Model-based task execution method

Abstract

The present description provides a model-based task execution method and apparatus, a device, and a readable storage medium. The method comprises: in response to an execution error of a target task for a target user request, acquiring a prompt word corresponding to the target user request, and acquiring a sub-task sequence which is generated by a task orchestration model performing logical reasoning on the target task on the basis of the prompt word; on the basis of the target user request and the sub-task sequence, determining a target factor corresponding to the execution error of the target task from among preset task execution error factors; adjusting the prompt word on the basis of a target adjustment strategy corresponding to the target factor, so as to obtain an adjusted new prompt word; and acquiring a new sub-task sequence which is generated by the task orchestration model performing logical reasoning on the target task on the basis of the new prompt word, and re-executing the target task on the basis of the new sub-task sequence.

Description

Model-based task execution method Cross-reference to related applications

[0001] This application claims priority to the Chinese patent application No. 202411272624.9, filed on September 11, 2024, with the Chinese Patent Office, the content of which is incorporated herein by reference in its entirety. TECHNICAL FIELD

[0002] Embodiments of the present application relate to, but are not limited to, the vehicle technical field, and in particular, but not limited to, a model-based task execution method. BACKGROUND

[0003] An intelligent cockpit system can refer to an in-vehicle integrated digital platform capable of integrating various IT and artificial intelligence technologies, which can provide intelligent experience for drivers and promote vehicle safety. SUMMARY

[0004] The following is a summary of the subject matter described in detail herein. This summary is not intended to limit the scope of the claims. The present application provides a model-based task execution method, which comprises: in response to a target task execution error requested by a target user, obtaining a prompt word corresponding to the target user request, and obtaining a sub-task sequence generated by a task arrangement model based on logical reasoning of the target task based on the prompt word; based on the target user request and the sub-task sequence, determining a target factor corresponding to the target task execution error from a preset task execution error factor; according to the target adjustment strategy corresponding to the target factor, adjusting the prompt word to obtain a new prompt word after adjustment; obtaining a new sub-task sequence generated by the task arrangement model based on logical reasoning of the target task based on the new prompt word, and re-executing the target task based on the new sub-task sequence.

[0005] In some embodiments, based on the target user request and the sub-task sequence, determining a target factor corresponding to the target task execution error from a preset task execution error factor comprises: inputting the target user request and the sub-task sequence into a classification model to perform classification calculation by the classification model to generate a category text corresponding to the target factor of the target task execution error; wherein the classification model is obtained by supervised learning training based on historical user requests and historical sub-task sequences corresponding to the historical user requests as training data, and the preset task execution error factor as a label.

[0006] In some embodiments, the preset task execution error factor includes at least one of the following factors: user request content error; parameter setting error in the prompt word; interface call error; task execution result does not meet user expectation.

[0007] In some embodiments, the adjustment strategy corresponding to the preset task execution error factor includes at least one of the following strategies: content verification is performed on the text corresponding to the user request, and the prompt word is adjusted based on the verification result; at least part of the parameters in the prompt word is adjusted; example guide information is added in the prompt word, and / or the parameters of the prompt word are adjusted according to the intent information corresponding to the user request; the intent information corresponding to the user request is added in the prompt word, and / or the parameters of the prompt word are adjusted according to the intent information.

[0008] In some embodiments, the task orchestration model is an LLM model; and the LLM model includes a Function Calling model.

[0009] In some embodiments, the task orchestration model is a model obtained by performing supervised fine-tuning training on a pre-trained task orchestration model, taking a preset user request and a prompt word sample corresponding to the preset user request as input variables, and taking a preset sub-task sequence corresponding to the preset user request as a label; the preset user request includes a request initiated for a service provided by an intelligent cockpit system; and the preset sub-task sequence includes at least one sub-task for calling the service provided by the intelligent cockpit system.

[0010] In some embodiments, before responding to a target task execution error of a target user request, the method further includes: obtaining feedback information corresponding to the target task during a target task execution process of the target user request; determining whether the feedback information contains indication information indicating a task execution error; and determining that the target task execution error occurs if the feedback information contains the indication information indicating the task execution error.

[0011] In some embodiments, the method is applied to an intelligent cockpit system of a vehicle, and the method further includes: performing intent recognition on the target user request to generate a prompt word corresponding to the target user request when the target user request is received, wherein the target user request is a request initiated for a service provided by the intelligent cockpit system; and inputting the prompt word and the target user request into the task orchestration model.

[0012] In some embodiments, the target task is a task generated by the intelligent cockpit system in response to the target user request.

[0013] The application also provides a model-based task execution device, which comprises: an acquisition unit configured to acquire a prompt word corresponding to a target user request in response to a target task execution error requested by the target user request, and acquire a subtask sequence generated by a task arrangement model based on logical reasoning of the target task based on the prompt word; a determination unit configured to determine a target factor corresponding to the target task execution error from preset task execution error factors based on the target user request and the subtask sequence; an adjustment unit configured to adjust the prompt word according to a target adjustment strategy corresponding to the target factor to obtain a new prompt word after adjustment; and an execution unit configured to acquire a new subtask sequence generated by the task arrangement model based on logical reasoning of the target task based on the new prompt word, and re-execute the target task based on the new subtask sequence.

