Dynamic task execution methods, devices, and electronic equipment based on large models

Abstract

This application provides a dynamic task execution method, apparatus, and electronic device based on a large model, belonging to the field of artificial intelligence technology. The method includes: inputting user request information for a user requesting to execute a task into a large model to obtain a target primary scene and a target secondary scene output by the large model; the target primary scene is determined from a pre-registered set of primary scenes; the target secondary scene is determined from a set of secondary scenes associated with the target primary scene; inputting the user request information and the target scene description into the large model to obtain a set of tools to be invoked for the secondary scene and the corresponding user task step description output by the large model; invoking the set of tools to be invoked for the secondary scene and executing the task steps corresponding to the user task step description. This application selects tools and executes task steps in the context of multi-level scene division, enabling precise scene matching and flexible tool scheduling and orchestration for task execution.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus and electronic device for dynamic task execution based on a large model. Background Technology

[0002] In implementing a large-model-based process and task orchestration and scheduling framework, a common technical solution is to select a series of tools to execute serially or in parallel based on user input information and large-model invocation, and return the results after all executions are completed.

[0003] Specifically, when a user query is received and input into the large model, the user query and all tools can be input into the large model. The large model can then output a series of tool calls that need to be executed to determine the response to the user query. Finally, several tasks can be invoked to execute the tasks requested by the user.

[0004] However, since both the tool execution plan and the tool execution are completed in one go, even if some parts of the tool execution fails, the user still needs to wait for the entire execution plan to be completed before the next process can be rearranged. This makes it impossible to dynamically adjust the execution plan and lacks reflection during the execution process. Summary of the Invention

[0005] This application provides a dynamic task execution method, apparatus, and electronic device based on a large model to solve the defect in the prior art that requires waiting for the completion of all execution plans before the process can be re-arranged.

[0006] This application provides a dynamic task execution method based on a large model, including: The user request information for the task to be executed is input into the large model to obtain the target primary scene and target secondary scene output by the large model; the target primary scene is determined from the pre-registered set of primary scenes; the target secondary scene is determined from the set of secondary scenes associated with the target primary scene; The user request information and target scene description are input into the large model to obtain the secondary scene toolset to be invoked and the corresponding user task step description output by the large model; the target scene description is determined based on the target primary scene description and the target secondary scene description. The secondary scenario toolset is invoked to execute the task steps corresponding to the user task step description.

[0007] According to the dynamic task execution method based on a large model provided in this application, the step of inputting user request information for a user requesting to execute a task into a large model to obtain a target primary scene and a target secondary scene output by the large model includes: inputting the user request information into the large model to obtain the target primary scene output by the large model; and inputting the user request information and the description of the target primary scene into the large model to obtain the target secondary scene output by the large model.

[0008] According to the dynamic task execution method based on a large model provided in this application, the secondary scenario toolkit to be invoked includes multiple tools to be invoked. Invoking the secondary scenario toolkit to be invoked and executing the task steps corresponding to the user task step description includes: inputting the secondary scenario toolkit to be invoked and the user task step description into the large model to obtain the target tool to be invoked output by the large model; the target tool to be invoked is determined from the multiple tools to be invoked in the secondary scenario toolkit; invoking the target tool to be invoked and executing the task steps corresponding to the user task step description to obtain tool execution results and tool execution information; inputting the tool execution information into the large model to obtain the... The process judgment result output by the large model is as follows: If the process judgment result is "process terminated" and the tool execution result is "execution failed", the tool execution information, the user request information, and the target primary scene description are input into the large model to obtain a new target secondary scene output by the large model. The user request information and the new target scene description are then input into the large model to obtain a new secondary scene tool set to be called and a corresponding user task step description output by the large model. The process then returns to the step of inputting the secondary scene tool set to be called and the user task step description into the large model. The new target scene description is determined based on the target primary scene description and the new target secondary scene description.

[0009] According to the dynamic task execution method based on a large model provided in this application, the step of calling the secondary scenario to be called toolset and executing the task steps corresponding to the user task step description further includes: when the process judgment result is the end of the process and the tool execution result is the execution success, performing information compliance processing on the tool execution information to obtain the task execution result; and returning the task execution result to the user.

[0010] According to the dynamic task execution method based on a large model provided in this application, the user request information is determined in the following manner: initial input information of the user requesting to execute the task is input into the large model to obtain the user's intent to be verified and its corresponding intent reliability output by the large model; when the intent reliability corresponding to the user's intent to be verified is less than a preset reliability threshold, subsequent input information of the user requesting to execute the task is obtained; based on the subsequent input information and the initial input information, complete dialogue information is determined; the complete dialogue information is input into the large model to obtain the user's complete intent output by the large model; based on the user's complete intent, the user request information is determined.

[0011] According to the dynamic task execution method based on a large model provided in this application, the step of inputting the user request information and the target scene description into the large model to obtain the secondary scene to be called toolset and the corresponding user task step description output by the large model includes: inputting the user request information, the target scene description and the user question mode into the large model to obtain the secondary scene to be called toolset and the corresponding user task step description output by the large model.

