Data processing method, device and equipment and readable storage medium

Abstract

The invention discloses a data processing method, device and equipment and a readable storage medium, and the method comprises the steps: carrying out the task planning processing of a business problem, and obtaining a plan list comprising K subtasks; performing scene detection on the environment information corresponding to the target subtask to obtain a scene detection result; if the scene detection result meets the multi-cue-word generation condition, generating an object-side cue word and a model-side cue word based on the scene detection result; in response to a control operation corresponding to the object side cue word, obtaining an object interaction result, and in a task tool of the target sub-task, executing a task instruction of the target sub-task according to the target context information of the target sub-task, the object interaction result and the model side cue word to obtain a sub-task execution result; and when sub-task execution results corresponding to the K sub-tasks are obtained, performing integration processing on the K sub-task execution results to obtain an answer result. By adopting the method, the accuracy and the completion rate of the answer result can be improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a data processing method, apparatus, device, and readable storage medium. Background Technology

[0002] Existing intelligent agents face multiple challenges when handling complex tasks. Due to limitations in their thinking, problems expressed in natural language are easily trapped in fixed mindsets, making it difficult to clearly describe specific needs and adapt to the multi-dimensional task requirements in complex task domains. When the limitation of the problem expression leads to the failure to generate an answer, existing intelligent agents will only provide vague feedback such as "task failed," "please try again," or "please provide more details." This requires human intervention to guess the reason for the failure, modify the input data and prompts, and wait for the intelligent agent to regenerate the answer. This not only wastes time but also increases the consumption of computing resources, resulting in high trial-and-error costs and a low completion rate of generating answer results. Summary of the Invention

[0003] This application provides a data processing method, apparatus, device, and readable storage medium that can improve the accuracy and completion rate of response results.

[0004] One embodiment of this application provides a data processing method, including: The process involves identifying a business problem, performing task planning on that problem, and generating a plan list. The plan list includes K subtasks, where K is a positive integer. Each subtask includes different task instructions and different task tools. Determine the target subtask from the K subtasks, obtain the environmental information corresponding to the target subtask, perform scene detection on the environmental information, and obtain the scene detection result; If the scene detection result meets the conditions for generating multiple prompt words, then object-side prompt words and model-side prompt words are generated based on the scene detection result; the object-side prompt words are used to guide the object to perform the control operation corresponding to the scene detection result. The system responds to control operations guided by object-side prompts, acquires object interaction results, and sends the object interaction results and model-side prompts to the task tool corresponding to the target subtask. Within the task tool, the system executes the task instructions corresponding to the target subtask based on the target context information, object interaction results, and model-side prompts, thus obtaining the subtask execution result. The model-side prompts are used to indicate the constraints on the task tool during the execution of the task instructions corresponding to the target subtask. When the execution results of the K subtasks are obtained, the execution results of the K subtasks are integrated and processed to obtain the answer to the business question.

[0005] One embodiment of this application provides a data processing apparatus, including: The task planning module is used to obtain business problems, perform task planning on the business problems, and obtain a plan list. The plan list includes K sub-tasks, where K is a positive integer. Each sub-task includes different task instructions and different task tools. The scene detection module is used to identify the target subtask among K subtasks, obtain the environmental information corresponding to the target subtask, perform scene detection on the environmental information, and obtain the scene detection result. The prompt word generation module is used to generate object-side prompt words and model-side prompt words based on the scene detection results if the scene detection results meet the conditions for multi-prompt word generation. The object-side prompt words are used to guide the object to perform the control operations corresponding to the scene detection results. The interaction processing module is used to respond to control operations guided by object-side prompts, obtain object interaction results, and send the object interaction results and model-side prompts to the task tool corresponding to the target subtask. In the task tool corresponding to the target subtask, the task instructions corresponding to the target subtask are executed according to the target context information, object interaction results, and model-side prompts to obtain the subtask execution result. The model-side prompts are used to indicate the constraints of the task tool corresponding to the target subtask during the execution of the task instructions corresponding to the target subtask. The result integration module is used to integrate the execution results of K subtasks when the execution results of the K subtasks are obtained, and to obtain the answer result corresponding to the business question.

[0006] In one possible implementation, the business problem is the (H+1)th session initiated by the object, where H is a positive integer; the business problem is associated with the H sessions preceding the initiation of the business problem; when the scene detection module obtains the environmental information corresponding to the target subtask, it is specifically used to perform the following operations: Obtain the session issues corresponding to each of the H round sessions, identify the subtasks in the plan list that are executed before the target subtask, and determine them as preceding subtasks. Based on the execution results of the preceding subtasks, the business issues, the plan list, and the session issues corresponding to each of the H round sessions, generate the target context information corresponding to the target subtask. Obtain object preference information and execution log information of the execution status of task tools in the tool library, and determine the object preference information and execution log information as the system log information corresponding to the target subtask; Based on the target context information and system log information corresponding to the target subtask, generate the environment information corresponding to the target subtask.

[0007] In one possible implementation, when the prompt generation module generates object-side prompts and model-side prompts based on scene detection results, it specifically performs the following operations: Retrieve object-side and model-side cue elements that match the scene detection results from the cue word library; both object-side and model-side cue elements are cue word templates that can be filled with key information. Based on the target context information and scene detection results corresponding to the target sub-task, generate the first key information for the object-side cue element and the second key information for the model-side cue element; Fill the first key information into the object-side prompt element to obtain the object-side prompt word, and fill the second key information into the model-side prompt element to obtain the model-side prompt word.

[0008] In one possible implementation, the prompt word generation module is also used to perform the following operations: Within the statistical period, obtain the trigger counts corresponding to W prompt word generation scenarios; prompt word generation scenario refers to the scenario detection results that meet the conditions for multiple prompt word generation; W is a positive integer; Scenarios where the number of triggers is greater than or equal to the optimization threshold are identified as scenarios to be optimized, and the execution log information of the scenarios to be optimized within the statistical period is obtained; the execution log information includes the object-side prompts, model-side prompts, and subtask execution results generated each time the multi-prompt-word generation conditions are met within the statistical period; Based on the execution log information, the object-side hint elements and model-side hint elements corresponding to the scenario to be optimized are optimized and updated.

[0009] In one possible implementation, the environmental information includes the target context information corresponding to the target subtask; the scene detection module is used to perform scene detection on the environmental information, and when obtaining the scene detection result, it is specifically used to perform the following operations: If the amount of data of the content to be processed in the target context information is greater than or equal to the data amount threshold, then the scene detection result corresponding to the target subtask is determined to be the content exceeding the limit result. If the scene detection result is a content exceeding the limit, then the scene detection result is determined to meet the conditions for generating multiple prompt words; Based on scene detection results, object-side and model-side prompts are generated, including: Based on the content exceeding the limit result, object-side prompts and model-side prompts are generated. Object-side prompts are used to guide the object to compress the amount of data in the content to be processed, while model-side prompts are used to instruct the task tool corresponding to the target subtask to generate the subtask execution result corresponding to the target subtask based on the compressed content of the content to be processed.

[0010] In one possible implementation, the environmental information includes the target context information corresponding to the target subtask; the scene detection module is used to perform scene detection on the environmental information, and when obtaining the scene detection result, it is specifically used to perform the following operations: If the object does not have the permission to process the content to be processed in the target context information, then the scene detection result corresponding to the target subtask is determined to be a permission-restricted result. If the scene detection result is a restricted access result, then the scene detection result is determined to meet the conditions for generating multiple prompt words; Based on scene detection results, object-side and model-side prompts are generated, including: Based on the permission restriction result, object-side prompts and model-side prompts are generated. Object-side prompts are used to guide objects to request processing permissions for the content to be processed in the target context information. Model-side prompts are used to instruct the task tool corresponding to the target subtask to trigger the execution of the task instruction corresponding to the target subtask when it obtains the actual content corresponding to the content to be processed in the target context information.

[0011] In one possible implementation, the environmental information includes the target context information corresponding to the target subtask; the scene detection module is used to perform scene detection on the environmental information, and when obtaining the scene detection result, it is specifically used to perform the following operations: If the content to be processed in the target context information is located in a sub-partition of the data table, and the task instruction corresponding to the target sub-task does not include a partition identifier for the sub-partition, then the scene detection result corresponding to the target sub-task is determined to be a retrieval-restricted result. If the scene detection result is a search-restricted result, then the scene detection result is determined to meet the conditions for generating multiple prompt words; Based on scene detection results, object-side and model-side prompts are generated, including: Based on the limited search results, object-side prompts and model-side prompts are generated. Object-side prompts are used to indicate to the object that the task instruction corresponding to the target subtask does not include the partition identifier. Model-side prompts are used to instruct the task tool corresponding to the target subtask to trigger the execution of the task instruction that includes the partition identifier when it obtains the task instruction that includes the partition identifier.

[0012] In one possible implementation, the environmental information includes the target context information corresponding to the target subtask; the scene detection module is used to perform scene detection on the environmental information, and when obtaining the scene detection result, it is specifically used to perform the following operations: If the operational risk level of the task instruction corresponding to the target subtask is greater than or equal to the risk level threshold, then the scenario detection result corresponding to the target subtask is determined as the risk confirmation result. If the scene detection result is a risk confirmation result, then the scene detection result is determined to meet the conditions for generating multiple prompt words; Based on scene detection results, object-side and model-side prompts are generated, including: Based on the risk confirmation results, object-side prompts and model-side prompts are generated. Object-side prompts are used to guide the object to confirm the operational risk level of the task instructions corresponding to the target sub-task, while model-side prompts are used to instruct the task tool corresponding to the target sub-task to trigger the execution of the task instructions corresponding to the target sub-task when it obtains the object's confirmation result on the operational risk level.

[0013] In one possible implementation, the scene detection module is used to obtain environmental information corresponding to the target subtask, perform scene detection on the environmental information, and when obtaining the scene detection result, it is specifically used to perform the following operations: If the target subtask is a subtask that has already been executed, and no execution result has been obtained for the subtask corresponding to the target subtask, then obtain the execution log information generated by the task tool corresponding to the target subtask during the execution of the task instruction corresponding to the target subtask; Based on the execution log information, generate the environment information corresponding to the target subtask, and determine the scenario detection result corresponding to the target subtask as the execution failure result; If the scene detection result is an execution failure, then the scene detection result is determined to meet the conditions for generating multiple prompt words.

[0014] In one possible implementation, the K subtasks include subtasks. Kazuko's Mission , where i is a positive integer less than or equal to K; subtask The execution order of subtasks Prior to the execution order; when the scene detection module is used to determine the target subtask among K subtasks, it is specifically used to perform the following operations: The subtasks that do not have data dependencies among the K subtasks are identified as the target subtasks; Or, if subtask With subtasks Data dependencies and obtaining subtasks The result of the subtask execution will then be the subtask This is identified as a target sub-task; Or through subtasks The corresponding task tool executes subtasks. If the corresponding task instruction fails, the subtask will be... This has been identified as a target sub-task.

[0015] In one possible implementation, the K subtasks are obtained by the planning agent through task planning; the scene detection module is used to obtain the environmental information corresponding to the target subtasks, perform scene detection on the environmental information, and specifically perform the following operations when obtaining the scene detection results: If the target subtask is a subtask that has already been executed, and the execution result of the subtask corresponding to the target subtask has been obtained, then the task execution accuracy of the target subtask execution result is generated by the planning agent, and the environmental information corresponding to the target subtask is generated based on the task execution accuracy. If the task execution accuracy in the environmental information is less than the accuracy threshold, then the scene detection result corresponding to the target subtask is determined to be an execution failure result. If the scene detection result is an execution failure, then the scene detection result is determined to meet the conditions for generating multiple prompt words.

[0016] In one possible implementation, the K sub-tasks are obtained by the planning agent through task planning; the prompt word generation module is used to generate object-side prompt words and model-side prompt words based on scene detection results, specifically for the following operations: Based on the execution failure result, object-side prompts and model-side prompts are generated; object-side prompts are used to guide the object to modify the task instructions of the target subtask. If the modification result of the object's task instruction for the target subtask is not obtained, the model-side prompt is used to instruct the planning agent to modify the target subtask to obtain the new task tool and new task instruction corresponding to the target subtask. If the modification result of the object's task instruction for the target subtask is obtained, the model-side prompt is used to instruct the planning agent to obtain the new task tool and new task instruction corresponding to the target subtask from the modification result. The new task tool and new task instruction are used to generate the subtask execution result corresponding to the target subtask.

[0017] In one possible implementation, the task planning module is used to perform task planning on business problems. When a list of plans is obtained, it is specifically used to perform the following operations: By planning an intelligent agent, the business problem is broken down into K sub-tasks, each corresponding to a task instruction; the K sub-tasks include sub-tasks. , where i is a positive integer less than or equal to K; Obtain the task domain type corresponding to each of the E candidate tools in the tool library, based on subtasks. The corresponding task instructions and the task domain types corresponding to the E candidate tools are used to determine the subtasks from the E candidate tools. The corresponding task tool; E is a positive integer; Once the task tools corresponding to the K sub-tasks are obtained, the task instructions and task tools corresponding to the K sub-tasks are determined as a plan list.

[0018] In one possible implementation, the K subtasks are obtained by the planning agent through task planning. The interaction processing module is used to execute the task instructions corresponding to the target subtask in the task tool corresponding to the target subtask, based on the target context information, object interaction results, and model-side prompts. When obtaining the subtask execution result, it is specifically used to perform the following operations: By executing the intelligent agent to call the task tool corresponding to the target subtask, the task tool corresponding to the target subtask executes the task instruction corresponding to the target subtask based on the target context information, object interaction results and model-side prompts, and obtains the subtask execution result corresponding to the target subtask. When the execution results of the K subtasks are obtained, the execution results of the K subtasks are integrated and processed to obtain the answer to the business question, including: If the communication mode corresponding to the business problem is the rapid analysis mode, then the context information of the next subtask of the target subtask is generated based on the execution result of the subtask corresponding to the target subtask. In the task tool corresponding to the next subtask, the task instructions corresponding to the next subtask are executed based on the context information obtained from the next subtask. When the execution results of the K subtasks are obtained, the planning agent integrates and processes the execution results of each subtask to obtain the answer to the business problem. If the communication mode corresponding to the business problem is the deep analysis mode, and the target execution result meets the plan modification conditions, then the plan list is modified to obtain an updated plan list. The new subtasks in the updated plan list are then executed until the execution results of each new subtask do not meet the plan modification conditions. At this point, the planning agent integrates the execution results of each subtask to obtain the answer to the business problem.

[0019] In one possible implementation, the communication mode corresponding to the business problem is a deep analysis mode, and the interaction processing module is also used to perform the following operations: The task execution accuracy is calculated by planning the intelligent agent and verifying the prompt words to generate the target sub-task execution results. If the task execution accuracy is less than the first accuracy threshold, it is determined that the target execution result meets the plan modification conditions. The update plan list includes a first update plan list and a second update plan list; by modifying the plan list, the updated plan list is obtained, which includes: If the task execution accuracy is less than the second accuracy threshold, the business problem is re-planned and processed in the planning agent using the plan list and modification prompts to obtain the first updated plan list; the new subtasks in the first updated plan list are different from the subtasks in the plan list; the first accuracy threshold is greater than the second accuracy threshold. If the task execution accuracy is greater than or equal to the second accuracy threshold, and the task execution accuracy is less than the first accuracy threshold, then the subtasks in the plan list that are executed before the target subtask are identified as preceding subtasks. The task instructions and execution tools in the target subtask and preceding subtask in the plan list are modified to obtain a second updated plan list. The second updated plan list includes subtasks in the plan list that are executed after the target subtask.

[0020] In one possible implementation, the interaction processing module is also used to perform the following operations: By matching keywords to business issues using a keyword database, guiding keywords that indicate the type of business issue are obtained; The text length of the business issue, the guiding keywords, and the number of subtasks in the plan list are weighted and summed to obtain the comprehensive score of the business issue; If the overall score is greater than or equal to the in-depth analysis threshold, then the communication mode corresponding to the business problem is determined to be the in-depth analysis mode. If the overall score is less than the in-depth analysis threshold, the communication mode corresponding to the business problem is determined to be the rapid analysis mode.

[0021] One embodiment of this application provides a computer device, including: a processor, a memory, and a network interface; The processor is connected to a memory and a network interface. The network interface is used to provide data communication functions, and the memory is used to store computer programs. When the computer program is executed by the processor, the computer device performs the method provided in the embodiments of this application.

