A digital employee operating method, system, device and medium

Abstract

The application discloses a digital employee operation method, system, device and medium. The method comprises the following steps: in response to a request of a to-be-processed task, a structured representation is constructed according to the to-be-processed task; an initial performance state of a digital employee is obtained; the digital employee comprises at least two task execution models; the task execution models in the digital employee have different task execution capabilities; the structured representation of the to-be-processed task is matched with the initial performance state of the digital employee to obtain a target operation scheme of the digital employee; and the digital employee is operated to process the structured representation based on the target operation scheme, so as to obtain an operation result of the digital employee. The structured representation can be constructed to accurately extract task entities, types and requirements, and the model execution capability difference and real-time load resource are combined to allocate tasks, so that the digital employee can be accurately and efficiently operated, and dynamic changes in the operation process can be coped with, and the processing requirements of enterprise-level tasks can be fully met.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, system, device and medium for operating a digital employee. Background Technology

[0002] With the deep integration of artificial intelligence and enterprise digital transformation, digital employees have become a core tool for enterprises to reduce costs and increase efficiency, and are widely used in areas such as intelligent customer service, production line process control in manufacturing, and user data processing and risk prediction in the financial industry. Currently, most digital employee systems are developed based on a single model, which can complete basic automated tasks. However, as enterprises increase their requirements for task processing efficiency, resource cost control, and scenario adaptability, the need for model-based collaborative digital employees is becoming increasingly urgent.

[0003] Existing digital employee technology relies on a single model architecture and lacks a dynamic collaboration mechanism for large, medium and small models. It cannot flexibly allocate model tasks according to task complexity and equipment resources, resulting in wasted computing resources and high response latency. Moreover, the performance of digital employees cannot be autonomously optimized as task data accumulates, making it difficult to adapt to changes in enterprise business in the long term and failing to meet the operational needs of digital employees in multiple scenarios for high efficiency, low cost and high adaptability. Summary of the Invention

[0004] The following is an overview of the subject matter described in detail herein. This overview is not intended to limit the scope of the claims.

[0005] The main objective of this disclosure is to propose a method, system, device, and storage medium for operating digital employees. By constructing a structured representation, it can accurately extract task entities, types, and requirements. Combined with differences in model execution capabilities and real-time load resource allocation, it can achieve accurate and efficient operation of digital employees and cope with dynamic changes during operation, fully meeting the processing needs of enterprise-level tasks.

[0006] A first aspect of this application provides a method for operating a digital employee, used in a central controller, the method comprising:

[0007] In response to a request for a pending task, a structured representation is constructed based on the pending task; the structured representation includes a target task, a target task type, and target task requirements; the target task is generated based on the pending task.

[0008] Obtain the initial performance state of the digital employee; the digital employee includes at least two task execution models; the task execution models in the digital employee have different capabilities in performing tasks;

[0009] The structured representation of the task to be processed is matched with the initial performance state of the digital employee to obtain the target operation plan of the digital employee; the target operation plan includes the target operation subject of the target task and the target operation sequence of the target operation subject;

[0010] Based on the target operation plan, the digital employee processes the structured representation to obtain the operation results of the digital employee.

[0011] In some embodiments of this application, constructing a structured representation based on the task to be processed includes:

[0012] Extract task entities and task operations from the tasks to be processed;

[0013] Based on the task entity and the task operation, the tasks to be processed are classified to obtain the task type of the tasks to be processed.

[0014] Based on the task type of the task to be processed, the task to be processed is divided into at least one target task to obtain the target task type of the target task;

[0015] Based on the target task and the target task type, determine the target task requirements;

[0016] The target task, target task attributes, and target task requirements are integrated to generate the structured representation.

[0017] In some embodiments of this application, dividing the task to be processed into at least one target task includes: dividing the task to be processed into a single target task and multiple target tasks.

[0018] In some embodiments of this application, the initial performance state includes the execution capability of each task execution model, and the step of matching the structured representation of the task to be processed with the initial performance state of the digital employee to obtain the target operation plan of the digital employee includes:

[0019] When the task to be processed is divided into a single target task, the task similarity is calculated with the execution capability of each task execution model according to the requirements of the target task; the target operation plan of the digital employee is determined according to the task similarity and the preset first similarity threshold.

[0020] When the task to be processed is divided into multiple target tasks, each target task is parsed to obtain the target task requirements of each target task and the correlation between each target task; the correlation represents the execution priority between each target task; according to the target task requirements of each target task, the corresponding task similarity is calculated with the execution capability of each task execution model; based on the correlation, the target operation plan of the digital employee is determined according to the task similarity and a preset second similarity threshold.

[0021] In some embodiments of this application, the initial performance state further includes the load resources of each task execution model, and determining the target operation plan of the digital employee based on the similarity of each task and a preset first similarity threshold includes:

[0022] The task similarity of each task execution model is compared with the preset first similarity threshold to obtain the corresponding comparison result;

[0023] If the comparison result shows that the task similarity of the task execution model is greater than or equal to the preset first similarity threshold, the task execution model is determined as the initial operating entity of the digital employee; the target operating entity of the digital employee is determined based on the load resources of the initial operating entity and the target task; and the target operating plan of the digital employee is determined based on the target operating entity and the target task.

[0024] If all the comparison results show that the task similarity of the task execution model is less than the preset first similarity threshold, the target operating entity of the digital employee is determined according to the load resources of each task execution model and the target task; and the target operating scheme of the digital employee is determined according to the target operating entity and the target task.

[0025] In some embodiments of this application, the structured representation further includes constraints on the target task, and the determination of the target operation plan for the digital employee based on the association relationship, according to the similarity of each task and a preset second similarity threshold, includes:

[0026] The task similarity of each task execution model is compared with the preset second similarity threshold to obtain the corresponding comparison result;

[0027] If the comparison result shows that the task similarity of the task execution model is greater than or equal to the preset second similarity threshold, the task execution model is determined as the initial running entity of the corresponding target task; if the comparison result shows that the task similarity of all the task execution models is less than the preset second similarity threshold, the initial running entity of the corresponding target task is determined according to the load resources of each task execution model and the target task corresponding to the task similarity.

[0028] Based on the aforementioned relationships, the initial operating sequence of each of the initial operating entities is determined;

[0029] Based on the running order, and according to the load resources of each initial running entity and the constraints of each target task, the target running entity of each target task and the target running order of each target running entity are determined.

[0030] Based on the target operating entities of each target task and the target operating sequence of the target operating entities, the target operating plan of the digital employee is generated.

[0031] In some embodiments of this application, the structured representation includes the constraints of the target task, and the step of running the digital employee to process the structured representation based on the target operation plan to obtain the operation result of the digital employee includes:

[0032] Based on the target running entity of the target task and the target running sequence of the target running entity, the digital employee processes the structured representation to obtain the real-time load resources of the target running entity corresponding to the target task in real time;

[0033] When the real-time load resources are less than a preset load threshold, the target running entity of the target task is adjusted according to the constraints of each target running entity to obtain the adjusted target running entity of the target task; and the target running order is updated according to the adjusted target running entity to obtain the updated target running order; and the digital employee is run according to the adjusted target running entity and the updated target running order to obtain the running result of the digital employee.

[0034] When the real-time load resource is greater than or equal to the preset load threshold, the digital employee is run according to the target running entity and the target running sequence of the target running entity, and the running result of the digital employee is obtained.

[0035] To achieve the above objectives, a second aspect of the present invention provides a digital employee operating system, the system comprising:

[0036] A response module is used to respond to a request from a task to be processed and to construct a structured representation based on the task to be processed; the structured representation includes a target task, a target task type, and target task requirements; the target task is generated based on the task to be processed.

