchase a software license to obtain the right to use the software product and, within the license period, implement the task execution method of the embodiments of this application through the software product.

[0116] The following example, using the field of task execution methods, illustrates the practical application of a data processing platform that implements the task execution method. This data processing platform has the functions of the task execution device described in subsequent embodiments. Specifically, the task execution device may be a terminal device or a server. For example, the terminal device or server may include a processor, which is used to execute the task execution method described in the first aspect above.

[0117] For example, a standard process may specifically include computer program code, and a task execution device may specifically be a code executor.

[0118] For example, the task execution device can specifically be an agent, which incorporates a large model. The agent can be used to execute the task execution method provided in the first aspect above.

[0119] Within the framework of intelligent agents, agents are endowed with the general capabilities of large models to solve one or more specific problems. Different types of intelligent agents can be automatically linked together to solve more complex problems.

[0120] In one exemplary application scenario, a user might want to transmit PCB layout defect information to a cloud server for storage, sharing, or further processing. The user uses computing frameworks such as TensorFlow and PyTorch to encode the PCB layout defect information using an AI model, and then sends the encoded bitstream via a transmission network between the mobile phone and the cloud server. This computing framework can be deployed on the client device or on the cloud server. After receiving the compressed PCB layout defect information, the cloud management platform utilizes cloud resources to process the information. For example, the cloud management platform might include a large language model, a detection tool library, and a PCB layout engine. The platform performs semantic understanding of the PCB layout defect information to obtain a defect description, uses a target detection tool to detect defects in the description, and derives optimization instructions based on the detected electronic components on the target PCB. Finally, the platform optimizes the layout of these electronic components based on these optimization instructions.

[0121] It is understood that the above example is only an exemplary introduction to one application scenario. The task execution method of this application embodiment can be applied to multiple fields, and is not limited here.

[0122] To address the current reliance on human experience to develop standard workflows, this application provides a task execution method and a corresponding task execution device. Specifically, the task execution method can be a process-oriented approach for automating tasks. For example, the task execution method provided in this application enables the efficient execution of various tasks based on standard workflows. The task execution method provided in this application is continuously improveable and evolves. Compared to task execution schemes that rely on human experience, this application can significantly improve the efficiency and results of task execution.

[0123] As shown in Figure 3, the task execution method device and the terminal device can interact, and the task execution method can include steps 301-305.

[0124] 301. The terminal device sends a task request to the task execution device.

[0125] In this embodiment, the terminal device is a user-operated terminal device that receives a task request from the user. For example, the task request may include a task to be performed. Depending on the user's needs, the task can be determined based on the domain. For instance, the task may be a problem the user hopes to solve, and it can belong to various different domains. For example, the task in this embodiment may be at least one of the following: analyzing the abnormal reasons for order numbers abcde, helping a user plan a travel itinerary to city A, providing a brief biography of historical figures A, B, C, or C, providing dietary advice for student ID 12345, etc. There are no limitations on the domain to which the task belongs or the content of the task. Furthermore, the task request may include a description of the task by the user, such as a textual or audio description of the task to be performed.

[0126] For example, as shown in Figure 2a, in this embodiment of the application, the terminal device can be a client, and the task execution device can be a data center. The client can interact with the data center and send task requests to the data center.

[0127] 302. The task execution device receives the task request.

[0128] The terminal device can generate a task request. The task execution device interacts with the terminal device. The task execution device receives the task request from the terminal device, parses the task request, and determines the task to be executed based on the task request.

[0129] For example, a user can control a terminal device and send task requests through the device. For instance, if a user needs dietary advice for a student with ID 12345, the task request sent by the terminal device could include tasks such as providing dietary advice for student ID 12345. Alternatively, the task execution device can parse the task request sent by the terminal device and obtain the task of providing dietary advice for student ID 12345.

[0130] This application does not limit the application scenario of the task request. For example, the task request can be a root cause localization task request in the field of operation and maintenance, that is, the terminal device can request the task execution device to execute the root cause localization task. As another example, the task request can be a user selection of a corresponding branch task request in the field of intelligent customer service, that is, the terminal device can request the task execution device to execute the branch task selected according to the different tasks of the user.

[0131] 303. The task execution device retrieves the first standard process from the standard process information database according to the task request.

[0132] In this embodiment, the standard process can be used to handle various types of problems. These problems are solved step by step in the form of a flowchart, and the entire task is ultimately completed through the standard process. The standard process provided in this embodiment can also be called a standard operating procedure, procedure, standard operating procedure (SOP), standardized workflow, etc. In the following examples, the standard process is used as an SOP for illustration. The standard process information database includes multiple standard processes, and the standard process information database can also be called SOP long-term memory.

[0133] In this application embodiment, different standard processes are required for different tasks to complete them. The task execution device in this application embodiment can be configured with a standard process information library, which can include multiple different standard processes that can be used to solve different tasks. The standard process information library stores multiple standard processes. For example, the standard process information library can include: standard process a, standard process b, and standard process c, where standard process a can be used to execute task 1, standard process b can be used to execute task 2, and standard process c can be used to execute task 3. There are no limitations on the domains associated with the multiple standard processes in the standard process information library or the content included in the multiple standard processes.

[0134] In the embodiments of this application, the standard process information database can be dynamically updated to continuously expand the standard process information database, so that the standard process information database can include more standard processes.

[0135] In this embodiment, after the task execution device obtains a task request, it uses the task request to retrieve a standard process information database to obtain a first standard process from the database. This first standard process is a standard process stored in the database. In this embodiment, the standard process information database is configured with multiple different standard processes to provide the first standard process corresponding to the task execution request, thus eliminating the need for manual configuration of standard processes for tasks and simplifying the method of providing standard processes for tasks.

[0136] For example, a task request could include providing dietary advice for student ID 12345. Based on this request, a standard process information database is retrieved, and a first standard process is obtained. This first standard process can provide dietary advice and specify a diet plan based on the student's student ID, gender, height, or deviation from the standard weight. As another example, a task request could include providing the biographical information of basketball stars abcde. This first standard process can search for the teams abcde played for and the major games they played in based on their name, providing their biographical information.

[0137] In this embodiment of the application, a standard process library can be retrieved based on the task request, thereby retrieving a standard process that matches the task request. The standard process can solve a type of task requested in the task request.

[0138] The standard process in this application embodiment may include multiple sub-standard processes. For each standard process in this application embodiment, multiple sub-standard processes may be included, and these multiple sub-standard processes are used to generate a complete standard process. For example, the first standard process is a standard process found in the standard process information database that matches the task request. Through the first standard process, the tasks included in the task request can be broken down into a series of multiple first sub-tasks. Each first sub-task has a specific execution method, and the jump logic between multiple first sub-tasks is linked through the execution results. The multiple sub-standard processes included in the first standard process can be used to execute the aforementioned multiple sub-tasks. For example, the first standard process may include: sub-standard process a1, sub-standard process a2, and sub-standard process a3. Sub-standard process a1 can be used to execute sub-task 1 and other sub-tasks that sub-task 1 jumps to; sub-standard process a2 can be used to execute sub-task 2 and other sub-tasks that sub-task 2 jumps to; and sub-standard process a3 can be used to execute sub-task 3 and other sub-tasks that sub-task 3 jumps to.

[0139] In some embodiments of this application, a sub-standard process information library can be configured for each standard process in the standard process information library. This sub-standard process information library can include multiple sub-standard processes corresponding to the standard process. For example, the first standard process may include sub-standard process a1, sub-standard process a2, and sub-standard process a3, wherein sub-standard process a1, sub-standard process a2, and sub-standard process a3 belong to sub-standard process information library 1. Other sub-standard process information libraries can also be configured for other standard processes in the standard process library besides the first standard process.

[0140] 304. The task execution device obtains the first sub-standard process from the first standard process according to the task request.

[0141] The first sub-standard process is a sub-standard process of the first standard process. The first sub-standard process is used to instruct multiple first subtasks to be executed sequentially based on jump relationships. The first sub-standard process includes a first root node, which is used to instruct the execution of the starting subtask among multiple first subtasks.

[0142] In this embodiment, the task execution device obtains a first standard flow, which includes multiple sub-standard flows. Based on a task request, a first sub-standard flow is obtained from the first standard flow. The multiple sub-standard flows include the first sub-standard flow, which is the sub-standard flow that matches the task request. The first sub-standard flow includes a first root node, which represents the first node in the data structure of the first sub-standard flow. Therefore, the first root node has no parent node in the first sub-standard flow; for example, the first sub-standard flow can be a tree data structure. Each node in the first sub-standard flow is used to execute a first subtask, and the first root node is used to execute the starting subtask of multiple first subtasks. The starting subtask can also be called the first subtask, the first subtask of the first subtask, or the root subtask; no limitation is made here.

[0143] For example, the first sub-standard process is used to execute the starting subtask indicated by the first root node and other subtasks that the starting subtask jumps to according to the jump relationship. The starting subtask can also be regarded as the task entry point of the first standard process.

[0144] In some embodiments of this application, the task execution device retrieves the sub-standard process information database corresponding to the first standard process according to the task request, determines the first sub-standard process, and the first sub-standard process corresponds to the task entry point.

[0145] The sub-standard process information library corresponding to the first standard process is obtained according to the first standard process. Based on the description of the sub-standard process information library in step 303, it can be known that the sub-standard process information library can include multiple sub-standard processes. The sub-standard process information library stores multiple sub-standard processes included in the first standard process, and the multiple sub-standard processes include the first sub-standard process. After the task execution device obtains the first standard process, it uses the task request to retrieve the sub-standard process information library corresponding to the first standard process to obtain the first sub-standard process from the sub-standard process information library. The first sub-standard process is a sub-standard process stored in the sub-standard process information library. The first root node included in the first sub-standard process indicates the starting subtask among multiple first subtasks to be executed. The starting subtask can also be regarded as the task entry point of the first standard process. For example, the task execution device retrieves the sub-standard process information library corresponding to the first standard process according to the task request to determine the first sub-standard process. The first sub-standard process corresponds to the task entry point.

[0146] In this embodiment, the substandard process information database is configured with a variety of different substandard processes to provide a task entry point for determining the first standard process. This allows for flexible selection of the starting subtask in the first standard process based on the task request, thus making it applicable to complex and dynamic tasks and improving the flexibility of task execution.

[0147] 305. The task execution device executes the plurality of first sub-tasks according to the first sub-standard process.

[0148] In this embodiment, after the task execution device obtains a first sub-standard process from the first standard process, it can obtain a starting subtask based on the first sub-standard process, and obtain other subtasks to which the starting subtask jumps according to the jump relationship between the first subtasks in the first sub-standard process. Then, according to the task content provided by the task request, it executes the starting subtask using the first sub-standard process. After the starting subtask is completed, it performs subtask jumps according to the first sub-standard process and executes other first subtasks to which the starting subtask jumps using the first sub-standard process. Not limited to this, the jump of the first subtask in this embodiment may include one jump or multiple jumps. When the termination condition of the first subtask is met, the execution of the first subtask stops, and the execution result for the task request is output.