[0022] One aspect of this application provides a computer-readable storage medium storing a computer program adapted to be loaded and executed by a processor, so that a computer device having the processor performs the method provided in this application.

[0023] One embodiment of this application provides a computer program product comprising a computer program stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium and executes the computer program, causing the computer device to perform the method provided in this application embodiment.

[0024] This application embodiment processes business problems through task planning, resulting in a plan list comprising K sub-tasks. Complex business problems can be broken down into K sub-tasks of appropriate granularity. Each sub-task includes specific task instructions and corresponding task tools. Through multi-task tool collaboration and multi-sub-task step-by-step reasoning, the solution process for complex business problems is decomposed into a progressive execution chain, reducing the processing pressure on individual task tools. Simultaneously, by leveraging the complementary functional characteristics of different task tools, the limitations of a single task tool in complex scenarios are avoided, improving the accuracy of sub-task execution. During the execution of the K sub-tasks, a target sub-task can be identified, and its corresponding environmental information can be obtained. Scene detection is performed on the environmental information to obtain scene detection results. If the scene detection results meet the conditions for multi-prompt generation, object-side prompts and model-side prompts can be generated based on the scene detection results. Object-side prompts explicitly guide the object to execute the control operations corresponding to the scene detection results, while model-side prompts implicitly inject constraints during the execution of the task instructions corresponding to the target sub-task by the task tool corresponding to the target sub-task. By distinguishing between object-side and model-side prompts, we can differentiate between the guidance needed by the user and the instructions required by the model. This allows us to optimize model constraints and the user's interaction environment, helping users express their needs more effectively and understand the model's capability boundaries. Responding to control operations guided by object-side prompts and obtaining object interaction results (such as supplementary user information, confirmed parameter options, and feedback on execution preferences) improves the coherence of the user conversation. Sending the object interaction results and model-side prompts to the task tool corresponding to the target subtask allows for targeted addition of constraints such as execution constraints, data specifications, or behavioral restrictions during the execution of the task instructions based on the target context information, object interaction results, and model-side prompts. This results in more accurate subtask execution results and ultimately improves the accuracy of the answer obtained by integrating the results of K subtasks. The dual-prompt word collaboration mechanism ensures that both the model and the user receive the most suitable guidance information. For users, it optimizes the retry process, reduces cognitive load, and allows for efficient task execution without requiring specialized technical skills, thus lowering the technical barrier. For the model, the injection of precise constraints improves execution efficiency and standardization, avoids invalid calculations, and saves computing resources. The synergistic effect of both effectively improves the interaction efficiency between users and the model, thereby increasing the completion rate of answers to complex business questions. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a schematic diagram of a network architecture provided in an embodiment of this application; Figure 2 This is a schematic diagram of a data processing scenario provided in an embodiment of this application. Figure 1 ; Figure 3 This is a flowchart illustrating a data processing method provided in an embodiment of this application. Figure 1 ; Figure 4 This is a schematic diagram of a data processing scenario provided in an embodiment of this application. Figure 2 ; Figure 5 This is a flowchart illustrating a data processing method provided in an embodiment of this application. Figure 2 ; Figure 6 This is a schematic diagram of a data processing scenario provided in an embodiment of this application. Figure 3 ; Figure 7 This is a flowchart illustrating a data processing method provided in an embodiment of this application. Figure 3 ; Figure 8 This is a flowchart illustrating a data processing method provided in an embodiment of this application. Figure 4 ; Figure 9 This is a flowchart illustrating a data processing method provided in an embodiment of this application. Figure 5 ; Figure 10 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application; Figure 11 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0027] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0028] It is understood that in the specific embodiments of this application, user (object or player) data, such as business issues and object preference information, are involved. When the above and below embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant regions.

[0029] If this application requires the collection of object data (such as user data), a prompt interface or pop-up window will be displayed before and during the collection process. This prompt interface or pop-up window is used to inform the user that certain data is being collected. The data acquisition steps will only begin after the user confirms the prompt interface or pop-up window; otherwise, the process will end. Furthermore, the acquired user data will be used in reasonable and legal scenarios or for legitimate purposes. Optionally, in scenarios where user data needs to be used but user authorization has not been obtained, authorization can be requested from the user, and the user data can only be used after authorization is granted.

[0030] Please see Figure 1 , Figure 1 This is a schematic diagram of a network architecture provided in an embodiment of this application. For example... Figure 1 As shown, the network architecture may include a service server 100 and a terminal device cluster. The terminal device cluster may include terminal devices 10a, 10b, ..., 10n. Each terminal device in the terminal device cluster may have a communication connection with the service server 100. For example, terminal device 10a may have a communication connection with the service server 100, and terminal device 10b may have a communication connection with the service server 100. The communication connection is not limited to a specific method. It may be a direct or indirect connection via wired communication, a direct or indirect connection via wireless communication, or other methods. This application does not impose any restrictions on these methods.

[0031] Each terminal device in the terminal device cluster can include: smartphones, tablets, laptops, desktop computers, intelligent voice interaction devices, smart home appliances (e.g., smart TVs), wearable devices, in-vehicle terminals, aircraft, and other intelligent terminals with data processing capabilities. In-vehicle terminals can be terminal devices used in intelligent transportation scenarios and assisted driving scenarios. It should be understood that, for example... Figure 1 Each terminal device in the terminal device cluster shown can be equipped with an application client that has data processing capabilities. When the application client runs on each terminal device, it can interact with the aforementioned... Figure 1 Data interaction is performed between the business servers 100 shown.

[0032] Specifically, the application client may include: smart agent client, vehicle client, smart home client, entertainment client (e.g., game client), multimedia client (e.g., video client), social client, and information client (e.g., news client), etc. In this embodiment, the application client may be integrated into a client (e.g., a social client) or may be a standalone client (e.g., a news client). This embodiment does not limit the type of application client.

[0033] The business server 100 can be the server corresponding to the application client. The business server 100 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. This embodiment of the application does not impose any restrictions.

[0034] like Figure 1 As shown, for ease of understanding, taking terminal device 10a in the terminal device cluster as an example, terminal device 10a can obtain the business question input by the object (user) in the application client. Terminal device 10a can send the business question to business server 100. Business server 100 can deploy an intelligent assistant system (Agent system). Business server 100 can perform task planning and processing on the business question through the Agent system to obtain a plan list. The plan list includes K sub-tasks derived from the business task, each sub-task including different task instructions and different task tools. The task tools can be Model Context Protocol (MCP) tools, such as artificial intelligence models.

[0035] The business server 100 can sequentially determine the target subtask from the K subtasks in the plan list according to the execution order (data dependencies), and obtain the corresponding environment information for the target subtask. This environment information may include the target context information and system log information. The target context information may include knowledge fragments retrieved from the knowledge base and memory information obtained from the memory system (e.g., business issues and the plan list). The system log information may include logs generated by the Agent system during the execution of the subtask, as well as object preference information and historical behavior information for that object. The memory system is responsible for persistently storing the interaction content of historical sessions and temporarily caching the interaction content of the current session.

[0036] The business server 100 can perform scene detection on environmental information and obtain scene detection results. If the scene detection results meet the conditions for generating multiple prompt words, the business server 100 can generate object-side prompt words and model-side prompt words based on the scene detection results. The business server 100 can return the object-side prompt words to the terminal device 10a, and display them to the object through the application client in the terminal device 10a, guiding the object to complete the control operation corresponding to the object-side prompt words. The object-side prompt words are used to guide the object to execute the control operation corresponding to the scene detection results.

[0037] Terminal device 10a can respond to control operations guided by object-side prompts, obtain the object interaction results corresponding to the control operations, and send the object interaction results to business server 100. Business server 100 can then send the object interaction results and model-side prompts to the task tool corresponding to the target subtask. In the task tool corresponding to the target subtask, the task instructions corresponding to the target subtask are executed based on the target context information, object interaction results, and model-side prompts, resulting in the subtask execution result. The model-side prompts are used to indicate the constraints of the task tool during the execution of the task instructions corresponding to the target subtask.

[0038] When the execution results of the K subtasks are obtained, the business server 100 can integrate and process the K subtask execution results to obtain the answer to the business question. The answer is then sent to the terminal device 10a.

[0039] Optionally, the terminal device 10a and the business server 100 can be integrated into one device, and this embodiment of the application does not impose any limitations on this.

[0040] This application's embodiments propose a dual-prompt mechanism consisting of object-side prompts and model-side prompts. This mechanism clearly separates the guidance information needed by the user from the execution instructions required by the model, achieving precise separation and bidirectional adaptation between "user interaction needs" and "model operation needs." On the one hand, it optimizes the model's constraint injection method, avoiding execution deviations caused by redundant or missing constraints and improving the standardization of model execution. On the other hand, it optimizes the user's interaction environment, guiding users to efficiently express their needs through concise and precise prompts, while clearly informing them of the model's capability boundaries and the execution requirements of the current task, helping users establish reasonable expectations and reducing ineffective interactions. The coordinated effect of the dual prompts effectively improves the interaction efficiency between the user and the model, reduces network communication costs, and increases the completion rate of answers to complex business questions.

[0041] Please see Figure 2 , Figure 2 This is a schematic diagram of a data processing scenario provided in an embodiment of this application. Figure 1 .like Figure 2 As shown, a computer device can acquire business questions input by an object (user). As illustrated in Scenario 101, the computer device can display a list of plans corresponding to the business questions through an application client. This list can include K subtasks, such as subtask 1, subtask 2, subtask 3, subtask 4, and subtask 5. The computer device can be as follows: Figure 1 The terminal device 10a and the business server 100 shown are integrated into one device.

[0042] Each subtask includes corresponding task instructions and tools. For example, subtask 1 could refer to the data preparation stage in the plan list, with task instructions of "1.1 Obtain the data table and its fields; 1.2 Determine the backtesting time; Determine the metrics to be calculated". Subtask 2 could refer to the strategy implementation stage in the plan list, with task instructions of "2.1 Define each metric; 2.2 Record the parameters required for each metric". Subtask 3 could refer to the backtesting execution stage in the plan list, with task instructions of "3.1 Calculate each metric; 3.2 Simulate each metric for backtesting". Subtask 4 could refer to the results analysis stage in the plan list, with task instructions of "...". Subtask 5 could refer to the results integration stage in the plan list, with task instructions of "...".

[0043] The computer device can execute the K subtasks sequentially according to the execution order among the K subtasks in the schedule (such as the order of data dependencies).

[0044] Taking the execution of a target subtask out of K subtasks as an example, the computer device can obtain the environmental information corresponding to the target subtask. This environmental information may include the target context information and system log information. If it is detected that an object does not have the permission to process the content to be processed in the target context information, the computer device can determine that the scene detection result for the target subtask is a permission-restricted result. If the scene detection result is a permission-restricted result, the computer device can determine that the scene detection result meets the conditions for generating multiple prompts. Based on the permission-restricted result, the computer device can generate object-side prompts and model-side prompts. The object-side prompts guide the object to request permission to process the content to be processed in the target context information, while the model-side prompts instruct the task tool corresponding to the target subtask to trigger the execution of the task instruction corresponding to the target subtask when it obtains the actual content corresponding to the content to be processed in the target context information.

[0045] As shown in Scenario 102, the computer device can display object-side prompts in area X. For example, the object-side prompts could be a message like "Data table retrieval failed, request read permission for the data table?" These prompts guide the object to request read permission. The computer device can respond to the control operation guided by the prompts, confirming the request for read permission. When the object obtains read permission, the computer device can define this permission as the result of the control operation. The computer device can then send the object interaction result and the model-side prompts to the task tool corresponding to the target subtask. In the task tool, based on the target context information, the object interaction result, and the model-side prompts, the task instructions for the target subtask are executed, resulting in the subtask execution result. The model-side prompts can instruct the task tool to trigger the execution of the task instructions for the target subtask when it obtains the actual content corresponding to the content to be processed in the target context information. For example, it could be "Note: Do not fabricate or create data."

[0046] When the execution results of the K subtasks are obtained, the computer device can integrate and process the execution results of the K subtasks to obtain the conclusion shown in Scenario 103.

[0047] The dual-hint mechanism, consisting of object-side and model-side hints, proposed in this application can be widely applied to various artificial intelligence models. Object-side hints clearly define the user's required guidance, reducing cognitive load and allowing efficient task execution without requiring specialized technical skills. Model-side hints enable precise constraint injection, improving execution efficiency and standardization, avoiding invalid computations, and allowing the model to reason within a controlled constraint environment, outputting more accurate and secure results. In the field of data analysis, object-side hints can alert users to syntax errors, data volume exceeding limits, insufficient permissions, etc., and provide transparent solution paths for users to choose from, lowering the technical threshold for data analysis, avoiding task execution failures or repeated retries, and improving the interactive efficiency of data analysis.

[0048] Please see Figure 3 , Figure 3 This is a flowchart illustrating a data processing method provided in an embodiment of this application. Figure 1 This data processing method can be executed by a computer device, which can be, for example, Figure 1 The terminal device 10a and the business server 100 shown are integrated into one device. The following description will use the example of this data processing method being executed by a computer device. This data processing method may include at least the following steps S101-S105: Step S101: Obtain the business problem, perform task planning processing on the business problem, and obtain a plan list; the plan list includes K sub-tasks, where K is a positive integer; each sub-task includes different task instructions and different task tools; Specifically, the computer device can obtain the business problem input by the user in the application client, decompose the business problem into tasks, and obtain the task instructions corresponding to K sub-tasks of the business problem. The task decomposition process can be to divide a complex task into smaller tasks to solve the complex problem. Task decomposition can be implemented through Chain of Thought (CoT), Hierarchical Task Network (HTN), Step-by-Step Prompting, etc., and this application embodiment does not limit it.

[0049] The computer device can acquire E candidate tools that are idle in the tool library. These E candidate tools can be AI models for different task domains, i.e., the task domain types corresponding to the E candidate tools. Based on the task instructions corresponding to K sub-tasks and the task domain types corresponding to the E candidate tools, the computer device can determine the task tools corresponding to the K sub-tasks from the E candidate tools. Once the task tools corresponding to the K sub-tasks are acquired, the task instructions and task tools corresponding to the K sub-tasks are compiled into a plan list. Each sub-task includes different task instructions and different task tools.

[0050] Step S102: Determine the target subtask from the K subtasks, obtain the environmental information corresponding to the target subtask, perform scene detection on the environmental information, and obtain the scene detection result; Specifically, a computer device can identify a target subtask from among K subtasks that do not have data dependencies, or identify a subtask as a target subtask when the preceding subtask of a certain subtask (which is an unexecuted subtask) has been completed.

[0051] Here, a preceding subtask refers to a subtask in the schedule that precedes the target subtask in the execution order. For ease of understanding, let's consider subtasks within a K-task list. Kazuko's Mission For example, if the subtask With subtasks Data dependencies (i.e., subtasks) For subtasks (preceding subtasks), and the subtasks have been acquired. If the result of the subtask execution is obtained, the computer device can then process the subtask. Identify it as the target subtask. Or, through subtasks... The corresponding task tool executes subtasks. If the corresponding task instruction fails, the subtask will be... This has been identified as a target sub-task.

[0052] Please see also Figure 4 , Figure 4 This is a schematic diagram of a data processing scenario provided in an embodiment of this application. Figure 2 .like Figure 4 As shown, the computer device can acquire environmental information corresponding to the target subtask. This environmental information may include target context information and system log information. The target context information may include knowledge fragments retrieved from the knowledge base and memory information obtained from the memory system (e.g., business problems and plan lists). The system log information may include logs generated by the Agent system during the execution of the subtask, as well as object preference information and historical behavior information for that object. The computer device can perform scene detection on the environmental information to obtain scene detection results. Scene detection can be used to determine the scene type of the environmental information. For example, if the environmental information includes environmental signals or control signals of a specific scene, the computer device can identify the corresponding characteristic scene as the scene detection result of the environmental information.

[0053] Step S103: If the scene detection result meets the conditions for generating multiple prompt words, then generate object-side prompt words and model-side prompt words based on the scene detection result; the object-side prompt words are used to guide the object to perform the control operation corresponding to the scene detection result. For details, please refer to the following: Figure 4 ,like Figure 4 As shown, the computer device can determine whether the scene detection result meets the conditions for generating multiple prompt words. Scene detection results that meet the conditions for generating multiple prompt words can include the following examples: 1. Content exceeding the limit result refers to the detection result when the amount of data of the content to be processed in the target context information detected in the environmental information is greater than or equal to the data amount threshold; 2. Restricted permission result refers to the detection result when the environment information detects that an object does not have the permission to process the content to be processed in the target context information; 3. Restricted search results refer to the detection results when the content to be processed in the target context information is located in a sub-partition of the data table in the environmental information, and the task instruction corresponding to the target sub-task does not include a partition identifier for the sub-partition; 4. Risk confirmation result refers to the detection result in the environmental information that the operational risk level of the task instruction corresponding to the target sub-task is greater than or equal to the risk level threshold. 5. Engine error results refer to the detection results when the syntax type of the task instruction corresponding to the target subtask is different from the database type of the content to be processed, as detected in the environment information.