[0012] This application also provides a dynamic task execution device based on a large model, comprising: The two-level scene determination module is used to input user request information for executing a task into a large model to obtain the target primary scene and target secondary scene output by the large model; the target primary scene is determined from a pre-registered set of primary scenes; the target secondary scene is determined from a set of secondary scenes associated with the target primary scene; The tool step determination module is used to input the user request information and the target scene description into the large model to obtain the secondary scene tool set to be called and the corresponding user task step description output by the large model; the target scene description is determined based on the target primary scene description and the target secondary scene description. The tool step execution module is used to call the secondary scenario tool set to be called and execute the task steps corresponding to the user task step description.

[0013] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the dynamic task execution method based on a large model as described above.

[0014] This application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the dynamic task execution method based on a large model as described above.

[0015] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the dynamic task execution method based on a large model as described above.

[0016] The dynamic task execution method, apparatus, and electronic device based on a large model provided in this application, by setting up multi-level scenarios, selects tools and executes task steps under the division of multi-level scenarios. For the task requested by the user, it identifies the first-level scenario and the second-level scenario, and then determines the set of tools to be called for the corresponding second-level scenario. Then, it can call tools to execute tasks with the second-level scenario as the basic unit. When an execution exception occurs, it does not need to wait for the entire execution plan to complete, but can dynamically adjust the execution plan when the task execution corresponding to the second-level scenario fails. It can achieve accurate scenario matching and flexible scheduling and orchestration of task execution. At the same time, defining multi-level scenarios can improve the accuracy of scenario selection when there are many scenarios. Furthermore, by defining multi-level scenarios and adding scenario description information and tool information for first-level and second-level scenarios, the large model can have the ability to schedule and orchestrate new tasks and new processes, and has good scalability. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the dynamic task execution method based on a large model provided in this application.

[0019] Figure 2 This is a flowchart illustrating the dynamic task execution method based on a large model and intelligent agent architecture provided in this application.

[0020] Figure 3 This is a flowchart illustrating the process of determining the primary and secondary scenarios provided in this application.

[0021] Figure 4 This is a schematic diagram of the structure of the dynamic task execution device based on a large model provided in this application.

[0022] Figure 5 This is a schematic diagram of the structure of the electronic device provided in this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions 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, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] The following is combined with Figures 1 to 5 This application describes a dynamic task execution method, apparatus, and electronic device based on a large model.

[0025] Figure 1 This is a flowchart illustrating the dynamic task execution method based on a large model provided in this application, such as... Figure 1 As shown, the dynamic task execution method based on the large model includes, but is not limited to, steps 101 to 103.

[0026] It should be noted that the execution subject of the dynamic task execution method based on large models provided in this application can be a server, computer device, such as a mobile phone, tablet computer, laptop computer, handheld computer, vehicle electronic device, wearable device, ultra-mobile personal computer (UMPC), netbook, or personal digital assistant (PDA), etc.

[0027] Step 101: Input the user request information for the user requesting to execute the task into the large model to obtain the target primary scene and target secondary scene output by the large model.

[0028] Among them, the target primary scene is determined from the pre-registered set of primary scenes; the target secondary scene is determined from the set of secondary scenes associated with the target primary scene.

[0029] Large model, also known as Language Large Model (LLM).

[0030] The large model is pre-registered with a set of first-level scenarios including multiple first-level scenarios and a set of second-level scenarios including multiple second-level scenarios. Each first-level scenario is pre-associated with several second-level scenarios, indicating that the second-level scenario is associated with a task mode used to execute the task corresponding to the first-level scenario. Different second-level scenarios represent different task modes used to execute the task corresponding to the associated first-level scenario.

[0031] User request information is information in which a user requests a large model to perform a certain task. Specifically, it can be any of the following: user input information input into the large model, multi-turn dialogue information consisting of user input information and guidance supplementary information obtained after guiding the user to supplement the large model for multiple rounds, or user intent information determined based on user input information or multi-turn dialogue information.

[0032] Specifically, user intent information refers to the user intent information and its corresponding intent reliability obtained by inputting user input information or multi-turn dialogue information into the large model.

[0033] Optionally, in addition to user input information, multi-turn dialogue information, and / or user intent information, the user request information may also include necessary prompts.

[0034] It should be noted that the specific tasks performed by the user request large model can be any of the task types in different communication, computer, life and production fields, such as communication network fault troubleshooting, server performance bottleneck optimization, database slow query root cause location, and abnormal vibration analysis of industrial equipment. This application does not impose any restrictions on this.

[0035] Specifically, the user inputs a user request to the large model to perform a certain task. The large model processes the user request information, first determining a target primary scene from the pre-registered set of primary scenes, then determining a target secondary scene from the set of secondary scenes associated with the target primary scene, and finally outputting the target primary scene and the target secondary scene. In the scene results output by the large model, each target secondary scene has the ability to complete the task corresponding to the primary scene.

