Information processing method and device based on large model, electronic equipment and storage medium

Abstract

The invention provides an information processing method and device based on a large model, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence, in particular to the technical fields of large models, smart questions and answers, smart search, smart medical treatment and the like. The information processing method based on the large model comprises the following steps: performing task arrangement on demand information related to a target object to obtain a task arrangement strategy containing a plurality of initial tasks; in the process of executing the task arrangement strategy by using the large model, updating at least one initial task in the task arrangement strategy according to the interactive operation of the target object to obtain a target task; and determining feedback information matched with the demand intention of the target object according to a target execution result determined by executing the target task by using the large model.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, and in particular to technical fields such as large models, intelligent question and answer, intelligent search, and intelligent medical care. Background Art

[0002] A Large Language Model (LLM) is a model built using deep learning algorithms that can perform generative tasks based on natural language understanding. It can be used to process user input questions and generate responses based on natural language expressions to meet user needs. Summary of the Invention

[0003] The present disclosure provides a large model-based information processing method, device, electronic device and storage medium.

[0004] According to one aspect of the present disclosure, a big model-based information processing method is provided, comprising: performing task scheduling on demand information related to a target object to obtain a task scheduling strategy containing multiple initial tasks; in the process of executing the task scheduling strategy using the big model, updating at least one initial task in the task scheduling strategy according to the interactive operation of the target object to obtain a target task; and determining feedback information that matches the demand intention of the target object based on the target execution result determined by executing the target task using the big model.

[0005] According to another aspect of the present disclosure, there is provided an information processing device based on a big model, comprising: a first obtaining module for performing task scheduling on demand information related to a target object to obtain a task scheduling strategy containing multiple initial tasks; a second obtaining module for updating at least one initial task in the task scheduling strategy according to the interactive operation of the target object in the process of executing the task scheduling strategy using the big model to obtain the target task; and a feedback information determination module for determining feedback information that matches the demand intention of the target object based on the target execution result determined by executing the target task using the big model.

[0006] According to another aspect of the present disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method provided according to an embodiment of the present disclosure.

[0007] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to execute the method provided according to an embodiment of the present disclosure.

[0008] According to another aspect of the present disclosure, a computer program product is provided, including a computer program, which implements the method provided according to the embodiment of the present disclosure when executed by a processor.

[0009] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The accompanying drawings are provided to facilitate a better understanding of the present invention and do not constitute a limitation of the present disclosure.

[0011] Figure 1 Schematically illustrates an exemplary system architecture to which a large model-based information processing method and apparatus can be applied according to an embodiment of the present disclosure;

[0012] Figure 2 Schematically shows a flow chart of a large model-based information processing method according to an embodiment of the present disclosure;

[0013] Figure 3 The following schematically illustrates a principle diagram of a task scheduling strategy according to an embodiment of the present disclosure;

[0014] Figure 4 The schematic diagram schematically shows the principle of the information processing method based on the large model according to the embodiment of the present disclosure;

[0015] Figure 5 The following schematically shows a system architecture diagram of an information processing system according to an embodiment of the present disclosure;

[0016] Figure 6 Schematically shows a flow chart of an information processing method based on a large model according to another embodiment of the present disclosure;

[0017] Figure 7 Schematically shows a block diagram of an information processing device based on a large model according to an embodiment of the present disclosure;

[0018] Figure 8 A block diagram schematically illustrates a structure of an artificial intelligence agent according to an embodiment of the present disclosure; and

[0019] Figure 9 A schematic block diagram of an example electronic device that can be used to implement the large model-based information processing method according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0020] The following description of exemplary embodiments of the present disclosure is made in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding. These details should be considered as merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0021] In the technical solution disclosed herein, the acquisition, storage and application of user personal information involved comply with the provisions of relevant laws and regulations, take necessary confidentiality measures, and do not violate public order and good morals.

[0022] The inventors discovered that when interacting with users through large models for question-and-answer sessions, the large model takes a long time to generate text, and the generated responses often differ significantly from the user's actual needs and intent. This requires the user to modify the question multiple times based on the response content, or to click "Regenerate" to instruct the large model to regenerate new responses. This results in long wait times for users during intelligent question-and-answer sessions with the large model, and can easily lead to a high level of redundant computing overhead on computing devices such as the server where the large model is deployed.

[0023] Embodiments of the present disclosure provide a large-scale model-based information processing method, apparatus, electronic device, and storage medium. The large-scale model-based information processing method includes: performing task scheduling on demand information related to a target object to obtain a task scheduling strategy containing multiple initial tasks; in the process of executing the task scheduling strategy using the large-scale model, updating at least one initial task in the task scheduling strategy based on the target object's interactive operation to obtain a target task; and determining feedback information that matches the target object's demand intent based on the target execution result determined by executing the target task using the large-scale model.