[0014] The application also provides an electronic device comprising a communication interface, a processor, a memory and a bus, wherein the communication interface, the processor and the memory are connected to each other through the bus; the memory stores machine-readable instructions, and the processor executes the model-based task execution method by calling the machine-readable instructions.

[0015] The application also provides a machine-readable storage medium storing machine-readable instructions, wherein the machine-readable instructions, when called and executed by a processor, implement the model-based task execution method.

[0016] The technical scheme provided by the application can at least have the following beneficial effects: through the above embodiments, the intelligent cockpit system can acquire a prompt word corresponding to a target user request in response to a target task execution error requested by the target user request, and acquire a subtask sequence generated by a task arrangement model based on logical reasoning of the target task based on the prompt word; a target factor corresponding to the target task execution error can be determined from preset task execution error factors based on the target user request and the subtask sequence; the prompt word can be adjusted according to a target adjustment strategy corresponding to the target factor to obtain a new prompt word after adjustment; a new subtask sequence generated by the task arrangement model based on logical reasoning of the target task based on the new prompt word can be acquired, and the target task can be re-executed based on the new subtask sequence. Since the target factor indicates the influencing factor of the task execution error, the prompt word can be adjusted according to the target adjustment strategy corresponding to the influencing factor, thereby improving the accuracy of the prompt word adjustment and the success rate of the task execution. Other aspects can be understood after reading and understanding the drawings and the detailed description. BRIEF DESCRIPTION OF DRAWINGS

[0017] The accompanying drawings are used to provide a further understanding of the technical solutions of the present application, constitute a part of the specification, and are used to explain the technical solutions of the present application together with the embodiments of the present application, and do not constitute a limitation on the technical solutions of the present application.

[0018] FIG. 1 is a flowchart illustrating a model-based task execution method according to an exemplary embodiment.

[0019] FIG. 2 is a flowchart illustrating another model-based task execution method according to an exemplary embodiment.

[0020] FIG. 3 is a schematic diagram illustrating fine-tuning training of a task orchestration model according to an exemplary embodiment.

[0021] FIG. 4 is a hardware structure diagram of an electronic device in which a model-based task execution apparatus is located according to an exemplary embodiment.

[0022] FIG. 5 is a block diagram of a model-based task execution apparatus according to an exemplary embodiment. DETAILED DESCRIPTION

[0023] The exemplary embodiments will be described in detail below with reference to the accompanying drawings. In the following description, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments are not meant to represent all implementations consistent with the present application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the present application as detailed in the appended claims.

[0024] It should be noted that the steps of the corresponding method are not necessarily performed in the order shown and described in the specification in other embodiments. In some other embodiments, the steps included in the method can be more or less than those described in the specification. In addition, a single step described in the specification can be divided into multiple steps for description in other embodiments, and multiple steps described in the specification can be combined into a single step for description in other embodiments.

[0025] At present, for a user request initiated by a user to an intelligent cockpit system, the task corresponding to the request can usually be processed based on a model, and when a task execution error occurs, the user request and the output result of the model can be analyzed, and the prompt word of the model can be adjusted based on the analysis result, and the task can be reprocessed based on the adjusted prompt word and the user request. Due to the lack of guidance information in the above analysis process, the accuracy of the adjustment of the prompt word is not high, so that the adaptation degree of the output result of the model to the task is not high, and thus the success rate of the task is not high.

[0026] Therefore, the present application aims to propose a model-based task execution method. The method determines the factors causing the target task execution error first, and then adjusts part of the prompt words in the prompt words based on the adjustment strategy corresponding to the factors, thereby improving the accuracy of prompt word adjustment.

[0027] In implementation, in response to a target task execution error for a target user request, a prompt word corresponding to the target user request is obtained, and a subtask sequence generated by a task arrangement model based on logical reasoning of the target task by the prompt word is obtained; based on the target user request and the subtask sequence, a target factor corresponding to the target task execution error is determined from preset task execution error factors; the prompt word is adjusted according to a target adjustment strategy corresponding to the target factor to obtain a new prompt word after adjustment; a new subtask sequence generated by the task arrangement model based on logical reasoning of the target task by the new prompt word is obtained, and the target task is re-executed based on the new subtask sequence.

[0028] Through the above embodiments, the intelligent cockpit system can obtain a prompt word corresponding to a target user request in response to a target task execution error for the target user request, and obtain a subtask sequence generated by a task arrangement model based on logical reasoning of the target task by the prompt word; based on the target user request and the subtask sequence, a target factor corresponding to the target task execution error can be determined from preset task execution error factors; the prompt word is adjusted according to a target adjustment strategy corresponding to the target factor to obtain a new prompt word after adjustment; a new subtask sequence generated by the task arrangement model based on logical reasoning of the target task by the new prompt word is obtained, and the target task is re-executed based on the new subtask sequence. Since the target factor indicates the cause of the task execution error, the target adjustment strategy corresponding to the target factor can be used to adjust the prompt word, thereby improving the accuracy of prompt word adjustment and the success rate of task execution.

[0029] The present application will be described below through specific embodiments and in combination with specific application scenarios.

[0030] Please refer to FIG. 1, which is a system architecture diagram of a model-based task execution method according to an exemplary embodiment. The intelligent cockpit system can be equipped with a task arrangement model, and the intelligent cockpit system can include a reflection subsystem built based on a ReAct framework. The reflection subsystem can be equipped with a classification model. The ReAct framework is a design framework for enhancing the reflection and self-adjustment capabilities of an intelligent agent. It combines the "reaction" (React) and "reflection" (Reflect) aspects, and can improve the performance of the intelligent agent in complex tasks by dynamically adjusting and optimizing the behavior of the intelligent agent. In this application, the framework can be used to optimize the prompt words of the task arrangement model.