[0054] 6. Execution failure result refers to the detection result when the system log information in the environment information indicates that the target subtask has been executed, and the target subtask failed to execute or the accuracy of the subtask execution result is too low.

[0055] If the scene detection result is as shown in the example above, the computer device can retrieve object-side and model-side cue elements that match the scene detection result from the cue word library. Both object-side and model-side cue elements are cue word templates that can be filled with key information. The computer device can generate first key information for the object-side cue element and second key information for the model-side cue element based on the target context information corresponding to the target subtask and the scene detection result. The first key information is filled into the object-side cue element to obtain the object-side cue word, and the second key information is filled into the model-side cue element to obtain the model-side cue word. For example, if the object lacks read permission for data table 1, the scene detection result is a restricted permission result. The retrieved object-side cue element could be "[First Key Information 1] Retrieval failed, please apply for [First Key Information 1]'s [First Key Information 2]", where first key information 1 could be "data table 1", and first key information 2 could be read permission and the path to apply for read permission, etc. Object-side prompts are used to guide objects to perform control operations corresponding to the scene detection results, while model-side prompts are used to indicate the constraints of the task tool corresponding to the target subtask during the execution of the task instructions corresponding to the target subtask.

[0056] Step S104: Respond to the control operation guided by the object-side prompt, obtain the object interaction result, and send the object interaction result and the model-side prompt to the task tool corresponding to the target subtask. In the task tool corresponding to the target subtask, execute the task instructions corresponding to the target subtask according to the target context information, object interaction result, and model-side prompt, and obtain the subtask execution result. The model-side prompt is used to indicate the constraints of the task tool corresponding to the target subtask during the execution of the task instructions corresponding to the target subtask. Specifically, computer devices can respond to control operations guided by prompts from the object side and obtain the object interaction results. Taking the scene detection result as a permission-restricted result as an example, the control operation is to confirm the request for read permission to the data table. When the object obtains read permission to the data table, the computer device can determine the read permission of the data table as the object interaction result corresponding to the control operation.

[0057] Computer devices can send the object interaction results and model-side prompts to the task tool corresponding to the target subtask. Within the task tool, the task instructions for the target subtask are executed based on the target context information, the object interaction results, and the model-side prompts, resulting in the subtask execution result. Taking a scene detection result with restricted permissions as an example, the model-side prompts can instruct the task tool corresponding to the target subtask to trigger the execution of the task instructions when it obtains the actual content corresponding to the content to be processed in the target context information. For example, it could be "Note: Do not fabricate or falsify data."

[0058] If the scene detection result does not meet the conditions for generating multiple prompt words, the computer device can also execute the task instructions corresponding to the target subtask in the task tool corresponding to the target subtask according to the target context information corresponding to the target subtask, and obtain the subtask execution result.

[0059] Step S105: When the execution results of the K subtasks are obtained, the execution results of the K subtasks are integrated to obtain the answer to the business question.

[0060] Specifically, after obtaining the execution results of the subtasks corresponding to the target subtask, the computer device can continue to execute the next subtask in the execution order after the target subtask in the plan list, until the execution results of the K subtasks are obtained. Then, the execution results of the K subtasks are integrated and processed to obtain the answer to the business question.

[0061] This application embodiment processes business problems through task planning, resulting in a plan list comprising K sub-tasks. Complex business problems can be broken down into K sub-tasks of appropriate granularity. Each sub-task includes specific task instructions and corresponding task tools. Through multi-task tool collaboration and multi-sub-task step-by-step reasoning, the solution process for complex business problems is decomposed into a progressive execution chain, reducing the processing pressure on individual task tools. Simultaneously, by leveraging the complementary functional characteristics of different task tools, the limitations of a single task tool in complex scenarios are avoided, improving the accuracy of sub-task execution. During the execution of the K sub-tasks, a target sub-task can be identified, and its corresponding environmental information can be obtained. Scene detection is performed on the environmental information to obtain scene detection results. If the scene detection results meet the conditions for multi-prompt generation, object-side prompts and model-side prompts can be generated based on the scene detection results. Object-side prompts explicitly guide the object to execute the control operations corresponding to the scene detection results, while model-side prompts implicitly inject constraints during the execution of the task instructions corresponding to the target sub-task by the task tool corresponding to the target sub-task. By distinguishing between object-side and model-side prompts, we can differentiate between the guidance needed by the user and the instructions required by the model. This allows us to optimize model constraints and the user's interaction environment, helping users express their needs more effectively and understand the model's capability boundaries. Responding to control operations guided by object-side prompts and obtaining object interaction results (such as supplementary user information, confirmed parameter options, and feedback on execution preferences) improves the coherence of the user conversation. Sending the object interaction results and model-side prompts to the task tool corresponding to the target subtask allows for targeted addition of constraints such as execution constraints, data specifications, or behavioral restrictions during the execution of the task instructions based on the target context information, object interaction results, and model-side prompts. This results in more accurate subtask execution results and ultimately improves the accuracy of the answer obtained by integrating the results of K subtasks. The dual-prompt word collaboration mechanism ensures that both the model and the user receive the most suitable guidance information. For users, it optimizes the retry process, reduces cognitive load, and allows for efficient task execution without requiring specialized technical skills, thus lowering the technical barrier. For the model, the injection of precise constraints improves execution efficiency and standardization, avoids invalid calculations, and saves computing resources. The synergistic effect of both effectively improves the interaction efficiency between users and the model, thereby increasing the completion rate of answers to complex business questions.

[0062] Please see Figure 5 , Figure 5 This is a flowchart illustrating a data processing method provided in an embodiment of this application. Figure 2 This data processing method can be executed by a computer device, which can be, for example, Figure 1 The terminal device 10a and the business server 100 shown are integrated into one device. The following description will use the example of this data processing method being executed by a computer device. This data processing method may include at least the following steps S201-S207: Step S201: Obtain the business problem, perform task planning processing on the business problem, and obtain a plan list; the plan list includes K sub-tasks, where K is a positive integer; each sub-task includes different task instructions and different task tools; Specifically, computer devices can acquire business questions input by users on application clients. These application clients can include intelligent assistant systems, or agent systems. Agent systems can include planning agents and execution agents. The planning agent is the "brain" or "mentor" of the agent system, responsible for formulating and monitoring high-level strategies. The execution agent is the "limbs" or "operator" of the agent system, responsible for executing specific tasks and providing feedback.

[0063] Computer devices can decompose business problems into tasks by planning intelligent agents, obtaining task instructions for K sub-tasks corresponding to the business problem. The task decomposition process can involve dividing complex tasks into smaller tasks to solve complex problems. Task decomposition can be implemented through hierarchical task networks based on thought chains, step-by-step prompting methods, etc., and the embodiments of this application are not limited thereto.

[0064] The computer device can acquire E candidate tools that are idle in the tool library. These E candidate tools can be AI models for different task domains, i.e., the task domain types corresponding to the E candidate tools. Based on the task instructions corresponding to K sub-tasks and the task domain types corresponding to the E candidate tools, the computer device can determine the task tools corresponding to the K sub-tasks from the E candidate tools. Once the task tools corresponding to the K sub-tasks are acquired, the task instructions and task tools corresponding to the K sub-tasks are compiled into a plan list. Each sub-task includes different task instructions and different task tools.

[0065] It is understood that the embodiments of this application can reasonably configure different models (task tools) for different sub-tasks, giving full play to the strengths of each model, thereby optimizing the execution effect of the plan list. For example, in the field of code generation, the embodiments of this application can assign Claude (Claude Series Large Language Models), GLM (General Language Model), KIMI-K2, etc., to the sub-tasks responsible for code generation. Among them, the Claude series models perform well in terms of code logic rigor and reasonable architecture design. The GLM model has unique advantages in understanding Chinese programming requirements and adapting to local development environments, while KIMI-K2 is superior in code executability and engineering practicality. In the field of text processing, the embodiments of this application can assign Deepseek series models to the sub-tasks responsible for text generation. The Deepseek series models perform in a balanced way in terms of text generation quality, semantic understanding depth, and instruction following accuracy, making them particularly suitable for scenarios requiring comprehensive capabilities. In the field of task execution, embodiments of this application can assign the KIMI-K2 model to the subtasks responsible for task execution. The KIMI-K2 model exhibits excellent controllability in task decomposition, step control, and result verification, and can ensure the reliable execution of complex tasks.

[0066] When assigning task tools to K sub-tasks, this application embodiment can implement a combination strategy framework, assigning task tools to the K sub-tasks according to the principle of "professional division of labor and complementary advantages." For example, in the requirements analysis phase: first, the task requirements are deeply analyzed to identify the core difficulties and key requirements of the task. In the model matching phase: the most suitable model combination is selected according to the task characteristics to ensure that each sub-task is handled by the model that is most proficient. In the process connection design: a clear data flow and result transmission mechanism is designed to ensure seamless connection of different model outputs. In the quality verification mechanism: a cross-model cross-checking and result verification process is established to ensure the reliability of the final output.

[0067] For example, in code development projects, this application's embodiment can employ a combination of models: "high-level models such as the Claude series are responsible for architecture design (i.e., planning-related content); models adapted for scheduling such as the GLM / KIMI series handle specific implementation; models adapted for testing such as the Deepseek series are used for testing and verification; and models with strong code generation capabilities, such as the GLM model, are responsible for code generation tools." This combination significantly improves code quality and development efficiency. For complex processing tasks, high-level models are responsible for generating core planning content, while other models focus on execution-level details. Their collaborative work ensures both content quality and successful task completion. This model combination strategy not only improves the effectiveness of individual tasks but, more importantly, establishes a scalable AI capability framework, laying a solid foundation for future integration of more specialized models.

[0068] Step S202: Determine the target subtask from the K subtasks; Specifically, a computer device can identify a target subtask from among K subtasks that do not have data dependencies, or identify a subtask as a target subtask when the preceding subtask of a certain subtask (which is an unexecuted subtask) has been completed.

[0069] Here, a preceding subtask refers to a subtask in the schedule that precedes the target subtask in the execution order. For ease of understanding, let's consider subtasks within a K-task list. Kazuko's Mission For example, if the subtask With subtasks Data dependencies (i.e., subtasks) For subtasks (preceding subtasks), and the subtasks have been acquired. If the result of the subtask execution is obtained, the computer device can then process the subtask. Identify it as the target subtask. Or, through subtasks... The corresponding task tool executes subtasks. If the corresponding task instruction fails, the subtask will be... This has been identified as a target sub-task.

[0070] Step S203: Obtain the environmental information corresponding to the target subtask, perform scene detection on the environmental information, and obtain the scene detection result; Specifically, the computer device can obtain the environmental information corresponding to the target subtask. This environmental information may include the target context information and system log information. Taking the H+1th round of the session initiated with a business problem as the object, and as an example where the business problem is associated with the H rounds of sessions preceding it, the target subtask is the A+1th dialogue in the K subtasks' respective dialogues. A session refers to an interaction process between the computer device and the Agent system in the application client. A dialogue refers to an interaction process between the intelligent agent in the Agent system and the task tool; that is, a dialogue is the process by which the intelligent agent calls the task tool to execute task processing. For example, by performing task planning on the business problem input in the H+1th session, K subtasks can be obtained, and then K dialogues can be constructed between the intelligent agent and the task tool. These H sessions can be associated with the business problem.

[0071] The agent system may also include a memory system, which can be a long short-term memory architecture. When the agent performs the Ath dialogue, its output may include the execution result of the subtask corresponding to the Ath dialogue and the context information used in the Ath dialogue. The computer device can input the output of the Ath dialogue into a high-efficiency summarization component. The high-efficiency summarization component extracts a summary of the output of the (A-1)th dialogue to obtain the unit memory information corresponding to the Ath dialogue, and stores the unit memory information in a context buffer as short-term memory information. The output of the Ath dialogue may include the execution result of the subtask output by the Ath dialogue, and the context information used in the Ath dialogue. The context information used in the Ath dialogue may include auxiliary prompts, example data, and execution description information. Auxiliary prompts refer to the auxiliary text generated by the agent for the invoked task tool in each dialogue; for example, it may be prompts for data that the prompting model cannot fabricate. Example data refers to the reference information provided by the agent for the invoked task tool; for example, it may be data in the output format of the prompting model. The execution description information includes pre-tool consideration content and tool scheduling content. The tool scheduling content is used to indicate the task tool called by the target subtask, and the pre-tool consideration content is the decision information for the policy agent to generate the tool scheduling content. The pre-tool consideration content is the reason and textual information explanation for the agent to allocate the task tool to the subtask.

[0072] The output of a completed dialogue can be extracted using a high-efficiency summarization component to obtain the corresponding unit memory information, also known as working memory information. Optionally, the unit memory information may also include execution description information, subtask execution results, auxiliary prompts, sample data, etc., which are not limited in this embodiment.

[0073] Taking the (A+1)th dialogue in the K dialogues within a business problem as an example, the short-term memory system can store working memory information, the business problem, and the plan list corresponding to dialogue A in the current session (H+1th session). The computer device can integrate the task instructions corresponding to the target subtask with the working memory information, business problem, and plan list corresponding to dialogue A to obtain the short-term memory information corresponding to the target subtask. It can be understood that before the current session (H+1th session) ends, the context buffer can store unit memory information of completed dialogues. When obtaining input content corresponding to dialogue A, taking the acquisition of short-term memory information corresponding to the target subtask as an example, dialogue A can include subtasks. In the corresponding dialogue, the computer device can retrieve subtasks from the context buffer. The corresponding unit memory information, i.e., obtaining subtasks. The corresponding execution description information, subtask execution results, auxiliary prompts, and sample data, or the results obtained after the above summary extraction. When the unit memory information corresponding to each of the A dialogues is obtained, the computer device can determine the unit memory information, business questions, and plan list corresponding to each of the A dialogues as the short-term memory information corresponding to the target subtask, that is, the short-term memory information used in the A+1th dialogue.

[0074] Long-term memory systems can store episodic memory, semantic memory, and embedding memory. Episodic memory refers to memories of specific scenarios or events, such as "a user consulted about a certain issue in January 2025." Specifically, it can be the input question and output result of a conversation, or the contextual information input and the output result of a subtask. Semantic memory refers to memories of concepts, rules, or knowledge, such as "the concept and definition of a certain term." Specifically, it can be further explanation of the conversation result and the subtask execution result. Embedding memory is a vector representation of episodic and semantic memory, used for vectorized retrieval and fast retrieval. Episodic, semantic, and embedding memory are all stored in the long-term memory system through a memory writing component after the agent's output (the conversation result of each round) is obtained.

[0075] In the memory system, the computer device can retrieve relevant memory information (including episodic memory, semantic memory, and vector memory) from the long-term memory system using the short-term memory information used in the (A+1)th dialogue. This retrieved information serves as the long-term memory information corresponding to the target sub-task, i.e., the long-term memory information used in the (A+1)th dialogue. This long-term memory information, along with the short-term memory information and knowledge fragments used in the (A+1)th dialogue, forms the initial context information for the (A+1)th dialogue. The agent then generates the execution result of the sub-task corresponding to the (A+1)th dialogue.

[0076] The agent system may also include a knowledge base, which stores candidate knowledge fragments for various task domains. The computer device can perform knowledge retrieval in the knowledge base based on the business problem and the target sub-task. For example, it can extract features from the business problem and the target sub-task separately to obtain a problem semantic feature vector and a sub-task semantic feature vector. The computer device can concatenate the problem semantic feature vector and the sub-task semantic feature vector to obtain a concatenated semantic feature vector. The similarity of this concatenated semantic feature vector with the candidate semantic feature vector corresponding to each candidate knowledge fragment is calculated to obtain the vector similarity for each candidate knowledge fragment. Candidate knowledge fragments with vector similarity greater than or equal to a similarity threshold are identified as the P knowledge fragments obtained through knowledge retrieval. The computer device can use the short-term memory information obtained from short-term memory, the long-term memory information obtained from long-term memory, and the P knowledge fragments obtained from the knowledge base to determine the initial context information corresponding to the target sub-task.