[0037] An acquisition module is used to acquire the initial performance status of the digital employee; the digital employee includes at least two task execution models; the task execution models in the digital employee have different capabilities in performing tasks;

[0038] A matching module is used to match the structured representation of the task to be processed with the initial performance state of the digital employee to obtain the target operation plan of the digital employee; the target operation plan includes the target operation subject of the target task and the target operation sequence of the target operation subject;

[0039] The running module is used to run the digital employee to process the structured representation based on the target running scheme, and obtain the running results of the digital employee.

[0040] To achieve the above objectives, a third aspect of the present invention provides an electronic device, comprising: at least one control processor and a memory for communicatively connecting to the at least one control processor; the memory stores instructions executable by the at least one control processor, the instructions being executed by the at least one control processor to enable the at least one control processor to perform the above-described method for operating a digital employee.

[0041] To achieve the above objectives, a fourth aspect of the present invention provides a computer-readable storage medium storing computer-executable instructions for causing a computer to perform the above-described method for operating a digital employee.

[0042] This application provides a method for operating a digital employee. In response to a request from a task to be processed, a structured representation is constructed based on the task. The initial performance state of the digital employee is obtained. The digital employee includes at least two task execution models. The task execution models within the digital employee have different capabilities in executing tasks. The structured representation of the task to be processed is matched with the initial performance state of the digital employee to obtain a target operating plan for the digital employee. Based on the target operating plan, the digital employee processes the structured representation to obtain the operating results. By constructing a structured representation, the method can accurately extract task entities, types, and requirements. Combined with differences in model execution capabilities and real-time load resource allocation, the method achieves accurate and efficient operation of the digital employee and can cope with dynamic changes during operation, fully meeting the processing needs of enterprise-level tasks.

[0043] It is understood that the beneficial effects of the second to fourth aspects compared with the related technologies are the same as the beneficial effects of the first aspect compared with the related technologies. Please refer to the relevant description in the first aspect above, which will not be repeated here. Attached Figure Description

[0044] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0045] Figure 1 This is a flowchart illustrating a method for operating a digital employee according to an embodiment of this application;

[0046] Figure 2 This is a schematic diagram of the operating system of a digital employee provided in an embodiment of this application;

[0047] Figure 3 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0048] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0049] In the description of this application, the use of terms such as "first," "second," etc., is for the purpose of distinguishing technical features only and should not be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated or the order of the technical features indicated.

[0050] In the description of this application, it should be understood that the orientation descriptions, such as up, down, etc., are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application.

[0051] In the description of this application, it should be noted that, unless otherwise explicitly defined, terms such as "setup," "installation," and "connection" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this application in conjunction with the specific content of the technical solution.

[0052] With the deep integration of artificial intelligence and enterprise digital transformation, digital employees have become a core tool for enterprises to reduce costs and increase efficiency, and are widely used in areas such as intelligent customer service, production line process control in manufacturing, and user data processing and risk prediction in the financial industry. Currently, most digital employee systems are developed based on a single model, which can complete basic automated tasks. However, as enterprises increase their requirements for task processing efficiency, resource cost control, and scenario adaptability, the need for model-based collaborative digital employees is becoming increasingly urgent.

[0053] Existing digital employee technology relies on a single model architecture and lacks a dynamic collaboration mechanism for large, medium and small models. It cannot flexibly allocate model tasks according to task complexity and equipment resources, resulting in wasted computing resources and high response latency. Moreover, the performance of digital employees cannot be autonomously optimized as task data accumulates, making it difficult to adapt to changes in enterprise business in the long term and failing to meet the operational needs of digital employees in multiple scenarios for high efficiency, low cost and high adaptability.

[0054] Based on this, embodiments of this application provide a method, system, electronic device, and medium for operating digital employees. The aim is to accurately extract task entities, types, and requirements by constructing structured representations, and to combine model execution capability differences and real-time load resource allocation tasks to achieve accurate and efficient operation of digital employees. Furthermore, it can cope with dynamic changes during operation and fully meet the processing needs of enterprise-level tasks.

[0055] The digital employee operation method, system, electronic device and medium provided in the embodiments of this application are specifically described through the following embodiments. First, the operation method of the digital employee in the embodiments of this application is described.

[0056] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0057] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0058] The digital employee operation method provided in this application relates to the field of artificial intelligence technology. This method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing the digital employee operation method, but is not limited to the above forms.

[0059] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0060] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirection to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data required for the proper functioning of these embodiments acquired.

[0061] Therefore, referring to Figure 1This application provides a method for operating a digital employee. This method is applied to a central controller, which can be a server, an electronic device, or a mobile terminal, etc. There are no specific limitations here. The method includes the following steps S110 to S140.

[0062] Step S110: In response to the request of the task to be processed, construct a structured representation based on the task to be processed; the structured representation includes the target task, the target task type, and the target task requirements; the target task is generated based on the task to be processed.

[0063] Step S120: Obtain the initial performance status of the digital employee; the digital employee includes at least two task execution models; the task execution capabilities of each task execution model in the digital employee are different;

[0064] Step S130: Match the structured representation of the task to be processed with the initial performance state of the digital employee to obtain the target operation plan of the digital employee; the target operation plan includes the target operation subject of the target task and the target operation sequence of the target operation subject;

[0065] Step S140: Based on the target operation plan, run the structured representation of digital employees to obtain the operation results of digital employees.

[0066] In this step, the process first responds to the request of the task to be processed, constructing a structured representation that includes the target task, target task type, and target task requirements. Preferably, natural language processing techniques are used to extract task entities and operations, or task elements are matched based on predefined templates, transforming the unstructured task request into a standardized data format containing the target task, target task type, and target task requirements. This provides a unified processing foundation for subsequent model matching and task allocation. The target task is generated based on the task to be processed, thereby breaking down the complex task into structured units with clear attributes, providing a quantifiable basis for subsequent model matching.

[0067] Furthermore, the initial performance status of the digital employee is obtained. The digital employee comprises at least two task execution models, each with different task execution capabilities. The initial performance status of the digital employee refers to the available resource status and execution capability indicators of each task execution model. This can be achieved by real-time monitoring of model computing power utilization, memory usage, and historical task success rate data collection, providing a basis for dynamically selecting the optimal execution entity.

[0068] Furthermore, the structured representation of the task to be processed is matched with the initial performance state of the digital employee to obtain the target operation plan of the digital employee. A similarity algorithm is preferably used to calculate the matching degree between task requirements and model capabilities, combined with a load balancing strategy to determine the execution order, thus solving the problem that a single model cannot adapt to complex tasks. The target operation plan includes the target operation entity of the target task and the target operation order of the target operation entity.

[0069] In one embodiment, digital employees include local small models deployed locally and large cloud models deployed in the cloud. Before running digital employees, a "system state vector" needs to be constructed to standardize and quantify the real-time status of local devices and cloud resources from 0 to 1, providing an objective basis for allocation decisions. The specific dimensions and calculation logic are as follows:

[0070] For the local small model status dimension, determine its model computing load rate, availability, and local data readiness rate. Among them, the computing load rate (0~1) is the weighted value of the current local GPU (e.g., 4090) memory usage and computing power utilization; the model availability (0 / 1) is whether the locally deployed target model (e.g., YOLOv8 visual model, Llama-2-7B text model) is in a normal operating state, with 1 for no faults and 0 for faults or updates; the local data readiness rate (0~1) is the percentage of local data (e.g., historical quality inspection samples, business templates) required by the task that has been loaded. For example, if 100 chip sample images are required and 80 have been loaded, the readiness rate is 0.8.