[0031] Based on this, when the intelligent cockpit system receives a target user request, the target user request can be subjected to intent recognition, a prompt word corresponding to the target user request can be generated, and the prompt word and the target user request can be input into the task arrangement model. The task arrangement model can arrange a target task for the target user request based on the prompt word and the target user request, and generate a subtask sequence. The subtask sequence can include at least one to-be-called interface and an interface parameter corresponding to the interface. The at least one to-be-called interface can correspond to at least one subtask in the subtask sequence. The subtask sequence can indicate an execution order corresponding to the at least one subtask.

[0032] The execution module equipped in the intelligent cockpit system can obtain the subtask sequence, and sequentially call the interfaces corresponding to the at least one subtask according to the execution order corresponding to the at least one subtask indicated by the subtask sequence, to execute the target task.

[0033] If the target task is executed, indication information indicating the task execution result can be generated, and the indication information can be returned to the user.

[0034] The intelligent cockpit system can obtain the feedback information of the user within a preset time period after the target task is executed. If the feedback information of the user indicates that the user is satisfied or no feedback information of the user is received within the preset time period, and no feedback information indicating that the target task execution has an error is received from other modules equipped, it can be indicated that the target task execution is successful. If any feedback information indicating that the task execution has an error is received, it can be indicated that the target task execution has an error.

[0035] In the target task execution process, when the target task execution fails, the intelligent cockpit system can obtain the prompt word corresponding to the target user request and the above-mentioned sub-task sequence; can determine the target factor corresponding to the target task execution failure from the pre-set task execution failure factors based on the target user request and the sub-task sequence; and can adjust the prompt word according to the target adjustment strategy corresponding to the target factor by the mounted optimization module to obtain the new prompt word after adjustment; can input the new prompt word and the above-mentioned target user request into the task arrangement model again, and the task arrangement model generates a new sub-task sequence by re-arranging the target task for the target user request based on the new prompt word and the target user request.

[0036] The intelligent cockpit system can obtain the new sub-task sequence, and can re-execute the target task based on the new sub-task sequence.

[0037] The intelligent cockpit system can input the target user request and the sub-task sequence into the classification model to generate a category text corresponding to the target task execution failure factor by classification calculation of the classification model; wherein the classification model can be obtained by supervised learning training based on historical user requests and historical sub-task sequences corresponding to the historical user requests as training data, and pre-set task execution failure factors as labels.

[0038] The above-mentioned target task execution failure factor can include at least one of the following factors: user request content error; parameter setting error in the prompt word; interface calling error; task execution result does not meet user expectation.

[0039] The adjustment strategy corresponding to the task execution failure factor includes at least one of the following strategies: content verification of the text corresponding to the user request, and adjustment of the prompt word based on the verification result; adjustment of at least part of the parameters in the prompt word; adding example guide information in the prompt word, and / or adjusting the parameters of the prompt word according to the intent information corresponding to the user request; adding the intent information corresponding to the user request in the prompt word, and / or adjusting the parameters of the prompt word according to the intent information.

[0040] The task arrangement model can be an LLM (Large Language Model) model, and the LLM model can include a Function Calling model.

[0041] The task arrangement model can be obtained by performing supervised fine-tuning training on a pre-trained task arrangement model, taking a preset user request and a prompt word sample corresponding to the preset user request as input variables, and taking a preset subtask sequence corresponding to the preset user request as a label. The preset user request includes a request initiated for a service provided by the intelligent cockpit system. The preset subtask sequence includes at least one subtask for calling the service provided by the intelligent cockpit system.

[0042] The intelligent cockpit system can further obtain feedback information corresponding to the target task during execution of the target task for the target user request before responding to the execution error of the target task for the target user request. It is determined whether the feedback information contains indication information indicating a task execution error. If the feedback information contains indication information indicating a task execution error, it can be determined that the target task execution error.

[0043] It should be noted that at least part of the above steps can also be performed by a model, which can be set according to actual needs, and the present application does not limit this.

[0044] FIG. 2 is a flowchart illustrating a model-based task execution method according to an exemplary embodiment. As shown in FIG. 2, the above-mentioned intelligent cockpit system can perform steps 202-208.

[0045] Step 202, in response to the target task execution error for the target user request, obtaining a prompt word corresponding to the target user request, and obtaining a subtask sequence generated by a task arrangement model based on logical reasoning of the target task based on the prompt word.

[0046] The above-mentioned intelligent cockpit system can provide several services to the user, and the above-mentioned user request can be a request initiated for the service provided by the intelligent cockpit system. The target task can refer to a task generated by the intelligent cockpit system in response to the user request.

[0047] The specific content of the above-mentioned several services can be set according to actual needs, and the present application does not limit this. For example, the service can refer to an entertainment service, a communication service, or a driving assistance service.

[0048] The above-mentioned task arrangement model can refer to a model that can perform logical reasoning on a target task based on a prompt word, split the target task into at least one subtask, and perform task arrangement on the at least one subtask to generate a subtask sequence containing at least one subtask.