[0036] Furthermore, in the large-model-based process and task orchestration and scheduling framework, by pre-registering and setting hierarchical first-level and second-level scenarios, the ability of the large model to adapt to new task orchestration can be greatly improved, exhibiting good scalability.

[0037] For example, if a large model needs to add the ability to orchestrate and schedule the execution of a new task or process, it is only necessary to add and register a new primary scenario and several new secondary scenarios associated with the new primary scenario. This includes adding scenario description information and tool registration information for the new primary and secondary scenarios, and adding the new primary and secondary scenarios to the primary scenario set and the secondary scenario set respectively. This will allow the new scenario to be incorporated into the overall task orchestration process.

[0038] Therefore, after the user requests to execute new tasks are input into the large model, the large model outputs the target first-level scene determined from the set of first-level scenes of newly added first-level scenes, and the target second-level scene determined from the set of second-level scenes of newly added second-level scenes.

[0039] Step 102: Input the user request information and target scene description into the large model to obtain the secondary scene tool set to be called and the corresponding user task step description output by the large model.

[0040] The target scene description is determined based on the target primary scene description and the target secondary scene description.

[0041] The target level 1 scene description is the scene description information of the target level 1 scene; the target level 2 scene description is the scene description information of the target level 2 scene; the scene description information is the description information added to the level 1 or level 2 scene during registration.

[0042] The set of tools to be invoked in a secondary scenario is a collection of tools identified from the tool registration information added during the registration of the secondary scenario. These tools are used to execute tasks requested by the user. Tools are functions or services outside the larger model. These services include, but are not limited to, several types of services such as translation services, professional knowledge bases, file processing tools, computer language programming tools, database query tools, and information retrieval tools. A single call to a function or service constitutes a tool invocation.

[0043] The user task step description is a description of several task steps involved in executing a user-requested task based on a secondary scenario's set of tools to be invoked. For the same task, the user task step description will differ depending on the specific secondary scenario's set of tools to be invoked.

[0044] Specifically, after the large model outputs the target primary scene and the target secondary scene, it is necessary to determine the set of tools to be called from the many tools associated with the target secondary scene.

[0045] When determining the set of tools to be invoked for the target secondary scenario, the target primary scenario description is determined based on the scenario description information of the target primary scenario to which the target secondary scenario belongs, the target secondary scenario description is determined based on the scenario description information of the target secondary scenario, and the target scenario description is determined based on the target primary scenario description and the target secondary scenario description.

[0046] Then, the user request information and the target scenario description are simultaneously input into the large model. The large model then determines some or all of the tools from the toolset added during the registration of the target secondary scenario, thus obtaining the toolset to be called in the secondary scenario and the corresponding user task step description.

[0047] Step 103: Invoke the secondary scenario toolkit to be invoked and execute the task steps corresponding to the user task step description.

[0048] Specifically, select some or all of the tools in the set of tools to be invoked in the secondary scenario to execute the task steps corresponding to the user task step description, thereby realizing task execution.

[0049] Understandably, when parallel invocation of the toolset to be invoked in the secondary scenario is supported, more than two toolsets to be invoked in the secondary scenario can be invoked in parallel.

[0050] More specifically, when calling the toolset to be called in the secondary scenario, the name and parameters of the tools to be called are identified according to the user task step description. The relevant tools are called and executed from the Model Context Protocol (MCP) server, thereby executing the task steps corresponding to the user task step description, obtaining the tool execution results (such as execution success, execution failure, etc.) and tool execution information (such as execution logs, etc.), and storing both the tool execution results and tool execution information in the context management module.

[0051] Figure 2 This is a flowchart illustrating the dynamic task execution method based on a large model and intelligent agent architecture provided in this application, as shown below. Figure 2 As shown, in implementing the dynamic task execution method based on a large model provided in this application on the basis of an end-to-end question-answering system framework built on a multi-agent architecture, the large model includes a process orchestration agent and a tool execution agent. The process orchestration agent orchestrates the target primary scenario and the target secondary scenario according to the user request information, and then orchestrates the secondary scenario to be invoked toolset and the corresponding user task step description according to the user request information and the target scenario description. The tool execution agent then invokes the secondary scenario to be invoked toolset, executes the task steps corresponding to the user task step description, and obtains the tool execution result and tool execution information.

[0052] The dynamic task execution method based on a large model provided in this application sets up multi-level scenarios, selects tools and executes task steps within these scenarios. For user-requested tasks, it identifies primary and secondary scenarios, determines the corresponding set of tools to be invoked for each secondary scenario, and then uses each secondary scenario as a basic unit to invoke tools for task execution. In case of execution errors, it eliminates the need to wait for the entire execution plan to complete; instead, it dynamically adjusts the execution plan when errors occur in the tasks corresponding to the secondary scenario. This enables precise scenario matching and flexible tool scheduling and orchestration. Furthermore, defining multiple scenarios improves the accuracy of scenario selection when there are many scenarios. By defining multiple scenarios and adding scenario descriptions and tool information for primary and secondary scenarios, the large model gains the ability to schedule and orchestrate new tasks and processes, exhibiting excellent scalability.