[0024] According to an embodiment of the present disclosure, by performing task scheduling based on the demand information of a target object, the execution results of multiple initial tasks in a task scheduling strategy can be made to meet the demand attributes represented by the demand information. In the process of executing multiple initial tasks in a task scheduling strategy using a large model, the initial tasks in the task scheduling strategy are synchronously updated according to the interactive operations of the target object. This can achieve the dynamic adjustment of the initial tasks in the task scheduling strategy according to the real-time demand represented by the user's interactive operations during the execution of the initial tasks based on the large model, so as to avoid the user having to wait for all the initial tasks in the task scheduling strategy to be completed before obtaining the reply content that meets the user's actual needs by updating the demand information. At the same time, the target execution result obtained by updating the intention of the interactive operation can meet the real-time needs of the target object, and then the target execution result that matches the actual needs of the target object can be determined by adjusting the task scheduling logic generated for the actual needs of the target object in real time. Thus, the feedback information determined based on the target execution result can meet the actual needs of the target object even if the task scheduling logic has not been executed, thereby improving the efficiency and quality of feedback information generation, reducing the redundant computing overhead of the computing equipment used to deploy the large model, improving user satisfaction and improving user experience.

[0025] Figure 1 An exemplary system architecture to which the large model-based information processing method and apparatus can be applied according to an embodiment of the present disclosure is schematically shown.

[0026] It should be noted that Figure 1 The examples shown are merely examples of system architectures to which the embodiments of the present disclosure may be applied, to help those skilled in the art understand the technical content of the present disclosure, but do not imply that the embodiments of the present disclosure may not be applied to other devices, systems, environments, or scenarios. For example, in another embodiment, an exemplary system architecture to which the large-model-based information processing method and apparatus may be applied may include a terminal device, but the terminal device may implement the large-model-based information processing method and apparatus provided by the embodiments of the present disclosure without interacting with a server.

[0027] like Figure 1 As shown, the system architecture 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 is used as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0028] A user may use a first terminal device 101, a second terminal device 102, or a third terminal device 103 to interact with a server 105 via a network 104 to receive or send messages, etc. Various communication client applications may be installed on the first terminal device 101, the second terminal device 102, or the third terminal device 103, such as knowledge reading applications, web browser applications, search applications, instant messaging tools, email clients, and / or social platform software (for example only).

[0029] The first terminal device 101 , the second terminal device 102 , and the third terminal device 103 may be various electronic devices having display screens and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

[0030] The server 105 may be a server that provides various services, such as a background management server (for example only) that supports content browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103. The background management server may analyze and process received data such as user requests, and feed back processing results (e.g., web pages, information, or data obtained or generated based on user requests) to the terminal devices.

[0031] It should be noted that the information processing method based on the large model provided in the embodiment of the present disclosure can generally be executed by the first terminal device 101, the second terminal device 102, or the third terminal device 103. Correspondingly, the information processing apparatus based on the large model provided in the embodiment of the present disclosure can also be set in the first terminal device 101, the second terminal device 102, or the third terminal device 103.

[0032] Alternatively, the information processing method based on the large model provided by the embodiment of the present disclosure may also be generally executed by the server 105. Accordingly, the information processing apparatus based on the large model provided by the embodiment of the present disclosure may generally be set in the server 105. The information processing method based on the large model provided by the embodiment of the present disclosure may also be executed by a server or server cluster that is different from the server 105 and can communicate with the first terminal device 101, the second terminal device 102, the third terminal device 103 and / or the server 105. Accordingly, the information processing apparatus based on the large model provided by the embodiment of the present disclosure may also be set in a server or server cluster that is different from the server 105 and can communicate with the first terminal device 101, the second terminal device 102, the third terminal device 103 and / or the server 105.

[0033] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is merely illustrative. Any number of terminal devices, networks and servers may be provided as required.

[0034] Figure 2 The flowchart of the information processing method based on the large model according to the embodiment of the present disclosure is schematically shown.

[0035] like Figure 2 As shown, the large model-based information processing method includes operations S210 to S230.

[0036] In operation S210 , task scheduling is performed on the demand information related to the target object to obtain a task scheduling strategy including a plurality of initial tasks.

[0037] In operation S220 , in the process of executing the task scheduling strategy using the large model, at least one initial task in the task scheduling strategy is updated according to the interactive operation of the target object to obtain a target task.