[0149] As illustrated by the foregoing embodiments, in this application embodiment, the standard process information library is configured with standard processes. This library provides a first standard process corresponding to a task request, eliminating the need for manual configuration of standard processes for each task request and simplifying the process. Furthermore, a first sub-standard process can be obtained from the first standard process based on the task request. This first sub-standard process instructs multiple first subtasks to be executed sequentially, and the first root node within the first sub-standard process indicates the starting subtask among the multiple first subtasks. Therefore, multiple first subtasks can be executed according to the first sub-standard process. Since the first sub-standard process can be obtained from the first standard process based on the task request, corresponding first sub-standard processes can be obtained for different task requests, making it applicable to complex and dynamic task requests, improving the flexibility of executing multiple first subtasks, and increasing the execution efficiency and success rate of multiple subtasks.

[0150] As illustrated by the examples in steps 301 to 305 above, for task requests from terminal devices, only the terminal device needs to provide the task request to be executed; the terminal device does not need to obtain the corresponding standard process planning method. The task execution device can configure a standard process information database. By searching the standard process information database, the understanding of standard process configuration information in actual scenarios is reduced, and the usability of the standard process is improved. At the same time, task retrieval can also improve task coverage.

[0151] In some embodiments of this application, the standard process information database is constructed in the following manner:

[0152] A1. The task execution device obtains subtask information of multiple second subtasks and subtask jump logic information between multiple second subtasks. The subtask jump logic information between multiple second subtasks is used to indicate the subtask jump method between multiple second subtasks. Multiple second subtasks are used to execute the jump of the second subtasks in sequence according to the jump relationship indicated by the sixth standard process.

[0153] A2. Jump to the second subtask according to the reverse jump method between multiple second subtasks, and add the subtask information of the second subtask after the jump to the node of the second subtask after the jump, to obtain the description information of the sixth standard process. The reverse jump method of the subtask is the opposite of the subtask jump method indicated by the subtask jump logic information.

[0154] A3. The task execution device constructs a standard process information database based on the description information of the sixth standard process, where the sixth standard process is any one of the standard processes in the standard process information database.

[0155] The sixth standard process is any standard process in the standard process information database. Steps A1 to A3 describe the construction method of any standard process in the standard process information database. Specifically, firstly, the subtask information of multiple second subtasks and the subtask jump logic information between multiple second subtasks are obtained. The subtask information of the second subtasks describes the task content of the second subtask. The subtask jump logic information between multiple second subtasks includes information on how multiple second subtasks jump sequentially according to the jump relationship indicated by the sixth standard process. The subtask jump logic information between multiple second subtasks indicates the jump method between multiple second subtasks. Multiple second subtasks can jump sequentially according to the jump relationship to complete a complete task. Then, the task execution device obtains the reverse subtask jump method, which is the opposite of the subtask jump method between multiple second subtasks. It then jumps to the second subtasks according to the reverse jump method and adds the subtask information of the second subtask after the jump to its node. For example, the second subtask jump can be based on a graphical structure, and the subtask information of the second subtask after the jump is added to its node within the graphical structure. In other words, the task execution device can obtain the description information of the sixth standard process by aggregating the subtask information of multiple second subtasks. For example, the subtask jump logic information indicates the second subtask at the starting position. The second subtask at the starting position sequentially searches for other second subtasks according to the jump relationship, and then aggregates the subtask information of multiple second subtasks to obtain the description information of the sixth standard process. The description information of the sixth standard process is obtained through steps A1-A2. These steps also allow obtaining the description information of multiple standard processes in the standard process information library. Therefore, the standard process information library is constructed through step A3. This application embodiment constructs a standard process information library by using description information of various standard processes, thereby providing a standard process information library for a variety of different tasks and simplifying the way standard processes are provided.

[0156] Furthermore, in some embodiments of this application, the A2 task execution device performs the jump to the second subtask according to the reverse jump method between multiple second subtasks, and adds the subtask information of the second subtask after the jump to the node of the second subtask after the jump, to obtain the description information of the sixth standard process, including:

[0157] A21. The task execution device jumps according to the reverse jump method between multiple second subtasks, and adds the subtask information of the second subtask after the jump to the node of the second subtask after the jump, so as to obtain the description information of multiple sub-standard processes of the sixth standard process.

[0158] A22. The task execution device generates the description information of the sixth standard process based on the description information of multiple sub-standard processes of the sixth standard process.

[0159] The task execution device can acquire subtask jump logic information between multiple second subtasks. This subtask jump logic information indicates the second subtask at the starting position. Starting from this second subtask, jumps are performed in reverse order, adding subtask information of the second subtask after the jump to its node. This allows the aggregation of subtask information from multiple second subtasks, ultimately yielding description information for multiple sub-standard processes of the sixth standard process. Finally, the task execution device generates the description information for the sixth standard process based on the description information of these multiple sub-standard processes. In this embodiment, reverse subtask jumps are used to aggregate information from multiple subtasks to obtain description information for multiple sub-standard processes. Therefore, the description information for the sixth standard process can be generated based on the description information of these multiple sub-standard processes, simplifying the method of generating the sixth standard process description information.

[0160] Furthermore, in some embodiments of this application, the A3 task execution device constructs a standard process information database based on the description information of the sixth standard process, including:

[0161] A31. The task execution device generates the first semantic representation corresponding to the sixth standard process based on the description information of the sixth standard process.

[0162] A32. The task execution device constructs a standard process information database based on the first semantic representation corresponding to the sixth standard process.

[0163] The task execution device can perform semantic parsing on the description information of the sixth standard process to obtain the first semantic representation corresponding to the sixth standard process. The semantic algorithm used for semantic parsing is not limited; for example, a Large Language Model (LLM) can be used to perform semantic parsing on the description information of the sixth standard process to obtain the first semantic representation. In this embodiment, the first semantic representation corresponding to the sixth standard process is one or more vector representations corresponding to the standard process. This first semantic representation can be sparse or dense. After generating the first semantic representation corresponding to the sixth standard process and obtaining the first semantic representations corresponding to multiple standard processes, the task execution device generates a standard process information database. In this embodiment, the first semantic representation corresponding to each standard process can be used for retrieval to achieve matching between tasks and standard processes.

[0164] Without limitation, in the embodiments of this application, the standard process information database stores multiple standard processes and their corresponding first semantic representations. In addition, the index values ​​of the corresponding standard processes can be generated based on the first semantic representations of the multiple standard processes. That is, the index values ​​of the multiple standard processes are used as keys, and the first semantic representations of the multiple standard processes are used as values. Key-value pairs are stored in the standard process information database.

[0165] In some embodiments of this application, the standard process information base includes: multiple standard processes and first semantic representations corresponding to the multiple standard processes.

[0166] For details on how to obtain the first semantic representation for each of the multiple standard processes, please refer to the description of step A31.

[0167] The following example illustrates the process of constructing the SOP information database, using the standard process as the SOP and sub-standard processes as sub-SOPs. For instance, this task request might include providing dietary advice for student ID 12345. The task request could be to provide dietary advice and specify a diet plan based on the student's student ID, gender, height, or deviation from the standard weight. The sub-SOP description information and the process of obtaining it are as follows: The sub-SOP corresponding to the second sub-task 3 is the description information of the second sub-task: "Provide dietary advice and specify a diet plan based on the student's deviation from the standard weight." The sub-SOP corresponding to the second sub-task 2, combined with the downstream sub-SOP (i.e., sub-SOP3) and the sub-task information of the second sub-task 2, yields the following sub-SOP: "Provide dietary advice and specify a diet plan based on the student's gender and height." Similarly, the sub-SOP corresponding to the second sub-task 1, combined with the downstream sub-SOP (i.e., sub-SOP2) and the task information of the second sub-task 1, yields the following sub-SOP: "Provide dietary advice and specify a diet plan based on the student's student ID." Therefore, the overall SOP description can be derived by combining the information from the sub-SOPs above. For example, the SOP description could be: "Provide dietary advice and specify a diet plan based on the student's student ID, gender, height, or deviation from the standard weight."

[0168] Figure 4 above illustrates the construction of the SOP information base using one task as an example. In this application embodiment, other SOPs in the SOP information base can also be expanded in a similar way, such as constructing an SOP for the biographical introduction of historical figures A, B, and C.

[0169] In some embodiments of this application, the 303 task execution device obtains a first standard process from a standard process information database according to a task request, including:

[0170] B1. The task execution device performs semantic parsing on the task request to obtain the target semantic representation;

[0171] B2. The task execution device performs semantic matching between the target semantic representation and the standard process information database to obtain a first semantic representation that matches the target semantic representation. The standard process information database includes: multiple standard processes and the first semantic representations corresponding to the multiple standard processes respectively.

[0172] B3. The task execution device obtains the first standard process from the standard process information database based on the first semantic representation that matches the target semantic representation.

[0173] The task execution device can perform semantic parsing on the task request to obtain a target semantic representation. The semantic algorithm used for semantic parsing is not limited; for example, a Large Language Model (LLM) can be used to perform semantic parsing on the task request to obtain the target semantic representation. Based on the aforementioned process of the standard process information database, it is known that the standard process information database stores multiple standard processes and their corresponding first semantic representations. Therefore, the task execution device performs semantic matching between the target semantic representation and the standard process information database to obtain a first semantic representation that matches the target semantic representation. Finally, the task execution device retrieves the first standard process from the standard process information database based on the first semantic representation that matches the target semantic representation. In this embodiment, by matching the target semantic representation with the first semantic representation, the first standard process can be quickly retrieved from the standard process information database.

[0174] In some embodiments of this application, the sub-standard process information database is constructed in the following manner:

[0175] C1. The task execution device obtains subtask information of multiple first subtasks and subtask jump logic information between multiple first subtasks. The subtask jump logic information between multiple first subtasks is used to indicate the subtask jump method between multiple first subtasks.

[0176] C2. The task execution device jumps to the first subtask in the reverse jump method between multiple first subtasks, and adds the subtask information of the first subtask after the jump to the position node of the first subtask, so as to obtain the description information of multiple sub-standard processes of the first standard process.

[0177] C3. The task execution device constructs a sub-standard process information database based on the description information of multiple sub-standard processes of the first standard process.