[0053] Based on the above embodiments, as an optional embodiment, the step of inputting the user request information for the user requesting to execute the task into the large model to obtain the target primary scene and target secondary scene output by the large model includes: The user request information is input into the large model to obtain the target primary scene output by the large model; The user request information and the target primary scene description are input into the large model to obtain the target secondary scene output by the large model.

[0054] Specifically, Figure 3 This application provides a flowchart illustrating the process of determining primary and secondary scenarios, such as... Figure 3 As shown, when performing orchestration processing for primary scenes, user request information is input into the large model. The large model then determines and outputs several target primary scenes from the pre-registered set of primary scenes based on the user request information. For example, scenes 1, 2, 3, and 4 are all selected target primary scenes.

[0055] Further, secondary scene orchestration is performed within each target primary scene. For each target primary scene, the target primary scene description is determined based on the scene description information of that target primary scene. The target primary scene description and the original user request information are input into the large model. The large model determines several target secondary scenes from the set of secondary scenes associated with that target primary scene based on the user request information and the target primary scene description, and outputs the target secondary scenes, thereby completing the orchestration of multi-level scenes.

[0056] Taking the orchestration of secondary scenes within Scene 1 as an example, the target primary scene description is determined based on the scene description information of Scene 1. The target primary scene description and the original user request information are input into the large model to obtain the target secondary scenes output by the large model, namely Scene 1.1, Scene 1.2 and Scene 1.3.

[0057] The dynamic task execution method based on a large model provided in this application improves the accuracy of scene matching and the flexibility of tool scheduling and orchestration by performing stepwise tool selection and execution in the case of dividing into multi-level scenes. First, a first-level scene orchestration is performed, and then a second-level scene orchestration is performed based on the selected first-level scene. This facilitates scene classification and improves the accuracy of scene selection when there are many scenes.

[0058] In another embodiment, the step of inputting the user request information and the target primary scene description into the large model to obtain the target secondary scene output by the large model includes: inputting the user request information into the large model to obtain the target primary scene output by the large model; and inputting the user request information into the large model to obtain the target secondary scene output by the large model.

[0059] Based on the above embodiments, as an optional embodiment, the secondary scenario toolkit to be invoked includes multiple tools to be invoked. Invoking the secondary scenario toolkit to be invoked and executing the task steps corresponding to the user task step description includes: The set of tools to be invoked in the secondary scenario and the description of the user task steps are input into the large model to obtain the target tools to be invoked output by the large model; the target tools to be invoked are determined from multiple tools to be invoked in the set of tools to be invoked in the secondary scenario. The target tool to be invoked is executed to perform the task steps corresponding to the user task step description, and the tool execution result and tool execution information are obtained. The tool execution information is input into the large model to obtain the process judgment result output by the large model; If the process judgment result is "process termination" and the tool execution result is "execution failure," the tool execution information, the user request information, and the target primary scene description are input into the large model to obtain a new target secondary scene output by the large model. The user request information and the new target scene description are then input into the large model to obtain a new set of tools to be called for the secondary scene and the corresponding user task step description output by the large model. The process then returns to the step of inputting the set of tools to be called for the secondary scene and the user task step description into the large model. The new target scene description is determined based on the target primary scene description and the new target secondary scene description.

[0060] Specifically, when each secondary scenario's toolset is invoked, the toolset and the user task step description are input into the large model. The large model then determines one or more target tools from the multiple tools in the toolset and outputs them, thus obtaining the target tools output by the large model. The tool is then invoked from the MCP server and executed according to the corresponding user task step description to perform the user task. The tool execution results and information are obtained and stored in the context management module.

[0061] The tool execution information is then input into the large model, which determines whether to continue the process (select the next tool to be called) or terminate the process, and outputs the process judgment result.

[0062] If the large model outputs a process judgment result indicating process termination and the tool execution information indicates execution failure, then the large model reselects a secondary scenario. The tool execution information, user request information, and the target primary scenario description are input into the large model to obtain a new target secondary scenario output by the large model. A new target secondary scenario description is determined based on the new target secondary scenario, and the new target scenario description is determined using the target primary scenario description and the new target secondary scenario description. The user request information and the new target scenario description are input into the large model to obtain a new set of tools to be invoked for the secondary scenario and the corresponding user task step description. Returning to the step of inputting the set of tools to be invoked for the secondary scenario and the user task step description into the large model, the task execution under the newly selected secondary scenario is performed.