[0071] For the cloud-based large model state dimension, determine its network latency, cloud quota remaining rate, and system state vector output. Network latency (mapped to 0~1) is the average latency from the local machine to the cloud server obtained through real-time ping testing, calculated as "latency mapping value = 1 - (actual latency / maximum acceptable latency)". Cloud quota remaining rate (0~1) is the remaining percentage of the cloud-based large model call quota under the current account; for example, if the total quota is 1000 times and 300 times have been used, the remaining rate is 0.7. Cloud queue length (mapped to 0~1) is the number of tasks currently waiting to be processed in the cloud, calculated as "queue mapping value = 1 - (actual queue number / maximum queue number)".

[0072] Furthermore, by integrating the above dimensions, a system state vector is generated, with the following format example: [Local load rate: 0.56, Local model availability: 1, Local data readiness rate: 0.8, Cloud latency mapping value: 0.6, Cloud quota remaining rate: 0.7, Cloud queue mapping value: 0.6].

[0073] Furthermore, based on the target operation plan, the digital employee processes the structured representation to obtain the operational results. These results refer to the structured data or operational instructions output by the digital employee after processing tasks according to the target plan. This can be achieved through model inference result aggregation or workflow engine scheduling log recording, ensuring the task execution process is traceable and meets business requirements. Therefore, this step, by constructing a dynamic matching mechanism between the task's structured representation and the performance status of multiple models, achieves collaborative scheduling based on task complexity and resource status, overcoming the limitations of a single model architecture and thus improving the task processing efficiency and resource utilization of the digital employee in complex scenarios.

[0074] In some embodiments, the construction of a structured representation based on the task to be processed in step S110 includes the following steps S210 to S250:

[0075] Step S210: Extract task entities and task operations from the tasks to be processed;

[0076] Step S220: Based on the task entity and task operation, classify the tasks to be processed to obtain the task type of the tasks to be processed;

[0077] Step S230: Based on the task type of the task to be processed, divide the task to be processed into at least one target task to obtain the target task type of the target task;

[0078] Step S240: Determine the target task requirements based on the target task and target task type;

[0079] Step S250: Integrate the target task, target task attributes, and target task requirements to generate a structured representation.

[0080] In this embodiment, task entities and task operations are extracted from the tasks to be processed. A task entity refers to a specific object or data element involved in the task, and a task operation refers to the behavior or processing method applied to the entity. Then, based on the task entities and task operations, the tasks to be processed are classified using a preset rule base or machine learning model to obtain the task type of the task to be processed.

[0081] Furthermore, based on the task type of the task to be processed, the task to be processed is divided into at least one target task, thus obtaining the target task type of the target task. Preferably, the task division adopts a dependency-based segmentation algorithm. Specifically, dividing the task to be processed into at least one target task includes: dividing the task to be processed into a single target task and multiple target tasks.

[0082] Preferably, based on the complexity and content of the tasks to be processed, the tasks are divided into single-target tasks and multiple-target tasks. The division method includes two scenarios: if the task involves only a single operation and the requirements are clearly defined, it is classified as a single-target task; if the task involves multi-stage operations or requires multi-model collaboration, it is broken down into multiple target tasks with independent requirements. The task division is based on the logical relationship between task entities, the independence of operation steps, and the divisibility of resource consumption.

[0083] Therefore, this embodiment avoids unnecessary splitting overhead by treating simple tasks directly as a single target task; for complex tasks, it improves the efficiency and quality of task processing by dividing them into multiple target tasks and utilizing the expertise of different task execution models in the digital employee system. This application achieves flexible division of tasks to be processed, which can adapt to different types and complexities of tasks to be processed, thereby improving the applicability and processing capabilities of the digital employee system.

[0084] Furthermore, based on the target task and its type, the target task requirements are determined, thereby integrating the target task, its attributes, and its requirements to generate a structured representation. Thus, this embodiment, by breaking down a complex task into multiple target tasks and clearly defining the type and requirements of each, enables digital employees to more accurately match suitable execution models, thereby optimizing resource allocation and improving overall task processing efficiency.

[0085] In one embodiment, the digital employee can be a local small model (or local computing processor) deployed on local computing power and a cloud large model deployed on cloud computing power. Before running the digital employee, the original task is transformed into a structured representation that can be understood by subsequent task allocation, providing a unified input for determining the operation of the digital employee. Moreover, the input of the original task can be multimodal. Specifically, for the multiple forms of input that the original task may contain, such as text, flow, and voice, unified parsing is achieved through "format adaptation + semantic alignment". The specific steps are as follows:

[0086] First, input format conversion is performed. The conversion methods for different modalities are as follows: Text input (such as natural language commands "generate monthly sales analysis report", business requirement documents): sentence preprocessing and stop word filtering are used to retain core semantic fragments; Workflow / process input (such as ERP system approval process, production quality inspection step diagram): nodes (such as "order review → inventory verification → shipment confirmation"), node dependencies (such as "inventory verification must be performed after order review") and execution constraints (such as "review time ≤ 2 hours") are extracted through flowchart parsing algorithms; Voice / multimodal input (such as meeting recording "optimize customer complaint handling process", labeled product defect diagram): speech is converted to text and then combined with NLP parsing. For image annotation, visual features (such as defect area coordinates, label "crack") are extracted and associated with text descriptions. Note that, in order to avoid ambiguity in the process of transforming domain knowledge, it is necessary to call domain knowledge graphs (such as the "product defect type library" in manufacturing and the "risk control indicator system" in finance) to complete the semantics of fuzzy inputs: for example, if the input is "process loan anomalies", the financial knowledge graph can be used to supplement implicit information such as "anomaly type (overdue / false information)" and "processing basis" to avoid semantic ambiguity.

[0087] Second, extract the core elements of the task. Specifically, by combining NLP with a rule engine, the task entities, constraints, and target metrics are accurately extracted to form a computable set of elements. The steps are as follows:

[0088] First, entity and attribute extraction is performed. Named Entity Recognition (NER) + attribute association model is used to extract the core task objects (e.g., "chip", "sales data"), operational actions (e.g., "detection", "analysis", "generation"), and association dimensions (e.g., "chip-surface-scratches", "sales data-2024 Q3-East China region"). For example, inputting "Analyze 2024 Q3 East China region home appliance sales revenue, error ≤ 5%", extracts the entity "home appliance sales revenue", attributes "time = 2024Q3" and "region = East China", and the constraint "error ≤ 5%".

[0089] Secondly, it is necessary to quantify the constraints and objectives in the core elements. This includes two categories: (1) explicit constraints: directly extract the numerical values ​​and logical conditions from the text (such as "accuracy ≥ 0.01 mm", "completion time ≤ 24 hours"). When there are no explicit numerical values, refer to the domain standards (such as "customer service response time" defaults to ≤ 30 seconds); (2) implicit objectives: deduce them through the task context, such as "optimize the supply chain" to deduce the objectives "inventory turnover rate increase ≥ 10%" and "logistics cost reduction ≤ 5%", and combine them with historical task data to quantify the thresholds.

[0090] Third, task type and complexity determination. Specifically, this involves determining whether to utilize local or cloud computing power. Based on an intent classification and complexity assessment model, tasks are assigned a "type label + complexity score" to support subsequent capability matching, as detailed below:

[0091] First, a multi-classification model (such as BERT fine-tuning) is used to accurately classify task types based on the domain scenario definition. In this embodiment, they are specifically divided into the following four categories: Execution-type tasks are standardized and repetitive tasks (such as "entering 100 customer information entries" or "generating a standardized quality inspection report"), characterized by clear rules and fixed steps; Decision-type tasks require judgment and evaluation (such as "determining whether a chip is qualified" or "approving a loan application"), characterized by reliance on thresholds and the need for risk / reward trade-offs; Analytical tasks require data processing and pattern mining (such as "analyzing user consumption behavior" or "predicting next month's sales"), characterized by large amounts of input data and the need for algorithmic model support; Creative tasks require innovative output (such as "designing promotional copy for new products" or "developing supply chain optimization plans"), characterized by the lack of fixed processes and reliance on experience / inspiration.