[0178] Steps C1 to C3 describe the construction method of multiple sub-standard processes in the sub-standard process information database. Specifically, firstly, the subtask information of multiple first subtasks and the subtask jump logic information between multiple first subtasks are obtained. The subtask information of the first subtasks describes the task content of the first subtask, and the subtask jump logic information between multiple first subtasks includes information on the sequential jump of multiple first subtasks according to the jump order. Then, the task execution device obtains the subtask reverse jump method, which is the opposite of the subtask jump method between multiple first subtasks, and performs the jump of the first subtask according to the subtask reverse jump method. The device also adds the subtask information of the first subtask after the jump to the node of the first subtask after the jump. For example, the jump of the first subtask can be based on a graph structure, and the subtask information of the first subtask after the jump is added to the node of the first subtask after the jump in the graph structure. That is, the task execution device can aggregate the subtask information of multiple first subtasks based on the subtask jump logic information between multiple first subtasks to obtain the description information of multiple sub-standard processes in the first standard process. For example, the subtask jump logic information indicates the first subtask at the starting position. This first subtask then sequentially searches for other first subtasks according to the jump relationship. The subtask information of these multiple first subtasks is then aggregated to obtain description information for multiple sub-standard processes of the first standard process. Since the description information for multiple sub-standard processes of the first standard process is obtained through steps C1-C2, a sub-standard process information library is constructed through step C3. This embodiment of the application constructs a sub-standard process information library using description information of various sub-standard processes, thereby providing a sub-standard process information library for various different tasks and simplifying the method of providing sub-standard processes.

[0179] Furthermore, in some embodiments of this application, the C3 task execution device constructs a sub-standard process information database based on the description information of multiple sub-standard processes of the first standard process, including:

[0180] C31. The task execution device generates second semantic representations corresponding to the multiple sub-standard processes of the first standard process based on the description information of the multiple sub-standard processes of the first standard process.

[0181] C32. The task execution device constructs a sub-standard process information database based on the second semantic representation corresponding to the multiple sub-standard processes of the first standard process.

[0182] The task execution device can perform semantic parsing on multiple sub-standard processes of the first standard process to obtain second semantic representations corresponding to the multiple sub-standard processes. The semantic algorithm used for semantic parsing is not limited; for example, a Large Language Model (LLM) can be used to perform semantic parsing on the descriptive information of the multiple sub-standard processes of the first standard process to obtain the second semantic representations. In this embodiment, the second semantic representations corresponding to the multiple sub-standard processes of the first standard process are one or more vector representations corresponding to the multiple sub-standard processes. These second semantic representations can be sparse or dense. After obtaining the second semantic representations corresponding to the multiple standard processes according to the generated second semantic representations of the multiple sub-standard processes of the first standard process, the task execution device generates a sub-standard process information database. In this embodiment, the second semantic representation corresponding to each sub-standard process can be used for retrieval to achieve matching between tasks and sub-standard processes.

[0183] In some embodiments of this application, the sub-standard process information database includes: multiple sub-standard processes of the first standard process and second semantic representations corresponding to the multiple sub-standard processes of the first standard process, respectively.

[0184] For details on how to obtain the second semantic representation corresponding to the various standard processes, please refer to the description of step C31.

[0185] In some embodiments of this application, the 304 task execution device obtains a first sub-standard process from a first standard process according to a task request, including:

[0186] D1. The task execution device obtains the sub-standard process information database according to the first standard process. The sub-standard process information database includes multiple sub-standard processes, and the multiple sub-standard processes belong to the first standard process.

[0187] D2. The task execution device retrieves the first sub-standard process from the sub-standard process information database according to the task request.

[0188] The task execution device establishes a sub-standard process information database. The sub-standard processes included in the sub-standard process information database belong to the first standard process. The task execution device can also obtain the first sub-standard process by searching the sub-standard process information database. The retrieval of the sub-standard process information database improves the task execution efficiency. The execution of the standard process starts from the starting subtask indicated by the first root node included in the first sub-standard process, reducing unnecessary calculations.

[0189] Furthermore, in some embodiments of this application, step D2, the task execution device obtains the first sub-standard process from the sub-standard process information database according to the task request, including:

[0190] D21. The task execution device performs semantic parsing on the task request to obtain the target semantic representation;

[0191] D22. The task execution device performs semantic matching between the target semantic representation and the sub-standard process information database to obtain a second semantic representation that matches the target semantic representation. The sub-standard process information database includes: multiple sub-standard processes of the first standard process and the second semantic representations corresponding to the multiple sub-standard processes of the first standard process respectively.

[0192] D23. The task execution device obtains the first sub-standard process from the sub-standard process information database based on the second semantic representation that matches the target semantic representation.

[0193] The task execution device can perform semantic parsing on the task request to obtain a target semantic representation. The semantic algorithm used for semantic parsing is not limited; for example, a Large Language Model (LLM) can be used to perform semantic parsing on the task request to obtain the target semantic representation. Based on the aforementioned process of the sub-standard process information database, it is known that the sub-standard process information database stores multiple sub-standard processes and corresponding second semantic representations. Therefore, the task execution device performs semantic matching between the target semantic representation and the sub-standard process information database to obtain a second semantic representation that matches the target semantic representation. Finally, the task execution device retrieves the first sub-standard process from the sub-standard process information database based on the second semantic representation that matches the target semantic representation. In this embodiment, the first sub-standard process can be quickly retrieved from the sub-standard process information database by matching the target semantic representation with the second semantic representation.

[0194] In this embodiment of the application, in addition to providing the task execution method shown in FIG3 above, the task execution device also provides another task execution method, which may include the following steps:

[0195] E1. The task execution device obtains the second standard procedure based on the task request. The second standard procedure is a standard procedure generated based on the task request and the experience database.

[0196] E2. The task execution device stores the second standard procedure in the standard procedure information database.

[0197] In this system, the task execution device interacts with the terminal device, receiving and parsing a task request. The task execution device can access an experience database, which provides expert experience information for generating standard procedures. Using the experience database and the task request, a second standard procedure can be generated. This second standard procedure is stored in a standard procedure information database for dynamic updates. In some optional implementations, the second standard procedure can also be stored in a sub-standard procedure information database for dynamic updates.

[0198] Furthermore, in some embodiments of this application, for example, when no first standard process is found in the standard process information database, the standard process information database can be dynamically updated through steps E1 and E2, so that the standard process information database can obtain the standard process and sub-standard process matching the task request during the next retrieval. Similarly, the standard process information database and the sub-standard process information database can be dynamically updated through steps E1 and E2, so that the standard process information database and the sub-standard process information database can obtain the standard process and sub-standard process matching the task request during the next retrieval.

[0199] In some embodiments of this application, the task execution method performed by the task execution device may include the following steps:

[0200] F1. The task execution device obtains first multimodal information according to the task request. The first multimodal information includes various types of information required to generate standard process information.

[0201] F2. The task execution device inputs the first multimodal information into the large model and outputs the third standard process through the large model. The large model is used to generate standard process information corresponding to various tasks.

[0202] F3. The task execution device stores the third standard process in the standard process information database and the sub-standard process information database.

[0203] The task execution device obtains first multimodal information according to the task request. For example, the first multimodal information includes various types of information required to generate standard process information. The multimodal type of the first multimodal information is not limited. For example, the first multimodal information can be voice data, image data, or text data, etc.

[0204] For example, the first type of multimodal information can include flowcharts, verbal descriptions, and structured data. The flowchart can be as shown in Figure 4. The verbal description information can include: obtaining detailed information such as gender, height, and weight based on the student's student ID; if such information exists, retrieving the corresponding standard weight based on the student's gender and height; and, if the standard weight is found, providing dietary suggestions and specifying a diet plan based on the student's weight deviation from the standard weight. Structured information can include structured configurations and other information obtained through other fields or other structural organization.

[0205] In this embodiment, the task execution device can train a large model and use the trained model to predict the first multimodal information to output a third standard process. The large model is used to generate standard process information corresponding to various tasks. When the first standard process is not found in the standard process information database, the third standard process is stored in the standard process information database to achieve dynamic updates to the standard process information database, so that the standard process information database can retrieve the standard process and sub-standard process matching the task request during the next retrieval. Alternatively, the third standard process can be stored in both the standard process information database and the sub-standard process information database to achieve dynamic updates to the standard process information database and the sub-standard process information database, so that the standard process information database and the sub-standard process information database can retrieve the standard process and sub-standard process matching the task request during the next retrieval.

[0206] In some embodiments of this application, the task execution method performed by the task execution device may include the following steps:

[0207] G1. The task execution device acquires the fourth standard procedure and stores the fourth standard procedure in the standard procedure information database. The fourth standard procedure is standard procedure information generated based on the experience database.

[0208] or,

[0209] H1. The task execution device acquires the second multimodal information, which includes various types of information required to generate standard process information.

[0210] H2. The task execution device inputs the second state information into the large model and outputs the fifth standard process through the large model. The large model is used to generate standard process information corresponding to various tasks.

[0211] H3. The task execution device stores the fifth standard process in the standard process information database and the sub-standard process information database.

[0212] The experience database can provide expert experience information for generating standard processes. When a task request is not required, a fourth standard process can be generated from the experience database and stored in a standard process information repository. This allows for dynamic updates to the standard process information repository, ensuring that the repository retrieves the standard process and sub-standard processes matching the task request during subsequent searches. Similarly, the fourth standard process can be stored in both the standard process information repository and the sub-standard process information repository, enabling dynamic updates to both repositories and ensuring that the repositories retrieve the standard process and sub-standard processes matching the task request during subsequent searches.

[0213] In this embodiment, the task execution device can train a large model and use the trained model to predict the second multimodal information to output a fifth standard process. The large model is used to generate standard process information corresponding to various tasks. The fifth standard process is stored in a standard process information database to achieve dynamic updates to the database, ensuring that the database and sub-standard process information databases retrieve the standard process and sub-standard process matching the task request during the next retrieval. Alternatively, the fifth standard process can be stored in both the standard process information database and the sub-standard process information database to achieve dynamic updates to both databases, ensuring that the databases retrieve the standard process and sub-standard process matching the task request during the next retrieval.

[0214] In some embodiments of this application, the task execution method executed by the task execution device may further perform anomaly detection on the initial subtask and the subtask to which the initial subtask jumps, in order to detect problems existing during the execution of the subtask. Specifically, the task execution method executed by the task execution device may include the following steps:

[0215] I1. During the execution of multiple first sub-tasks through the action model, the task execution device obtains the action sequence corresponding to the first sub-standard process. Multiple actions in the action sequence are used to execute multiple first sub-tasks or sub-task jumps. The multiple actions in the action sequence are implemented based on the sub-task information or sub-task jump logic information of multiple first sub-tasks.

[0216] I2. The task execution device performs anomaly evaluation on the action sequence based on the execution results of multiple actions in the action sequence, the subtask information of multiple first subtasks, and the subtask jump logic information, through the first evaluation model, so as to obtain suspicious actions from the action sequence.

[0217] I3. The task execution device performs anomaly detection on the suspicious action based on the execution result of the suspicious action and the subtask information of the first subtask corresponding to the suspicious action, or based on the execution result of the suspicious action and the subtask jump logic information of the first subtask corresponding to the suspicious action, through the second evaluation model, so as to determine that the suspicious action is an abnormal action.