[0063] If the large model outputs a process judgment result indicating that the process should continue, then the large model will reselect the tool to be invoked within the same secondary scenario. The tool execution information, the set of tools to be invoked in the secondary scenario, and the user task step description are input into the large model to obtain a new target tool to be invoked. The model then returns to the step of invoking the target tool, and, with the secondary scenario unchanged, uses other tools to execute the task.

[0064] Taking a communication network troubleshooting task with the user request message "My computer can't connect to Wi-Fi, the internet speed is very slow, please help me troubleshoot the problem" as an example, assuming the pre-registered primary scenario set includes the "Wireless Network" primary scenario and the "Wired Network" primary scenario, and the secondary scenario set associated with the "Wireless Network" primary scenario includes the "Signal Strength Check" secondary scenario and the "Network Coverage Check" secondary scenario, after analyzing the user request information, the large model outputs the target primary scenario as "Wireless Network" and the target secondary scenario as "Signal Strength Check".

[0065] In the target secondary scenario of "signal strength investigation", the user request information, the target secondary scenario description determined based on the "signal strength investigation" secondary scenario, and the target primary scenario description determined based on the "wireless network" primary scenario are input into the large model to obtain the secondary scenario tool set to be called output by the large model, that is, the signal detection tool set including multiple signal detection tools, and the corresponding user task step description, that is, the current device signal strength detection step description.

[0066] The signal detection toolset and the description of the current device signal strength detection steps are further input into the large model. This yields some or all of the signal detection tools output by the large model, which are then used as the target tools to be invoked. The target tools are invoked, and the computer signal strength is detected according to the current device signal strength detection steps. The detected computer signal strength is then used as tool execution information, and the tool execution result is determined based on this information. Furthermore, the tool execution information is input into the large model to obtain a process judgment result (continue or terminate the process) output by the large model.

[0067] If the process judgment result is to terminate the process and the tool execution result is to fail, then the large model is used to reselect the secondary scenario. For example, if the target secondary scenario re-determined by the large model is "network coverage investigation", then under the target secondary scenario of "network coverage investigation", the user request information, the target secondary scenario description determined based on the "network coverage investigation" secondary scenario, and the target primary scenario description determined based on the "wireless network" primary scenario are input into the large model to obtain a new set of tools to be called for the secondary scenario output by the large model. That is, a hotspot scanning tool set including multiple surrounding hotspot scanning tools is obtained, as well as the corresponding user task step description, that is, the scanning steps of the network coverage of the current device location.

[0068] Then, return to the step of inputting the secondary scenario tool set to be called and the user task step description into the large model, and further input the hot spot scanning tool set and the scanning steps of the network coverage of the current device location into the large model. The hot spot scanning tools output by the large model are used as the target tool to be called. The target tool to be called is called and the network coverage is detected according to the scanning steps of the network coverage of the current device location.

[0069] Optionally, if the process judgment result is to continue the process, the tool execution information, the secondary scenario tool set to be called, and the user task step description are input into the large model to obtain a new target tool to be called output by the large model, and then return to the step of calling the target tool to be called.

[0070] The dynamic task execution method based on a large model provided in this application improves the system's fault tolerance by using a multi-level dynamic orchestration and reflection mechanism. If a secondary scenario under a primary scenario fails to execute, the system will reflect on the tool execution information from the context management module and reselect a secondary scenario to execute the task. This avoids orchestration failures caused by the model repeatedly selecting the same scenario. Furthermore, by introducing the reflection function, the process orchestration can be dynamically adjusted in real time, avoiding the waste of time and model resources.

[0071] Based on the above embodiments, as an optional embodiment, the step of invoking the secondary scenario toolset to be invoked and executing the task steps corresponding to the user task step description further includes: If the process judgment result is that the process is terminated and the tool execution result is that the execution is successful, the tool execution information is processed for information compliance to obtain the task execution result; The task execution result will be returned to the user.

[0072] Specifically, in combination Figure 2 As shown, in implementing the dynamic task execution method based on a large model provided in this application on the basis of an end-to-end question-answering system framework built on a multi-agent architecture, the large model also includes a process verification agent and a result generation agent.

[0073] When the execution result of the tool corresponding to the toolset to be invoked in any secondary scenario is successful, the process verification agent performs process verification. Based on the successful tool execution information, it determines whether the task execution process requested by the user has ended. If the process verification agent determines that the process judgment result is that the process has ended (i.e., the process verification has passed), the process verification agent sends the successful tool execution information to the result generation agent. The result generation agent then performs information compliance processing based on the successful tool execution information, obtains the task execution result, and returns the task execution result to the user, ensuring that the information returned to the user meets the requirements.

[0074] On the other hand, if the process verification agent determines that the process judgment result is to continue the process (i.e., the process verification fails), then the process orchestration agent is returned to continue the process orchestration for the next step.

[0075] The dynamic task execution method based on a large model provided in this application can reduce the large model's misjudgment of the process completion by performing process verification after successfully executing a task in any secondary scenario's toolkit, thus avoiding returning unprocessed execution results to the user and improving the user experience.