[0049] The intelligent cockpit system can execute a target task corresponding to the target user request based on a task scheduling model. Specifically, the intelligent cockpit system can perform intent recognition on the target user request, generate a prompt word corresponding to the target user request, and use the prompt word to assist in executing the target task. The intelligent cockpit system can input the prompt word and the target user request into the task scheduling model, perform logical reasoning on the prompt word and the target user request by the task scheduling model, and generate a subtask sequence.

[0050] The subtask sequence can include at least one interface to be called and an interface parameter corresponding to the interface. The subtask sequence can also indicate an execution order corresponding to the at least one subtask. Therefore, the intelligent cockpit system can obtain the subtask sequence, and sequentially call the interfaces corresponding to the at least one subtask according to the execution order corresponding to the at least one subtask indicated by the subtask sequence, to execute the target task.

[0051] During the task execution process, when an error in executing the target task is monitored, the intelligent cockpit system can obtain the prompt word and the subtask sequence in response to the error in executing the target task, to facilitate subsequent adjustment of the prompt word.

[0052] The specific form of the model can be set according to actual needs, which is not limited in the present application.

[0053] In an embodiment, the task scheduling model can be an LLM model. The LLM model can include a Function Calling model.

[0054] An LLM model (Large Language Model) is a machine learning model capable of processing and generating natural language text, usually composed of a large-scale neural network. The LLM base model has been pre-trained and can understand and generate human language text, including syntax structure, semantic content, and context relevance. It can accurately simulate human language capabilities and perform tasks such as conversation, article generation, and question answering.

[0055] A Function Calling model is a programming model that allows a system to call specific functions or methods at runtime. It allows the encapsulation and reuse of functions by defining function interfaces. The model supports passing parameters, executing function logic, and returning results, making the code more modular and easier to maintain.

[0056] In the present application, the subtask sequence generated by the task scheduling model can include at least one interface to be called and a calling parameter corresponding to the interface, so that the intelligent cockpit system can directly call the interface to execute the target task.

[0057] In an embodiment shown, the task arrangement model is a model obtained by performing supervised fine-tuning training on a pre-trained task arrangement model, taking a preset user request and a prompt word sample corresponding to the preset user request as input variables, and taking a preset subtask sequence corresponding to the preset user request as a label. The preset user request includes a request initiated for a service provided by the intelligent cockpit system. The preset subtask sequence includes at least one subtask for invoking the service provided by the intelligent cockpit system.

[0058] Since the task arrangement model can be used to process tasks corresponding to services provided by the intelligent cockpit system, in order to improve the accuracy of logical reasoning of the task arrangement model, the task arrangement model can be fine-tuned.

[0059] Specifically, a preset user request initiated for a service provided by the intelligent cockpit system and a prompt word sample corresponding to the preset user request are taken as input variables, and a preset subtask sequence corresponding to the preset user request and containing at least one subtask for invoking the service provided by the intelligent cockpit system is taken as a label, and supervised fine-tuning training is performed on a pre-trained task arrangement model.

[0060] For example, referring to FIG. 3, which is a schematic diagram of fine-tuning training of a task arrangement model according to an exemplary embodiment. The preset user request can include user request_1, user request_2, and user request_3, the prompt word corresponding to the user request_1 can be prompt word_1, and the corresponding subtask sequence can be subtask sequence_1, the prompt word corresponding to the user request_2 can be prompt word_2, and the corresponding subtask sequence can be subtask sequence_2, the prompt word corresponding to the user request_3 can be prompt word_3, and the corresponding subtask sequence can be subtask sequence_3. The pre-trained task arrangement model can be fine-tuned based on the three groups of training data and labels.

[0061] For example, when the user request is "I want to check my vehicle maintenance records", the prompt word corresponding to the user request can be "invoke the getMaintenanceRecords function, parameter: vehicle number = 123ABC, date = 2024", and the corresponding sub-task sequence can include "getMaintenanceRecords(vehicleid='123ABC', date='2024')"; when the user request is "start navigation to the nearest gas station", the prompt word corresponding to the user request can be "invoke the startNavigation function, parameter: destination = nearest gas station", and the corresponding sub-task sequence can include "startNavigation(destination='nearest gas station')"; when the user request is "book the first train from Shanghai to Beijing tomorrow", the prompt word corresponding to the user request can be "invoke the bookFirstTrain function, parameter: origin = Shanghai, destination = Beijing, date = 2024-07-03", and the corresponding sub-task sequence can include "bookFirstTrain(origin='Shanghai', destination='Beijing', date='2024-07-03')".

[0062] In this way, the accuracy of logical reasoning of the task orchestration model on the task corresponding to the service provided by the intelligent cockpit system can be improved.

[0063] In an embodiment shown, before responding to the target task execution error for the target user request, the intelligent cockpit system can obtain feedback information corresponding to the target task during the target task execution process for the target user request; determine whether the feedback information contains indication information indicating a task execution error; and if the feedback information contains indication information indicating a task execution error, determine that the target task execution error.

[0064] The target task execution process refers to a process from receiving a user request to receiving feedback information indicating user satisfaction or no feedback information from the user within a preset period after the target task is executed, that is, the target task for the target user request is successfully executed or the execution time exceeds the preset time to indicate the end of the task.