[0218] The task execution device can detect the process of an action model executing multiple first sub-tasks. This action model can be used to execute multiple sub-actions in a first sub-standard process. The task execution device obtains the action sequence corresponding to the first sub-standard process. Multiple actions in the action sequence are used to execute corresponding sub-tasks or sub-task jumps respectively. The multiple actions in the action sequence are implemented based on the sub-task information or sub-task jump logic information of multiple first sub-tasks. Executing multiple actions in the action sequence can obtain the execution results of multiple actions in the action sequence. For example, the execution results of multiple actions in the action sequence can be stored in the SOP short-term memory. The task execution device can train a first evaluation model. Based on the execution results of multiple actions in the action sequence and the sub-task information and sub-task jump logic information of multiple first sub-tasks, the trained first evaluation model is used to perform anomaly evaluation and obtain suspicious actions from the action sequence. Anomaly evaluation refers to evaluating whether there are anomalies in the action sequence as a whole through the first evaluation model. When anomalies are found in the action sequence, suspicious actions are obtained from the action sequence. The task execution device can train a second evaluation model. The task execution device obtains the execution result of the suspicious action. Based on the execution result of the suspicious action and the subtask information of the first subtask corresponding to the suspicious action, or based on the execution result of the suspicious action and the subtask jump logic information of the first subtask corresponding to the suspicious action, the trained second evaluation model is used to perform anomaly detection on the suspicious action. Anomaly detection refers to evaluating whether the suspicious action is abnormal from a local perspective through the second evaluation model. When the suspicious action is abnormal, it is determined to be an abnormal action.

[0219] In this embodiment of the application, the first evaluation model and the second evaluation model can be large models. The first evaluation model can perform overall anomaly evaluation, and the second evaluation model can perform local anomaly detection. Through the coarse-to-fine two-stage evaluation method, problems in the process can be efficiently discovered, taking into account both the efficiency and effectiveness of process execution.

[0220] Furthermore, in some embodiments of the application, the task execution method executed by the task execution device may include the following steps in addition to performing the aforementioned steps I1 to I3:

[0221] J1. The task execution device optimizes the abnormal action by using an optimization model based on the execution result of the abnormal action and the subtask information of the subtask corresponding to the abnormal action, or based on the execution result of the abnormal action and the subtask jump logic information of the first subtask corresponding to the abnormal action, to obtain the execution result of the optimized action.

[0222] J2. The task execution device evaluates the execution result of the optimized action through the second evaluation model and obtains the first evaluation result.

[0223] The task execution device can train an optimization model. Based on the execution result of the abnormal action and the subtask information or subtask jump logic information of the corresponding subtask, the trained optimization model is used to optimize the action and obtain the optimized action execution result. For example, the optimization model can be a large model.

[0224] The task execution device calls the second evaluation model again for local evaluation. The second evaluation model is used to evaluate the execution result of the optimized action to obtain the first evaluation result. In this embodiment, by optimizing the abnormal action using the optimization model and evaluating the execution result of the optimized action using the second evaluation model, it is possible to detect whether the optimized action still has problems, thereby determining whether the optimization of the abnormal action was successful.

[0225] In other embodiments of this application, the task execution method executed by the task execution device may include the following steps in addition to performing the aforementioned steps I1 to I3 and J1 to J2:

[0226] K1. When the first evaluation result indicates that there is a problem with the execution result of the optimized action, the task execution device uses the backtracking model to backtrack based on the execution result of the optimized action and the subtask information and subtask jump logic information of multiple first subtasks, so as to obtain the first related action that caused the problem from the action sequence and obtain historical backtracking information.

[0227] K2, the task execution device optimizes from the first associated action that caused the problem through the optimization model until the abnormal action is reached, and obtains the execution result of the first associated action after optimization;

[0228] K3, the task execution device evaluates the execution result of the optimized first associated action through the second evaluation model and obtains the second evaluation result.

[0229] The task execution device can train a backtracking model, and perform backtracking based on the optimized action execution results and the subtask information and subtask jump logic information of multiple first subtasks. Using the trained backtracking model, action backtracking is performed to obtain the first related action that caused the problem from the action sequence and to obtain historical backtracking information. For example, this backtracking model can be a large model.

[0230] The task execution device uses a backtracking model to identify the first associated action that caused the problem from the action sequence. It then calls the optimization model again to optimize from this first associated action until an abnormal action is reached, obtaining the execution result of the optimized first associated action. Next, a second evaluation model is called to evaluate the execution result of the optimized first associated action, obtaining a second evaluation result. In this embodiment, by optimizing from the first associated action to the abnormal action using the optimization model, and by evaluating the execution result of the optimized first associated action using the second evaluation model, it is possible to detect whether the action still has problems after further optimization, thus determining whether the optimization from the first associated action to the abnormal action was successful.

[0231] In other embodiments of this application, the task execution method performed by the task execution device may include the following steps in addition to performing the aforementioned steps I1 to I3, J1 to J2, and K1 to K3:

[0232] L1. When the second evaluation result indicates that there is a problem with the execution result of the optimized first associated action, the task execution device uses the backtracking model to backtrack based on historical backtracking information, the execution result of the optimized first associated action, the subtask information of multiple first subtasks, and the subtask jump logic information, so as to obtain the second associated action that caused the problem from the action sequence.

[0233] L2. The task execution device optimizes from the second associated action through the optimization model until the abnormal action is reached, and obtains the execution result of the optimized second associated action.

[0234] L3. The task execution device evaluates the execution result of the optimized second associated action through the second evaluation model to obtain the third evaluation result.

[0235] In this embodiment, when the second evaluation result indicates a problem with the execution result of the optimized first associated action, the task execution device can call the backtracking model again. Based on historical backtracking information, the execution result of the optimized first associated action, and the subtask information and subtask jump logic information of multiple first subtasks, backtracking is performed to obtain the problematic second associated action from the action sequence. The optimization model is then called again to optimize from the problematic second associated action until an abnormal action is reached, obtaining the execution result of the optimized second associated action. Then, the second evaluation model is called again to evaluate the execution result of the optimized second associated action, obtaining a third evaluation result. In this embodiment, by optimizing the second associated action using the optimization model until an abnormal action is reached, and by evaluating the execution result of the optimized second associated action using the second evaluation model, it is possible to detect whether the action still has problems after further optimization, thus determining whether the optimization of the second associated action up to the abnormal action was successful.

[0236] It is understandable that if the third evaluation result indicates that there is a problem with the execution result of the optimized second associated action, steps L1 to L3 can be executed again until the evaluation result output by the second evaluation model indicates that there is no problem with the execution result of the optimized associated action. In some other embodiments, when the aforementioned steps L1 to L3 are executed multiple times, a termination condition can be set, and the optimized action can be output when the termination condition is met.

[0237] Furthermore, in some embodiments of the application, when the first evaluation result indicates that there are no problems with the execution result of the optimized action, it means that the optimization of the abnormal action is successful. The abnormal action is optimized by the optimization model, and the problem of the action is overcome.

[0238] M1. When the first evaluation result indicates that there are no problems with the execution result of the optimized action, the method provided in this application embodiment further includes:

[0239] The task execution device updates the standard process information database using the optimized actions;

[0240] And / or,

[0241] M2, the task execution device uses the optimized actions to train at least one of the following models: action model, first evaluation model, and second evaluation model.

[0242] In this embodiment, when the first evaluation result indicates that there are no problems with the execution result of the optimized action, the task execution device can update the information base and / or optimize the model using the optimized action. The information base may include a standard process information base, or the updated database may include a standard process information base and a sub-standard process information base. The model is then trained: an action model, a first evaluation model, and a second evaluation model. By updating the information base and / or optimizing the model, the information base and model can be used for subsequent task execution, improving the efficiency and success rate of task execution.

[0243] Furthermore, in some embodiments of the application, when the second evaluation result indicates that there are no problems with the execution result of the optimized first associated action, it means that the optimization of the first associated action up to the abnormal action is successful. The optimization model is used to optimize the first associated action up to the abnormal action, thus overcoming the problems existing in the action.

[0244] When the second evaluation result indicates that there is no problem with the execution result of the optimized first associated action, the method provided in this application embodiment further includes:

[0245] N1. The task execution device updates the standard process information database using the optimized first associated action.

[0246] And / or,

[0247] N2. The task execution device trains at least one of the following models using the optimized first associated action: action model, first evaluation model, second evaluation model, and optimization model.

[0248] In this embodiment, when the second evaluation result indicates that the execution result of the optimized first associated action has no problems, the task execution device can update the information base and / or optimize the model using the optimized action. The information base may include a standard process information base, or the updated database may include a standard process information base and a sub-standard process information base. The model is then trained: an action model, a first evaluation model, a second evaluation model, and an optimization model. By updating the information base and / or optimizing the model, the information base and model can be used for subsequent task execution, improving the efficiency and success rate of task execution.

[0249] Furthermore, in some embodiments of the application, when the third evaluation result indicates that there are no problems with the execution result of the optimized second associated action, it means that the optimization of the second associated action up to the abnormal action is successful. The optimization of the second associated action up to the abnormal action through the optimization model overcomes the problems existing in the action.

[0250] When the third evaluation result indicates that there are no problems with the execution result of the optimized second associated action, the method provided in this application embodiment further includes:

[0251] P1. The task execution device updates the standard process information database using the optimized second associated action;

[0252] And / or,

[0253] P2. The task execution device trains at least one of the following models using the optimized second associated action: action model, first evaluation model, second evaluation model, backtracking model, and optimization model.

[0254] In this embodiment, when the third evaluation result indicates that there are no problems with the execution result of the optimized second associated action, the task execution device can update the information base and / or optimize the model using the optimized action. The information base may include a standard process information base, or the updated database may include a standard process information base and a sub-standard process information base. The model is then trained as follows: an action model, a first evaluation model, a second evaluation model, a backtracking model, and an optimization model. By updating the information base and / or optimizing the model, the information base and model can be used for subsequent task execution, improving the efficiency and success rate of task execution.

[0255] This application also provides a task execution method implemented through an intelligent agent, which is a proxy capable of perceiving the environment and taking actions to achieve a specific goal. The intelligent agent can be software, hardware, or a system, possessing autonomy, adaptability, and interactivity. By perceiving changes in the environment, the intelligent agent makes judgments and decisions based on its learned knowledge and algorithms, and then executes actions to influence the environment or achieve predetermined goals. Intelligent agents are widely used in the field of artificial intelligence; they can learn autonomously and continuously evolve to better complete tasks.

[0256] Specifically, the task execution method implemented by the intelligent agent may include the following steps:

[0257] S1. The intelligent agent receives a task request from the terminal device.