[0077] Computer devices can compress the initial context information to obtain the target context. Context compression methods can include attention methods such as sliding window attention and sparse attention, or compression methods such as hierarchical memory compression (e.g., key-value cache (KV Cache)) or layered memory compression (e.g., reflection tree, associative network). This application embodiment does not impose any limitations on these methods.

[0078] Computer devices can acquire object preference information (such as business problems, identity preferences, object historical behavior, etc.) and execution log information (such as available tools, execution status, system logs, etc.) for the execution status of task tools in the tool library. The object preference information and execution log information are then used to determine the system log information corresponding to the target subtask. Based on the target context information and system log information corresponding to the target subtask, the environment information corresponding to the target subtask is generated.

[0079] Computer devices can perform scene detection on environmental information and obtain scene detection results. If the target subtask is a subtask that has already been executed, and the execution result for the corresponding subtask has been obtained, the computer device can generate the task execution accuracy rate of the target subtask execution result through a planning agent, and generate the environmental information corresponding to the target subtask based on the task execution accuracy rate. The task execution accuracy rate can be included in the execution status information in the execution log. If the target subtask is a subtask that has already been executed, and the execution result for the corresponding subtask has not been obtained, the computer device can obtain the execution log information generated by the task tool corresponding to the target subtask during the execution of the task instructions corresponding to the target subtask, and generate the environmental information corresponding to the target subtask based on the execution log information.

[0080] Step S204: If the scene detection result meets the conditions for generating multiple prompt words, then generate object-side prompt words and model-side prompt words based on the scene detection result; the object-side prompt words are used to guide the object to perform the control operation corresponding to the scene detection result. For details, please refer to the following: Figure 6 , Figure 6 This is a schematic diagram of a data processing scenario provided in an embodiment of this application. Figure 3 .like Figure 6 As shown, computer devices can use dynamic detection components to determine whether scene detection results meet the conditions for generating multiple prompt words. Scene detection results that meet the conditions for generating multiple prompt words can include the following examples: 1. Content exceeding the limit result refers to the detection result when the amount of data of the content to be processed in the target context information detected in the environmental information is greater than or equal to the data amount threshold; 2. Restricted permission result refers to the detection result when the environment information detects that an object does not have the permission to process the content to be processed in the target context information; 3. Restricted search results refer to the detection results when the content to be processed in the target context information is located in a sub-partition of the data table in the environmental information, and the task instruction corresponding to the target sub-task does not include a partition identifier for the sub-partition; 4. Risk confirmation result refers to the detection result in the environmental information that the operational risk level of the task instruction corresponding to the target sub-task is greater than or equal to the risk level threshold. 5. Engine error result refers to the detection result when the database type of the task instruction corresponding to the target subtask is different from the database type of the content to be processed, as detected in the environment information.

[0081] 6. Execution failure result refers to the detection result when the system log information in the environment information indicates that the target subtask has been executed, and the target subtask failed to execute or the accuracy of the subtask execution result is too low.

[0082] If the scene detection result is as shown in the example above, the computer device can retrieve object-side and model-side cue elements that match the scene detection result from the cue word library. Both object-side and model-side cue elements are cue word templates that can be populated with key information. The computer device can generate first key information for the object-side cue element and second key information for the model-side cue element based on the target context information corresponding to the target subtask and the scene detection result. The first key information is filled into the object-side cue element to obtain the object-side cue word, and the second key information is filled into the model-side cue element to obtain the model-side cue word. The object-side cue word can be explicitly presented to the user, while the model-side cue word is implicitly injected into the model to add constraints.

[0083] Taking the content exceeding limit result as an example, the environmental information can include the target context information corresponding to the target subtask. If the amount of data of the content to be processed in the target context information is greater than or equal to the data volume threshold detected in the environmental information, the computer device can determine that the scene detection result corresponding to the target subtask is a content exceeding limit result. The content to be processed in the target context information can refer to SQL data, data tables, tool content (such as sample data input to the task tool, auxiliary prompts, tool pre-thinking content in the execution description information, and tool scheduling content, etc.). The data volume threshold can be a threshold set according to the context window of the task tool.

[0084] If the scene detection result is a content exceeding limit result, the computer device can determine that the scene detection result meets the conditions for generating multiple prompt words. Based on the content exceeding limit result, object-side prompt words and model-side prompt words are generated. The object-side prompt words guide the object to compress the amount of data to be processed, while the model-side prompt words instruct the task tool corresponding to the target subtask to generate the subtask execution result based on the compressed content of the content to be processed. For example, when there are too many SQL data rows, the object-side prompt word could be "Too many SQL data rows, please adjust the number of data rows," and the model-side prompt word could be "Please use compressed content to generate the subtask execution result corresponding to the target subtask." The compressed content can be manually adjusted content input by the object, or it can be obtained by the Agent system compressing the content to be processed; this embodiment does not impose any restrictions. When the data table information is too large, the object-side prompt word could be "The table information is too long, causing the model to fail to load; please manually reduce the amount of table information introduced," and the model-side prompt word could be "Important table information is missing; please summarize and provide feedback to the user based on the current information."

[0085] Optionally, when the number of SQL data rows is too large, the object-side prompt can be "Due to data volume threshold limitations, this step only returns 10,000 rows of data.", while the model-side prompt can be "Time filtering must be added, the number of data rows processed should not exceed 10,000, filtering is prohibited, and indexed columns should be used preferentially." This allows for the direct return of processing results based on quantitative data.

[0086] Taking a restricted access result as an example, if the environment information detects that an object lacks the processing permission for the content to be processed in the target context information, the computer device can determine that the scene detection result corresponding to the target subtask is a restricted access result. Processing permissions can include creation permission, editing permission, deletion permission, read permission, and write permission. If the scene detection result is a restricted access result, the computer device can determine that the scene detection result meets the conditions for generating multiple prompt words, and generate object-side prompt words and model-side prompt words based on the restricted access result. The object-side prompt word guides the object to apply for processing permission for the content to be processed in the target context information, while the model-side prompt word instructs the task tool corresponding to the target subtask to trigger the execution of the task instruction corresponding to the target subtask when it obtains the actual content corresponding to the content to be processed in the target context information. For example, if the object lacks read permission for data table 1, the object-side prompt word could be "Failed to obtain data table 1, do you want to apply for read permission for data table 1?", and the model-side prompt word could be "Note: Do not fabricate or create data."

[0087] Taking restricted search results as an example, if the content to be processed in the target context information is detected in the environmental information as being located in a sub-partition of the data table, and the task instruction corresponding to the target sub-task does not include a partition identifier for the sub-partition, then the computer device can determine that the scene detection result corresponding to the target sub-task is a restricted search result. The data table can include several sub-partitions, each storing different content. Sub-partitions require specifying the data table and partition identifier to be accessed. If the scene detection result is a restricted search result, the computer device can determine that the scene detection result meets the conditions for generating multiple prompt words. The computer device can generate object-side prompt words and model-side prompt words based on the restricted search result. Object-side prompt words are used to inform the object that the task instruction corresponding to the target sub-task does not include a partition identifier, while model-side prompt words are used to instruct the task tool corresponding to the target sub-task to trigger the execution of the task instruction including the partition identifier when it receives a task instruction including the partition identifier. For example, when a business problem or task instruction does not include a partition identifier, the object-side prompt word could be "Please specify the partition identifier," and the model-side prompt word could be "Search based on the partition identifier entered by the object."

[0088] Taking risk confirmation results as an example, computer equipment can calculate the operational risk level of the task instructions corresponding to the target sub-task. The operational risk level can be pre-set based on the type of task instruction. If the operational risk level of the task instruction corresponding to the target sub-task is detected in the environmental information to be greater than or equal to the risk level threshold, the computer equipment can determine that the scenario detection result corresponding to the target sub-task is a risk confirmation result. For example, task instructions for designing permissions such as create, edit, delete, and write permissions can be task instructions with an operational risk level greater than or equal to the risk level threshold. If the scenario detection result is a risk confirmation result, the computer equipment can determine that the scenario detection result meets the conditions for generating multiple prompt words, and generate object-side prompt words and model-side prompt words based on the risk confirmation result. The object-side prompt words are used to guide the object to confirm the operational risk level of the task instructions corresponding to the target sub-task, while the model-side prompt words are used to instruct the task tool corresponding to the target sub-task to trigger the execution of the task instructions corresponding to the target sub-task when it obtains the object's confirmation result regarding the operational risk level. For example, when a task instruction includes deleting a data table, the object-side prompt could be "This involves a data table deletion operation, please confirm again," while the model-side prompt could be "Execute the task instruction corresponding to the target subtask based on the object's confirmation result."

[0089] Taking engine error results as an example, the database can include multiple data engines, each using different syntax for task instructions. Therefore, if the syntax type of the task instruction corresponding to the target subtask is detected in the environment information to be different from the database type of the content to be processed, the computer device can determine that the scene detection result corresponding to the target subtask is an engine error result. If the scene detection result is an engine error result, the computer device can determine that the scene detection result meets the conditions for generating multiple prompt words. The computer device can generate object-side prompt words and model-side prompt words based on the engine error result. Among them, the object-side prompt words are used to guide the object to confirm the syntax type of the task instruction corresponding to the target subtask, and the model-side prompt words are used to instruct the task tool corresponding to the target subtask to trigger the execution of the task instruction corresponding to the target subtask when it obtains the object's confirmation result for the operation risk level.

[0090] Taking execution failure as an example, if the environmental information detects that the target subtask is a previously executed subtask, and no execution result for the corresponding subtask is obtained, the computer device can determine that the scene detection result for the target subtask is an execution failure. Alternatively, if the environmental information detects that the task execution accuracy is less than the accuracy threshold, the computer device can also determine that the scene detection result for the target subtask is an execution failure. If the scene detection result is an execution failure, the computer device can determine that the scene detection result meets the conditions for generating multiple prompt words. The computer device can generate object-side prompt words and model-side prompt words based on the execution failure result. Object-side prompt words guide the object to modify the task instructions for the target subtask; model-side prompt words instruct the planning agent to generate new task tools and new task instructions corresponding to the target subtask. For example, the object-side prompt word could be "SQL query failed, the reason is: (brief error information categorized and organized according to error information), please adjust the task instructions according to the reason." The model-side prompt word could be "Please generate the subtask execution result according to the new task tools and new task instructions obtained by the object."

[0091] Optionally, when the scenario detection result is an execution failure, the Agent system can also automatically repair known issues. In this case, the object-side prompt can be "SQL query failed, the reason is: (brief error information categorized according to the error information), intelligent repair is being performed for you", and the model-side prompt can be "Last error: [specific error information], known issues must be repaired before retrying".

[0092] Optionally, after displaying the prompt "SQL query failed, reason: (brief error message categorized according to error information), please adjust the task instructions according to the reason" to the object, if no modification result of the object's task instructions for the target subtask is obtained, the model-side prompt is used to instruct the planning agent to modify the target subtask to obtain the new task tools and new task instructions corresponding to the target subtask; if the modification result of the object's task instructions for the target subtask is obtained, the model-side prompt is used to instruct the planning agent to obtain the new task tools and new task instructions corresponding to the target subtask from the modification result. The new task tools and new task instructions are used to generate the subtask execution result corresponding to the target subtask. Step S205: Respond to the control operation guided by the object-side prompt, obtain the object interaction result, and send the object interaction result and the model-side prompt to the task tool corresponding to the target subtask. In the task tool corresponding to the target subtask, execute the task instructions corresponding to the target subtask according to the target context information, object interaction result, and model-side prompt, and obtain the subtask execution result. The model-side prompt is used to indicate the constraints of the task tool corresponding to the target subtask during the execution of the task instructions corresponding to the target subtask. Specifically, computer devices can respond to control operations guided by prompts from the object side and obtain the object interaction results. Taking the scene detection result as a permission-restricted result as an example, the control operation is to confirm the request for read permission to the data table. When the object obtains read permission to the data table, the computer device can determine the read permission of the data table as the object interaction result corresponding to the control operation.

[0093] The computer device can send the object interaction results and model-side prompts to the task tool corresponding to the target subtask. In the task tool corresponding to the target subtask, the task instructions corresponding to the target subtask are executed according to the target context information, object interaction results and model-side prompts, and the subtask execution results are obtained.

[0094] Please see also Figure 6 ,like Figure 6 As shown, if the scene detection result does not meet the conditions for generating multiple prompt words, the computer device can also execute the task instructions corresponding to the target subtask in the task tool corresponding to the target subtask according to the target context information corresponding to the target subtask, and obtain the subtask execution result.

[0095] Step S206: When the execution results of the K subtasks are obtained, the execution results of the K subtasks are integrated to obtain the answer to the business question.

[0096] Specifically, after obtaining the execution results of the subtasks corresponding to the target subtask, the computer device can continue to execute the next subtask in the execution order after the target subtask in the plan list, until the execution results of the K subtasks are obtained. Then, the execution results of the K subtasks are integrated and processed to obtain the answer to the business question.

[0097] Step S207: Within the statistical period, obtain the trigger counts corresponding to the W prompt word generation scenarios; a prompt word generation scenario refers to the scenario detection result that meets the multi-prompt word generation conditions; W is a positive integer; determine the prompt word generation scenarios with trigger counts greater than or equal to the optimization threshold as scenarios to be optimized, and obtain the execution log information of the scenarios to be optimized within the statistical period; the execution log information includes the object-side prompt words, model-side prompt words, and subtask execution results generated each time the multi-prompt word generation conditions are met within the statistical period; optimize and update the object-side prompt elements and model-side prompt elements corresponding to the scenarios to be optimized based on the execution log information.

[0098] Specifically, the computer device can obtain the trigger counts for W prompt word generation scenarios within each statistical period. A prompt word generation scenario refers to the scenario detection results that meet the multi-prompt word generation conditions. That is, the computer device can count the trigger counts for content exceeding limits, permission restrictions, retrieval restrictions, risk confirmation, engine errors, and execution failures within the statistical period. The computer device can identify prompt word generation scenarios with trigger counts greater than or equal to the optimization threshold as scenarios to be optimized and obtain the execution log information of the scenarios to be optimized within the statistical period. The execution log information includes the object-side prompt words, model-side prompt words, and subtask execution results generated each time the multi-prompt word generation conditions are met within the statistical period. Based on the execution log information, the computer device can optimize and update the object-side prompt elements and model-side prompt elements corresponding to the scenarios to be optimized, forming a closed-loop system for continuous learning and optimization.

[0099] This application embodiment processes business problems through task planning, resulting in a plan list comprising K sub-tasks. Complex business problems can be broken down into K sub-tasks of appropriate granularity. Each sub-task includes specific task instructions and corresponding task tools. Through multi-task tool collaboration and multi-sub-task step-by-step reasoning, the solution process for complex business problems is decomposed into a progressive execution chain, reducing the processing pressure on individual task tools. Simultaneously, by leveraging the complementary functional characteristics of different task tools, the limitations of a single task tool in complex scenarios are avoided, improving the accuracy of sub-task execution. During the execution of the K sub-tasks, a target sub-task can be identified, and its corresponding environmental information can be obtained. Scene detection is performed on the environmental information to obtain scene detection results. If the scene detection results meet the conditions for multi-prompt generation, object-side prompts and model-side prompts can be generated based on the scene detection results. Object-side prompts explicitly guide the object to execute the control operations corresponding to the scene detection results, while model-side prompts implicitly inject constraints during the execution of the task instructions corresponding to the target sub-task by the task tool corresponding to the target sub-task. By distinguishing between object-side and model-side prompts, we can differentiate between the guidance needed by the user and the instructions required by the model. This allows us to optimize model constraints and the user's interaction environment, helping users express their needs more effectively and understand the model's capability boundaries. Responding to control operations guided by object-side prompts and obtaining object interaction results (such as supplementary user information, confirmed parameter options, and feedback on execution preferences) improves the coherence of the user conversation. Sending the object interaction results and model-side prompts to the task tool corresponding to the target subtask allows for targeted addition of constraints such as execution constraints, data specifications, or behavioral restrictions during the execution of the task instructions based on the target context information, object interaction results, and model-side prompts. This results in more accurate subtask execution results and ultimately improves the accuracy of the answer obtained by integrating the results of K subtasks. The dual-prompt word collaboration mechanism ensures that both the model and the user receive the most suitable guidance information. For users, it optimizes the retry process, reduces cognitive load, and allows for efficient task execution without requiring specialized technical skills, thus lowering the technical barrier. For the model, the injection of precise constraints improves execution efficiency and standardization, avoids invalid calculations, and saves computing resources. The synergistic effect of both effectively improves the interaction efficiency between users and the model, thereby increasing the completion rate of answers to complex business questions.