[0065] For example, when the task is not completed, the target task is in execution, the intelligent cockpit system can obtain feedback information corresponding to the target task, which can be feedback information sent by a module carried by the intelligent cockpit system to indicate that the target task execution fails, can be feedback information indicating that the user is not satisfied, and can also include other feedback information indicating that the task execution fails, which is not limited in the present application. The intelligent cockpit system can determine whether the feedback information contains indication information indicating that the task execution fails, and if the feedback information contains indication information indicating that the task execution fails, it can be determined that the target task execution fails.

[0066] In this way, subsequent remedial measures can be taken when the task execution fails, thereby improving the success rate of task execution and improving the user's experience.

[0067] Step 204, based on the target user request and the sub-task sequence, determining a target factor corresponding to the target task execution failure from the pre-set task execution failure factors.

[0068] When it is determined that the target task execution fails, the target user request and the output result of the model can be analyzed to determine which factor of the pre-set task execution failure factors the influencing factor of the target task execution failure belongs to, and the prompt word can be improved in subsequent targeted.

[0069] The specific way of determining the target factor corresponding to the target task execution failure can be set according to actual needs, which is not limited in the present application.

[0070] In an embodiment shown, the intelligent cockpit system can input the target user request and the sub-task sequence into a classification model to perform classification calculation by the classification model to generate a category text corresponding to the target factor of the target task execution failure; wherein the classification model is obtained by supervised learning training based on historical user requests and historical sub-task sequences corresponding to the historical user requests as training data, and the pre-set task execution failure factors as labels.

[0071] The intelligent cockpit system can determine the target factor corresponding to the target task execution failure based on the classification model. The classification model can be obtained by supervised learning training based on historical user requests and historical sub-task sequences corresponding to the historical user requests as training data, and the pre-set task execution failure factors as labels, so as to accurately classify and calculate the target factor corresponding to the target task execution failure.

[0072] For example, the target user request can be "find and book the nearest flight", the prompt word corresponding to the user request can be "call the searchFlights function, parameters: departure place = Shanghai, destination = Beijing, date = 2024-05-30", and the sub-task sequence corresponding to the user request can include "searchFlights ('Shanghai', 'Beijing', '2024-05-30')". Based on the target user request and the sub-task sequence, the category text corresponding to the target task execution error factor can be generated, and the target factor can be "the result is not as expected".

[0073] In this way, the accuracy of determining the target factor corresponding to the target task execution error can be improved.

[0074] The specific factors included in the preset task execution error factor can be set according to actual needs, and the present application does not limit this.

[0075] In an embodiment shown, the task execution error factor can include at least one of the following factors: user request content error; parameter setting error in the prompt word; interface call error; task execution result not meeting user expectations.

[0076] The user request content error can refer to incomplete or inaccurate information in the user request.

[0077] For example, the target user request can be "turn on the air conditioner", and at this time, the user request does not include the temperature parameter information of the air conditioner, and the information of the target user request is incomplete.

[0078] The parameter setting error in the prompt word can refer to inaccurate interface parameter setting of the callable interface included in the prompt word.

[0079] For example, the target user request can be "find and book the nearest flight", and the prompt word corresponding to the user request can be "call the searchFlights function, parameters: departure place = Shanghai, destination = Beijing, date = 2024-05-30". Since the date is set to a specific date, and the user request expresses the nearest flight, the interface parameter setting is inaccurate.

[0080] The interface call error can refer to an error occurring in the process of calling the interface to be called included in the sub-task sequence by the intelligent cockpit system based on the sub-task sequence.

[0081] For example, the intelligent cockpit system reports an error when searching for flight information according to the output result of the task scheduling model.

[0082] The task execution result does not meet the user's expectation, which can refer to a result output by the intelligent cockpit system after completing the interface call does not meet the expectation.

[0083] For example, the target user request can be "find and book the nearest flight", the prompt word corresponding to the user request can be "call the searchFlights function, parameters: departure city=Shanghai, destination city=Beijing, date=2024-05-30", and the task execution result can be "no flight meeting the conditions can be found", which does not meet the user's expectation.

[0084] In this way, the influencing factors of task execution error can be subdivided, so that subsequent targeted optimization can be facilitated, thereby improving the success rate of task execution.

[0085] At step 206, the prompt word is adjusted according to the target adjustment strategy corresponding to the target factor to obtain a new adjusted prompt word.

[0086] The intelligent cockpit system determines the influencing factors of the target task execution error, and can adjust according to the target adjustment strategy corresponding to the target factor. Further, the prompt word can be adjusted to obtain a new adjusted prompt word, so that subsequent logical reasoning by the task arrangement model based on the new prompt word can be facilitated, thereby improving the success rate of task execution.

[0087] The intelligent cockpit system can determine the part of the prompt word to be optimized based on the target factor, and can adjust the part of the prompt word according to the target adjustment strategy corresponding to the target factor.

[0088] For example, the target user request can be "find and book the nearest flight", the prompt word corresponding to the user request can be "call the searchFlights function, parameters: departure city=Shanghai, destination city=Beijing, date=2024-05-30", and when the target factor is "the result does not meet the expectation", the part of the prompt word to be optimized can be determined as the setting parameter corresponding to the date, so the setting parameter of the date can be adjusted.

[0089] In an embodiment shown, the adjustment strategy corresponding to the task execution error factor can include at least one of the following strategies: content verification is performed on the text corresponding to the user request, and the prompt word is adjusted based on the verification result; at least part of the parameters in the prompt word is adjusted; example guide information is added in the prompt word, and / or the parameters of the prompt word are adjusted according to the intent information corresponding to the user request; intent information corresponding to the user request is added in the prompt word, and / or the parameters of the prompt word are adjusted according to the above-mentioned intent information.