[0076] Based on the above embodiments, as an optional embodiment, the user request information is determined in the following way: The initial input information of the user's request to execute the task is input into the large model to obtain the user's intent to be verified and its corresponding intent reliability output by the large model. When the reliability of the intent corresponding to the user's intent to be verified is less than a preset reliability threshold, the subsequent input information of the user's request to execute the task is obtained; Based on the subsequent input information and the initial input information, the complete dialogue information is determined; The complete dialogue information is input into the large model to obtain the complete user intent output by the large model. Based on the user's complete intent, the user request information is determined.

[0077] Furthermore, this application improves the accuracy of multi-level scene recognition by using user intent recognition verification and user input information guidance and supplementation.

[0078] Specifically, in combination Figure 2 As shown, in implementing the dynamic task execution method based on a large model provided in this application on the basis of an end-to-end question-answering system framework built on a multi-agent architecture, the large model also includes a user translation agent and a scene recognition agent.

[0079] After the initial input information of the user requesting to execute the task is input into the large model, the user translation agent of the large model performs intent recognition based on the initial input information. The recognized intent and the corresponding confidence level are used as the user's intent to be verified and its corresponding intent reliability. Furthermore, it is determined whether the intent reliability of the user's intent to be verified is less than a preset reliability threshold.

[0080] If the user translation agent determines that the reliability of the user's intent to be verified is greater than or equal to the preset reliability threshold, then the user's intent to be verified is directly used as the user request information and sent to the scene recognition agent. The scene recognition agent then identifies the target primary scene and the target secondary scene based on the user request information.

[0081] If the user translation agent determines that the reliability of the user's intent to be verified is less than a preset reliability threshold, the large model returns user input guidance information to guide the user to supplement the input with relevant information about the task they requested to perform, until the reliability of the user's intent is greater than or equal to the preset reliability threshold. After the user has completed inputting the supplementary information, the large model obtains subsequent input information.

[0082] The user translation agent further determines the complete dialogue information obtained after multiple rounds of dialogue based on the initial input information and subsequent input information. It then performs intent recognition based on the complete dialogue information to obtain the user's complete intent. Finally, the user's complete intent is used as user request information and sent to the scene recognition agent. The scene recognition agent then identifies the target primary scene and target secondary scene based on the user request information.

[0083] In one embodiment, obtaining subsequent input information for the user's request to perform a task includes: Return to user input guidance supplementary information, which is used to guide the user to supplement the input with relevant information about the task they requested to perform; Receive a subsequent input from the user, and determine the user's intent to be verified and its corresponding reliability based on the subsequent input and the initial input. When the reliability of the intent corresponding to the user's intent to be verified is less than the preset reliability threshold, return to the step of returning the user's input guidance supplementary information until the reliability of the intent corresponding to the user's intent to be verified is greater than or equal to the preset reliability threshold. Determine subsequent input information based on all subsequent input information provided by the user.

[0084] The dynamic task execution method based on a large model provided in this application obtains complete dialogue information supplemented by the user by combining the reliability of the user input information's intent, and performs intent recognition on the complete dialogue information to obtain the user's complete intent. This enables the translation of user questions across first-level scenarios, facilitating the identification and processing of subsequent processes. It also improves the accuracy of scenario recognition by identifying subsequent first-level and second-level scenarios based on the user's complete intent.

[0085] Based on the above embodiments, as an optional embodiment, the step of inputting the user request information and target scene description into the large model to obtain the secondary scene tool set to be invoked and the corresponding user task step description output by the large model includes: The user request information, the target scene description, and the user question pattern are input into the large model to obtain the secondary scene tool set to be called and the corresponding user task step description output by the large model.

[0086] Among them, the user questioning mode is a combination of professional language and common questioning methods used by users when inputting information into the large model.

[0087] Professional terminology refers to the specific professional terms, jargon, and standardized expressions used by users when inputting information. Examples include "VLAN" and "SSL application blocking" in communication networks, "chief complaint" and "differential diagnosis" in the medical field, and "cause of action" and "burden of proof" in the legal field.

[0088] Common question formats are different ways of expressing the same thing when users from different regions and ages input information, such as specific dialects.

[0089] Specifically, when the large model outputs the target primary scene and the target secondary scene, and it is necessary to determine the toolset to be called in the secondary scene, on the one hand, the user request information is determined based on the user input information or the user intent information identified from the user input information; on the other hand, the target scene description is determined based on the scene description information of the target primary scene and the target secondary scene; and on the third hand, the user questioning pattern is determined from the user input information. Finally, the user request information, the target scene description, and the user questioning pattern are all input into the large model to obtain the toolset to be called in the secondary scene and the user task step description output by the large model.

[0090] The dynamic task execution method based on large models provided in this application supports the selection of toolsets to be invoked in secondary scenarios based on user questioning patterns based on user dialogue and common questioning methods. This enables the acquisition of the optimal toolset solution, while ensuring that the target tool exists in the toolset. It also makes the toolset carried in the Prompt of the large model as concise as possible during the process orchestration, effectively controlling the scope of tools and thus improving the accuracy of the large model.