[0038] In operation S230 , feedback information matching the demand intention of the target object is determined based on the target execution result determined by executing the target task using the large model.

[0039] According to embodiments of the present disclosure, demand information may be information representing the target subject's needs, such as question text or query information input by the target subject. Demand information may include any data type, such as text, images, or audio. Embodiments of the present disclosure do not limit the specific data type of demand information.

[0040] According to embodiments of the present disclosure, task orchestration based on demand information may include processing the demand information based on a deep learning model to generate a task orchestration strategy comprising multiple initial tasks and dependencies between the initial tasks. Initial tasks may include any type of data processing task, such as information search, image recognition, or data calculation. Dependencies may represent the execution logic between different initial tasks.

[0041] According to an embodiment of the present disclosure, a large model can be used to execute multiple initial tasks based on the dependency relationship in the task scheduling strategy. For example, the large model can execute the first initial task "image segmentation task" to perform image segmentation on the image in the demand information to obtain the segmented image area of ​​Building A. The image area of ​​Building A can be the initial execution result of the first initial task "image segmentation task". After obtaining the image area of ​​Building A, the second initial task "target recognition task" having a dependency relationship with the first initial task can be determined based on the dependency relationship in the task scheduling strategy. The second initial task is executed by calling the target recognition tool using the large model to determine that the object type obtained by identifying the image area of ​​Building A is "Gymnasium A".

[0042] In some embodiments, the task orchestration strategy can be represented based on a topological structure such as a task chain or task tree containing multiple initial tasks. The topological structure representing the task orchestration strategy can include task nodes representing the initial tasks and edge relationships representing dependency relationships. This allows the large model to more clearly execute the multiple initial tasks according to the execution logic indicated by the dependency relationships in the task orchestration strategy, realizing the execution of task orchestration logic using the large model.

[0043] According to embodiments of the present disclosure, the large model may be a trained large language model that can process multimodal data such as text, speech, and images based on a relatively large number of model parameters to perform the initial task. It should be noted that the number of large models involved in embodiments of the present disclosure may be one or more, and the embodiments of the present disclosure do not limit the number of large models.

[0044] According to an embodiment of the present disclosure, executing task orchestration logic using a large model may include utilizing a computing unit, to which model parameters of the large model are deployed, to execute multiple initial tasks according to dependencies indicated by the task orchestration logic. The process of executing the task orchestration logic using the large model may include the computing unit performing data processing on the initial tasks.

[0045] In some embodiments, during the execution of the task orchestration logic using the large model, feedback information matching the target object's needs has not yet been output. The target object can perform any type of interactive operation, such as an input operation, during the execution of the initial task using the large model. Operational information such as the type of interactive operation and the input information carried by the interactive operation can represent the real-time demand attributes of the target object during the execution of the task orchestration logic.

[0046] In some embodiments, updating at least one initial task in the task orchestration logic based on an interactive operation with a target object may include updating task attribute parameters of the initial task based on the operational information carried by the interactive operation with the target object, thereby obtaining the target task. For example, tool attribute parameters such as the tool name and tool parameters of the tool resources involved in executing the initial task may be updated, but this is not limited to such. Task attribute parameters may also include reference knowledge domains related to the initial task, style attributes, output result data format, and other parameters used to control the execution results of the task.

[0047] For example, updating the initial task may include updating the prompt words for performing the initial task "information retrieval" using the large model based on the interactive operation input "Please refer to medical field knowledge", and performing the target search task based on the new prompt words "Search for medical field knowledge".

[0048] In some embodiments, updating at least one initial task in the task scheduling logic according to the interactive operation of the target object may also include updating the dependency relationship between multiple initial tasks in the task scheduling strategy. For example, the serial execution logic of initial task 1 to initial task 6 in the task scheduling strategy may be modified to a first task chain of initial task 1 to initial task 4 and a second task chain of initial task 5 to initial task 6 in parallel through the operation information indicated by the interactive operation. In this way, the dependency relationship in the task scheduling logic can be modified based on the interactive operation to adjust the execution logic and execution order between multiple initial tasks, so that the large model can execute multiple tasks in multiple task chains based on the parallel model to improve the generation efficiency of feedback information. It should be understood that the initial task after the dependency relationship is updated can be understood as the target task. Feedback information can be determined based on the target execution result of the target task and the initial execution result of the initial task.

[0049] In some embodiments, updating at least one initial task in the task scheduling logic according to the interactive operation of the target object may also include updating the task attribute parameters of the initial task and the dependency relationship between the initial tasks in the task scheduling logic. The embodiments of the present disclosure will not be repeated here.