[0092] Secondly, a multi-dimensional evaluation system (0-1 standardized score, with 1 being the highest complexity) is constructed to quantitatively evaluate the complexity. The multi-dimensional evaluation system can be found in Table 1.

[0093] Table 1

[0094]

[0095] The final complexity score for each task type is the weighted sum of each dimension (the weights are set by domain experts, such as the higher weight for "data dependency" in analytical tasks).

[0096] Fourth, perform task layering and decomposition, and generate structured descriptions. Specifically, refer to the task decomposition logic of deep reinforcement learning, break down complex tasks into independently executable atomic subtasks, and clarify dependencies.

[0097] First, determine the hierarchical decomposition rules. There are two types of decomposition methods. The first type is based on "objective-step-action": for example, "chip quality inspection" (overall objective) → "appearance inspection," "performance inspection" (first-level sub-task) → "scratch inspection," "deformation inspection" (second-level sub-task, atomic level). The second type is based on "data flow": for example, "generate sales analysis report" → "data acquisition (ERP+CRM)" → "data cleaning (missing value handling)" → "indicator calculation (sales revenue / gross profit margin)" → "visualization generation." Each sub-task must clearly define its input (e.g., "data acquisition" input "time range 2024Q3") and output (e.g., "cleaned data table").

[0098] Secondly, determine the dependencies and the ID of each task. Assign a unique ID to each subtask (e.g., "Chip Quality Inspection - Appearance Inspection - Scratch Inspection" is denoted as T1-1-1), and record the dependencies through a directed graph: T1-1-1 (scratch inspection) must be executed before T1-1-4 (multi-defect fusion judgment), and T1-2 (performance inspection) and T1-1 (appearance inspection) can be executed in parallel to avoid resource waste.

[0099] Next, a structured representation output is generated. By integrating all the information from the above steps, a structured task description in JSON format required for this embodiment is generated, ensuring that the subsequent task allocation module can directly parse it and achieve a precise association between "task-capability-operating entity".

[0100] In some embodiments, in step S130, the structured representation of the task to be processed is matched with the initial performance state of the digital employee to obtain the target operation plan of the digital employee, including the following steps S310 to S320:

[0101] Step S310: When the tasks to be processed are divided into single target tasks, calculate the corresponding task similarity with the execution capability of each task execution model according to the target task requirements; determine the target operation plan of digital employees based on the task similarity and the preset first similarity threshold.

[0102] Step S320: When the task to be processed is divided into multiple target tasks, each target task is analyzed to obtain the target task requirements of each target task and the correlation between each target task; the correlation represents the execution priority between each target task; according to the target task requirements of each target task, the corresponding task similarity is calculated with the execution capability of each task execution model; based on the correlation, according to the task similarity and the preset second similarity threshold, the target operation plan of the digital employee is determined.

[0103] In this embodiment, the initial performance state of the digital employee includes the execution capabilities of each task execution model. When the tasks to be processed are divided into single target tasks, the task similarity is calculated based on the target task requirements and the execution capabilities of each task execution model. Then, based on the task similarities and a preset first similarity threshold, the target operation plan of the digital employee is determined. When the tasks to be processed are divided into multiple target tasks, each target task is first analyzed to obtain the target task requirements and the relationships between the target tasks. Then, based on the target task requirements of each target task, the task similarity is calculated based on the execution capabilities of each task execution model. Finally, based on the relationships, the target operation plan of the digital employee is determined based on the task similarities and a preset second similarity threshold. Specifically, the relationships represent the execution priorities between the target tasks; the first and second similarity thresholds can be dynamically adjusted based on historical task execution data.

[0104] For example, when a task to be processed is broken down into multiple target tasks, the requirements and relationships of each target task are first analyzed. For instance, task A needs to be executed after task B is completed. Then, the similarity between the requirements of each target task and the execution capabilities of each model is calculated. If the similarity is higher than a second similarity threshold, the corresponding model is selected as the initial running entity; if the similarity is lower than the second similarity threshold, the initial running entity is selected based on the model's load resources. Then, based on the execution priority determined by the relationships, the order of the initial running entities is adjusted. For example, if the initial running entity for task A is model X, and the initial running entity for task B is model Y, and task A needs to be executed first, then model X's execution order is placed before model Y. Finally, combining the real-time load resources of each model and task constraints, such as task processing time limits or resource usage limits, the target running entities and their order are dynamically adjusted to generate the final execution plan.

[0105] Therefore, this embodiment ensures that the execution order of multiple subtasks conforms to logical dependencies by considering the relationships between tasks. At the same time, by introducing relationship constraints and multi-threshold matching mechanisms, it ensures that multiple target tasks are coordinated and optimized in terms of resource allocation and execution order, thereby improving processing efficiency and reducing the risk of resource conflicts. This enhances the ability and efficiency of digital employees in handling complex tasks, while also optimizing the utilization of computing resources.

[0106] In one embodiment, the digital employee includes a local small model deployed locally and a cloud-based large model deployed in the cloud. Before running the digital employee, after constructing a "system state vector," a single-task allocation decision is made. For simple tasks that do not require decomposition, a three-dimensional matching model of "capability requirement - model capability - system state" is used to output the allocation result. The specific process includes: calculating the cosine similarity between the task capability requirement vector in the structured representation and the baseline capability vectors of the local small model and the cloud-based large model, respectively, to initially determine that the cloud model is more suitable for the capability requirements. Then, a "system state weight coefficient" is introduced to correct the initial matching result, realizing a weighted correction of the system state. Finally, the target operation plan of the digital employee is determined based on the corrected similarity.

[0107] For the allocation of subtasks in complex tasks (decomposed tasks), a hierarchical allocation strategy of "subtask capability matching - dependency constraints - resource load balancing" is adopted. The specific process is as follows:

[0108] Extract a separate "subtask capability requirement vector" for each subtask and match it with the model's capabilities to achieve differentiated matching of subtask capability requirements: Example subtask 1 "Scratch detection": Capability requirement vector [0.2, 0.95, 0.3, 0.4, 0.1, 0.7], similarity to the local visual model (YOLOv8) is 0.82, similarity to the cloud model is 0.65 → local model is preferred; Example subtask 2 "Multi-defect fusion judgment": It needs to integrate the detection results of three types: "scratch / deformation / color difference", capability requirement vector [0.6, 0.3, 0.9, 0.7, 0.2, 0.5], similarity to the large cloud model is 0.88, similarity to the local model is 0.52 → cloud model is preferred.

[0109] By constraining dependencies and execution order, the allocation is ensured to conform to the logical sequence of tasks based on the subtask dependency graph. For example, if subtask A (e.g., "scratch detection") is a prerequisite for subtask B (e.g., "multi-defect fusion judgment"), then: if A is allocated locally and B is allocated in the cloud: after A is executed, the results of A (e.g., scratch size data) must be encrypted and transmitted to the cloud before starting B; if both A and B are allocated locally / in the cloud: they are executed directly in sequence without cross-platform transmission, reducing latency.

[0110] In the example, in the process of “scratch detection (local) → deformation detection (local) → color difference detection (local) → multi-defect fusion judgment (cloud),” the results need to be packaged and transmitted to the cloud to execute the fourth sub-task after the first three sub-tasks are completed.

[0111] If, after initial allocation, a "single-end load too high" occurs (e.g., when three sub-tasks are allocated locally simultaneously, the load rate rises from 0.56 to 0.92, exceeding the threshold of 0.8), load balancing adjustments are triggered. Specifically, this includes: selecting "portable sub-tasks": sub-tasks with no hard threshold requirements for model capabilities (e.g., "color difference detection" requires a visual accuracy of 0.85, which is met both locally and in the cloud), and with a small amount of cross-end data transmission; migration strategy: migrating "color difference detection" from local to the cloud. After migration, the local load rate drops to 0.68, and the cloud queue mapping value drops from 0.6 to 0.55 (still within an acceptable range).