[0091] In one embodiment, combined with Figure 2 As shown, in implementing the dynamic task execution method based on a large model provided in this application, the large model includes a plan initialization agent, which is invoked when an initial execution plan needs to be formulated in the scenario. This initial execution plan serves as input information for subsequent process orchestration, allowing the process orchestration agent to select the next execution tool based on both the initial plan and the current tool's execution results.

[0092] Overall, the dynamic task execution method based on a large model provided in this application divides tasks into multi-level scenarios within the workflow scheduling and orchestration based on the large model, and gradually selects and executes tools. This multi-level dynamic orchestration scheme facilitates scenario classification, improves the accuracy of scenario selection when there are many scenarios, and has good scalability. By introducing reflection functions at different levels, the workflow orchestration is dynamically adjusted in real time, avoiding the waste of time and model resources. At the same time, by selecting the toolset based on user dialogue and common questioning methods, the scope of tools is effectively controlled, improving the accuracy of model answers. By using a multi-agent architecture to build an end-to-end question-answering system framework, user input information is processed more effectively and accurately, and system execution results are returned.

[0093] Figure 4 This is a schematic diagram of the structure of the dynamic task execution device based on a large model provided in this application, such as... Figure 4As shown, the dynamic task execution device based on the large model includes, but is not limited to, a two-level scene determination module 401, a tool step determination module 402, and a tool step execution module 403.

[0094] The two-level scene determination module 401 is used to input the user request information of the user requesting to execute the task into the large model to obtain the target first-level scene and the target second-level scene output by the large model; the target first-level scene is determined from the pre-registered set of first-level scenes; the target second-level scene is determined from the set of second-level scenes associated with the target first-level scene.

[0095] The tool step determination module 402 is used to input the user request information and the target scene description into the large model to obtain the secondary scene to be called toolset and the corresponding user task step description output by the large model; the target scene description is determined based on the target primary scene description and the target secondary scene description.

[0096] The tool step execution module 403 is used to call the secondary scenario tool set to be called and execute the task steps corresponding to the user task step description.

[0097] It should be noted that the dynamic task execution device based on a large model provided in this application can execute the dynamic task execution method based on a large model described in any of the above embodiments during actual operation, which will not be elaborated in this embodiment.

[0098] The dynamic task execution device based on a large model provided in this application selects tools and executes task steps by setting up multi-level scenarios. For user-requested tasks, it identifies primary and secondary scenarios, then determines the corresponding set of tools to be invoked for each secondary scenario. This allows for task execution by invoking tools based on the secondary scenario as the basic unit. In case of execution errors, it eliminates the need to wait for the entire execution plan to complete; instead, it dynamically adjusts the execution plan when an error occurs in the task corresponding to the secondary scenario. This enables precise scenario matching and flexible tool scheduling and orchestration. Furthermore, defining multiple scenarios improves the accuracy of scenario selection when there are many scenarios. By defining multiple scenarios and adding scenario descriptions and tool information for primary and secondary scenarios, the large model gains the ability to schedule and orchestrate new tasks and processes, exhibiting excellent scalability.

[0099] Figure 5 This is a schematic diagram of the structure of the electronic device provided in this application, such as... Figure 5As shown, the electronic device may include: a processor 510, a communications interface 550, a memory 530, and a communication bus 540, wherein the processor 510, the communications interface 550, and the memory 530 communicate with each other through the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute the dynamic task execution method based on a large model provided in any of the above embodiments. The dynamic task execution method based on a large model includes, but is not limited to, the following steps: inputting user request information requesting the user to execute a task into the large model to obtain a target primary scene and a target secondary scene output by the large model; the target primary scene is determined from a pre-registered set of primary scenes; the target secondary scene is determined from a set of secondary scenes associated with the target primary scene; inputting the user request information and the target scene description into the large model to obtain a set of secondary scene tools to be called and a corresponding user task step description output by the large model; the target scene description is determined based on the target primary scene description and the target secondary scene description; calling the set of secondary scene tools to be called to execute the task steps corresponding to the user task step description.

[0100] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0101] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the dynamic task execution method based on a large model provided in any of the above embodiments. The dynamic task execution method based on a large model includes, but is not limited to, the following steps: inputting user request information requesting a task to be executed into a large model to obtain a target primary scene and a target secondary scene output by the large model; the target primary scene is determined from a pre-registered set of primary scenes; the target secondary scene is determined from a set of secondary scenes associated with the target primary scene; inputting the user request information and the target scene description into the large model to obtain a set of secondary scene tools to be invoked and a corresponding user task step description output by the large model; the target scene description is determined based on the target primary scene description and the target secondary scene description; invoking the set of secondary scene tools to be invoked to execute the task steps corresponding to the user task step description.