[0050] According to an embodiment of the present disclosure, by updating at least one initial task in the task scheduling logic according to the interactive operation of the target object, the updated target task can meet the actual needs indicated by the interactive operation. Thus, the actual needs of the target object can be met by using the large model to execute the target task in the task scheduling logic to obtain the target execution result, so that the feedback information determined according to the target execution result can match the real-time needs of the target object. At the same time, it can also avoid the target object inputting the modified demand information to reuse the large model to reply to the target object after the task scheduling logic is fully executed, thereby reducing the waiting time for the target object to obtain satisfactory feedback information, improving the user experience and reducing the computing overhead of the computing device deploying the large model.

[0051] In some embodiments, the large model-based information processing method may further include: pushing at least one of a task scheduling strategy and an initial execution result of an initial task to a target object.

[0052] In one example, a task orchestration strategy can be pushed to a target object, and a directed acyclic graph representing multiple initial tasks and dependency relationships between the multiple initial tasks in the task orchestration strategy can be displayed, so that the target object can perform interactive operations on the task nodes representing the initial tasks and the edge relationships representing the dependency relationships in the directed acyclic graph to input operation information.

[0053] In one example, the execution results obtained by executing the initial task in the task orchestration strategy using the large model can also be pushed to the target object. For example, multiple scenic images of City A retrieved by the initial "image retrieval task" can be pushed to the target object. The target object can remove specific scenic images based on the multiple scenic images displayed to reduce the amount of data processing generated by the execution of subsequent initial tasks or target tasks in the task orchestration strategy, lower the computational overhead of generating feedback information, and reduce the degree of match between the generated feedback information and the actual needs of the target object due to processing scenic images not needed by the target object, thereby improving the user experience.

[0054] In some embodiments, the initial execution results and task scheduling strategies can also be pushed to the target object so that the target object can flexibly and dynamically adjust and schedule the tasks that the large model needs to perform, so that the task scheduling strategy can more accurately meet the needs of the target object, improve the matching degree between the feedback information and the target object's needs and intentions, and improve the user experience.

[0055] According to an embodiment of the present disclosure, the first obtaining module may include: using a large model to perform semantic understanding on demand information to obtain demand intent; and performing task scheduling based on the demand intent to obtain a task scheduling strategy.

[0056] According to an embodiment of the present disclosure, the demand intent can represent the demand attributes of the target object. For example, the demand intent can be unstructured natural language description information "generate a press release with the theme of campus sports meeting". For another example, the demand intent can also be structured data. For example, the demand intent can be multiple intent keywords represented based on a table. By using a large model to understand the intent of the demand information, it is possible to correct the semantically ambiguous parts of the demand information or complete the missing parts of the information, so that the obtained demand intent can more accurately represent the actual demand attributes of the target object, so as to improve the accuracy of the task scheduling strategy determined subsequently, reduce the reduction in the demand matching between the feedback information and the target object due to the missing tasks, and thus improve the user experience.

[0057] In some embodiments, using a large model to perform semantic understanding on demand information to obtain demand intent may also include: determining the initial tasks corresponding to each intent content based on multiple intent contents in the structured demand intent, and then determining the task scheduling strategy based on preset dependencies.

[0058] In some embodiments, using a big model to perform semantic understanding of demand information and obtain demand intentions can also include: using a big model to process demand information and object attribute information of the target object, and then based on the big model, the preferences, habits and other target object attributes represented by the demand information and object attribute information can be integrated to understand the demand intentions, and obtain demand intentions that can meet the actual needs of the target object to improve user experience.

[0059] In some embodiments, using a large model to perform semantic understanding on demand information to obtain demand intent may also include: based on an object attribute knowledge graph, using a large model to perform semantic understanding on demand information to obtain demand intent.

[0060] According to an embodiment of the present disclosure, an object attribute node in an object attribute knowledge graph represents the historical interaction information or object preference attributes of a target object. Historical interaction information may represent, but is not limited to, the target object's browsing interaction behavior, collection interaction behavior, like interaction behavior, search interaction behavior, etc. in a historical period. Historical interaction information may also include the contextual content input by the target object, historical conversation content with the large model, etc. Object preference attributes may include attribute information determined based on the target object's preference selection operation, such as sports preference attributes, medical knowledge preference attributes, etc. The embodiments of the present disclosure do not limit the specific type and generation method of object preference attributes.