[0112] The final output is a subtask allocation table, formatted as shown in Table 2:

[0113] Table 2

[0114]

[0115] In some embodiments, step S310 determines the target operation plan for digital employees based on the similarity of each task and a preset first similarity threshold, including the following steps S410 to S430:

[0116] Step S410: Compare the task similarity of each task execution model with the preset first similarity threshold to obtain the corresponding comparison results;

[0117] Step S420: If the comparison result shows that the task similarity of the task execution model is greater than or equal to the preset first similarity threshold, the task execution model is determined as the initial operating entity of the digital employee; based on the load resources and target tasks of the initial operating entity, the target operating entity of the digital employee is determined; based on the target operating entity and target tasks, the target operating plan of the digital employee is determined.

[0118] Step S430: If all comparison results show that the task similarity of the task execution model is less than the preset first similarity threshold, determine the target operating entity of the digital employee based on the load resources and target tasks of each task execution model; determine the target operating plan of the digital employee based on the target operating entity and target tasks.

[0119] In this embodiment, the initial performance state includes the load resources of each task execution model. First, the task similarity of each task execution model is compared with a preset first similarity threshold to obtain the corresponding comparison results. The task similarity comparison process can filter out potentially suitable models through the preset threshold, and the load resource assessment further optimizes model selection.

[0120] Furthermore, when the comparison result shows that the task similarity of the task execution model is greater than or equal to a preset first similarity threshold, the initial operating entity of the digital employee is determined by combining task adaptability and the current resource usage status of the model. Then, based on the load resources and target task of the initial operating entity, the target operating entity of the digital employee is determined. Finally, based on the target operating entity and target task, the target operation plan of the digital employee is determined.

[0121] When all comparison results indicate that the task similarity of the task execution models is less than a preset first similarity threshold, the target operating entity of the digital employee is determined based on the load resources and target tasks of each task execution model. Then, based on the target operating entity and target tasks, the target operating plan of the digital employee is determined.

[0122] For example, when the task similarity meets the threshold, the model with high adaptability is preferentially selected as the initial candidate, but it is necessary to further verify whether its current load allows the execution of the new task. If the remaining resources of the initial running entity are sufficient to handle the target task, it is directly determined as the target running entity; if its load is too high, a model with a lower load is reselected based on the remaining resources. When none of the models have reached the similarity threshold, tasks are allocated based on load resources, for example, selecting the model with the highest idle rate to execute the task. This achieves a dynamic balance between task adaptability and resource utilization, avoiding response delays caused by overloaded high-load models, while ensuring reasonable task allocation in low-adaptability scenarios and improving overall operating efficiency.

[0123] Therefore, this embodiment dynamically selects the most suitable task execution model as the running entity based on task similarity and load resources, thereby improving the task processing efficiency and resource utilization of the digital employee and enabling it to more intelligently adapt to task requirements of different types and complexities. Simultaneously, by setting a similarity threshold, unsuitable models are prevented from being selected as the running entity, ensuring the quality of task execution. Furthermore, by considering load resource factors, overloading of a single model is prevented, achieving a balanced allocation of resources within the digital employee.

[0124] In some embodiments, in step S320, based on the association relationship, the target operation plan for the digital employee is determined according to the similarity of each task and a preset second similarity threshold, including the following steps S510 to S550:

[0125] Step S510: Compare the task similarity of each task execution model with the preset second similarity threshold to obtain the corresponding comparison results;

[0126] Step S520: If the comparison result shows that the task similarity of the task execution model is greater than or equal to the preset second similarity threshold, the task execution model is determined as the initial running entity of the corresponding target task; if the comparison result shows that the task similarity of all task execution models is less than the preset second similarity threshold, the initial running entity of the corresponding target task is determined according to the load resources of each task execution model and the target task corresponding to the task similarity.

[0127] Step S530: Determine the initial running sequence of each initial running entity based on the association relationship;

[0128] Step S540: Based on the running sequence, determine the target running entity and the target running sequence of each target task according to the load resources of each initial running entity and the constraints of each target task.

[0129] Step S550: Generate the target operation plan for digital employees based on the target operation subject and the target operation sequence of each target task.

[0130] In this embodiment, the structured representation includes the constraints of the target task. Based on the association relationship, the steps for determining the target operation plan of the digital employee according to the similarity of each task and a preset second similarity threshold include: firstly, comparing the task similarity of each task execution model with the preset second similarity threshold to obtain the corresponding comparison results.

[0131] Specifically, if the comparison result shows that the task similarity of the task execution model is greater than or equal to a preset second similarity threshold, then that task execution model is determined as the initial running entity for the corresponding target task. If the comparison result shows that the task similarity of all task execution models is less than the preset second similarity threshold, then the initial running entity for the corresponding target task is determined based on the load resources of each task execution model and the target task corresponding to the task similarity.

[0132] Furthermore, based on the relationships, the initial operating sequence of each initial operating entity is determined. Then, based on the operating sequence, and according to the load resources of each initial operating entity and the constraints of each target task, the target operating entity and the target operating sequence of each target operating entity are determined. Finally, based on the target operating entity and the target operating sequence of each target task, the target operating plan for digital employees is generated.

[0133] Preferably, the load resources of the task execution model may include CPU utilization, memory utilization, etc.; the constraints of the target task may include task priority, task completion time limit, etc.; the initial running order can be determined according to the dependencies between target tasks; the determination of the target running entity can comprehensively consider the load resources of the initial running entity and the constraints of the target task to select the most suitable task execution model; the determination of the target running order can be adjusted according to the actual situation of the target running entity based on the initial running order.

[0134] Therefore, this embodiment can flexibly select the initial running subject based on task similarity and preset threshold, and determine the final target running subject and running order in combination with load resources and constraints, thereby improving the utilization of computing resources, reducing response latency, enhancing the adaptability and running efficiency of digital employees in multiple scenarios, and realizing dynamic collaboration of multiple models of digital employees.

[0135] In some embodiments, in step S140, based on the target operating scheme, the structured representation of the digital employee is processed to obtain the operating results of the digital employee, including the following steps S610 to S630:

[0136] Step S610: Based on the target running entity of the target task and the target running sequence of the target running entity, run the digital employee processing structured representation to obtain the real-time load resources of the target running entity corresponding to the target task in real time;

[0137] Step S620: When the real-time load resources are less than the preset load threshold, adjust the target running entity of the target task according to the constraints of each target running entity to obtain the adjusted target running entity of the target task; and update the target running order according to the adjusted target running entity to obtain the updated target running order; run the digital employee according to the adjusted target running entity and the updated target running order to obtain the running result of the digital employee.

[0138] Step S630: When the real-time load resources are greater than or equal to the preset load threshold, run the digital employees according to the target running entity and the target running sequence of the target running entity, and obtain the running results of the digital employees.

[0139] In this embodiment, firstly, based on the target running entity of the target task and the target running sequence of the target running entity, a structured representation of digital employees is processed to obtain the real-time load resources of the target running entity corresponding to the target task. The real-time load resources are quantified through the computational resource utilization or memory usage of the task execution model.

[0140] Specifically, when the real-time load resources are less than a preset load threshold, the target running entities of the target tasks are adjusted according to the constraints of each target running entity, resulting in the adjusted target running entities of the target tasks. Then, based on the adjusted target running entities, the target execution order is updated, resulting in the updated target execution order. The updated target execution order must satisfy the execution priority in the association relationship, and ensure that the adjusted target running entities have the execution capabilities required to execute the target tasks.