[0090] For example, the content verification on the text corresponding to the user request and the adjustment of the prompt word based on the verification result can mean that the integrity or accuracy of the target user request is verified, and the temperature parameter information can be added in the prompt word based on the verification result. Adding example guide information in the prompt word can mean adding example information related to the user request in the prompt word. Adding intent information corresponding to the user request in the prompt word can mean adding information for detailed description of the intent of the user request in the prompt word.

[0091] It should be noted that, since the error reasons corresponding to different task execution error factors are different, for different task execution error factors, more targeted adjustment strategies can be used to adjust the prompt word, so as to improve the accuracy of the prompt word adjustment.

[0092] For example, the adjustment strategy corresponding to the above user request content error can be: content verification on the text corresponding to the user request, and adjustment of the prompt word based on the verification result; the adjustment strategy corresponding to the parameter setting error in the above prompt word can be: adjusting at least part of the parameters in the prompt word; the adjustment strategy corresponding to the above interface calling error can be: adding example guide information in the prompt word, and / or adjusting the parameters of the above prompt word according to the intent information corresponding to the user request; the adjustment strategy corresponding to the above task execution result not meeting the user's expectation can be: adding intent information corresponding to the user request in the prompt word, and / or adjusting the parameters of the above prompt word according to the intent information.

[0093] Since the user request content error can mean that the information in the user request is incomplete or inaccurate, the content of the text corresponding to the user request can be verified, and the prompt word can be adjusted based on the verification result.

[0094] For example, the target user request can be "turn on the air conditioner", at this time, the user request does not contain the temperature parameter information of the air conditioner, and the information of the target user request is incomplete. The integrity of the target user request can be verified, and the temperature parameter information can be added in the prompt word based on the verification result.

[0095] Since the parameter setting error in the prompt word can mean that the interface parameter setting of the callable interface contained in the prompt word is inaccurate, at least part of the parameters in the prompt word can be adjusted.

[0096] For example, the target user request can be "find and book the nearest flight", and the prompt word corresponding to the user request can be "call the searchFlights function, parameters: departure city = Shanghai, destination city = Beijing, date = 2024-05-30". Since the date is set to a specific date, and the user request expresses the nearest flight, the setting of the interface parameter is not accurate. In contrast, the setting parameter corresponding to the date can be adjusted to "date = 2024-05-30 to 2024-06-02".

[0097] Among them, since the interface call error can refer to the error of the intelligent cockpit system in calling the interface to be called contained in the sub-task sequence based on the sub-task sequence. Therefore, example guidance information can be added to the prompt word, and / or the parameters of the prompt word can be adjusted according to the above intention information.

[0098] For example, the target user request can be "find and book the nearest flight", and the prompt word corresponding to the user request can be "call the searchFlights function, parameters: departure city = Shanghai, destination city = Beijing, date = 2024-05-30". When the target factor corresponding to the target task execution error is "interface call error", the prompt word can be adjusted to: "call the searchFlights function, parameters: departure city = Shanghai, destination city = Beijing, date = 2024-05-30 to 2024-06-02, flight type = direct flight. The user wants to find the nearest direct flight, and can accept any flight within the date range. Example output: 'Flight number CA123, Shanghai-Beijing, 2024-05-30, direct flight'".

[0099] Among them, the task execution result does not meet the user's expectation, which can refer to the result output by the intelligent cockpit system after completing the call of the interface. Therefore, intention information corresponding to the user request can be added to the prompt word, and / or the parameters of the prompt word can be adjusted according to the above intention information.

[0100] For example, the target user request can be "find and book the nearest flight", and the prompt word corresponding to the user request can be "call the searchFlights function, parameters: departure city = Shanghai, destination city = Beijing, date = 2024-05-30". When the target factor corresponding to the target task execution error is "the result does not meet the expectation", the prompt word can be adjusted to: "call the searchFlights function, parameters: departure city = Shanghai, destination city = Beijing, date = 2024-05-30 to 2024-06-02, flight type = direct flight. The user wants to find the nearest direct flight, and can accept any flight within the date range".

[0101] In this way, the influence factors of task execution error can be optimized according to the subdivided tasks, thereby improving the success rate of task execution.

[0102] In step 208, a new sub-task sequence generated by the task arrangement model based on the new prompt word for logical reasoning of the target task is obtained, and the target task is re-executed based on the new sub-task sequence.

[0103] The intelligent cockpit system can input the new prompt word and the target user request into the task arrangement model, and the task arrangement model can generate a new sub-task sequence based on the new prompt word for logical reasoning of the target task. The intelligent cockpit system can obtain the new sub-task sequence, and can re-execute the target task based on the new sub-task sequence.

[0104] It should be noted that when the intelligent cockpit system is still executing based on the new sub-task sequence and the target task execution error still exists, the above steps 202 to 208 can be re-executed to adjust the new prompt word until the task execution is successful, or the task execution time exceeds the preset time, thereby improving the success rate of task execution.

[0105] Corresponding to the embodiment of the model-based task execution method, the present specification also provides an embodiment of a model-based task execution device.