[0061] According to an embodiment of the present disclosure, by constructing an object indication graph based on historical interaction information and object preference attributes, the personalized preferences or habits of the target object can be represented based on entities, and the association relationship between multiple different types of personalized preferences or habits can be represented based on edges. In this way, the personalized potential demand attributes of the target object can be structured based on the object attribute knowledge graph, and the large model can be used to perform semantic understanding of the demand information through the object knowledge graph. The large model can more clearly understand the potential demand attributes of the target object based on the structured personalized preferences, and thus use the large model to understand the demand intentions contained in the demand information by combining the personalized preference attributes, so that the demand intentions can more accurately represent the actual intentions of the target object, thereby improving the rationality and adaptability of the subsequent task scheduling strategy, and improving the user experience by executing the updated task scheduling logic.

[0062] In some embodiments, the demand intent includes an intended time intent that characterizes the target object's desire to obtain feedback information. The intended time intent can represent an intention related to the time of obtaining feedback information, such as the desired moment, duration, or time period. For example, the intended time intent can be "20 minutes later" or "11:00 am on x month y day, 2025." The demand intent containing the intended time intent can be obtained by using the first model to understand the intent of the demand information.

[0063] In some embodiments, performing task scheduling based on demand intent to obtain a task scheduling strategy may include: performing task scheduling on the demand content of the demand intent using a first large model based on delay prompt information determined based on the expected time intention to obtain a task scheduling strategy.

[0064] According to an embodiment of the present disclosure, the delay prompt information can be used to prompt the large model to execute the task scheduling logic to output the feedback information required for the strategy execution time, which needs to meet the time or duration when the target object expects to obtain the feedback information.

[0065] For example, the time delay prompt information may be "output the PPT file related to this month's work report within 20 minutes."

[0066] In some embodiments, a first large model is used to perform task orchestration based on demand intent, and a second large model is used to execute the task orchestration strategy. The first large model and the second large model can be different large models with different model parameters. By collaborating based on multiple different large models to perform task orchestration and task orchestration strategy execution respectively, different types of tasks can be executed based on the model capabilities of each large model, thereby improving the efficiency and quality of task orchestration strategy generation and task orchestration strategy execution based on the collaboration of multiple large models, thereby improving the matching degree between feedback information and target objects, and thus improving the user experience.

[0067] In some embodiments, the second largest model executes the task scheduling strategy to obtain feedback information, and the strategy execution duration matches the expected time intention. For example, the second largest model obtains the time required for feedback information by executing the task scheduling strategy of the first largest model based on the delay prompt information to output prompts, which matches the time requirement attributes such as the moment, duration or time period of the expected feedback information expressed by the expected time intention. Therefore, the first largest model can output a task scheduling strategy adapted to the second largest model to output feedback information according to the expected time intention, so as to meet the time requirement of the target object to obtain feedback information, meet the diverse actual needs of users, and thus improve the user experience.

[0068] In some embodiments, the information processing method based on the large model may also include: based on the task attribute information of the initial task, using the first large model to perform semantic understanding according to the model description information of each of multiple candidate large models, and determining the second large model for performing the initial task from the multiple candidate large models.

[0069] According to an embodiment of the present disclosure, the model description information represents the model performance attributes of the candidate large model. For example, the model performance attributes may be model performance indicators of the candidate large model, such as accuracy, average latency, and computational overhead cost. In addition, the model performance attributes may also include the fields to which the candidate large model is adapted, such as representing the candidate large model's adaptability to various types of fields, such as medical fields, education fields, and engineering construction fields. For another example, the model performance attributes may also represent the scenarios to which the candidate large model is adapted, such as representing the candidate large model's adaptability to intelligent search scenarios, script editing scenarios, and marketing copywriting scenarios.

[0070] In some embodiments, the model performance attribute may also represent a model performance indicator such as the scale of model parameters of the candidate large model, which is used to evaluate the computational overhead of the candidate large model.

[0071] According to the embodiments of the present disclosure, by semantically understanding the model description information of each of the multiple candidate large models based on the task attribute information of the multiple initial tasks, the first large model can fully understand the degree of match between the model performance of the candidate large model and the performance requirements required to execute the initial tasks, thereby enabling the second model determined from the candidate large model for each initial task to be more accurately adapted to the task requirement scenario, task execution parameters and other task attribute information of the initial task in the task scheduling strategy. This can further improve the ability of the execution results output by the task scheduling strategy to adapt to the demand intent represented by the demand information, thereby reducing the frequency of the target object performing interactive operations, improving the degree of match between the feedback information and the target object's demand intent, and thus improving the user experience.

[0072] In one embodiment, multiple initial tasks in the orchestration strategy can be executed based on different second large models. Therefore, the task orchestration strategy output by the first large model can be used to schedule multiple different large models to execute different initial tasks respectively, thereby improving the execution efficiency of the initial tasks.