[0141] Furthermore, based on the adjusted target operating entity and the updated target operating sequence, the digital employees are run to obtain their operating results. The preset load threshold is dynamically set according to the maximum resource capacity of the task execution model. The adjustment process for the target operating entity includes reassigning the target tasks to other task execution models with load resources below the preset threshold, and re-determining the operating sequence based on task priority or time limits in the constraints.

[0142] Furthermore, when the real-time load resources are greater than or equal to a preset load threshold, the digital employees are run according to the target operating entity and its target operating sequence to obtain the operating results of the digital employees. Wherein;

[0143] Therefore, this embodiment monitors the load of the target operating entity in real time, adjusts task allocation in a timely manner when the load is low, makes full use of the computing resources of each model, and makes adjustments while taking constraints into account, thus ensuring the quality of task processing. It realizes the dynamic adjustment and optimized allocation of digital employee resources, improves the overall operating efficiency of digital employees, avoids resource waste, and enables digital employees to respond more flexibly to different task requirements.

[0144] In one embodiment, digital employees include local small models deployed on-premises and cloud-based large models deployed in the cloud. Before running digital employees, a "feasibility verification" is performed on the target operating plan, with the following specific steps:

[0145] Step 1: Capability threshold verification to ensure that the assigned model meets the task's hard capability requirements (e.g., "scratch detection" visual accuracy requirement ≥ 0.95, local YOLOv8 accuracy 0.96, verification passed).

[0146] Step 2, Resource Constraint Verification: Ensure that no single-end resource exceeds the threshold after allocation (e.g., local load rate ≤ 0.8, cloud queue mapping value ≥ 0.5, verification passed).

[0147] If verification fails (e.g., a sudden failure of the local model reduces availability to 0), the running method will be redefined.

[0148] Furthermore, the triggering conditions are adjusted in real time. During task execution, the process is monitored continuously, and allocations are dynamically adjusted in case of special circumstances. These special circumstances include: sudden increases in local load, cloud network outages, and changes in task requirements. Specifically, when local load suddenly increases (e.g., due to a sudden increase in other high-priority tasks, with the load rate rising from 0.68 to 0.95), unexecuted local subtasks (such as "deformation detection") are paused and migrated to the cloud. When the cloud network is interrupted (e.g., the latency mapping value drops to 0), subtasks already allocated to the cloud (such as "multi-defect fusion judgment") are temporarily stored and executed after the network is restored, or a simplified local backup model (with slightly lower accuracy but usable in emergencies) is activated. When task requirements change (the user temporarily increases accuracy requirements), the originally allocated local "scratch detection" is upgraded to a large cloud model to ensure accuracy meets the standards.

[0149] Furthermore, the complete operation process is recorded, including input parameters (task capability vector, system state vector), calculation process (similarity score, correction coefficient), basis for determining the operation plan (threshold, priority), and adjustment records, forming a traceable "decision log" for subsequent strategy optimization (such as adjusting the similarity weight coefficient based on the log).

[0150] In one embodiment, the digital employee includes a local small model deployed locally and a cloud-based large model deployed in the cloud. The digital employee is run according to the target execution plan. The specific execution process includes two types: split execution (local small model / cloud-based large model), where local and cloud results do not need to be merged; and merging and processing results from multiple subtasks (complex task scenarios).

[0151] Among them, the split-end execution control (local small model / cloud large model) adopts differentiated execution control logic for different allocation objects, covering the entire process of "pre-execution preparation - in-execution monitoring - post-execution verification" to ensure that the task proceeds as expected. Taking the "scratch detection" task as an example, for the local small model, before execution, firstly, the "local data requirements" in the structured representation (such as chip sample image, quality inspection accuracy threshold ≥0.01mm) are called, and the target data stored locally is loaded by the "local data scheduling module"; secondly, the running status of the local small model (such as YOLOv8 vision model) is checked: memory usage (≤80%), model parameter integrity, and after confirming that there are no abnormalities, the model configuration is loaded (such as setting the detection confidence threshold to 0.7 to match the task accuracy requirements); finally, the "local execution instruction" is generated, with the following format example: {Instruction ID: E20240905001, Model: YOLOv8-chip-detect, Input data: chip_sample_001.png, Output format: JSON (including defect location / size / confidence), Timeout threshold: 5s}.

[0152] Furthermore, during execution, the "local resource monitoring module" collects execution data in real time: GPU computing power utilization (≥50% to ensure normal model inference), inference progress (such as stage feedback of "image preprocessing completed → feature extraction in progress → target detection in progress"), and intermediate results (such as the coordinates of the initially identified scratch area). If an anomaly occurs (such as computing power utilization suddenly dropping to 0, indicating that the model is stuck), the "local anomaly recovery mechanism" is triggered: the model is restarted and re-executed from the most recent normal stage, while an anomaly log is recorded.

[0153] Furthermore, after execution, the system first performs a "result format validation": checking whether the output JSON contains all required fields (such as "defect_type", "position_x", "position_y", "size_mm", "confidence"); then it performs a "precision threshold validation": determining whether "size_mm" (e.g., the calculated scratch length is 0.008mm) meets the task requirements (≥0.01mm is a defect that needs to be labeled, <0.01mm is acceptable), and also checking whether "confidence" (e.g., 0.85) is ≥ the model configuration threshold (0.7); if the validation passes, a "local execution result package" is generated; if the validation fails (e.g., the "size_mm" field is missing), a "retry mechanism" is triggered (reloading data and executing, retrying a maximum of 2 times), and if it still fails, a "local execution exception" is reported and cloud support is requested.

[0154] Taking the "multi-defect fusion judgment" task as an example, for the large cloud model, before execution, the "cross-end data transmission" is first processed: if the input data comes from local sub-tasks (such as the results of "scratch detection" and "deformation detection"), the "encrypted transmission module" uses the SSL / TLS protocol to package the data and upload it to the cloud, while recording the transmission progress (such as "8MB of 10MB of data has been uploaded, progress 80%)". Secondly, a "cloud execution request" is sent to the large cloud model (such as the GPT-4-turbo-industry model), carrying key parameters: task ID (associated with local sub-task ID), capability requirements (logical reasoning dimension ≥ 0.9), and output requirements (must include "comprehensive judgment result", "judgment basis", and "risk level"). Finally, after receiving the request, the cloud checks the large model resources: the current call queue length (must be ≤ 30) and the model load rate (must be ≤ 70%). After confirming availability, it returns an "execution ready response", which the local machine receives and waits for the result.

[0155] During execution, the execution status is obtained in real time through the "cloud status synchronization module": the "stage feedback" returned by the cloud (such as "subtask result parsing in progress → multi-defect logic association in progress → judgment rule matching in progress") and the estimated remaining time (such as "complete in 2 seconds"); if a network interruption occurs (transmission timeout > 10 seconds), the "breakpoint resume mechanism" is triggered: based on the uploaded partial data, a "breakpoint resume interface" is requested from the cloud to avoid data re-uploading; if the cloud returns "insufficient resources" (such as queue length exceeding the threshold), the task is added to the "cloud waiting queue" and the queue status is checked every 3 seconds until the queue is idle or the local timeout threshold is reached (such as 30 seconds, if timeout occurs, switch to the local backup model).

[0156] After the results are returned from the cloud, a "data integrity check" is performed first: checking for missing data (such as a missing "risk level" field) or format errors (such as a judgment result of "qualified" without any basis for judgment); secondly, a "logical consistency check" is performed: comparing the correlation between the cloud results and the local subtask results (such as whether the cloud's judgment of "chip unqualified" includes the locally detected "scratch length 0.012mm" and "deformation 0.005mm"); if the check passes, the results are decrypted and stored; if the check fails, a "result regeneration request" is sent to the cloud, carrying the reason for the check failure (such as "missing judgment basis"), requesting the cloud to re-execute and supplement the information.