[0258] In this process, the intelligent agent interacts with the terminal device, receives task requests from the terminal device, and parses the task requests.

[0259] In this embodiment, the task request can be a task from various business scenarios, and is not limited here. Specifically, the task request can be from the fields of smart home, autonomous driving, gaming, intelligent question answering, and communication. For example, in the smart home field, the task request could be a user initiating fault detection of smart home appliances, or a user controlling a robot vacuum cleaner to customize a route throughout the house. In the autonomous driving field, the task request could be a user initiating automatic obstacle avoidance of a smart vehicle, or controlling a smart vehicle to update maps. In the gaming field, the task request could be a game player controlling a game character to automatically find a path, or controlling one game character to perform multi-task allocation and execution with other game characters. In the communication field, the task request could specifically be evaluating the signal quality of attached terminals in a base station cell.

[0260] The application scenarios for task requests in this application embodiment are not limited. Specifically, they can be determined by combining the business scenarios involved in the tasks sent by the terminal device with the domain in which the intelligent agent framework is applied.

[0261] S2. The agent retrieves the first standard process from the standard process information database based on the task request.

[0262] In this embodiment, the intelligent agent can be configured with a standard process information library, which may include multiple different standard processes that can be used to solve different tasks. The standard process information library stores multiple standard processes, including a first standard process.

[0263] In this embodiment of the application, the intelligent agent can dynamically update the standard process information database to continuously expand the standard process information database, so that the standard process information database can include more standard processes.

[0264] In this embodiment, after the intelligent agent obtains a task request, it uses the task request to retrieve a standard process information database to obtain a first standard process from the database. This first standard process is a standard process stored in the database. In this embodiment, the standard process information database is configured with multiple different standard processes to provide the first standard process corresponding to the task request in the execution task request. This eliminates the need for manual configuration of standard processes for tasks, simplifying the method of providing standard processes for tasks.

[0265] S3. The agent obtains the first sub-standard process from the first standard process according to the task request. The first sub-standard process is a sub-standard process of the first standard process.

[0266] The first sub-standard process is used to instruct multiple first subtasks to be executed sequentially based on jump relationships. The first sub-standard process includes a first root node, which is used to instruct the execution of the starting subtask among the multiple first subtasks.

[0267] The standard process in this application embodiment may include multiple sub-standard processes. For each standard process, there are multiple sub-standard processes, which are used to generate a complete standard process. For example, the first standard process is used to execute a task request. The task request can be broken down into a series of multiple sub-tasks through the first standard process. Each sub-task has a clear execution method, and the jump logic between multiple sub-tasks is linked by the execution result.

[0268] For example, an intelligent agent can configure a sub-standard process information library for each standard process in the standard process information library. The sub-standard process information library can include multiple sub-standard processes corresponding to the standard process. Searching the sub-standard process information library can find the first sub-standard process.

[0269] In this embodiment, after the intelligent agent obtains a first standard process for executing a task request, it obtains a sub-standard process information library corresponding to the first standard process. This sub-standard process information library may include multiple sub-standard processes. The sub-standard process information library stores multiple sub-standard processes included in the first standard process. After obtaining the first standard process, the intelligent agent uses the task request to retrieve the sub-standard process information library corresponding to the first standard process to obtain the first sub-standard process from the sub-standard process information library. The first sub-standard process is a sub-standard process stored in the sub-standard process information library. The first root node included in the first sub-standard process indicates the starting subtask among multiple first subtasks. This starting subtask can also be considered as the task entry point of the first standard process. For example, the intelligent agent retrieves the sub-standard process information library corresponding to the first standard process according to the task request, determines the first sub-standard process, and the first sub-standard process corresponds to the task entry point.

[0270] In this embodiment, the substandard process information database is configured with a variety of different substandard processes to provide a task entry point for determining the first standard process. This allows for flexible selection of the starting subtask in the first standard process based on the task request, thus making it applicable to complex and dynamic tasks and improving the flexibility of executing task requests.

[0271] S4. The agent executes multiple first sub-tasks according to the first sub-standard process.

[0272] After the agent obtains the first sub-standard process, it can obtain the starting subtask and other first subtasks that the starting subtask jumps to based on the first sub-standard process. Then, according to the task content provided in the task request, it executes the starting subtask using the first sub-standard process. After the starting subtask is completed, it jumps to other first subtasks according to the first sub-standard process and executes them using the first sub-standard process. It is not limited that, in this embodiment, the subtask jump can be one hop or multiple hops. When the termination condition of a first subtask is met, the execution of the first subtask stops, and the execution result for the task request is output.

[0273] According to the task execution method provided in the embodiments of this application, the intelligent agent is configured with standard processes in a standard process information library. This library provides a first standard process corresponding to a task request, eliminating the need for manual configuration of standard processes for each task request and simplifying the process of providing standard processes for task requests. Furthermore, the intelligent agent can obtain a first sub-standard process from the first standard process based on the task request. This first sub-standard process instructs multiple first sub-tasks to be executed sequentially, and the first root node included in the first sub-standard process indicates the starting sub-task among the multiple first sub-tasks. Therefore, multiple first sub-tasks can be executed according to the first sub-standard process. Since the first sub-standard process can be obtained from the first standard process based on the task request, corresponding first sub-standard processes can be obtained for different task requests, making it applicable to complex and dynamic different task requests, improving the flexibility of executing multiple first sub-tasks, and increasing the execution efficiency and success rate of multiple sub-tasks.

[0274] In the embodiments of S1 to S4 described above, the method steps performed by the intelligent agent can also refer to the contents of the embodiment shown in Figure 3 above, and see the examples in steps 302 to S305 above for details.

[0275] To facilitate a better understanding and implementation of the above-described solutions in the embodiments of this application, specific examples of corresponding application scenarios are provided below.

[0276] Next, we will use the standard process as the SOP and the sub-standard process as the sub-SOP for example.

[0277] To illustrate the real-world business scenarios provided in this application, the large-model intelligent agent framework for process task automation can be used in various real-world business scenarios. For example, in the root cause localization scenario in the operations and maintenance field, it is necessary to gradually locate the underlying root cause using a predefined fault tree or flowchart. The selection and execution of sub-tasks involved can be improved through a large model to enhance the accuracy and generalization ability of the tasks. As another example, in the intelligent customer service assistant scenario, different branches need to be selected based on different user tasks to achieve increasingly refined solutions. Therefore, a large-model-based intelligent agent framework is also needed to assist users in making decisions during the intermediate processes.

[0278] The following five examples will illustrate the points.

[0279] The following Example 1 illustrates the process of dynamic retrieval based on the SOP information base and sub-SOP information base. This embodiment of the application can improve the efficiency of SOP process execution. Specifically, SOP description information and subtask description information are constructed using SOP information, and corresponding indexes are generated; suitable SOPs and task entry points are dynamically retrieved based on user task requests. Through dynamic retrieval of SOPs and task entry points based on user task requests, the coverage of user task requests and the efficiency of process execution are improved.

[0280] The construction process of the SOP information base and sub-SOP information base will be explained in the following embodiment 2. By expanding the application scenarios of SOP planning through the automatic production method of SOP, the embodiments of this application can improve the task coverage of SOP. Specifically, SOP information is generated using multimodal data to form an SOP information base, which can be used for retrieval of user task requests.

[0281] The SOP evaluation and feedback process is illustrated in Example 3, which integrates coarse-to-fine granular process evaluation and adopts a coarse-to-fine anomaly localization method, balancing execution efficiency and effectiveness. This application's embodiment can improve SOPs; specifically, it uses a coarse-to-fine evaluation approach to locate abnormal actions.

[0282] The following Example 4 illustrates the process of improving SOP execution, integrating reflective and adaptive problem backtracking mechanisms to meet the complexity of action dependencies within the process and enhance the ability to improve problems.

[0283] The iterative optimization process of SOPs will be illustrated in Example 5. This application's embodiments enable long-term SOP optimization. Specifically, through problem assessment, backtracking, and reflective improvement, experiential data can be accumulated to optimize the execution of the SOP planning process. By integrating various experiential data utilization methods, the evaluation, improvement, and execution capabilities of the SOP process are enhanced, achieving self-evolution of the method.

[0284] Figure 5 shows a flowchart illustrating a task execution method applied in an embodiment of this application, which mainly includes:

[0285] S1. Receive task requests sent by terminal devices.

[0286] S2. Retrieve SOPs from the SOP information database based on the task request.

[0287] For details, please refer to the following embodiment for an explanation of the SOP generation process in the SOP information database.

[0288] S3. Determine if there is an SOP for executing the task.

[0289] S4. If no SOP for performing the task is found in the SOP information database, an SOP is generated based on the task and the expert experience database. The generated SOP is then stored in the SOP information database, which can also be called the SOP long-term memory. For example, an SOP can specifically be code.

[0290] See the example of generating SOP in the following Example 2.

[0291] S5. If an SOP for performing the task is retrieved from the SOP information database, the sub-SOP in the sub-SOP information database corresponding to the SOP is determined according to the task, and the SOP planning and execution is performed. The execution results of the action sequence generated during the SOP planning and execution process are stored in the SOP short-term memory.

[0292] S6. Conduct SOP evaluation and feedback on the SOP planning and execution process.

[0293] See the example of SOP evaluation feedback in the following Example 2.

[0294] S7. Determine if any issues were found in the SOP assessment feedback.

[0295] See Example 3 for a detailed explanation of the problems identified from SOP evaluation feedback.

[0296] S8. If a problem is found, an SOP will be improved. After the improvement is completed, S5 will be re-executed.

[0297] See the examples of SOP improvements in the following embodiments four and five.

[0298] If no problems are found, the execution will end.

[0299] This application's embodiments target real-world user task requests online. By retrieving SOPs and task entry points, suitable SOPs and starting subtasks for resolving the task are located. If no suitable SOP is found, corresponding SOP information is automatically generated and stored in an SOP information database for future retrieval. After the SOP information database is formed, the retrieved or generated SOPs are executed step-by-step according to the process logic. The execution process interacts with historical execution results to ensure accurate utilization of contextual information. After execution, a coarse-grained and fine-grained evaluation and feedback mechanism is used to locate potential anomalies. If problems are found, adaptive problem improvement is automatically implemented to optimize process execution. Finally, after accumulating execution data from multiple processes, the overall framework's ability to solve SOP planning tasks can be further improved through configuration optimization or model optimization.

[0300] The task execution method provided in the embodiments of this application will be described in detail below:

[0301] Example 1 of this application:

[0302] As shown in Figure 6, this embodiment includes the following steps:

[0303] Step S01: Summarize the SOP information to obtain complete SOP description information.

[0304] In step S01, the summarization of SOP information can draw on methods such as text summarization and graph learning to aggregate the description information of subtasks organized in flowchart form to obtain the description information of the entire SOP. For example, node aggregation can be used to continuously aggregate the information of downstream subtasks from bottom to top. The aggregation method can utilize a large model for summarization to obtain the summary of the entire SOP information.