[0073] In some embodiments, the orchestration strategy includes multiple task topologies having a dependency relationship with a specified initial task, and the interaction operation is associated with a first task topology among the multiple task topologies.

[0074] According to embodiments of the present disclosure, a task topology can be represented as any topological form, such as a task link or a task tree. Multiple initial tasks in a task topology have dependencies, and the task execution process can be represented by the dependencies between each task topology and a specified initial task.

[0075] Figure 3 The following schematically illustrates a principle diagram of a task scheduling strategy according to an embodiment of the present disclosure.

[0076] like Figure 3As shown, the task scheduling strategy in this embodiment can be a file task scheduling strategy 300 for generating a presentation file. By using the first large model to perform task scheduling on the target object's input requirement information "generate a presentation file introducing attractions in city A", the file task scheduling strategy 300 is obtained. The file task scheduling strategy 300 can include multiple initial tasks, and the multiple initial tasks can be based on Figure 3 The circular elements shown in the figure represent the initial tasks. The initial tasks are the first to sixth initial tasks. The first initial task is the presentation outline generation task, the second initial task is the image material retrieval task, the third initial task is the image material editing task, the fourth initial task is the scenic spot introduction text retrieval task, the fifth initial task is the scenic spot introduction text semantic fusion task, and the sixth initial task is the presentation file creation task.

[0077] The file task scheduling strategy 300 may include two different task topologies having a dependency relationship with the first initial task, namely a first task topology 310 and a second task topology 320. Both the first task topology 310 and the second task topology 320 have a dependency relationship with the first initial task.

[0078] In some embodiments, updating at least one initial task in the task orchestration strategy according to the interactive operation of the target object may include: using a large model to understand the intent of the operation information of the interactive operation, and updating the first initial task related to the operation information in the first task topology.

[0079] For example, if the first initial task is a page search task, the operation information may include input information for "Website A." The task execution parameters of the page search task can be modified based on "Website A" so that, by executing the target task, a page search for "Website A" is added, and the target execution results obtained include pages on Website A that are relevant to the demand intent.

[0080] According to an embodiment of the present disclosure, feedback information is determined based on the initial execution result and target execution result of the second initial task in the second task topology. The second initial task in the second task topology among multiple task topologies is not affected by the interactive operation. In the process of updating the first initial task using the operation information, the large model can continue to be used to execute multiple second initial tasks in the second task topology, and the initial execution result and target execution result of the second initial task can be used to generate feedback information. This can reduce the number of updated initial tasks in the task scheduling strategy, avoid the redundant computing overhead and waiting time caused by the full adjustment of the task scheduling strategy that needs to be executed for the large model, and improve the response efficiency of providing feedback information to the target object, so as to improve the user experience.

[0081] In some embodiments, using a large model to perform semantic understanding on the operation information of the interactive operation, updating the first initial task related to the operation information in the first task topology may include: using the large model to perform semantic understanding on the operation information and requirement information to obtain a semantic understanding result; and updating multiple first initial tasks in the first task topology based on the semantic understanding result.

[0082] According to an embodiment of the present disclosure, the semantic understanding result may indicate the semantic relevance between the demand information and the operation information. For example, the semantic understanding result may indicate the semantic difference between the operation information and the demand information, the change in demand intention, the potential demand intention, etc.

[0083] For example, the large model can perform semantic understanding on the demand information "basketball game results" and the interactive operation information "yesterday's A football league game results" to determine that the semantic understanding result is expressed as "the target object is relatively interested in the basketball game results and also wants to know the results of yesterday's A football league game." Based on this semantic understanding result, the target object's demand intent can be expressed as a change from the demand for basketball game results to the demand for A football game results. The potential demand intent can also indicate a strong interest in multiple popular games.

[0084] The large model can then use the semantic understanding results as prompt information to update the task attribute parameters of the first initial task in the first task topology, allowing the resulting target task to call a search tool to obtain this week's basketball game results and yesterday's A-League football game results from an authorized match results website. At the same time, a football match results analysis task can be added as a target task in the first task topology. By executing the target task using the large model, the output analysis content can include text analysis of the impact of yesterday's football match results on the league rankings of teams in multiple leagues.

[0085] According to an embodiment of the present disclosure, a large model is used to perform semantic understanding on demand information and operation information, so that the obtained semantic understanding result represents the semantic relevance between the demand information and the operation information. In this way, the large model is used to synchronously update multiple first initial tasks in the same task topology based on the potential demand intentions, demand intention changes, and other intentions related to the actual needs of the target object expressed by the semantic understanding results, so as to improve the update accuracy and update efficiency of the initial tasks. At the same time, during the process of updating the first task topology, the second initial task in the second task topology can be synchronously executed to avoid redundant computing overhead and redundant waiting time caused by the regeneration of the task scheduling strategy.