[0106] Please refer to FIG. 4, which is a hardware structure diagram of an electronic device in which a model-based task execution device is located according to an exemplary embodiment. At the hardware level, the device includes a processor 402, an internal bus 404, a network interface 406, a memory 408, and a non-volatile memory 410, and of course, it can also include other hardware required by the business. One or more embodiments of the present specification can be implemented in a software manner, such as reading the corresponding computer program from the non-volatile memory 410 into the memory 408 by the processor 402 and then running. Of course, in addition to the software implementation, one or more embodiments of the present specification do not exclude other implementation manners, such as logic devices or a combination of software and hardware, etc., that is, the execution subject of the following processing flow is not limited to each logical unit, but can also be hardware or a logic device.

[0107] Please refer to FIG. 5, which is a block diagram of a model-based task execution apparatus according to an exemplary embodiment. The model-based task execution apparatus can be applied to the electronic device shown in FIG. 4 to implement the technical solutions of the present specification. Among them, the apparatus can include: an acquisition unit 502 configured to, in response to a target task execution error requested by a target user, acquire a prompt word corresponding to the target user request, and acquire a subtask sequence generated by a task arrangement model based on logical reasoning of the target task based on the prompt word; a determination unit 504 configured to determine a target factor corresponding to the target task execution error from a preset task execution error factor based on the target user request and the subtask sequence; an adjustment unit 506 configured to adjust the prompt word according to a target adjustment strategy corresponding to the target factor to obtain a new prompt word after adjustment; and an execution unit 508 configured to acquire a new subtask sequence generated by the task arrangement model based on logical reasoning of the target task based on the new prompt word, and re-execute the target task based on the new subtask sequence.

[0108] In the present embodiment, the determination unit 504 can be configured to: input the target user request and the subtask sequence into a classification model to generate a category text corresponding to the target factor of the target task execution error by classification calculation of the classification model; wherein the classification model is obtained by supervised learning training based on historical user requests and historical subtask sequences corresponding to the historical user requests as training data, and preset task execution error factors as labels.

[0109] In the present embodiment, the task execution error factor can include at least one of the following factors: user request content error; parameter setting error in the prompt word; interface calling error; task execution result not meeting user expectations.

[0110] In the present embodiment, the adjustment strategy corresponding to the task execution error factor can include at least one of the following strategies: content verification of the text corresponding to the user request, and adjustment of the prompt word based on the verification result; adjustment of at least part of the parameters in the prompt word; adding example guide information in the prompt word, and / or adjusting the parameters of the prompt word according to the intent information corresponding to the user request; adding intent information corresponding to the user request in the prompt word, and / or adjusting the parameters of the prompt word according to the intent information.

[0111] In the present embodiment, the task arrangement model can be an LLM model; and the LLM model can include a Function Calling model.

[0112] In the embodiment, the task arrangement model can be obtained by performing supervised fine-tuning training on the pre-trained task arrangement model, taking the preset user request and the prompt word sample corresponding to the preset user request as input variables, and taking the preset subtask sequence corresponding to the preset user request as a label. The preset user request can include a request initiated for a service provided by the intelligent cockpit system. The preset subtask sequence can include at least one subtask for calling the service provided by the intelligent cockpit system.

[0113] In the embodiment, the obtaining unit 502 is further configured to: before responding to a target task execution error for a target user request, obtain feedback information corresponding to the target task in a target task execution process for the target user request; determine whether the feedback information contains indication information indicating a task execution error; and if the feedback information contains the indication information indicating the task execution error, determine that the target task has an execution error.

[0114] The functions and effects of the units in the device are specifically described in the implementation process of the corresponding steps in the method, and will not be described here.

[0115] For the device embodiment, since it basically corresponds to the method embodiment, the related parts can be referred to the part of the method embodiment. The device embodiments described above are only illustrative, and the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., they can be located in one place or distributed on multiple network units. Some or all modules can be selected to achieve the purpose of the scheme of the present application according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0116] The system, device, module or unit illustrated in the embodiments can be specifically implemented by a computer chip or entity, or by a product with certain functions. A typical implementation device is a computer, and the specific form of the computer can be a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email transceiver device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0117] In a typical configuration, a computer includes one or more processors (CPUs), input / output interfaces, network interfaces, and memories.

[0118] Memory can include, without being limited to, non- transitory storage in computer-readable media, random access memory (RAM), and / or read-only memory (ROM), such as flash memory, etc. Memory is an example of computer-readable media.

[0119] Computer-readable media includes permanent and non- permanent, removable and non-removable media implemented in any method or technology for storage of information such as computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, without being limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disks (DVDs) or other optical storage, magnetic cassettes, magnetic disks storage, quantum memory, graphene-based storage media, or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition provided herein, a computer-readable medium does not include transitory computer-readable media, such as modulated data signals and carrier waves.

[0120] It is also important to note that the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises... a" does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0121] The specific embodiments described herein have been chosen for purposes of illustration. Other embodiments are within the scope of the following claims. In some instances, no actions or steps can be performed, even if they have been described as actions or steps within exemplary embodiments. In some instances, the order of actions or steps can be different from the order in which the actions or steps are described. Moreover, not all processes or steps can be required to achieve the desired result. In certain embodiments, multitasking and parallel processing can be advantageous.