[0305] For example, SOP information should at least include: subtask information for each subtask and jump information between subtasks. The following is an example of the information configuration file for a dietary recommendation scenario:

[0306] Step S02: For each subtask in the SOP, summarize the sub-SOP information corresponding to each subtask to obtain the description information required when the subtask is used as the task entry point.

[0307] In step S02, for each sub-SOP corresponding to a sub-task, the sub-SOP information can be aggregated as task entry information using a method similar to that in S01. Specifically, a dynamic programming approach can be used to construct the sub-SOP description information from the bottom up, improving the efficiency of generating SOP description information. Dynamic programming refers to continuously aggregating sub-SOPs from the bottom up, gradually constructing the sub-SOP description information.

[0308] Step S03: Construct the index of SOP and the index of task entry based on the SOP description information and the sub-SOP description information of task entry, respectively.

[0309] In step S03, based on the description information obtained in steps S01 and S02 above, sparse or dense semantic representations can be used to map the information to one or more vector representations, and key-value pairs are constructed based on this to form the SOP information base and sub-SOP information base respectively.

[0310] For example, as shown in Figure 6, the SOPs stored in the long-term memory of the SOPs i It includes multiple sub-SOPs. For example, the starting subtasks corresponding to multiple sub-SOPs are A, B, C, D, E, F, G, H and I. Each sub-SOP is used to execute different subtasks and the subtasks that the subtasks jump to.

[0311] Step S04: For user task requests, locate the relevant SOPs and task entry points based on the retrieval method.

[0312] In step S04, for the user task request, sparse or dense semantic representations are constructed in a corresponding modal form. A semantic matching retrieval method is used to retrieve the SOP configuration, locating the most relevant SOP from the constructed SOP information database. i And using semantic matching retrieval methods to retrieve task entry points from the sub-SOP information database. ij For example, if the starting subtask of the retrieved sub-SOP is B, then B will be used as the task entry point to execute the user's task request.

[0313] Semantic parsing parses the task request into a semantic representation, such as segmenting the task into words using a pre-defined vocabulary and mapping them to corresponding identifiers (IDs). Semantic representation uses the model to process the semantic representation obtained from task parsing and generate feature vectors, such as using a model that generates sentence vectors to generate feature vectors based on the ID sequence mapped to the task.

[0314] As described in Embodiment 1 of this application, for actual task requests, users do not need to obtain the corresponding SOP planning details to determine whether a specific task can be resolved. The retrieval method reduces the need to understand SOP configuration information in real-world scenarios, improving SOP usability. Simultaneously, task retrieval also improves task coverage; repeated calls to the retrieval method during task resolution ensure that various user task requirements are met. Finally, task entry point retrieval improves execution efficiency, focusing the process solely on the user's task request and reducing unnecessary computation.

[0315] Example 2 of this application:

[0316] This embodiment consists of the following steps:

[0317] Step S11: Generate SOP planning based on expert experience.

[0318] Based on the task request, in the absence of relevant SOPs, SOP information is generated using expert experience.

[0319] For example, for a task involving dietary recommendations, the SOP information database does not store the corresponding SOP for that task. It is necessary to combine the task "Give dietary recommendations to student ID 12345, etc." with the experience database. For example, the experience database can provide tools to execute subtasks 1, 2, and 3 respectively. Fill in the required field information in the SOP information configuration file as shown above.

[0320] Step S12: Automatic generation of SOP planning based on large model.

[0321] Based on the task request, and combining multimodal information such as flowcharts, language descriptions, and structured data, SOP configuration information is automatically generated through a large model.

[0322] In step S12, multimodal generation technology automatically maps business-related SOP information to the space of structured configuration or code. The technologies involved include: multimodal data retrieval based on user queries, providing examples, and output format prompting. For example, multimodal data retrieval refers to retrieving available multimodal databases using user task requests, similar to the retrieval of SOP information databases. Providing examples is a method in large-scale model prompting, i.e., providing similar input and output examples so that the large model can better understand the output requirements. For example, providing multimodal data from different tasks and the configuration information generated under that multimodal data as examples. Output format refers to the required output format information. An example of an output format is as follows: You are an expert in generating JSON format data. Please refer to and understand the provided flowchart to generate the corresponding structured configuration information, where the dictionary key corresponds to the subtask index; the value must contain "task" and "next," i.e., the definition of the subtask and the next subtask. Similar SOP configuration data is as follows: ..., please generate configuration information that meets the requirements.

[0323] For example, a multimodal large-scale model can be trained based on massive amounts of web page data and manually collected question-and-answer data. Taking GPT-4V as an example, GPT-4V is a vision-enabled GPT-4 that can automatically generate configuration information based on multimodal information, such as the flowchart shown in Figure 4, and the organization of information, such as the meaning of the sub-task objective (task) and next-hop sub-task (next) fields. It can automatically generate the corresponding structured configuration. The language description information could be: obtaining detailed information such as gender, height, and weight based on the student's student ID; if the student's gender and height exist, retrieving the corresponding standard weight based on the student's gender and height; and providing dietary suggestions and specifying a diet plan based on the student's weight deviation from the standard weight if the standard weight is found. The above multimodal information is provided as input to the large-scale model, and the output requirements are provided in text form, such as generating a structured configuration. The large-scale model can then generate the aforementioned structured configuration.

[0324] Step S13: Construct a Standard Operating Procedure (SOP) information database.

[0325] The generated SOP information can be used to build an SOP information database; depending on the different domain tasks, the SOP generation methods in steps S12 and S13 can also be used to expand the SOP information database in the offline stage.

[0326] In step S13, the SOP information database is expanded using SOP information automatically generated from an expert database or a large model, and an SOP index is constructed. In the offline phase, even without a user request, SOP information can be generated from multimodal data, and the SOP information database can be constructed offline. For example, even without a user task request, SOP information can be generated using prompt word engineering techniques by providing examples or output format requirements.

[0327] As can be seen from the description of Embodiment 2 of this application, SOP generation can reduce or even eliminate the need for human experts to write related SOP information or code, which can greatly reduce the workload of manual labor and improve the usability of the framework. At the same time, utilizing multimodal information accumulated in business can also make the generated SOP information more accurate, ensuring that problem-solving is carried out according to the expected business process.

[0328] Example 3 of this application:

[0329] This embodiment consists of the following steps:

[0330] Step S21: For the action sequence in the SOP execution process, use overall coarse evaluation to locate the problematic action.

[0331] In step S21, based on the action sequence during the SOP execution process, combined with the subtask description information defined in the SOP and the jump logic relationship between subtasks, the large model is used to determine whether there are problematic actions, including but not limited to: using the large model based on prompt word engineering to provide evaluation based on SOP information and execution information, constructing a positive and negative sample learning evaluator, and direct matching of action results with expected information, etc.

[0332] For example, a prompt could be: "The sequence of actions to be performed is as follows:...". Please determine whether there are any errors in the assistant's responses during the above dialogue. If so, identify the earliest error. This prompt and the provided sequence of actions allow for a rough evaluation of the overall action sequence.

[0333] For example, if the prompt words from step S21 are input into the large model, the model may output the following evaluation: There are content errors in the dialogue process, and an error occurred in the response in the 4th round.

[0334] For example, positive samples can be action sequences corresponding to tasks that are completed normally, while negative samples are action sequences corresponding to tasks that have problems during execution. Using positive and negative samples, binary classifiers and sequence labeling models can be trained to determine whether a problem exists and where the problem occurs for any action sequence to be tested.

[0335] For example, expected information mainly refers to subtask information where the execution of certain actions in the action sequence does not meet the subtask definition. For instance, if the subtask is to search for the biography of a person on the Internet, but the search results are empty, then it does not match the requirements of the subtask and can be considered to have a problem.

[0336] Step S22: For the specific action that has been located, use local detailed evaluation to examine the problem of the action.

[0337] In step S22, based on the action sequence during the SOP execution process and the specific subtask description information, the large model is used to determine whether there is a problem in the execution of the current subtask. The specific method is the same as that listed in step S21.

[0338] For example, local evaluation can be called fine evaluation. Fine evaluation is to evaluate whether there are any problems with the action based on the subtask information, subtask jump logic information and action execution results of the located subtask. For example, the following prompt words can be used to input into the large model for action evaluation: For the current subtask information and subtask jump logic information: ..., the preliminary action output is: ..., please analyze in detail whether the preliminary action output results meet the requirements.

[0339] For example, a task request might include providing dietary advice for student ID 12345. The execution result of subtask 1 might be: "Student ID 12343's height is xxx, weight is xxx." However, the execution result has a problem because the output ID 12343 does not match the ID 12345 in the task request.

[0340] Step S23: If the detailed assessment and localization are successful, return to the abnormal action; otherwise, use the current feedback information to perform a rough localization of the problem again to determine whether there are any other problems.

[0341] In step S23, if the fine localization result finds no anomalies, the coarse localization in step S21 is repeated. This re-performation includes additional feedback information, and the specific method remains the same as in step S21. Other issues may be abnormal actions different from those previously identified, i.e., other actions in the action sequence.

[0342] As can be seen from the description of Embodiment 3 of this application, if the action evaluation method only adopts overall evaluation, it cannot accurately locate the problem. If only a local fine evaluation method is used, the problem location method is relatively inefficient. For SOP process execution, the embodiment of this application adopts a coarse-to-fine two-stage evaluation method, which can efficiently discover problems in the process and take into account both the efficiency and effectiveness of process execution.

[0343] Example 4 of this application

[0344] As shown in Figure 7, this embodiment consists of the following steps:

[0345] Step S31: Utilize feedback information from the positioning process to reflect on and improve the execution results of problematic actions.

[0346] In step S31, based on the aforementioned anomaly location information, subtask information, and subtask jump logic information, the execution optimization of subtasks is carried out using the large model based on prompt word engineering, thereby improving the problem.

[0347] For example, based on the aforementioned anomaly location information, subtask information, and subtask jump logic information, the execution optimization of subtasks can be carried out using a large model based on prompt words, thereby improving the problem. For example, prompt words can be constructed as follows: In response to the evaluation feedback information: ..., re-execute the ... action to avoid the problem indicated in the feedback information.

[0348] Step S32: If the improvement fails, use a dynamic backtracking model to locate the relevant actions for improvement.

[0349] In step S32, if problems still exist after the improved result is evaluated, the actions that may have caused the current problem are backtracked based on the execution results of the preceding actions, subtask information, and subtask jump logic information. A dynamic backtracking model is used to locate relevant actions for improvement, including but not limited to using a large model to identify the relevant actions that caused the problem based on evaluation feedback, SOP information, execution results, and other information. After the problem is backtracked, the backtracked actions are re-executed up to the subsequent actions of the problematic action.