[0086] In some embodiments, using a large model to understand the intent of the operation information of the interactive operation, updating the first initial task related to the operation information in the first task topology can include: based on the tool description prompt word, using the first large model to understand the tool call intent of the operation information to obtain the tool call task in the first task topology.

[0087] According to embodiments of the present disclosure, a tool description prompt represents tool description information related to a preset tool. A preset tool can be a service resource preset in the system or an external service resource that can be called under authorization conditions. The tool description information describes the preset tool's parameter range, applicable scenarios, output result type, and other information used to describe the preset tool's performance.

[0088] In some embodiments, the first large model can understand the tool call intention contained in the operation information based on the tool performance of the preset tool indicated by the tool description prompt, and generate a tool call task that matches the tool call intention as the target task based on the tool call intention to update the first initial task in the first task topology. This can make the first task topology in the updated task scheduling strategy more accurately adapt to the actual tool call requirements of the target object, so as to improve the matching degree between the target execution result and the requirement intention, and further improve the matching degree between the feedback information and the requirement intention, thereby improving the user experience.

[0089] According to an embodiment of the present disclosure, the tool calling task instructs the second large model to call the target tool in the preset tool by executing the tool calling task, execute at least one specified task in the first task topology, and obtain the target tool execution result for determining feedback information.

[0090] For example, a tool call task can be a code editing task. The second model is used to call a code editing tool as the target tool to perform code editing to execute the specified task. A code script that matches the task attribute information of the tool call task as the specified task is output, and the code execution script is used as the target tool execution result. Thus, the second model can be used to execute a code execution task that has a dependency on the code editing task in the first task topology. The code execution tool is called as the target tool by the second model to execute the code script, and the code execution result is obtained as the target tool execution result.

[0091] According to an embodiment of the present disclosure, by utilizing the first large model to update the initial task in the task orchestration strategy, and utilizing the second large model to execute the tool call task as the target task to call the corresponding target tool to perform the specified task according to the task attribute information of the tool call task, task orchestration and tool call can be performed through multiple different large models, and combined with utilizing the second large model to synchronously execute the second initial task in the second task topology, so as to improve the response efficiency for interactive operations, and improve the overall execution efficiency of the task orchestration strategy and the information quality of the target execution results, thereby achieving the improvement of the matching degree between the feedback information and the demand intention of the target object, and achieving the technical effect of improving the user experience.

[0092] In some embodiments, using a large model to understand the intent of the operation information of the interactive operation, updating the first initial task related to the operation information in the first task topology can also include: based on the tool description prompt words, using the first large model to understand the execution time intent of the operation information to obtain a scheduled call task.

[0093] According to an embodiment of the present disclosure, the scheduled calling task is used to instruct the second large model to call the target tool in the preset tool to perform the specified task at a specified time or a specified period that matches the execution time intention represented by the operation information.

[0094] In some embodiments, the execution time intention represented by the operation information can be determined based on key information content such as the keyword "8 o'clock tonight" carried by the operation information, and the key information content is related to a specified time or a specified time period.

[0095] In some embodiments, execution time intent can also be determined by semantically understanding the operation information using the first large model. By using the large model to understand the execution time intent implicit in the operation information, the tool's call time attributes can be determined, and the target tool can be identified from the preset tools based on the tool's capabilities as indicated by the tool description information. This allows the first large model to generate a target tool that will be called at a specified time or time period that matches the specified time intent, causing the target tool to perform a specified task, thereby generating a scheduled call task.

[0096] For example, the operation information could be "Compute the results of the match between Team A and Team B, and create a score sheet." By using the first model to understand the execution time intent of the operation information, we can determine that the execution time intent in the understanding result is "The match between Team A and Team B starts at 8:00 PM tonight, and ends between 10:00 PM and 10:20 PM." Therefore, the scheduled call task can be determined to call the search tool to obtain the match results between Team A and Team B from the authorized information interface between 10:00 PM and 10:20 PM tonight.

[0097] According to the embodiments of the present disclosure, by semantically understanding the operation information and tool description information of the interactive operation during the execution of the editing task strategy, a scheduled calling task is determined, and the tool adapted to the needs of the target object can be called according to the time requirements to obtain the tool calling result, so that the feedback information can further more accurately meet the diverse real-time needs of the target object, avoiding the delay in obtaining feedback information generated by the target object interacting with the large language model after a specified time or a specified period, and intelligently and automatically responding to the delay requirements of the target object through the scheduled calling task to improve the user experience.