[0102] In another aspect, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the dynamic task execution method based on a large model provided in any of the above embodiments. The dynamic task execution method based on a large model includes, but is not limited to, the following steps: inputting user request information requesting a user to execute a task into a large model to obtain a target primary scene and a target secondary scene output by the large model; the target primary scene is determined from a pre-registered set of primary scenes; the target secondary scene is determined from a set of secondary scenes associated with the target primary scene; inputting the user request information and the target scene description into the large model to obtain a set of secondary scene tools to be invoked and a corresponding user task step description output by the large model; the target scene description is determined based on the target primary scene description and the target secondary scene description; invoking the set of secondary scene tools to be invoked to execute the task steps corresponding to the user task step description.

[0103] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0104] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A dynamic task execution method based on a large model, characterized in that, include: The user request information for the user requesting to execute the task is input into the large model to obtain the target primary scene and target secondary scene output by the large model; The target primary scene is determined from a pre-registered set of primary scenes; the target secondary scene is determined from a set of secondary scenes associated with the target primary scene. The user request information and target scene description are input into the large model to obtain the secondary scene toolset to be invoked and the corresponding user task step description output by the large model; the target scene description is determined based on the target primary scene description and the target secondary scene description. The secondary scenario toolset is invoked to execute the task steps corresponding to the user task step description.

2. The dynamic task execution method based on a large model according to claim 1, characterized in that, The process of inputting user request information for executing a task into a large model to obtain the target primary scene and target secondary scene output by the large model includes: The user request information is input into the large model to obtain the target primary scene output by the large model; The user request information and the target primary scene description are input into the large model to obtain the target secondary scene output by the large model.

3. The dynamic task execution method based on a large model according to claim 2, characterized in that, The secondary scenario toolkit includes multiple tools to be invoked. Invoking the secondary scenario toolkit and executing the task steps corresponding to the user task step description includes: The set of tools to be invoked in the secondary scenario and the description of the user task steps are input into the large model to obtain the target tools to be invoked output by the large model; the target tools to be invoked are determined from multiple tools to be invoked in the set of tools to be invoked in the secondary scenario. The target tool to be invoked is executed to perform the task steps corresponding to the user task step description, and the tool execution result and tool execution information are obtained. The tool execution information is input into the large model to obtain the process judgment result output by the large model; If the process judgment result is "process termination" and the tool execution result is "execution failure," the tool execution information, the user request information, and the target primary scene description are input into the large model to obtain a new target secondary scene output by the large model. The user request information and the new target scene description are then input into the large model to obtain a new set of tools to be called for the secondary scene and the corresponding user task step description output by the large model. The process then returns to the step of inputting the set of tools to be called for the secondary scene and the user task step description into the large model. The new target scene description is determined based on the target primary scene description and the new target secondary scene description.

4. The dynamic task execution method based on a large model according to claim 3, characterized in that, The step of invoking the secondary scenario's toolset and executing the task steps corresponding to the user task step description further includes: If the process judgment result is that the process is terminated and the tool execution result is that the execution is successful, the tool execution information is processed for information compliance to obtain the task execution result; The task execution result will be returned to the user.

5. The dynamic task execution method based on a large model according to any one of claims 1-4, characterized in that, The user request information is determined based on the following method: The initial input information of the user's request to execute the task is input into the large model to obtain the user's intent to be verified and its corresponding intent reliability output by the large model. When the reliability of the intent corresponding to the user's intent to be verified is less than a preset reliability threshold, the subsequent input information of the user's request to execute the task is obtained; Based on the subsequent input information and the initial input information, the complete dialogue information is determined; The complete dialogue information is input into the large model to obtain the complete user intent output by the large model. Based on the user's complete intent, the user request information is determined.

6. The dynamic task execution method based on a large model according to any one of claims 1-4, characterized in that, The step of inputting the user request information and target scene description into the large model to obtain the secondary scene tool set to be invoked and the corresponding user task step description output by the large model includes: The user request information, the target scene description, and the user question pattern are input into the large model to obtain the secondary scene tool set to be called and the corresponding user task step description output by the large model.

7. A dynamic task execution device based on a large model, characterized in that, include: The two-level scene determination module is used to input the user request information of the user requesting to execute the task into the large model, and obtain the target first-level scene and the target second-level scene output by the large model; The target primary scene is determined from a pre-registered set of primary scenes; the target secondary scene is determined from a set of secondary scenes associated with the target primary scene. The tool step determination module is used to input the user request information and the target scene description into the large model to obtain the secondary scene tool set to be called and the corresponding user task step description output by the large model; the target scene description is determined based on the target primary scene description and the target secondary scene description. The tool step execution module is used to call the secondary scenario tool set to be called and execute the task steps corresponding to the user task step description.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the dynamic task execution method based on a large model as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the dynamic task execution method based on a large model as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the dynamic task execution method based on a large model as described in any one of claims 1 to 6.

Images