[0157] The multi-subtask result merging process is designed for complex task scenarios, specifically complex tasks assigned to multiple execution endpoints (such as "chip full-process quality inspection" which includes 4 subtasks, 3 local executions and 1 cloud execution). This requires the integration of multi-source outputs through a "result merging module." Specifically, this includes: First, standardizing the output formats of different models. The "scratch detection results" (JSON format, including coordinates and dimensions) and "deformation detection results" (XML format, including deformation area and angle) output by local small models need to be converted to a unified JSON format using a "format conversion plugin." Second, standardizing data units: for example, if the local "deformation detection" output deformation amount is "5μm," it needs to be converted to the task-required "mm" (i.e., 0.005mm) to ensure data dimensional consistency. Finally, adding "result metadata": labeling each subtask result with "execution endpoint" (local / cloud), "confidence level" (e.g., scratch detection confidence level 0.85), and "execution time" (e.g., 1.2s) to facilitate weight allocation during subsequent merging.

[0158] Furthermore, the matching and execution of merging rules are specifically based on the "merging requirements" in the structured representation (e.g., "multi-defect fusion judgment requires comprehensive analysis of all sub-task results, taking defects with a confidence level ≥ 0.7 as the judgment criterion"), matching the corresponding merging rules. These merging rules include logical association rules, weight superposition rules, and conflict resolution rules.

[0159] Specifically, logical association rules are applicable to tasks involving "multi-defect fusion judgment." For example, if subtask 1 (scratch detection) determines "defect exists (confidence 0.85)," subtask 2 (deformation detection) determines "no defect (confidence 0.9)," and subtask 3 (color difference detection) determines "defect exists (confidence 0.75)," then during merging, the "scratch + color difference" defect information is extracted as input to the cloud-based large model's "comprehensive judgment." Weighting rules are applicable to tasks involving "multi-dimensional scoring" (such as "customer satisfaction analysis" broken down into "service attitude score," "response speed score," and "resolution rate score"). For example, local subtask 1 (service attitude...) The output scores are as follows: 8 points for local subtask 1 (response speed), 9 points for local subtask 2 (response speed), and 7 points for cloud subtask 3 (resolution rate), with a combined score of 8 × 0.3 + 9 × 0.2 + 7 × 0.5 = 7.7 points. The conflict resolution rule states that if there is a conflict between the results of multiple subtasks (e.g., subtask 1 determines "chip qualified" while subtask 3 determines "chip unqualified"), the "high confidence result" will be used first (e.g., subtask 1 has a confidence of 0.6 and subtask 3 has a confidence of 0.85, so subtask 3 will be the standard). If the confidence levels are close (difference < 0.1), the "supplementary verification subtask" will be called (e.g., re-detecting key areas).

[0160] Furthermore, after the merge is completed, the first step is to perform an "integrity check": check whether all subtask results are covered (e.g., "multi-defect fusion" must include three types of results: scratches, deformation, and color difference, none of which can be missing); the second step is to perform a "reasonableness check": determine whether the merged result conforms to common sense in the domain (e.g., if the merged result is determined to be "chip qualified", but the scratch size of 0.015mm is greater than the threshold of 0.01mm, it is determined to be unreasonable); if the check passes, a "merge result report" is generated; if the check fails, the source of the problem is located (e.g., the result of subtask 3 was omitted) and the merge is re-executed, while the merge rules are optimized (e.g., adding a "mandatory subtask result check" step).

[0161] Furthermore, based on the task type and allocation results, two output modes are adopted to ensure that the results meet user expectations and are easy to use subsequently. These output models include direct output and merged output.

[0162] Specifically, the direct output is for single tasks / subtasks that do not need to be merged. The triggering conditions include: first, a single task (such as "extracting order sales revenue" which only needs to be executed by a local small model); second, multiple subtasks, where the results of some subtasks need to be output separately (such as in "chip quality inspection", where the user requires "scratch detection results to be displayed in real time without waiting for other subtasks").

[0163] Furthermore, the output format and content are specified. Specifically, it is output in a dual format of "structured + visual". The structured format (JSON / CSV) facilitates subsequent system calls, while the visual format (such as image annotations) allows users to view it intuitively.

[0164] Furthermore, send "results ready notifications" (such as local pop-ups or SMS reminders) to users, including result previews (such as thumbnails of visualized images) and viewing paths; if users have questions about the results (such as believing that the marked scratches are stains), support "results feedback entry", where users can mark the questioned areas and submit them, and the feedback will be associated with the task ID for subsequent model optimization (such as updating the YOLOv8 training samples).

[0165] Furthermore, the report is output in the form of a "comprehensive report," integrating three parts: "summary of execution at each end," "core information of the merged results," and "supporting evidence," as shown in Table 3 (Comprehensive Chip Quality Inspection Report).

[0166] Table 3

[0167]

[0168] Simultaneously, a machine-readable version (JSON format) is output, containing the original results of all subtasks and the merging calculation process, facilitating integration with the enterprise ERP system. The final merged result is stored in a dual backup system: local storage for quick viewing and encrypted cloud storage for cross-departmental sharing (e.g., from the quality control department to the production department), with access control for sharing. For example, the production department can only view the "Overall Conclusion" and "Recommended Measures," but not the detailed subtask data (to protect model details), while the quality control department can view the complete report.

[0169] like Figure 2 As shown in some embodiments of this application, a digital employee operating system is provided. The system includes a response module 210, an acquisition module 220, a matching module 230, and an operating module 240. Specifically:

[0170] The response module 210 is configured to respond to a request from a task to be processed by constructing a structured representation based on the task to be processed; the structured representation includes a target task, a target task type, and target task requirements; the target task is generated based on the task to be processed.

[0171] The acquisition module 220 is used to acquire the initial performance status of the digital employee; the digital employee includes at least two task execution models; the task execution models in the digital employee have different capabilities in performing tasks;

[0172] The matching module 230 is used to match the structured representation of the task to be processed with the initial performance state of the digital employee to obtain the target operation plan of the digital employee; the target operation plan includes the target operation subject of the target task and the target operation sequence of the target operation subject;

[0173] The running module 240 is used to run the digital employee to process the structured representation based on the target running scheme, and obtain the running results of the digital employee.

[0174] It should be noted that the digital employee operating system provided in this embodiment is based on the same inventive concept as the digital employee operating method described above. Therefore, the relevant content of the digital employee operating method described above also applies to the content of the digital employee operating system, and will not be repeated here.

[0175] To address this, the system responds to requests from pending tasks by constructing a structured representation based on those tasks; it obtains the initial performance state of the digital employee; each digital employee comprises at least two task execution models; the task execution capabilities of each model within the digital employee differ; the system matches the structured representation of the pending task with the initial performance state of the digital employee to obtain the target execution plan for the digital employee; based on the target execution plan, the system runs the digital employee to process the structured representation, obtaining the execution results. By constructing the structured representation, the system can accurately extract task entities, types, and requirements, and allocate tasks based on differences in model execution capabilities and real-time load resources. In this way, the system achieves accurate and efficient operation of the digital employee, can handle dynamic changes during operation, and fully meets the processing needs of enterprise-level tasks.

[0176] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method for operating digital employees.

[0177] like Figure 3 , Figure 3 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. The electronic device includes:

[0178] At least one battery;

[0179] At least one memory;

[0180] At least one processor;

[0181] At least one program;

[0182] The program is stored in memory, and the processor executes at least one program to implement the above-described method for operating a digital employee in accordance with this disclosure.

[0183] This electronic device can be any smart terminal, including mobile phones, tablets, personal digital assistants (PDAs), and in-vehicle computers.

[0184] The electronic devices according to embodiments of this application will now be described in detail.

[0185] The processor 1600 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this disclosure.