[0100] Meanwhile, by optimizing model inference and user interaction through a dual-prompt word architecture, dialogue breaks and misunderstandings can be reduced, which is beneficial for user understanding and the smooth flow of the dialogue system, thus improving dialogue coherence. User-side prompt words are dynamically generated based on real-time user context, enabling the system to adapt to different users' expression habits and knowledge backgrounds, providing a truly personalized interactive experience and improving the personalized interaction experience. By statistically analyzing the trigger frequency of detection results in various scenarios and optimizing and updating the prompt word library, effective execution cases and error correction information from actual business scenarios can be accumulated into execution log information. Optimizing and updating prompt elements can help the model gradually adapt to the personalized needs of specific scenarios, correct inference biases, and make the model's inference logic more in line with actual business rules, further improving the accuracy and reliability of the answer results.

[0101] Please see Figure 7 , Figure 7 This is a flowchart illustrating a data processing method provided in an embodiment of this application. Figure 3 This data processing method can be executed by a computer device, which can be, for example, Figure 1 The terminal device 10a and the business server 100 shown are integrated into one device. The following description will use the example of this data processing method being executed by a computer device. This data processing method may include at least the following steps S301-S318: Step S301, input the business problem; Step S302: Obtain the context information corresponding to the business problem; Step S303: Obtain memory information; Step S304: Obtain knowledge fragments; Specifically, the computer device can acquire the business question input by the user in the application client and distribute it to the planning agent. The planning agent can obtain the context information corresponding to the business question, which may include memory information obtained from the memory system and knowledge fragments obtained from the knowledge base. The method of obtaining the context information corresponding to the business question can be the same as obtaining the aforementioned... Figure 5 The method of obtaining the initial context information in step S203 of the corresponding embodiment is the same, and will not be repeated here in the embodiments of this application.

[0102] Step S305: Generate a plan list; Step S306: Issue the plan list; Specifically, the planning agent can decompose the business problem into tasks based on the context information corresponding to the business problem, obtaining the task instructions corresponding to K sub-tasks of the business problem. The task decomposition process can be to divide a complex task into smaller tasks to solve the complex problem. Task decomposition can be implemented through thinking chains, hierarchical task networks, step-by-step prompting methods, etc., and this application embodiment does not impose any limitations.

[0103] The planning agent can obtain E candidate tools that are idle in the tool library. These E candidate tools can be AI models for different task domains, i.e., the task domain types corresponding to the E candidate tools. Based on the task instructions for each of the K sub-tasks and the task domain types corresponding to the E candidate tools, the planning agent determines the corresponding task tools for each of the K sub-tasks from the E candidate tools. Once the task tools for each of the K sub-tasks are obtained, the task instructions and task tools for each of the K sub-tasks are compiled into a plan list. Each sub-task includes different task instructions and different task tools. The planning agent can then distribute the plan list to the execution agent.

[0104] The prompts used to instruct the planning agent to generate a plan list could be as follows: "You are a planning expert, and your underlying model is an intelligent assistant system. You excel at analyzing problems and creating plans for complex issues. Task tools are merely for reference; do not use scheduling tools. Just create plans. Output requirements: Plan list: Plans or solutions devised for the current problem, breaking down complex problems into multiple sub-tasks (controlling the total number of steps to no more than 5 and the fewer the better; the total text content to no more than 500 words; prompting for diagrams whenever possible; utilizing knowledge base content for planning); the prompts indicate that analyses of different dimensions must be performed separately." Alternatively, it could be as follows: "You are a professional project planning assistant, skilled in decomposing tasks and creating action steps. Note that the task tool is only for your reference; do not use the scheduling tool. Task: Business problem. Please create an action step to complete this task based on the system's capabilities, including a description of each step and its dependencies. Note that for very simple dialogue questions, simply output one action step. The output format should be a JSON array, with each element containing a step description and a list of indices of dependent steps. Output requirements: For the current task, decompose complex problems into multiple sub-problems that require the use of other capabilities to solve (only provide the problems, not the answers). Note that analyses from different dimensions must be performed separately. Ensure that the step logic is reasonable, the dependencies are correct, and the number of steps is between 1 and 4. For simple problems, only one sub-step is needed. Note that only JSON format must be output." This application embodiment does not limit the content of the prompts used to instruct the planning agent to perform task planning.

[0105] Optionally, in data analysis scenarios, the toolkit includes the following task tools: 1. Natural Language to SQL Tool: Leveraging the semantic parsing capabilities of large models and agent support, it converts user queries (such as "find the highest-selling products in the last 7 days") into compliant SQL query statements. The tool automatically validates syntax and table lineage to ensure the generated SQL is securely executable. 2. SQL Execution Tool: Connects to data warehouses (such as ClickHouse, MySQL, Presto, etc.), executes the generated SQL, and returns structured results. It supports result caching, pagination, and timeout control to avoid putting pressure on the production environment. 3. Python Code Execution Tool: Provides a sandbox environment, allowing users to run custom Python code for data cleaning, statistical analysis, or complex calculations. For example, users can use Pandas for data aggregation or Scikit-learn for machine learning predictions. 4. Visualization Chart Generation Tool: Integrates libraries such as Matplotlib and Plotly to automatically generate visual charts such as line charts and bar charts based on data. The tool supports customized chart styles and can be embedded in reports or dashboards. 5. Search Tool: A tool built using a search engine to retrieve external information to assist in generating data analysis conclusions. 6. Chart Relationship Query Component: Enter the URL of a report or chart, and the tool automatically parses the underlying data tables it depends on.

[0106] Optionally, in data asset management scenarios, the toolkit includes the following task tools: 1. Data Table Asset Assessment: This tool comprehensively measures the quality of data tables through a quantitative score (0-100 points). Assessment dimensions include metadata completeness, data freshness, access frequency, security compliance, and storage cost. Low-scoring tables trigger improvement suggestions, such as adding field comments or optimizing storage strategies. 2. Output Task Detection Tool: This tool lists the dependent tasks of the data table (such as Spark jobs and scheduled tasks), including task status, runtime, and the last successful time. If a task fails, a notification message is displayed. 3. Partition Information Management Tool: This tool displays the table's partition structure, partition size, number of records, and lifecycle status. It supports filtering data by partition or cleaning up historical partitions to free up storage space. 4. Storage Cost Analysis Tool: Based on data volume, access patterns, and storage type, this tool calculates monthly storage costs and provides optimization suggestions (such as migrating cold data to low-cost storage). 5. Query Popularity Analysis Tool: This tool analyzes the frequency and type of queries performed on data tables, identifying popular and idle tables, providing a basis for resource allocation and architecture optimization.

[0107] It is understood that the above-mentioned task tools can be added or removed at any time according to the characteristics of the MCP protocol, so as to be applied to general neighborhoods or vertical fields. This application embodiment does not impose any restrictions here.

[0108] Step S307: Determine the target sub-task; For details, please refer to the above. Figure 5 The specific content of step S202 in the corresponding embodiment will not be repeated here.

[0109] Step S308: Perform task processing on the target subtask; Step S309: Return the execution result of the subtask corresponding to the target subtask to the executing agent; Specifically, the computer device can obtain the environmental information corresponding to the target subtask. The environmental information corresponding to the target subtask may include the target context information and system log information corresponding to the target subtask. The executing agent can obtain the initial context information of the target subtask. The initial context information may include memory information obtained from the memory bank and knowledge fragments obtained from the knowledge base.

[0110] Computer devices can compress the initial context information to obtain the target context. The compression methods described below can be attention methods such as sliding window attention and sparse attention, or compression methods such as hierarchical memory compression (e.g., KVCache) or layered memory compression (e.g., reflection tree, associative network), etc. The embodiments of this application are not limited here.

[0111] System log information can include logs generated by the agent system during the execution of subtasks, as well as object preference information and historical behavior information for that object. Computer devices can perform scene detection on environmental information and obtain scene detection results. Scene detection can be used to determine the scene type of the environmental information. For example, if the environmental information includes environmental signals or control signals specific to a particular scene, the computer device can identify the corresponding characteristic scene as the scene detection result.

[0112] The computer device can determine whether the scene detection results meet the conditions for generating multiple prompt words. If the scene detection results meet the conditions, the computer device can generate object-side prompt words and model-side prompt words based on the scene detection results. The methods for generating object-side and model-side prompt words can be found above. Figure 5 The specific description of step S204 in the corresponding embodiment will not be repeated here.

[0113] The computer device can send the object interaction results and model-side prompts to the task tool corresponding to the target subtask. In the task tool corresponding to the target subtask, the task instructions corresponding to the target subtask are executed according to the target context information, object interaction results and model-side prompts, and the subtask execution results are obtained.

[0114] If the scene detection result does not meet the conditions for generating multiple prompt words, the computer device can also execute the task instructions corresponding to the target subtask in the task tool corresponding to the target subtask according to the target context information corresponding to the target subtask, and obtain the subtask execution result.

[0115] Step S310: If the communication mode corresponding to the business problem is the rapid analysis mode, then update the working memory information; Specifically, the executing agent can determine the communication mode for a business problem based on its complexity. The agent can perform keyword matching on the business problem using a keyword database to obtain guiding keywords indicating the problem type. For example, guiding keywords can be words representing different thinking needs, such as "why," "what," "analysis," and "comparison." The executing agent can then perform a weighted summation of the text length of the business problem, guiding keywords, and the number of subtasks in the plan list to obtain a comprehensive score for the business problem. If the comprehensive score is greater than or equal to the deep analysis threshold, the executing agent can determine that the communication mode corresponding to the business problem is the deep analysis mode; if the comprehensive score is less than the deep analysis threshold, the computer can determine that the communication mode corresponding to the business problem is the rapid analysis mode.

[0116] If the communication mode corresponding to the business problem is a rapid analysis mode, the executing agent can generate context information for the next subtask of the target subtask based on the execution result of the subtask corresponding to the target subtask. In the task tool corresponding to the next subtask, the task instructions corresponding to the next subtask are executed again using the context information obtained from the next subtask, thus obtaining the execution result of the next subtask. The computer device can store the execution result of the subtask corresponding to the target subtask in the context buffer of the short-term memory system, thereby updating the working memory information.

[0117] Step S311: If the communication mode corresponding to the business problem is the deep analysis mode, then return the execution result of the subtask corresponding to the target subtask to the planning agent. Step S312: Determine whether the conditions for plan modification are met; Specifically, if the communication mode corresponding to the business problem is the deep analysis mode, the execution agent can send the execution results of the subtasks corresponding to the target subtasks to the planning agent. The planning agent can determine whether the target execution results meet the plan modification conditions. The planning agent can generate the task execution accuracy of the target subtask execution results by verifying the prompt words. If the task execution accuracy is less than the first accuracy threshold, the planning agent can determine that the target execution results meet the plan modification conditions.

[0118] Understandably, the validation prompt could be something like this: "You are a task completion and accuracy checker. Please determine whether the task has been perfectly completed based on the issues and completion feedback. If perfectly completed and correct, enter 'Yes'; otherwise, enter 'No' and include suggested modifications. Return format: {"Yes or No", "Modification Suggestion": "Step Status"}, please only return JSON format."

[0119] Step S313: If the execution result of the subtask corresponding to the target subtask meets the plan modification conditions, then the plan list is modified. Specifically, if the execution result of the subtask corresponding to the target subtask meets the conditions for plan modification, the planning agent can modify the plan list. That is, it can decide whether to modify the task instructions and task tools of the target subtask or re-plan the business problem based on the task execution accuracy.

[0120] If the task execution accuracy is less than the second accuracy threshold, the planning agent can re-plan the task based on the plan list and modification prompts to obtain a first updated plan list. The new subtasks in the first updated plan list are distinct from the subtasks in the plan list, and the first accuracy threshold is greater than the second accuracy threshold. If the task execution accuracy is greater than or equal to the second accuracy threshold, but less than the first accuracy threshold, the computer can designate the subtasks in the plan list that precede the target subtask in the execution order as preceding subtasks. It then modifies the task instructions and execution tools in the target subtask and preceding subtasks in the plan list to obtain a second updated plan list. The second updated plan list includes subtasks in the plan list that follow the target subtask in the execution order.

[0121] The modification prompt can be understood as follows: "You are an intelligent plan modification assistant. Please make reasonable modifications to the existing plan based on user feedback or new requirements. Important constraints: You can only modify and adjust the unexecuted part of the plan; executed steps must remain unchanged! Current plan information: Executed steps (cannot be modified): Unexecuted steps (can be modified): User feedback or new requirements: Please modify the plan based on the feedback. You can: 1. Modify the description of unexecuted steps; 2. Add new steps (try not to add unnecessary steps, including previously completed steps, the total number of steps should not exceed 5); 3. Delete unnecessary unexecuted steps; 4. Adjust the dependencies of unexecuted steps; 5. Reorder the unexecuted steps. Notes: Added or modified steps cannot affect the integrity of executed steps; consider the constraints of executed steps when adjusting dependencies; please only return the modified unexecuted steps part, in JSON array format."

[0122] Step S314: Issue the update plan list; Step S315: Process the new subtasks in the update plan list; Specifically, after obtaining the update plan list (either the first or second update plan), the planning agent can distribute the update plan list to the execution agent. The execution agent can then continue to execute new subtasks in the update plan list. For example, to obtain the first update plan list by replanning a business problem, the execution agent can process all subtasks in the first update plan sequentially. Similarly, to obtain the second update plan list by modifying the target subtasks in the plan list, the execution agent can reprocess the new subtasks corresponding to the target subtasks in the second update plan list.

[0123] Step S316: Return the execution results of K subtasks; Step S317: Integrate the execution results of the K subtasks; Step S318: Return the answer result.

[0124] Specifically, when the computer device obtains or updates the execution results of all subtasks in the plan list (taking the execution results of K subtasks as an example), it can integrate and process the execution results of each subtask through an intelligent agent to obtain the answer to the business question. The answer to the business question is then returned to the user through the application client.

[0125] Optionally, the Agent system may also include a side-branch verification agent. This agent can reuse the architecture of the planning agent and the execution agent. The computer device can respond to side-branch verification operations on the answer results, inputting the business question and the answer result into the side-branch verification agent. Within the agent, problem diagnosis is performed on the business question and answer result using side-branch verification prompts, resulting in a diagnostic result for the answer result. This diagnostic result includes unit diagnostic information corresponding to the execution results of each of the U sub-tasks in the answer result. Each unit diagnostic information indicates abnormal execution information and key correction points for a sub-task. For example, a side-branch verification prompt could be: "Please reflect on whether the execution results of each sub-task in the answer result for the business question are correct. If incorrect, please point out the incorrect content and the correction plan." The computer device can respond to correction operations on the answer result, correcting the answer result based on the unit diagnostic information corresponding to the execution results of the U sub-tasks, resulting in a corrected answer result.

[0126] The side-support verification agent is used to assist in judging and correcting the content of the conversation. If the answer to the business question is incorrect, the side-support verification agent can correct the answer to the business question after the user triggers (i.e. responds to the side-support verification operation for the answer) or automatically triggers it, and returns the corrected answer to the user, thereby further improving the accuracy of the answer.

[0127] It is understood that this application proposes two communication modes: a rapid analysis mode and a deep analysis mode. Rapid analysis mode: The Mentor first decomposes the plan and then passes it to the Operator for execution iteration. After the Mentor generates the plan, it is directly given to the Operator module for operation, reducing unnecessary interaction between the Mentor and Operator—adapting to the need for quick answers to simple questions, reducing unnecessary interactions, achieving rapid response, and meeting users' immediate needs for answers to simple questions. Deep analysis: Decomposition and execution are interleaved + dynamic changes to sub-plans. After the Mentor generates a plan that can be decomposed into sub-tasks, the first sub-task without data dependencies is given to the Operator for execution. After the sub-task is completed, the execution result is sent back to the Mentor. The Mentor then verifies whether to modify the plan. If so, it submits the updated plan list to the Operator; otherwise, it submits the next new target sub-task without data dependencies to the Operator for execution. This supports plan adjustment, task verification, and user intervention, ensuring the accuracy and controllability of the analysis process, and ultimately generating a highly reliable deep analysis report. The rapid analysis mode uses shallow agent communication to reduce computational overhead and improve response speed and efficiency for simple tasks by minimizing interactions between agents. The deep communication mode, on the other hand, employs multiple interactions to ensure the effectiveness of information communicated by each agent. Both modes support rapid responses to simple problems and slower but more accurate and controllable responses to difficult problems. They can also be strongly matched to business problems, allowing users to choose the communication mode and intervene in the mentor process for control. Through multiple rounds of verification and dynamic plan adjustments, the mode avoids the accumulation of errors caused by blind execution, significantly improving the accuracy and reliability of complex tasks.