[0350] For example, the following prompt can be constructed: Based on the execution sequence of actions: ..., please analyze which round's result is most likely to have caused the abnormal action in round 4.

[0351] As shown in Figure 7, after the optimization model improves action D in the action sequence, the improved action D is input into the evaluation model. The evaluation model evaluates the improved D. If the evaluation result indicates that the improvement was unsuccessful, the backtracking model is used to backtrack and determine that the related action B of the problematic action D has a problem. Then, action B is input into the optimization model and improved through the optimization model.

[0352] Among them, the optimization model, evaluation model and backtracking model shown in Figure 7 can be large models.

[0353] Step S33: If the problem is still not fixed, backtrack again.

[0354] In step S33, if the problematic action is still not fixed, the historical backtracking information is used to perform a new backtracking process. The backtracking method is the same as in step S32, but historical backtracking information is added as additional information. For example, the historical backtracking information can be the analysis results output by the large model based on the constructed prompt words.

[0355] As described in Embodiment 4 of this application, current problem improvement is achieved through reflective improvement in a fixed form, that is, the triggering conditions and dependency paths for reflection are determined, making the improvement method too rigid. This embodiment of the application integrates a reflective and adaptive problem backtracking mechanism, satisfying the complexity and flexibility of action dependencies within the process, and improving the ability to improve problems.

[0356] Example 5 of this application:

[0357] This embodiment consists of the following steps:

[0358] Step S41: Optimize the SOP information database using the data generated during the evaluation, retrospection, and improvement process. The SOP information database stores SOP information.

[0359] In step S41, the evaluation and backtracking data can be used to configure key checkpoints and backtracking dependency paths. The SOP improvement data can be used to fine-tune prompt words and optimize task description information, as well as provide reference examples.

[0360] Step S42: Update the optimization model, evaluation model and backtracking model shown in Figure 7 using the data generated during the evaluation, backtracking and improvement process.

[0361] In step S42, the evaluated, backtracked, and improved data can be used to generate corresponding training samples and to specifically optimize the corresponding model capabilities.

[0362] The following is an example of a key checkpoint. The historical execution results of the current SOP show that the action problem always occurs at subtask 3. Subsequent evaluations can focus on checking the action results of subtask 3.

[0363] The following is an example of a backtracking path. The historical execution results of the current SOP show that the problem of subtask 3 is always caused by subtask 1. Therefore, the backtracking of the action problem of subtask 3 can focus on checking whether there is a problem with the action of subtask 1.

[0364] Examples of fine-tuning prompt words are illustrated below. By adjusting the prompt words, the output of the large model can be made to meet expectations. For example, adjusting the prompt words in the above evaluation can make the evaluation results more accurate. There are no restrictions on the specific adjustment methods for the prompt words.

[0365] The following is an example of optimizing task description information. Optimization task information refers to the information of subtasks. The information of subtask 1 in the SOP information is "to obtain detailed information such as gender, height, and weight based on the student's student ID". Optimization can obtain "to obtain detailed information such as gender, height, and weight based on the student's ID", which can help the current subtask to be executed better. Because the evaluation may also affect the evaluation by utilizing the information of the subtask.

[0366] The following example illustrates how successful subtask execution can be provided as additional input to the model to improve the success rate of subtask execution.

[0367] As can be seen from the description of Embodiment 5 of this application, by accumulating historical experience data obtained during the improvement process, the SOP information base and model capabilities can be optimized, so that subsequent user task requests can be met more efficiently and accurately.

[0368] Based on the foregoing examples, this application provides dynamic retrieval of SOPs and task entry points. SOP retrieval improves the coverage of user requests, while task entry point retrieval enhances process execution efficiency. This application also provides SOP generation, expanding the application scenarios of SOP planning. Furthermore, this application provides SOP evaluation feedback, integrating coarse-to-fine granular process evaluation and employing a coarse-to-fine anomaly localization method to balance execution efficiency and effectiveness. Finally, this application provides SOP adaptive improvement, integrating reflective and adaptive problem backtracking mechanisms to meet the complexity and flexibility of action dependencies within the process, enhancing problem improvement capabilities. Finally, this application provides SOP iterative optimization, integrating various methods of utilizing empirical data to improve the evaluation, improvement, and execution capabilities of the SOP process, achieving self-evolution of the method.

[0369] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0370] To facilitate better implementation of the above-described solutions in the embodiments of this application, related apparatus for implementing the above-described solutions is also provided below.

[0371] Please refer to Figure 8. An embodiment of this application provides a task execution device 800, which may include: a receiving module 801, a standard process acquisition module 802, a sub-standard process acquisition module 803, and a task execution module 804.

[0372] The receiving module is used to receive task requests;

[0373] The standard process acquisition module is used to acquire a first standard process from the standard process information database according to the task request;

[0374] The sub-standard process acquisition module is used to acquire a first sub-standard process from the first standard process according to the task request. The first sub-standard process is a sub-standard process of the first standard process. The first sub-standard process is used to instruct multiple first subtasks to be executed sequentially based on a jump relationship. The first sub-standard process includes a first root node, which is used to instruct the execution of the starting subtask among the multiple first subtasks.

[0375] The task execution module is used to execute the plurality of first sub-tasks according to the first sub-standard process.

[0376] In this embodiment, the module is an example of a software functional unit, and the data processing device may include code running on a computing instance. The computing instance may be at least one of a physical host (computing device), a virtual machine, a container, or other computing devices. Further, the aforementioned computing device may be one or more. For example, the data processing device may include code running on multiple hosts / virtual machines / containers. It should be noted that the multiple hosts / virtual machines / containers used to run the application may be distributed in the same region or in different regions. The multiple hosts / virtual machines / containers used to run the code may be distributed in the same available zone (AZ) or in different AZs, each AZ including one or more geographically proximate data centers. Typically, a region may include multiple AZs.

[0377] Similarly, multiple hosts / virtual machines / containers used to run this code can be distributed within the same Virtual Private Cloud (VPC) or across multiple VPCs. Typically, a VPC is set up within a single region. Communication between two VPCs within the same region, and between VPCs in different regions, requires a communication gateway to be set up within each VPC to enable interconnection between VPCs.

[0378] As an example of a hardware functional unit, a data processing device may include at least one computing device, such as a server. Alternatively, the data processing device may also be a device implemented using an application-specific integrated circuit (ASIC) or a programmable logic device (PLD). The aforementioned PLD may be implemented using a complex PLD (CPLD), a field-programmable gate array (FPGA), generic array logic (GAL), or any combination thereof.

[0379] The data processing unit comprises multiple computing devices that can be distributed within the same region or in different regions. Similarly, the multiple computing devices can be distributed within the same Availability Zone (AZ) or in different AZs. Likewise, the multiple computing devices can be distributed within the same Virtual Private Cloud (VPC) or multiple VPCs. These multiple computing devices can be any combination of computing devices such as servers, ASICs, PLDs, CPLDs, FPGAs, and GALs.

[0380] This application also provides a computing device 130. As shown in FIG9, the computing device 130 includes a bus 132, a processor 134, a memory 136, and a communication interface 138. The processor 134, the memory 136, and the communication interface 138 communicate with each other via the bus 132. The computing device 130 may be a server or a terminal device. It should be understood that this application does not limit the number of processors and memories in the computing device 130.

[0381] Bus 132 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, only one line is used in Figure 9, but this does not imply that there is only one bus or one type of bus. Bus 134 can include pathways for transmitting information between various components of computing device 130 (e.g., memory 136, processor 134, communication interface 138).

[0382] The processor 134 may include any one or more processors such as a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor (MP), or a digital signal processor (DSP).

[0383] Memory 136 may include volatile memory, such as random access memory (RAM). Processor 134 may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid state drive (SSD).

[0384] The memory 136 stores executable program code, and the processor 134 executes the executable program code to implement the functions of the aforementioned acquisition module and training module, thereby realizing the data processing method applied to the computing device cluster in the above embodiments. That is, the memory 136 stores instructions for executing the data processing method applied to the computing device cluster in the above embodiments.

[0385] The communication interface 138 uses transceiver modules such as, but not limited to, network interface cards and transceivers to enable communication between the computing device 130 and other devices or communication networks.

[0386] This application also provides a computing device cluster. The computing device cluster includes at least one computing device. The computing device can be a server, such as a central server, an edge server, or a local server in a local data center. In some embodiments, the computing device can also be a terminal device such as a desktop computer, a laptop computer, or a smartphone.

[0387] As shown in Figure 10, the computing device cluster includes at least one computing device 130. The memory 136 of one or more computing devices 130 in the computing device cluster may store the same instructions for performing data processing methods.

[0388] In some possible implementations, the memory 136 of one or more computing devices 130 in the computing device cluster may also store partial instructions for executing data processing methods. In other words, a combination of one or more computing devices 130 can jointly execute instructions for executing data processing methods.

[0389] It should be noted that the memory 136 in different computing devices 130 within the computing device cluster can store different instructions, each used to execute a portion of the data processing method's functions. That is, the instructions stored in the memory 136 of different computing devices 130 can implement one or more functions of the acquisition module.

[0390] In some possible implementations, one or more computing devices in a computing device cluster can be connected via a network. This network can be a wide area network (WAN) or a local area network (LAN), etc. Figure 10 illustrates one possible implementation. As shown in Figure 10, two computing devices 130A and 130B are connected via a network. Specifically, they are connected to the network through communication interfaces in each computing device. In this type of possible implementation, the memory 136 in computing device 130A can store instructions for executing the functions of a first processing module. Simultaneously, the memory 136 in computing device 130B can store instructions for executing the functions of a second processing module. Alternatively, the memory 136 in computing device 130A can store instructions for executing some functions of the second processing module, while the memory 136 in computing device 130B can store instructions for executing another part of the functions of the second processing module, and so on.

[0391] It should be understood that the functions of computing device 130A shown in Figure 11 can also be performed by multiple computing devices 130. Similarly, the functions of computing device 130B can also be performed by multiple computing devices 130.

[0392] This application embodiment also provides another computing device cluster. The connection relationship between the computing devices in this computing device cluster can be similar to the connection method of the computing device cluster in Figure 10. The difference is that the memory 136 of one or more computing devices 130 in this computing device cluster can store the same instructions for executing data processing methods.

[0393] In some possible implementations, the memory 136 of one or more computing devices 130 in the computing device cluster may also store partial instructions for executing data processing methods. In other words, a combination of one or more computing devices 130 can jointly execute instructions for executing data processing methods.

[0394] It should be noted that the memory 136 in different computing devices 130 within the computing device cluster can store different instructions for executing parts of the data processing methods. That is, the instructions stored in the memory 136 of different computing devices 130 can implement one or more functions of the processing module.

[0395] This application also provides a computer program product containing instructions. The computer program product may be a software or program product containing instructions, capable of running on a computing device or stored on any usable medium. When the computer program product is run on at least one computing device, it causes the at least one computing device to perform a data processing method.