[0122] The terminology used in this description of one or more embodiments should not be interpreted as limiting of the description of one or more embodiments. The description of one or more embodiments and the accompanying drawings that form a part of this detailed description are given by way of illustration only and not of limitation. The use of the terms "a" or "an" and "the" or "said" preceding articles are intended to be interpreted as including both the singular and the plural, unless otherwise indicated by context. It will also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0123] It will be understood that, although the terms first, second, third, etc. can be used herein to describe various information, these terms are not intended to denote a temporal or chronological order. Rather, these terms are used only as a shorthand notation to distinguish one item from another. For example, in the absence of a departing from the scope of one or more embodiments, a first item can also be termed a second item, and similarly, a second item can also be termed a first item. Depending on the context, the word "if' as used herein can be interpreted to mean "when" or "in response to determining" or "in response to a determination".

[0124] The above description of one or more embodiments is not intended to limit the description of one or more embodiments. Any modifications, equivalents, improvements, and the like made during the spirit and principles of one or more embodiments should be included within the scope of protection of one or more embodiments.

[0125] The user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation portal for user to choose authorization or refusal.

Claims

1. A model-based task execution method, comprising: in response to a target task execution error requested by a target user, obtaining a prompt word corresponding to the target user request, and obtaining a sub-task sequence generated by a task arrangement model based on logical reasoning of the target task according to the prompt word; determining a target factor corresponding to the target task execution error from a preset task execution error factor based on the target user request and the sub-task sequence; adjusting the prompt word according to a target adjustment strategy corresponding to the target factor to obtain a new prompt word after adjustment; obtaining a new sub-task sequence generated by the task arrangement model based on logical reasoning of the target task according to the new prompt word, and re-executing the target task based on the new sub-task sequence.

2. The method of claim 1, wherein, determining a target factor corresponding to the target task execution error from a preset task execution error factor based on the target user request and the sub-task sequence, comprising: inputting the target user request and the sub-task sequence into a classification model to generate a category text corresponding to the target factor of the target task execution error by classification calculation of the classification model; wherein the classification model is obtained by supervised learning training based on historical user requests and historical sub-task sequences corresponding to the historical user requests as training data, and preset task execution error factors as labels.

3. The method of claim 1 or 2, wherein, The preset task execution error factor includes at least one of the following factors: user request content error; parameter setting error in the prompt word; interface calling error; task execution result does not meet user expectation.

4. The method of claim 3, wherein, The adjustment strategy corresponding to the preset task execution error factor includes at least one of the following strategies: content verification is performed on the text corresponding to the user request, and the prompt word is adjusted based on the verification result; adjust at least part of the parameters in the prompt word; adding example guide information in the prompt word, and / or adjusting the parameters of the prompt word according to the intent information corresponding to the user request; adding intent information corresponding to the user request in the prompt word, and / or adjusting the parameters of the prompt word according to the intent information.

5. The method according to any one of claims 1 to 4, wherein, The task arrangement model is an LLM model; the LLM model includes a Function Calling model.

6. The method according to any one of claims 1 to 5, wherein, The task arrangement model is a model obtained by supervised fine-tuning training on a pre-trained task arrangement model, taking a preset user request and a prompt word sample corresponding to the preset user request as input variables, and taking a preset sub-task sequence corresponding to the preset user request as a label; wherein the preset user request includes a request initiated for a service provided by an intelligent cockpit system; the preset sub-task sequence includes at least one sub-task for calling the service provided by the intelligent cockpit system.

7. The method according to any one of claims 1 to 6, wherein, Before responding to the target task execution error requested by the target user, the method further comprises: obtaining feedback information corresponding to the target task during the target task execution process for the target user request; determine whether the feedback information contains indication information indicating a task execution error; if the feedback information contains the indication information indicating the task execution error, determine that the target task execution error exists.

8. The method according to any one of claims 1 to 7, wherein, The method is applied to an intelligent cockpit system of a vehicle, and the method further includes: performing intent recognition on the target user request to generate a prompt word corresponding to the target user request when the target user request is received, wherein the target user request is a request initiated for a service provided by the intelligent cockpit system; inputting the prompt word and the target user request into the task orchestration model.

9. The method of claim 8, wherein, The target task refers to a task generated by the intelligent cockpit system in response to the target user request.

10. A model-based task execution apparatus, comprising: an acquisition unit configured to, in response to a target task execution error for a target user request, acquire a prompt word corresponding to the target user request, and acquire a subtask sequence generated by a task orchestration model based on logical reasoning of a target task by using the prompt word; a determination unit configured to determine, based on the target user request and the subtask sequence, a target factor corresponding to the target task execution error from preset task execution error factors; an adjustment unit configured to adjust the prompt word according to a target adjustment strategy corresponding to the target factor to obtain a new prompt word after adjustment; an execution unit configured to acquire a new subtask sequence generated by the task orchestration model based on logical reasoning of the target task by using the new prompt word, and re-execute the target task based on the new subtask sequence.

11. An electronic device, comprising a communication interface, one or more processors, a memory, and a bus, the communication interface, the one or more processors, and the memory being connected to each other through the bus; the memory stores machine-readable instructions, and the one or more processors execute the model-based task execution method according to any one of claims 1 to 9 by invoking the machine-readable instructions.

12. A machine-readable storage medium, wherein, The machine-readable storage medium stores machine-readable instructions, and the machine-readable instructions, when invoked and executed by one or more processors, implement the model-based task execution method according to any one of claims 1 to 9.

Images