[0128] Furthermore, the embodiments of this application utilize a multi-model collaboration mechanism to fully leverage the strengths of different models (task tools). This design achieves efficient synergy between model capabilities and the intelligent agent framework, ultimately completing the leap from "passive data querying" to "proactive planning, execution, and iterative decision-making insights," forming an end-to-end intelligent analysis closed loop.

[0129] Please see Figure 8 , Figure 8 This is a flowchart illustrating a data processing method provided in an embodiment of this application. Figure 4This data processing method can be executed by a computer device, which can be, for example, Figure 1 The terminal device 10a shown is an integrated device of the second server 200. The following description will use the example of this data processing method being executed by a computer device. This data processing method may include at least the following steps S401-S411: Step S401: Obtain the business issue; Specifically, a smart assistant system (Agent system) can be deployed in the computer device. The computer device can obtain the business question input by the user in the application client, break down the business question into tasks, and obtain the task instructions corresponding to K sub-tasks for the business question. The task decomposition process can be to divide a complex task into smaller tasks, thereby solving the complex problem.

[0130] The computer device can acquire E candidate tools that are idle in the tool library. These E candidate tools can be AI models for different task domains, i.e., the task domain types corresponding to the E candidate tools. Based on the task instructions corresponding to K sub-tasks and the task domain types corresponding to the E candidate tools, the computer device can determine the task tools corresponding to the K sub-tasks from the E candidate tools. Once the task tools corresponding to the K sub-tasks are acquired, the task instructions and task tools corresponding to the K sub-tasks are compiled into a plan list. Each sub-task includes different task instructions and different task tools.

[0131] Step S402: Obtain memory information; Step S403: Obtain knowledge fragments; Specifically, the computer device can send a list of plans corresponding to business problems to the context engineering component, which is used to create textual information that the model can easily understand. For ease of understanding, let's take the target subtask among K subtasks as an example. The computer device can determine the target subtask from the K subtasks; the method for determining the target subtask can be found above. Figure 3 The specific description of step S102 in the corresponding embodiment will not be repeated here.

[0132] The context engineering component can obtain the initial context information of the target subtask. It can also initiate the collection component to execute three data collection tasks in parallel: 1) Historical dialogue query: obtaining the context information (i.e., memory information) of the current session; 2) Knowledge base retrieval: searching for relevant knowledge fragments; 3) Object feature acquisition: obtaining object preference information and object historical behavior information. Executing these three data collection tasks in parallel can significantly reduce input / output latency and improve response speed. The initial context information can include memory information obtained from the memory bank and knowledge fragments obtained from the knowledge base. The content of the memory information and knowledge fragments can be found above. Figure 5 The specific description of step S203 in the corresponding embodiment will not be repeated here. After the collection component is completed, the context engineer can perform format conversion on the initial context information to unify the data format for easier subsequent processing.

[0133] Step S404: Send initial context information; Step S405: Supplement knowledge fragments; Specifically, the computer device can send the target subtask and its initial context information to the model invocation component. This component, which can be a large language model invocation layer, acts as a control and decision-making center, responsible for understanding and generating the session content. The model invocation component can then use the initial context information to retrieve knowledge from the knowledge base, further supplementing the knowledge fragments in the initial context information.

[0134] Step S406: Invoke the task tool; Step S407: Return the subtask execution result; Specifically, the model invocation component can invoke the task tool corresponding to the target subtask through the tool management component to generate the subtask execution result. The tool management component is responsible for tool registration and scheduling. The method for generating the subtask execution result for the target subtask can be found above. Figure 7 The descriptions of steps S308 and S309 in the corresponding embodiments will not be repeated here.

[0135] Step S408: Determine whether the plan list has been completed; Step S409: If the plan list is not completed, execute the next subtask in the plan list; Specifically, the model invocation component can obtain the plan list, determine whether the plan list has been completed, and if the plan list has not been completed, the model invocation component can execute the next subtask in the plan list whose execution order is next to the target subtask.

[0136] Step S410: If the plan list is completed, then integrate the execution results of the K subtasks. Step S411: Update memory information.

[0137] Specifically, if the plan list is completed, the model invocation component can integrate the execution results of the K subtasks obtained from the execution plan list to obtain the answer to the business question. After obtaining the answer to the business question, the intelligent assistant system can generate a memory summary information for this session through the efficient summarization component and update the memory information in the memory system. The memory system can serve as a "working memory area" or "shared blackboard" connecting the planning agent and the execution agent, storing task states, historical records, and intermediate results to ensure collaborative work within a unified cognitive context. The efficient summarization component can perform context compression. Context compression methods can include attention methods such as sliding window attention and sparse attention, or compression methods such as hierarchical memory compression (e.g., KV Cache) or layered memory compression (e.g., reflection trees, associative networks). This embodiment of the application does not impose limitations on these methods.

[0138] The memory system proposed in this application can include the interaction content of the current session or dialogue and working memory information in short-term memory information, and can persistently store memory summary information in historical sessions in long-term memory information. By compressing and injecting memory information into the initial context information, information loss caused by brute-force truncation can be avoided, the model context can be prevented from exceeding the threshold, and the core information can be effectively balanced to ensure the coherence and logical consistency of multi-turn dialogues.

[0139] Please see Figure 9 , Figure 9 This is a flowchart illustrating a data processing method provided in an embodiment of this application. Figure 5 This data processing method can be executed by a computer device, which can be, for example, Figure 1 The terminal device 10a shown is an integrated device of the second server 200. The following description will use the example of this data processing method being executed by a computer device. This data processing method may include at least the following steps S501-S517: Step S501, input the business problem; Step S502, forwarding service issues; For details, please refer to the above. Figure 5 The specific description of step S201 in the corresponding embodiment will not be repeated here.

[0140] Step S503: Collect environmental information; Step S504: Return environmental information; Specifically, the context engineering component can obtain target context information and system log information to form environment information. The target context information corresponding to the target subtask can include knowledge fragments retrieved from the knowledge base and memory information obtained from the memory system (e.g., business issues and plan lists). System log information can include log information generated by the Agent system during the execution of the subtask, as well as object preference information and object historical behavior information for that object. For more information on environment information, please refer to the above. Figure 5 The specific description of step S203 in the corresponding embodiment will not be repeated here.

[0141] Step S505, Scene detection; Step S506: Forward the scene detection results; Step S507: Determine whether the conditions for generating multiple prompt words are met; Specifically, the dynamic prompt word component can perform scene detection on environmental information, obtain scene detection results, and send these results to the trigger condition detection component. The trigger condition detection component can determine whether the scene detection results meet the conditions for generating multiple prompt words. Scene detection results that meet the conditions for generating multiple prompt words can include the following examples: content exceeding limits, permission restrictions, search restrictions, risk confirmation, engine error, and execution failure. If the scene detection results are among the above, the computer device can determine that the scene detection results meet the conditions for generating multiple prompt words. For details on determining whether the scene detection results meet the conditions for generating multiple prompt words, please refer to the above. Figure 5 The specific description of step S204 in the corresponding embodiment will not be repeated here.

[0142] Step S508, return to the triggering scene; Step S509: Obtain the prompting element; Specifically, the trigger condition detection component can return trigger scenarios, i.e., scenario detection results that meet the conditions for generating multiple prompt words. The computer device can then retrieve object-side and model-side prompt elements that match the scenario detection results from the prompt word library. Both object-side and model-side prompt elements are prompt word templates that can be filled with key information.

[0143] Step S510: Generate object-side prompts and model-side prompts; Step S511: Display the prompt words on the target side; Specifically, the computer device can generate first key information for object-side cue elements and second key information for model-side cue elements based on the target context information and scene detection results corresponding to the target sub-task. The first key information is filled into the object-side cue element to obtain the object-side cue word, and the second key information is filled into the model-side cue element to obtain the model-side cue word. The object-side cue word can be explicitly presented to the user, while the model-side cue word is implicitly injected into the model to add constraints.

[0144] Step S512: Obtain the optimization results; Specifically, computer devices can respond to control operations guided by prompts from the object side and obtain the object interaction results. Taking the scene detection result as a permission-restricted result as an example, the control operation is to confirm the request for read permission to the data table. When the object obtains read permission to the data table, the computer device can determine the read permission of the data table as the object interaction result corresponding to the control operation.

[0145] Step S513: Send model-side prompts and optimization results; Step S514: Generate the subtask execution result; Specifically, the computer device can send the object interaction results and model-side prompts to the task tool corresponding to the target subtask. In the task tool corresponding to the target subtask, the task instructions corresponding to the target subtask are executed according to the target context information, object interaction results and model-side prompts, and the subtask execution results are obtained.

[0146] If the scene detection result does not meet the conditions for generating multiple prompt words, the computer device can also execute the task instructions corresponding to the target subtask in the task tool corresponding to the target subtask according to the target context information corresponding to the target subtask, and obtain the subtask execution result.

[0147] Step S515: Return the answer result; Specifically, after obtaining the execution results of the subtasks corresponding to the target subtask, the computer device can continue to execute the next subtask in the execution order after the target subtask in the plan list, until the execution results of the K subtasks are obtained. Then, the execution results of the K subtasks are integrated and processed to obtain the answer to the business question.

[0148] Step S516: Count the number of triggers; Step S517, optimization and update.

[0149] Specifically, the computer device can obtain the trigger counts for W prompt word generation scenarios within each statistical period. A prompt word generation scenario refers to the scenario detection results that meet the multi-prompt word generation conditions. That is, the computer device can count the trigger counts for content exceeding limits, permission restrictions, retrieval restrictions, risk confirmation, engine errors, and execution failures within the statistical period. The computer device can identify prompt word generation scenarios with trigger counts greater than or equal to the optimization threshold as scenarios to be optimized and obtain the execution log information of the scenarios to be optimized within the statistical period. The execution log information includes the object-side prompt words, model-side prompt words, and subtask execution results generated each time the multi-prompt word generation conditions are met within the statistical period. Based on the execution log information, the computer device can optimize and update the object-side prompt elements and model-side prompt elements corresponding to the scenarios to be optimized, forming a closed-loop system for continuous learning and optimization.

[0150] This application embodiment processes business problems through task planning, resulting in a plan list comprising K sub-tasks. Complex business problems can be broken down into K sub-tasks of appropriate granularity. Each sub-task includes specific task instructions and corresponding task tools. Through multi-task tool collaboration and multi-sub-task step-by-step reasoning, the solution process for complex business problems is decomposed into a progressive execution chain, reducing the processing pressure on individual task tools. Simultaneously, by leveraging the complementary functional characteristics of different task tools, the limitations of a single task tool in complex scenarios are avoided, improving the accuracy of sub-task execution. During the execution of the K sub-tasks, a target sub-task can be identified, and its corresponding environmental information can be obtained. Scene detection is performed on the environmental information to obtain scene detection results. If the scene detection results meet the conditions for multi-prompt generation, object-side prompts and model-side prompts can be generated based on the scene detection results. Object-side prompts explicitly guide the object to execute the control operations corresponding to the scene detection results, while model-side prompts implicitly inject constraints during the execution of the task instructions corresponding to the target sub-task by the task tool corresponding to the target sub-task. By distinguishing between object-side and model-side prompts, we can differentiate between the guidance needed by the user and the instructions required by the model. This allows us to optimize model constraints and the user's interaction environment, helping users express their needs more effectively and understand the model's capability boundaries. Responding to control operations guided by object-side prompts and obtaining object interaction results (such as supplementary user information, confirmed parameter options, and feedback on execution preferences) improves the coherence of the user conversation. Sending the object interaction results and model-side prompts to the task tool corresponding to the target subtask allows for targeted addition of constraints such as execution constraints, data specifications, or behavioral restrictions during the execution of the task instructions based on the target context information, object interaction results, and model-side prompts. This results in more accurate subtask execution results and ultimately improves the accuracy of the answer obtained by integrating the results of K subtasks. The dual-prompt word collaboration mechanism ensures that both the model and the user receive the most suitable guidance information. For users, it optimizes the retry process, reduces cognitive load, and allows for efficient task execution without requiring specialized technical skills, thus lowering the technical barrier. For the model, the injection of precise constraints improves execution efficiency and standardization, avoids invalid calculations, and saves computing resources. The synergistic effect of both effectively improves the interaction efficiency between users and the model, thereby increasing the completion rate of answers to complex business questions.

[0151] Meanwhile, by optimizing model inference and user interaction through a dual-prompt word architecture, dialogue breaks and misunderstandings can be reduced, which is beneficial for user understanding and the smooth flow of the dialogue system, thus improving dialogue coherence. User-side prompt words are dynamically generated based on real-time user context, enabling the system to adapt to different users' expression habits and knowledge backgrounds, providing a truly personalized interactive experience and improving the personalized interaction experience. By statistically analyzing the trigger frequency of detection results in various scenarios and optimizing and updating the prompt word library, effective execution cases and error correction information from actual business scenarios can be accumulated into execution log information. Optimizing and updating prompt elements can help the model gradually adapt to the personalized needs of specific scenarios, correct inference biases, and make the model's inference logic more in line with actual business rules, further improving the accuracy and reliability of the answer results.

[0152] Please see Figure 10 , Figure 10 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application. Figure 10 As shown, the data processing device 1 includes a task planning module 710, a scene detection module 720, a prompt word generation module 730, an interaction processing module 740, and a result integration module 750.

[0153] The task planning module 710 is used to acquire business problems, perform task planning on the business problems, and obtain a plan list. The plan list includes K sub-tasks, where K is a positive integer. Each sub-task includes different task instructions and different task tools. The scene detection module 720 is used to determine the target subtask among K subtasks, obtain the environmental information corresponding to the target subtask, perform scene detection on the environmental information, and obtain the scene detection result. The prompt word generation module 730 is used to generate object-side prompt words and model-side prompt words based on the scene detection results if the scene detection results meet the conditions for generating multiple prompt words; the object-side prompt words are used to guide the object to perform the control operations corresponding to the scene detection results. The interaction processing module 740 is used to respond to control operations guided by object-side prompts, obtain object interaction results, and send the object interaction results and model-side prompts to the task tool corresponding to the target subtask. In the task tool corresponding to the target subtask, the task instructions corresponding to the target subtask are executed according to the target context information, object interaction results, and model-side prompts to obtain the subtask execution results. The model-side prompts are used to indicate the constraints of the task tool corresponding to the target subtask during the execution of the task instructions corresponding to the target subtask. The result integration module 750 is used to integrate the execution results of the K subtasks when the execution results of the K subtasks are obtained respectively, and to obtain the answer result corresponding to the business question.

[0154] In one possible implementation, the business problem is the (H+1)th session initiated by the object, where H is a positive integer; the business problem is associated with the H sessions preceding the initiation of the business problem; when the scene detection module 720 obtains the environmental information corresponding to the target subtask, it is specifically used to perform the following operations: Obtain the session issues corresponding to each of the H round sessions, identify the subtasks in the plan list that are executed before the target subtask, and determine them as preceding subtasks. Based on the execution results of the preceding subtasks, the business issues, the plan list, and the session issues corresponding to each of the H round sessions, generate the target context information corresponding to the target subtask. Obtain object preference information and execution log information of the execution status of task tools in the tool library, and determine the object preference information and execution log information as the system log information corresponding to the target subtask; Based on the target context information and system log information corresponding to the target subtask, generate the environment information corresponding to the target subtask.

[0155] In one possible implementation, when the prompt word generation module 730 generates object-side prompt words and model-side prompt words based on scene detection results, it is specifically used to perform the following operations: Retrieve object-side and model-side cue elements that match the scene detection results from the cue word library; both object-side and model-side cue elements are cue word templates that can be filled with key information. Based on the target context information and scene detection results corresponding to the target sub-task, generate the first key information for the object-side cue element and the second key information for the model-side cue element; Fill the first key information into the object-side prompt element to obtain the object-side prompt word, and fill the second key information into the model-side prompt element to obtain the model-side prompt word.