[0396] This application also provides a computer-readable storage medium. The computer-readable storage medium can be any available medium that a computing device can store, or a data storage device such as a data center that includes one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive). The computer-readable storage medium includes instructions that instruct the computing device to perform a data processing method.

[0397] This application also provides a chip system including a processor for implementing the steps performed by the aforementioned computing device cluster. In one possible design, the chip system may further include a memory for storing necessary program instructions and data. This chip system may be composed of chips or may include chips and other discrete devices.

[0398] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0399] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0400] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0401] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0402] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A task execution method characterized by, include: Receive task requests; The first standard process is retrieved from the standard process information database according to the task request; According to the task request, a first sub-standard process is obtained from the first standard process. The first sub-standard process is a sub-standard process of the first standard process. The first sub-standard process is used to instruct multiple first subtasks to be executed sequentially based on a jump relationship. The first sub-standard process includes a first root node, which is used to instruct the execution of the starting subtask among the multiple first subtasks. The plurality of first sub-tasks are executed according to the first sub-standard process.

2. The method of claim 1, wherein, The step of obtaining the first standard process from the standard process information database according to the task request includes: The task request is semantically parsed to obtain a target semantic representation; Semantic matching is performed between the target semantic representation and the standard process information database to obtain a first semantic representation that matches the target semantic representation. The standard process information database includes: multiple standard processes and first semantic representations corresponding to the multiple standard processes respectively. The first standard process is obtained from the standard process information database based on the first semantic representation that matches the target semantic representation.

3. The method according to claim 1 or 2, characterized in that, The step of obtaining the first sub-standard process from the first standard process according to the task request includes: A sub-standard process information database is obtained according to the first standard process. The sub-standard process information database includes multiple sub-standard processes, and the multiple sub-standard processes belong to the first standard process. The first substandard process is retrieved from the substandard process information database according to the task request.

4. The method of claim 3, wherein, The step of retrieving the first sub-standard process from the sub-standard process information database according to the task request includes: The task request is semantically parsed to obtain a target semantic representation; Semantic matching is performed between the target semantic representation and the sub-standard process information base to obtain a second semantic representation that matches the target semantic representation. The sub-standard process information base includes: multiple sub-standard processes of the first standard process and second semantic representations corresponding to the multiple sub-standard processes of the first standard process, respectively. The first substandard process is obtained from the substandard process information database based on the second semantic representation that matches the target semantic representation.

5. The method according to any one of claims 1 to 4, characterized in that, The method further includes: When the first standard process is not obtained from the standard process information database, a second standard process is obtained according to the task request. The second standard process is a standard process generated based on the task request and the experience database. The second standard process is stored in the standard process information database.

6. The method according to any one of claims 1 to 5, characterized in that, The method further includes: When the first standard process is not obtained from the standard process information database, the first multimodal information is obtained according to the task request. The first multimodal information includes various types of information required to generate the standard process. The first multimodal information is input into the large model, and the large model outputs the third standard process. The large model is used to generate standard processes corresponding to various tasks. The third standard process is stored in the standard process information database.

7. The method according to any one of claims 1 to 6, characterized in that, The method further includes: Obtain the fourth standard process and store the fourth standard process in the standard process information database. The fourth standard process is a standard process generated based on an experience database. or, Acquire second multimodal information, which includes various types of information required to generate the standard process; The second state information is input into the large model, and the large model outputs the fifth standard process. The large model is used to generate standard process information corresponding to various tasks. The fifth standard process is stored in the standard process information database.

8. The method according to any one of claims 1 to 7, characterized in that, The method further includes: During the execution of the multiple first sub-tasks through the action model, an action sequence corresponding to the first sub-standard process is obtained. Multiple actions in the action sequence are used to execute the multiple first sub-tasks or sub-task jumps. The multiple actions in the action sequence are implemented based on the sub-task information or sub-task jump logic information of the multiple first sub-tasks. Based on the execution results of multiple actions in the action sequence, the subtask information of the multiple first subtasks, and the subtask jump logic information, the action sequence is evaluated for anomalies using a first evaluation model to obtain suspicious actions from the action sequence. Based on the execution result of the suspicious action and the subtask information of the first subtask corresponding to the suspicious action, or based on the execution result of the suspicious action and the subtask jump logic information of the first subtask corresponding to the suspicious action, the suspicious action is subjected to anomaly detection by the second evaluation model to determine that the suspicious action is an abnormal action.

9. The method of claim 8, wherein, The method further includes: Based on the execution result of the abnormal action and the subtask information of the first subtask corresponding to the abnormal action, or based on the execution result of the abnormal action and the subtask jump logic information of the first subtask corresponding to the abnormal action, the abnormal action is optimized by an optimization model to obtain the execution result of the optimized action. The execution result of the optimized action is evaluated using the second evaluation model to obtain the first evaluation result.

10. The method of claim 9, wherein, The method further includes: When the first evaluation result indicates that there is a problem with the execution result of the optimized action, the backtracking model is used to backtrack based on the execution result of the optimized action, the subtask information of the multiple first subtasks and the subtask jump logic information, so as to obtain the first associated action that caused the problem from the action sequence and obtain historical backtracking information. The optimization model is used to optimize the first associated action that caused the problem until the abnormal action is reached, and the execution result of the optimized first associated action is obtained. The second evaluation model is used to evaluate the execution result of the optimized first associated action to obtain a second evaluation result.

11. The method of claim 10, wherein, The method further includes: When the second evaluation result indicates that there is a problem with the execution result of the optimized first associated action, the backtracking model is used to backtrack based on the historical backtracking information, the execution result of the optimized first associated action, the subtask information of the multiple first subtasks, and the subtask jump logic information, so as to obtain the second associated action that caused the problem from the action sequence. The optimization model is used to optimize the second associated action until the abnormal action is reached, thereby obtaining the execution result of the optimized second associated action. The execution result of the optimized second associated action is evaluated using the second evaluation model to obtain a third evaluation result.

12. The method of claim 9, wherein, When the first evaluation result indicates that there are no problems with the execution result of the optimized action, the method further includes: The optimized actions are used to update the standard process information database; And / or, The optimized action is used to train at least one of the following models: an action model, a first evaluation model, and a second evaluation model.

13. The method of claim 10, wherein, When the second evaluation result indicates that there is no problem with the execution result of the optimized first associated action, the method further includes: The optimized first associated action is used to update the standard process information database; And / or, The optimized first associated action is used to train at least one of the following models: an action model, a first evaluation model, a second evaluation model, and an optimized model.

14. The method of claim 11, wherein, When the third evaluation result indicates that there are no problems with the execution result of the optimized second associated action, the method further includes: The optimized second associated action is used to update the standard process information database; And / or, The optimized second associated action is used to train at least one of the following models: action model, first evaluation model, second evaluation model, backtracking model, and optimization model.

15. The method according to any one of claims 1 to 14, characterized in that, The standard process information database is constructed in the following manner: Obtain subtask information of multiple second subtasks and subtask jump logic information between the multiple second subtasks. The subtask jump logic information between the multiple second subtasks is used to indicate the subtask jump method between the multiple second subtasks. The multiple second subtasks are used to execute the jump of the second subtasks sequentially according to the jump relationship indicated by the sixth standard process. The second subtask is jumped according to the reverse jump method between the multiple second subtasks, and the subtask information of the second subtask after the jump is added to the node of the second subtask after the jump, so as to obtain the description information of the sixth standard process. The reverse jump method of the subtask is opposite to the jump method of the subtask indicated by the subtask jump logic information. The standard process information database is constructed based on the description information of the sixth standard process, wherein the sixth standard process is any one of the standard processes in the standard process information database.

16. The method of claim 15, wherein, The step of performing the jump to the second subtask according to the reverse jump method between the multiple second subtasks, and adding the subtask information of the second subtask after the jump to the node of the second subtask after the jump, to obtain the description information of the sixth standard process, includes: Jump according to the reverse jump method between the multiple second subtasks, and add the subtask information of the second subtask after the jump to the node of the second subtask after the jump, to obtain the description information of multiple sub-standard processes of the sixth standard process. The description information of the sixth standard process is generated based on the description information of multiple sub-standard processes of the sixth standard process.

17. The method according to claim 15 or 16, characterized in that, The step of constructing the standard process information database based on the description information of the six standard processes includes: Generate a first semantic representation corresponding to the sixth standard process based on the description information of the sixth standard process; The standard process information database is constructed based on the first semantic representation corresponding to the sixth standard process.

18. The method according to any one of claims 3 to 17, characterized in that, The sub-standard process information database is constructed in the following manner: Obtain subtask information of multiple first subtasks and subtask jump logic information between the multiple first subtasks. The subtask jump logic information between the multiple first subtasks is used to indicate the subtask jump method between the multiple first subtasks. The first subtask is jumped according to the reverse jump method between the multiple first subtasks, and the subtask information of the first subtask after the jump is added to the position node of the first subtask after the jump, so as to obtain the description information of multiple sub-standard processes of the first standard process. The sub-standard process information database is constructed based on the description information of multiple sub-standard processes of the first standard process.

19. The method of claim 18, wherein, The step of constructing the sub-standard process information database based on the description information of multiple sub-standard processes of the first standard process includes: Based on the description information of the multiple sub-standard processes of the first standard process, generate second semantic representations corresponding to the multiple sub-standard processes of the first standard process respectively. The sub-standard process information database is constructed based on the second semantic representations corresponding to the multiple sub-standard processes of the first standard process.

20. A task execution apparatus characterized by comprising: include: The receiving module is used to receive task requests; The standard process acquisition module is used to acquire a first standard process from the standard process information database according to the task request; The sub-standard process acquisition module is used to acquire a first sub-standard process from the first standard process according to the task request. The first sub-standard process is a sub-standard process of the first standard process. The first sub-standard process is used to instruct multiple first subtasks to be executed sequentially based on a jump relationship. The first sub-standard process includes a first root node, which is used to instruct the execution of the starting subtask among the multiple first subtasks. The task execution module is used to execute the plurality of first sub-tasks according to the first sub-standard process.

21. An electronic device comprising a processor and a memory coupled to the processor, the processor being configured to perform the method of any one of claims 1 to 19.

22. A computer-readable storage medium comprising instructions which, when executed on a computer, cause the computer to perform the method of any one of claims 1 to 19.

23. A computer program product comprising instructions that, when run on a computer, cause the computer to perform the method as described in any one of claims 1 to 19.

24. A chip, characterized by An electronic device comprising one or more interface circuits and one or more processors; the interface circuit is configured to receive a signal from a memory of the electronic device, and send the signal to the processor, the signal comprising computer instructions stored in the memory; when the processor executes the computer instructions, the electronic device is caused to perform the method of any one of claims 1-19.

Images