[0186] The memory 1700 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1700 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1700 and is called and executed by the processor 1600 to perform a digital employee operation method according to an embodiment of this disclosure.

[0187] The input / output interface 1800 is used to implement information input and output.

[0188] The communication interface 1900 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0189] Bus 2000 transmits information between various components of the device (e.g., processor 1600, memory 1700, input / output interface 1800, and communication interface 1900);

[0190] The processor 1600, memory 1700, input / output interface 1800 and communication interface 1900 are connected to each other within the device via bus 2000.

[0191] This disclosure also provides a storage medium, which is a computer-readable storage medium storing computer-executable instructions for causing a computer to perform the above-described method for operating a digital employee.

[0192] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0193] The embodiments described in this disclosure are for the purpose of more clearly illustrating the technical solutions of this disclosure and do not constitute a limitation on the technical solutions provided by this disclosure. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by this disclosure are also applicable to similar technical problems.

[0194] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this disclosure, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0195] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0196] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0197] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification 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 data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any related variations, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0198] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, 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 (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0199] In the several embodiments provided in this application, it should be understood that the disclosed apparatus 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; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0200] 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.

[0201] 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.

[0202] 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 multiple instructions to cause an electronic 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 programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0203] The above is a detailed description of the preferred embodiments of this application. However, the embodiments of this application are not limited to the above-described implementation methods. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the embodiments of this application. All such equivalent modifications or substitutions are included within the scope defined by the claims of the embodiments of this application.

[0204] The embodiments of this application have been described in detail above with reference to the accompanying drawings. However, this application is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of this application.

Claims

1. A method for operating a digital employee, characterized in that, The method includes: In response to a request for a pending task, a structured representation is constructed based on the pending task; the structured representation includes a target task, a target task type, and target task requirements; the target task is generated based on the pending task. Obtain the initial performance state of the digital employee; the digital employee includes at least two task execution models; the task execution models in the digital employee have different capabilities in performing tasks; The structured representation of the task to be processed is matched with the initial performance state of the digital employee to obtain the target operation plan of the digital employee; the target operation plan includes the target operation subject of the target task and the target operation sequence of the target operation subject; Based on the target operation plan, the digital employee processes the structured representation to obtain the operation results of the digital employee; The construction of a structured representation based on the task to be processed includes: Extract task entities and task operations from the tasks to be processed; Based on the task entity and the task operation, the tasks to be processed are classified to obtain the task type of the tasks to be processed. Based on the task type of the task to be processed, the task to be processed is divided into at least one target task to obtain the target task type of the target task; Based on the target task and the target task type, determine the target task requirements; The target task, target task attributes, and target task requirements are integrated to generate the structured representation.

2. The method for operating a digital employee according to claim 1, characterized in that, The step of dividing the task to be processed into at least one target task includes: dividing the task to be processed into a single target task and multiple target tasks.

3. The method for operating a digital employee according to claim 2, characterized in that, The initial performance state includes the execution capabilities of each task execution model. Matching the structured representation of the task to be processed with the initial performance state of the digital employee to obtain the target operation plan for the digital employee includes: When the task to be processed is divided into a single target task, the task similarity is calculated with the execution capability of each task execution model according to the requirements of the target task; the target operation plan of the digital employee is determined according to the task similarity and the preset first similarity threshold. When the task to be processed is divided into multiple target tasks, each target task is parsed to obtain the target task requirements of each target task and the correlation between each target task; the correlation represents the execution priority between each target task; according to the target task requirements of each target task, the corresponding task similarity is calculated with the execution capability of each task execution model; based on the correlation, the target operation plan of the digital employee is determined according to the task similarity and a preset second similarity threshold.

4. The method for operating a digital employee according to claim 3, characterized in that, The initial performance state also includes the load resources of each task execution model. Determining the target operation plan for the digital employee based on the similarity of each task and a preset first similarity threshold includes: The task similarity of each task execution model is compared with the preset first similarity threshold to obtain the corresponding comparison result; If the comparison result shows that the task similarity of the task execution model is greater than or equal to the preset first similarity threshold, the task execution model is determined as the initial operating entity of the digital employee; the target operating entity of the digital employee is determined based on the load resources of the initial operating entity and the target task; and the target operating plan of the digital employee is determined based on the target operating entity and the target task. If all the comparison results show that the task similarity of the task execution model is less than the preset first similarity threshold, the target operating entity of the digital employee is determined according to the load resources of each task execution model and the target task; and the target operating scheme of the digital employee is determined according to the target operating entity and the target task.

5. The method for operating a digital employee according to claim 3, characterized in that, The structured representation also includes the constraints of the target task. The determination of the target operational plan for the digital employee based on the association, according to the similarity of each task and a preset second similarity threshold, includes: The task similarity of each task execution model is compared with the preset second similarity threshold to obtain the corresponding comparison result; If the comparison result shows that the task similarity of the task execution model is greater than or equal to the preset second similarity threshold, the task execution model is determined as the initial running entity of the corresponding target task; if the comparison result shows that the task similarity of all the task execution models is less than the preset second similarity threshold, the initial running entity of the corresponding target task is determined according to the load resources of each task execution model and the target task corresponding to the task similarity. Based on the aforementioned relationships, the initial operating sequence of each of the initial operating entities is determined; Based on the running order, and according to the load resources of each initial running entity and the constraints of each target task, the target running entity of each target task and the target running order of each target running entity are determined. Based on the target operating entities of each target task and the target operating sequence of the target operating entities, the target operating plan of the digital employee is generated.

6. The method for operating a digital employee according to claim 1, characterized in that, The structured representation includes the constraints of the target task. The process of running the digital employee to process the structured representation based on the target operation plan, and obtaining the operation results of the digital employee, includes: Based on the target running entity of the target task and the target running sequence of the target running entity, the digital employee processes the structured representation to obtain the real-time load resources of the target running entity corresponding to the target task in real time; When the real-time load resources are less than a preset load threshold, the target running entity of the target task is adjusted according to the constraints of each target running entity to obtain the adjusted target running entity of the target task; and the target running order is updated according to the adjusted target running entity to obtain the updated target running order; and the digital employee is run according to the adjusted target running entity and the updated target running order to obtain the running result of the digital employee. When the real-time load resource is greater than or equal to the preset load threshold, the digital employee is run according to the target running entity and the target running sequence of the target running entity, and the running result of the digital employee is obtained.

7. A digital employee operating system, characterized in that, The system includes: A response module is used to respond to a request from a task to be processed and to construct a structured representation based on the task to be processed; the structured representation includes a target task, a target task type, and target task requirements; the target task is generated based on the task to be processed. An acquisition module is used to acquire the initial performance status of the digital employee; the digital employee includes at least two task execution models; the task execution models in the digital employee have different capabilities in performing tasks; A matching module is used to match the structured representation of the task to be processed with the initial performance state of the digital employee to obtain the target operation plan of the digital employee; the target operation plan includes the target operation subject of the target task and the target operation sequence of the target operation subject; The running module is used to run the digital employee to process the structured representation based on the target running scheme, and obtain the running results of the digital employee; The construction of a structured representation based on the task to be processed includes: Extract task entities and task operations from the tasks to be processed; Based on the task entity and the task operation, the tasks to be processed are classified to obtain the task type of the tasks to be processed. Based on the task type of the task to be processed, the task to be processed is divided into at least one target task to obtain the target task type of the target task; Based on the target task and the target task type, determine the target task requirements; The target task, target task attributes, and target task requirements are integrated to generate the structured representation.

8. An electronic device, characterized in that, It includes at least one control processor and a memory for communicatively connecting to the at least one control processor; the memory stores instructions executable by the at least one control processor to enable the at least one control processor to perform a method for operating a digital employee as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for causing a computer to perform a method of operating a digital employee as described in any one of claims 1 to 7.

Images