[0156] In one possible implementation, the prompt word generation module 730 is also used to perform the following operations: Within the statistical period, obtain the trigger counts corresponding to W prompt word generation scenarios; prompt word generation scenario refers to the scenario detection results that meet the conditions for multiple prompt word generation; W is a positive integer; Scenarios where the number of triggers is greater than or equal to the optimization threshold are identified as scenarios to be optimized, and the execution log information of the scenarios to be optimized within the statistical period is obtained; the execution log information includes the object-side prompts, model-side prompts, and subtask execution results generated each time the multi-prompt-word generation conditions are met within the statistical period; Based on the execution log information, the object-side hint elements and model-side hint elements corresponding to the scenario to be optimized are optimized and updated.

[0157] In one possible implementation, the environmental information includes the target context information corresponding to the target subtask; the scene detection module 720 is used to perform scene detection on the environmental information, and when obtaining the scene detection result, it is specifically used to perform the following operations: If the amount of data of the content to be processed in the target context information is greater than or equal to the data amount threshold, then the scene detection result corresponding to the target subtask is determined to be the content exceeding the limit result. If the scene detection result is a content exceeding the limit, then the scene detection result is determined to meet the conditions for generating multiple prompt words; Based on scene detection results, object-side and model-side prompts are generated, including: Based on the content exceeding the limit result, object-side prompts and model-side prompts are generated. Object-side prompts are used to guide the object to compress the amount of data in the content to be processed, while model-side prompts are used to instruct the task tool corresponding to the target subtask to generate the subtask execution result corresponding to the target subtask based on the compressed content of the content to be processed.

[0158] In one possible implementation, the environmental information includes the target context information corresponding to the target subtask; the scene detection module 720 is used to perform scene detection on the environmental information, and when obtaining the scene detection result, it is specifically used to perform the following operations: If the object does not have the permission to process the content to be processed in the target context information, then the scene detection result corresponding to the target subtask is determined to be a permission-restricted result. If the scene detection result is a restricted access result, then the scene detection result is determined to meet the conditions for generating multiple prompt words; Based on scene detection results, object-side and model-side prompts are generated, including: Based on the permission restriction result, object-side prompts and model-side prompts are generated. Object-side prompts are used to guide objects to request processing permissions for the content to be processed in the target context information. Model-side prompts are used to instruct the task tool corresponding to the target subtask to trigger the execution of the task instruction corresponding to the target subtask when it obtains the actual content corresponding to the content to be processed in the target context information.

[0159] In one possible implementation, the environmental information includes the target context information corresponding to the target subtask; the scene detection module 720 is used to perform scene detection on the environmental information, and when obtaining the scene detection result, it is specifically used to perform the following operations: If the content to be processed in the target context information is located in a sub-partition of the data table, and the task instruction corresponding to the target sub-task does not include a partition identifier for the sub-partition, then the scene detection result corresponding to the target sub-task is determined to be a retrieval-restricted result. If the scene detection result is a search-restricted result, then the scene detection result is determined to meet the conditions for generating multiple prompt words; Based on scene detection results, object-side and model-side prompts are generated, including: Based on the limited search results, object-side prompts and model-side prompts are generated. Object-side prompts are used to indicate to the object that the task instruction corresponding to the target subtask does not include the partition identifier. Model-side prompts are used to instruct the task tool corresponding to the target subtask to trigger the execution of the task instruction that includes the partition identifier when it obtains the task instruction that includes the partition identifier.

[0160] In one possible implementation, the environmental information includes the target context information corresponding to the target subtask; the scene detection module 720 is used to perform scene detection on the environmental information, and when obtaining the scene detection result, it is specifically used to perform the following operations: If the operational risk level of the task instruction corresponding to the target subtask is greater than or equal to the risk level threshold, then the scenario detection result corresponding to the target subtask is determined as the risk confirmation result. If the scene detection result is a risk confirmation result, then the scene detection result is determined to meet the conditions for generating multiple prompt words; Based on scene detection results, object-side and model-side prompts are generated, including: Based on the risk confirmation results, object-side prompts and model-side prompts are generated. Object-side prompts are used to guide the object to confirm the operational risk level of the task instructions corresponding to the target sub-task, while model-side prompts are used to instruct the task tool corresponding to the target sub-task to trigger the execution of the task instructions corresponding to the target sub-task when it obtains the object's confirmation result on the operational risk level.

[0161] In one possible implementation, the scene detection module 720 is used to obtain environmental information corresponding to the target subtask, perform scene detection on the environmental information, and when obtaining the scene detection result, it is specifically used to perform the following operations: If the target subtask is a subtask that has already been executed, and no execution result has been obtained for the subtask corresponding to the target subtask, then obtain the execution log information generated by the task tool corresponding to the target subtask during the execution of the task instruction corresponding to the target subtask; Based on the execution log information, generate the environment information corresponding to the target subtask, and determine the scenario detection result corresponding to the target subtask as the execution failure result; If the scene detection result is an execution failure, then the scene detection result is determined to meet the conditions for generating multiple prompt words.

[0162] In one possible implementation, the K subtasks include subtasks. Kazuko's Mission , where i is a positive integer less than or equal to K; subtask The execution order of subtasks Prior to the execution order; when the scene detection module 720 is used to determine the target subtask among K subtasks, it is specifically used to perform the following operations: The subtasks that do not have data dependencies among the K subtasks are identified as the target subtasks; Or, if subtask With subtasks Data dependencies and obtaining subtasks The result of the subtask execution will then be the subtask This is identified as a target sub-task; Or through subtasks The corresponding task tool executes subtasks. If the corresponding task instruction fails, the subtask will be... This has been identified as a target sub-task.

[0163] In one possible implementation, the K subtasks are obtained by the planning agent through task planning; the scene detection module 720 is used to acquire environmental information corresponding to the target subtasks, perform scene detection on the environmental information, and specifically perform the following operations when obtaining the scene detection results: If the target subtask is a subtask that has already been executed, and the execution result of the subtask corresponding to the target subtask has been obtained, then the task execution accuracy of the target subtask execution result is generated by the planning agent, and the environmental information corresponding to the target subtask is generated based on the task execution accuracy. If the task execution accuracy in the environmental information is less than the accuracy threshold, then the scene detection result corresponding to the target subtask is determined to be an execution failure result. If the scene detection result is an execution failure, then the scene detection result is determined to meet the conditions for generating multiple prompt words.

[0164] In one possible implementation, the K sub-tasks are obtained by the planning agent through task planning; the prompt word generation module 730 is used to generate object-side prompt words and model-side prompt words based on scene detection results, specifically for performing the following operations: Based on the execution failure result, object-side prompts and model-side prompts are generated; object-side prompts are used to guide the object to modify the task instructions of the target subtask. If the modification result of the object's task instruction for the target subtask is not obtained, the model-side prompt is used to instruct the planning agent to modify the target subtask to obtain the new task tool and new task instruction corresponding to the target subtask. If the modification result of the object's task instruction for the target subtask is obtained, the model-side prompt is used to instruct the planning agent to obtain the new task tool and new task instruction corresponding to the target subtask from the modification result. The new task tool and new task instruction are used to generate the subtask execution result corresponding to the target subtask.

[0165] In one possible implementation, the task planning module 710 is used to perform task planning processing on business problems. When a plan list is obtained, it is specifically used to perform the following operations: By planning an intelligent agent, the business problem is broken down into K sub-tasks, each corresponding to a task instruction; the K sub-tasks include sub-tasks. , where i is a positive integer less than or equal to K; Obtain the task domain type corresponding to each of the E candidate tools in the tool library, based on subtasks. The corresponding task instructions and the task domain types corresponding to the E candidate tools are used to determine the subtasks from the E candidate tools. The corresponding task tool; E is a positive integer; Once the task tools corresponding to the K sub-tasks are obtained, the task instructions and task tools corresponding to the K sub-tasks are determined as a plan list.

[0166] In one possible implementation, the K subtasks are obtained by the planning agent through task planning. The interaction processing module 740 is used to execute the task instructions corresponding to the target subtask in the task tool corresponding to the target subtask, based on the target context information, object interaction results, and model-side prompts. When obtaining the subtask execution result, it is specifically used to perform the following operations: By executing the intelligent agent to call the task tool corresponding to the target subtask, the task tool corresponding to the target subtask executes the task instruction corresponding to the target subtask based on the target context information, object interaction results and model-side prompts, and obtains the subtask execution result corresponding to the target subtask. When the execution results of the K subtasks are obtained, the execution results of the K subtasks are integrated and processed to obtain the answer to the business question, including: If the communication mode corresponding to the business problem is the rapid analysis mode, then the context information of the next subtask of the target subtask is generated based on the execution result of the subtask corresponding to the target subtask. In the task tool corresponding to the next subtask, the task instructions corresponding to the next subtask are executed based on the context information obtained from the next subtask. When the execution results of the K subtasks are obtained, the planning agent integrates and processes the execution results of each subtask to obtain the answer to the business problem. If the communication mode corresponding to the business problem is the deep analysis mode, and the target execution result meets the plan modification conditions, then the plan list is modified to obtain an updated plan list. The new subtasks in the updated plan list are then executed until the execution results of each new subtask do not meet the plan modification conditions. At this point, the planning agent integrates the execution results of each subtask to obtain the answer to the business problem.

[0167] In one possible implementation, the communication mode corresponding to the business problem is a deep analysis mode, and the interaction processing module 740 is also used to perform the following operations: The task execution accuracy is calculated by planning the intelligent agent and verifying the prompt words to generate the target sub-task execution results. If the task execution accuracy is less than the first accuracy threshold, it is determined that the target execution result meets the plan modification conditions. The update plan list includes a first update plan list and a second update plan list; by modifying the plan list, the updated plan list is obtained, which includes: If the task execution accuracy is less than the second accuracy threshold, the business problem is re-planned and processed in the planning agent using the plan list and modification prompts to obtain the first updated plan list; the new subtasks in the first updated plan list are different from the subtasks in the plan list; the first accuracy threshold is greater than the second accuracy threshold. If the task execution accuracy is greater than or equal to the second accuracy threshold, and the task execution accuracy is less than the first accuracy threshold, then the subtasks in the plan list that are executed before the target subtask are identified as preceding subtasks. The task instructions and execution tools in the target subtask and preceding subtask in the plan list are modified to obtain a second updated plan list. The second updated plan list includes subtasks in the plan list that are executed after the target subtask.

[0168] In one possible implementation, the interaction processing module 740 is also used to perform the following operations: By matching keywords to business issues using a keyword database, guiding keywords that indicate the type of business issue are obtained; The text length of the business issue, the guiding keywords, and the number of subtasks in the plan list are weighted and summed to obtain the comprehensive score of the business issue; If the overall score is greater than or equal to the in-depth analysis threshold, then the communication mode corresponding to the business problem is determined to be the in-depth analysis mode. If the overall score is less than the in-depth analysis threshold, the communication mode corresponding to the business problem is determined to be the rapid analysis mode.

[0169] This application embodiment processes business problems through task planning, resulting in a plan list comprising K sub-tasks. Complex business problems can be broken down into K sub-tasks of appropriate granularity. Each sub-task includes specific task instructions and corresponding task tools. Through multi-task tool collaboration and multi-sub-task step-by-step reasoning, the solution process for complex business problems is decomposed into a progressive execution chain, reducing the processing pressure on individual task tools. Simultaneously, by leveraging the complementary functional characteristics of different task tools, the limitations of a single task tool in complex scenarios are avoided, improving the accuracy of sub-task execution. During the execution of the K sub-tasks, a target sub-task can be identified, and its corresponding environmental information can be obtained. Scene detection is performed on the environmental information to obtain scene detection results. If the scene detection results meet the conditions for multi-prompt generation, object-side prompts and model-side prompts can be generated based on the scene detection results. Object-side prompts explicitly guide the object to execute the control operations corresponding to the scene detection results, while model-side prompts implicitly inject constraints during the execution of the task instructions corresponding to the target sub-task by the task tool corresponding to the target sub-task. By distinguishing between object-side and model-side prompts, we can differentiate between the guidance needed by the user and the instructions required by the model. This allows us to optimize model constraints and the user's interaction environment, helping users express their needs more effectively and understand the model's capability boundaries. Responding to control operations guided by object-side prompts and obtaining object interaction results (such as supplementary user information, confirmed parameter options, and feedback on execution preferences) improves the coherence of the user conversation. Sending the object interaction results and model-side prompts to the task tool corresponding to the target subtask allows for targeted addition of constraints such as execution constraints, data specifications, or behavioral restrictions during the execution of the task instructions based on the target context information, object interaction results, and model-side prompts. This results in more accurate subtask execution results and ultimately improves the accuracy of the answer obtained by integrating the results of K subtasks. The dual-prompt word collaboration mechanism ensures that both the model and the user receive the most suitable guidance information. For users, it optimizes the retry process, reduces cognitive load, and allows for efficient task execution without requiring specialized technical skills, thus lowering the technical barrier. For the model, the injection of precise constraints improves execution efficiency and standardization, avoids invalid calculations, and saves computing resources. The synergistic effect of both effectively improves the interaction efficiency between users and the model, thereby increasing the completion rate of answers to complex business questions.

[0170] Meanwhile, by optimizing model inference and user interaction through a dual-prompt word architecture, dialogue breaks and misunderstandings can be reduced, which is beneficial for user understanding and the smooth flow of the dialogue system, thus improving dialogue coherence. User-side prompt words are dynamically generated based on real-time user context, enabling the system to adapt to different users' expression habits and knowledge backgrounds, providing a truly personalized interactive experience and improving the personalized interaction experience. By statistically analyzing the trigger frequency of detection results in various scenarios and optimizing and updating the prompt word library, effective execution cases and error correction information from actual business scenarios can be accumulated into execution log information. Optimizing and updating prompt elements can help the model gradually adapt to the personalized needs of specific scenarios, correct inference biases, and make the model's inference logic more in line with actual business rules, further improving the accuracy and reliability of the answer results.

[0171] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0172] Please see Figure 11 , Figure 11 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Figure 11 As shown, the computer device 1000 may include a processor 1001, a network interface 1004, and a memory 1005. Furthermore, the computer device 1000 may also include a user interface 1003 and at least one communication bus 1002. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen and a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as at least one disk storage device. Optionally, the memory 1005 may also be at least one storage device located remotely from the processor 1001. Figure 11 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and a device control application.

[0173] In such Figure 11In the computer device 1000 shown, the network interface 1004 provides network communication elements; the user interface 1003 is mainly used to provide an input interface for the user; and the processor 1001 can be used to call the device control application stored in the memory 1005 to achieve: The process involves identifying a business problem, performing task planning on that problem, and generating a plan list. The plan list includes K subtasks, where K is a positive integer. Each subtask includes different task instructions and different task tools. Determine the target subtask from the K subtasks, obtain the environmental information corresponding to the target subtask, perform scene detection on the environmental information, and obtain the scene detection result; If the scene detection result meets the conditions for generating multiple prompt words, then object-side prompt words and model-side prompt words are generated based on the scene detection result; the object-side prompt words are used to guide the object to perform the control operation corresponding to the scene detection result. The system responds to control operations guided by object-side prompts, acquires object interaction results, and sends the object interaction results and model-side prompts to the task tool corresponding to the target subtask. Within the task tool, the system executes the task instructions corresponding to the target subtask based on the target context information, object interaction results, and model-side prompts, thus obtaining the subtask execution result. The model-side prompts are used to indicate the constraints on the task tool during the execution of the task instructions corresponding to the target subtask. When the execution results of the K subtasks are obtained, the execution results of the K subtasks are integrated and processed to obtain the answer to the business question.

[0174] It should be understood that the computer device 1000 described in the embodiments of this application can execute the foregoing text. Figure 3 , Figure 5 , Figure 7 , Figure 8 and Figure 9 The description of the data processing method in any corresponding embodiment will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated.

[0175] Furthermore, it should be noted that this application embodiment also provides a computer-readable storage medium, which stores a computer program. When the processor executes the computer program, it can execute the aforementioned... Figure 3 , Figure 5 , Figure 7 , Figure 8 and Figure 9The description of the data processing method in any corresponding embodiment is already provided, and therefore will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated. For technical details not disclosed in the computer-readable storage medium embodiments related to this application, please refer to the description of the method embodiments of this application.

[0176] The aforementioned computer-readable storage medium can be an internal storage unit of the data processing apparatus or computer device provided in any of the foregoing embodiments, such as a hard disk or memory of the computer device. The computer-readable storage medium can also be an external storage device of the computer device, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., provided on the computer device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of the computer device. The computer-readable storage medium is used to store the computer program and other programs and data required by the computer device. The computer-readable storage medium can also be used to temporarily store data that has been displayed or will be displayed.

[0177] Furthermore, it should be noted that this application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. The processor of a computer device reads the computer program from the computer-readable storage medium and executes the computer program, causing the computer device to perform the aforementioned... Figure 3 , Figure 5 , Figure 7 , Figure 8 and Figure 9 The method provided in any of the corresponding embodiments.

[0178] The terms "first," "second," etc., in the specification, claims, and drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the term "comprising," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or device that includes a series of steps or units is not limited to the listed steps or modules, but may optionally include steps or modules not listed, or may optionally include other step units inherent to these processes, methods, apparatuses, products, or devices.

[0179] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the foregoing description as a network element. Whether these network elements are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can implement the described network elements using different methods for each specific application, but such implementation should not be considered beyond the scope of this application.

[0180] The methods and related apparatus provided in this application are described with reference to the method flowcharts and / or structural diagrams provided in this application. Specifically, each block of the method flowchart and / or structural diagram, as well as combinations of blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable device to create a machine, such that the instructions, which execute via the processor of the computer or other programmable device, generate instructions for implementing the process. Figure 1 A schematic diagram of one or more processes and / or structures. Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 A schematic diagram of one or more processes and / or structures. Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable device to cause a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable device for implementing the process. Figure 1 A process or multiple processes and / or structures illustrate the steps of the functions specified in one or more boxes.

[0181] The steps in the method of this application embodiment can be adjusted, combined, or deleted according to actual needs.

[0182] The modules in the device of this application embodiment can be merged, divided, and deleted according to actual needs.

[0183] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application shall still fall within the scope of this application.

Claims

1. A data processing method, characterized in that, include: The business problem is identified, and task planning is performed on the business problem to obtain a plan list. The plan list includes K sub-tasks, where K is a positive integer. Each sub-task includes different task instructions and different task tools. Among the K subtasks, a target subtask is determined, environmental information corresponding to the target subtask is obtained, scene detection is performed on the environmental information, and scene detection results are obtained. If the scene detection result meets the conditions for generating multiple prompt words, then object-side prompt words and model-side prompt words are generated based on the scene detection result; The object-side prompts are used to guide the object to perform the control operations corresponding to the scene detection results; In response to the control operation guided by the object-side prompt, the system obtains the object interaction result and sends the object interaction result and the model-side prompt to the task tool corresponding to the target subtask. Within the task tool, the system executes the task instructions corresponding to the target subtask based on the target context information, the object interaction result, and the model-side prompt, thereby obtaining the subtask execution result. The model-side prompt is used to indicate the constraints in the process of the task tool executing the task instructions corresponding to the target subtask. When the execution results of the K subtasks are obtained, the execution results of the K subtasks are integrated to obtain the answer to the business question.

2. The method according to claim 1, characterized in that, The business problem is the (H+1)th session initiated by the object, where H is a positive integer; the business problem is associated with the H sessions preceding the initiation of the business problem. The step of obtaining the environmental information corresponding to the target subtask includes: Obtain the session issues corresponding to the H round sessions respectively, determine the subtasks in the plan list that are executed before the target subtask as the preceding subtasks, and generate the target context information corresponding to the target subtask based on the execution results of the subtasks corresponding to the preceding subtasks, the business issues, the plan list and the session issues corresponding to the H round sessions respectively; Obtain object preference information and execution log information of the execution status of task tools in the tool library, and determine the object preference information and the execution log information as the system log information corresponding to the target subtask; Based on the target context information and system log information corresponding to the target subtask, generate the environment information corresponding to the target subtask.

3. The method according to claim 1, characterized in that, The generation of object-side prompts and model-side prompts based on the scene detection results includes: Obtain object-side and model-side prompt elements that match the scene detection results from the prompt word library; both object-side and model-side prompt elements are prompt word templates that can be filled with key information; Based on the target context information and scene detection results corresponding to the target sub-task, generate first key information for the object-side cue element and second key information for the model-side cue element; The first key information is filled into the object-side prompt element to obtain the object-side prompt word, and the second key information is filled into the model-side prompt element to obtain the model-side prompt word.

4. The method according to claim 3, characterized in that, Also includes: Within the statistical period, obtain the number of triggers corresponding to the scenarios in which W prompt words are generated; The prompt word generation scenario refers to the scenario detection result that satisfies the multiple prompt word generation conditions; W is a positive integer; Scenarios where the number of triggers is greater than or equal to the optimization threshold are identified as scenarios to be optimized. The execution log information of the scenarios to be optimized within the statistical period is obtained. The execution log information includes the object-side prompts, model-side prompts, and subtask execution results generated each time the multi-prompt generation conditions are met within the statistical period. Based on the execution log information, the object-side prompt element and the model-side prompt element corresponding to the scenario to be optimized are optimized and updated.

5. The method according to claim 1, characterized in that, The environmental information includes the target context information corresponding to the target sub-task; the step of performing scene detection on the environmental information to obtain scene detection results includes: If the amount of data of the content to be processed in the target context information is greater than or equal to the data amount threshold, then the scene detection result corresponding to the target subtask is determined to be the content exceeding the limit result. If the scene detection result is that the content exceeds the limit, then the scene detection result is determined to meet the conditions for generating multiple prompt words; The generation of object-side prompts and model-side prompts based on the scene detection results includes: Based on the content exceeding the limit result, object-side prompts and model-side prompts are generated; the object-side prompts are used to guide the object to compress the amount of data of the content to be processed, and the model-side prompts are used to instruct the task tool corresponding to the target subtask to generate the subtask execution result corresponding to the target subtask based on the compressed content of the content to be processed when it obtains the compressed content of the content to be processed.

6. The method according to claim 1, characterized in that, The environmental information includes the target context information corresponding to the target sub-task; the step of performing scene detection on the environmental information to obtain scene detection results includes: If the object does not have the permission to process the content to be processed in the target context information, then the scene detection result corresponding to the target subtask is determined to be a permission-restricted result. If the scene detection result is the permission-restricted result, then the scene detection result is determined to meet the conditions for generating multiple prompt words; The generation of object-side prompts and model-side prompts based on the scene detection results includes: Based on the permission restriction result, object-side prompts and model-side prompts are generated; the object-side prompts are used to guide the object to apply for processing permissions for the content to be processed in the target context information, and the model-side prompts are used to instruct the task tool corresponding to the target sub-task to trigger the execution of the task instruction corresponding to the target sub-task when it obtains the actual content corresponding to the content to be processed in the target context information.

7. The method according to claim 1, characterized in that, The environmental information includes the target context information corresponding to the target sub-task; the step of performing scene detection on the environmental information to obtain scene detection results includes: If the content to be processed in the target context information is located in a sub-partition of the data table, and the task instruction corresponding to the target sub-task does not include a partition identifier for the sub-partition, then the scene detection result corresponding to the target sub-task is determined to be a retrieval-restricted result. If the scene detection result is the search-restricted result, then the scene detection result is determined to meet the multi-prompt word generation conditions; The generation of object-side prompts and model-side prompts based on the scene detection results includes: Based on the restricted search results, the object-side prompt and the model-side prompt are generated; the object-side prompt is used to prompt the object that the task instruction corresponding to the target subtask does not include the partition identifier, and the model-side prompt is used to instruct the task tool corresponding to the target subtask to trigger the execution of the task instruction including the partition identifier when it obtains the task instruction including the partition identifier.

8. The method according to claim 1, characterized in that, The environmental information includes the target context information corresponding to the target sub-task; the step of performing scene detection on the environmental information to obtain scene detection results includes: If the operational risk level of the task instruction corresponding to the target subtask is greater than or equal to the risk level threshold, then the scene detection result corresponding to the target subtask is determined to be a risk confirmation result. If the scene detection result is the risk confirmation result, then the scene detection result is determined to meet the multi-prompt word generation conditions; The generation of object-side prompts and model-side prompts based on the scene detection results includes: Based on the risk confirmation result, the object-side prompt and the model-side prompt are generated; the object-side prompt is used to guide the object to confirm the operational risk level of the task instruction corresponding to the target sub-task, and the model-side prompt is used to instruct the task tool corresponding to the target sub-task to trigger the execution of the task instruction corresponding to the target sub-task when it obtains the object's confirmation result for the operational risk level.

9. The method according to claim 1, characterized in that, The step of obtaining environmental information corresponding to the target sub-task, performing scene detection on the environmental information, and obtaining scene detection results includes: If the target subtask is a subtask that has already been executed, and no execution result has been obtained for the subtask corresponding to the target subtask, then obtain the execution log information generated by the task tool corresponding to the target subtask during the execution of the task instruction corresponding to the target subtask; Based on the execution log information, generate the environmental information corresponding to the target subtask, and determine that the scene detection result corresponding to the target subtask is an execution failure result; If the scene detection result is the execution failure result, then it is determined that the scene detection result meets the conditions for generating multiple prompt words.

10. The method according to claim 1, characterized in that, The K subtasks include subtasks Kazuko's Mission , where i is a positive integer less than or equal to K; the subtask The execution order of the subtasks Before the execution order; The step of determining the target subtask among the K subtasks includes: The subtasks among the K subtasks that do not have data dependencies are identified as the target subtasks; Or, if the subtask Having for the subtask The data dependencies, and the acquisition of the subtasks. If the result of the subtask execution is obtained, then the subtask will be... This is identified as a target sub-task; Or through the subtask The corresponding task tool executes the subtask. If the corresponding task instruction fails, the subtask will be... This has been identified as a target sub-task.

11. The method according to claim 1, characterized in that, The K sub-tasks are obtained by the planning agent through task planning; the step of obtaining environmental information corresponding to the target sub-task, performing scene detection on the environmental information, and obtaining scene detection results includes: If the target subtask is a subtask that has already been executed, and the execution result of the subtask corresponding to the target subtask has been obtained, then the planning agent generates the task execution accuracy of the target subtask execution result, and generates the environmental information corresponding to the target subtask based on the task execution accuracy. If the task execution accuracy in the environmental information is less than the accuracy threshold, then the scene detection result corresponding to the target subtask is determined to be an execution failure result. If the scene detection result is the execution failure result, then it is determined that the scene detection result meets the conditions for generating multiple prompt words.

12. The method according to claim 10 or 11, characterized in that, The K sub-tasks are obtained by the planning agent through task planning; the generation of object-side prompts and model-side prompts based on the scene detection results includes: Based on the execution failure result, object-side prompts and model-side prompts are generated; the object-side prompts are used to guide the object to modify the task instructions of the target subtask. If no modification result of the object's task instruction for the target subtask is obtained, the model-side prompt is used to instruct the planning agent to modify the target subtask to obtain a new task tool and new task instruction corresponding to the target subtask. If a modification result of the object's task instruction for the target subtask is obtained, the model-side prompt is used to instruct the planning agent to obtain the new task tool and new task instruction corresponding to the target subtask from the modification result. The new task tool and the new task instruction are used to generate the subtask execution result corresponding to the target subtask.

13. The method according to claim 1, characterized in that, The task planning process for the business problem yields a plan list, including: The business problem is broken down into K sub-tasks by a planning intelligent agent, resulting in task instructions corresponding to each sub-task; the K sub-tasks include sub-tasks. , where i is a positive integer less than or equal to K; Obtain the task domain type corresponding to each of the E candidate tools in the tool library, based on the subtask. The corresponding task instructions and the task domain types corresponding to the E candidate tools are used to determine the subtask from the E candidate tools. The corresponding task tool; E is a positive integer; When the task tools corresponding to the K sub-tasks are obtained, the task instructions and task tools corresponding to the K sub-tasks are determined as a plan list.

14. The method according to claim 1, characterized in that, The K sub-tasks are obtained by the planning agent through task planning; the task tool corresponding to the target sub-task executes the task instructions corresponding to the target sub-task based on the target context information, the object interaction results, and the model-side prompts to obtain the sub-task execution results, including: By executing the intelligent agent to call the task tool corresponding to the target subtask, the task tool corresponding to the target subtask executes the task instruction corresponding to the target subtask according to the target context information, the object interaction result and the model-side prompt, and obtains the subtask execution result corresponding to the target subtask. When the execution results of the K subtasks are obtained, the execution results of the K subtasks are integrated to obtain the answer to the business question, including: If the communication mode corresponding to the business problem is the rapid analysis mode, then the context information of the next subtask of the target subtask is generated based on the execution result of the subtask corresponding to the target subtask. In the task tool corresponding to the next subtask, the task instruction corresponding to the next subtask is executed based on the context information obtained by the next subtask until the execution results of the subtasks corresponding to the K subtasks are obtained. Then, the planning agent integrates and processes the execution results of each subtask to obtain the answer result corresponding to the business problem. If the communication mode corresponding to the business problem is the deep analysis mode, and the target execution result meets the plan modification conditions, then the plan list is modified to obtain an updated plan list. The new sub-tasks in the updated plan list are then executed until the execution results of each new sub-task do not meet the plan modification conditions. Then, the planning agent integrates the execution results of each sub-task to obtain the answer to the business problem.

15. The method according to claim 14, characterized in that, The communication mode corresponding to the business problem is the in-depth analysis mode, and the method also includes: The task execution accuracy rate of the target subtask execution result is generated by the planning agent and the verification prompt words. If the task execution accuracy rate is less than the first accuracy rate threshold, it is determined that the target execution result meets the plan modification conditions. The update plan list includes a first update plan list and a second update plan list; the process of modifying the plan list to obtain the update plan list includes: If the task execution accuracy is less than the second accuracy threshold, then in the planning agent, the business problem is re-planned using the plan list and modification prompts to obtain a first updated plan list; the new subtasks in the first updated plan list are different from the subtasks in the plan list; the first accuracy threshold is greater than the second accuracy threshold. If the task execution accuracy is greater than or equal to the second accuracy threshold, and the task execution accuracy is less than the first accuracy threshold, then the subtasks in the plan list that are executed before the target subtask are determined as preceding subtasks. The task instructions and execution tools in the target subtask and the preceding subtasks in the plan list are modified to obtain a second updated plan list. The second updated plan list includes the subtasks in the plan list that are executed after the target subtask.

16. The method according to claim 14, characterized in that, Also includes: The business problem is matched with keywords from a keyword database to obtain guiding keywords that indicate the problem type. The comprehensive score of the business problem is obtained by weighted summation of the text length of the business problem, the guiding keywords, and the number of subtasks in the plan list; If the comprehensive score is greater than or equal to the in-depth analysis threshold, then the communication mode corresponding to the business problem is determined to be the in-depth analysis mode; If the overall score is less than the depth analysis threshold, then the communication mode corresponding to the business problem is determined to be the rapid analysis mode.

17. A data processing apparatus, characterized in that, include: The task planning module is used to acquire business problems, perform task planning processing on the business problems, and obtain a plan list; the plan list includes K sub-tasks, where K is a positive integer; each sub-task includes different task instructions and different task tools. The scene detection module is used to determine the target subtask among the K subtasks, obtain the environmental information corresponding to the target subtask, perform scene detection on the environmental information, and obtain the scene detection result. The prompt word generation module is used to generate object-side prompt words and model-side prompt words based on the scene detection results if the scene detection results meet the multi-prompt word generation conditions; the object-side prompt words are used to guide the object to perform the control operation corresponding to the scene detection results; An interaction processing module is used to respond to the control operation guided by the object-side prompt, obtain the object interaction result, and send the object interaction result and the model-side prompt to the task tool corresponding to the target subtask. In the task tool corresponding to the target subtask, the task instruction corresponding to the target subtask is executed according to the target context information corresponding to the target subtask, the object interaction result, and the model-side prompt, to obtain the subtask execution result. The model-side prompt is used to indicate the constraints of the task tool corresponding to the target subtask during the execution of the task instruction corresponding to the target subtask. The result integration module is used to integrate the execution results of the K subtasks when the execution results of the K subtasks are obtained respectively, so as to obtain the answer result corresponding to the business question.

18. A computer device, characterized in that, include: Processor, memory, and network interface; The processor is connected to the memory and the network interface, wherein the network interface is used to provide data communication functions, the memory is used to store computer programs, and the processor is used to call the computer programs so that the computer device executes the method according to any one of claims 1-16.

19. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded and executed by a processor to cause a computer device having the processor to perform the method of any one of claims 1-16.

20. A computer program product, characterized in that, The computer program product includes a computer program stored in a computer-readable storage medium and adapted to be read and executed by a processor to cause a computer device having the processor to perform the method of any one of claims 1-16.

Images