[0098] In some embodiments, the task scheduling policy may indicate different processing modes for multiple initial tasks or multiple target tasks. The processing modes shown in Table 1.1 and Table 1.2 exemplify the processing modes that the task scheduling policy may indicate.

[0099] Table 1.1

[0100]

[0101] Table 1.2

[0102]

[0103] It should be noted that the processing modes shown in Table 1.1 and Table 1.2 are only used to illustrate the execution process or application scenarios of the large model execution task scheduling strategy, and are not used to limit the specific scope of the task scheduling strategy provided in the embodiments of this disclosure.

[0104] In some embodiments, the information processing method based on the big model may further include: based on the operation information of the interactive operation, using the computing unit where the big model is deployed to stop executing at least one initial task in the task scheduling strategy.

[0105] According to an embodiment of the present disclosure, the computing unit may include any type of device or component such as a central processing unit (CPU) or a graphics processing unit (GPU) that can perform computing tasks based on model parameters of a large model.

[0106] In some embodiments, after receiving the operation information, the computing unit can be paused to execute the initial task according to the model parameters, and the initial task indicated by the interactive operation in the task scheduling strategy can be updated according to the task attribute information represented by the operation information to obtain an updated target task. After obtaining the target task, the computing unit is used to execute the target task according to the model parameters of the large model to obtain the target execution result. This can avoid the computational overhead redundancy and long delay waiting time caused by re-scheduling the task according to the operation information and context after receiving the operation information of the interactive operation to obtain a new task scheduling strategy. By stopping the computing unit to stop executing the initial task, and resuming the use of the computing unit to continue executing the task scheduling strategy after the target task is updated, it is possible to flexibly control the computational load of the computing unit in the process of responding to the interaction problem of the target object and save computational overhead.

[0107] In some embodiments, the orchestration strategy includes a plurality of task topologies having a dependency relationship with a specified initial task, and a stopped state topology in the plurality of task topologies includes the initial task whose execution is stopped;

[0108] According to an embodiment of the present disclosure, the information processing method based on the big model may further include: in the process of stopping the execution of at least one initial task in the task scheduling strategy using the computing unit deploying the big model, using the computing unit to execute the initial task in the associated topology.

[0109] According to an embodiment of the present disclosure, the associated topology is other task topologies except the stop state topology among the multiple task topologies.

[0110] For example, the associated topology may be a second task topology among the multiple task topologies that has not been updated based on the operation information.

[0111] For another example, the associated topology may be the first task topology among the multiple task topologies that has been updated in a historical period.

[0112] By using the computing unit to stop executing the initial task in the stopped state topology, the redundant computing overhead incurred by the computing unit due to the execution of the stopped initial task can be avoided. Furthermore, while the computing unit is stopping the stopped topology, the execution efficiency of the task scheduling policy is improved by continuing to execute at least one of the initial task and the target task in the associated topology that is unrelated to the operation information of the current interactive operation. This improves the user experience by increasing the response speed of output feedback information.

[0113] Figure 4 The schematic diagram schematically shows the principle of the information processing method based on the large model according to the embodiment of the present disclosure.

[0114] like Figure 4As shown, the task scheduling strategy in this embodiment can be a manuscript task scheduling strategy 400 for generating a competition press release. The manuscript task scheduling strategy 400 can include multiple initial tasks, and the multiple initial tasks can be based on Figure 4 The circular elements shown in the figure represent the initial tasks, which are the first to sixth initial tasks. The first initial task is the manuscript outline generation task, the second initial task is the match image material retrieval task, the third initial task is the match image material editing task, the fourth initial task is the match process text retrieval task, the fifth initial task is the player data retrieval task, and the sixth initial task is the manuscript semantic fusion task.

[0115] Among them, when the target object performs an interactive operation on the 4th initial task and inputs the operation information, the second computing unit stops executing the 4th initial task. The task topology corresponding to the 4th initial task can be a stopped state task topology 420. In the process of updating the 4th initial task based on the operation information, the first computing unit can continue to execute the initial task in the associated topology 410 to retrieve the game image material in real time and edit the game image material. In this way, the calculation process for the task topology can be dynamically stopped based on multiple different computing units to improve the execution efficiency of the manuscript task scheduling strategy 400. It also avoids the additional computing overhead caused by re-tasking and improves the user experience by reducing the computing overhead and feedback information delay.

[0116] In some embodiments, in the process of executing the task orchestration strategy using the big model, updating at least one initial task in the task orchestration strategy according to the interactive operation of the target object may also include: in the process of executing the task orchestration strategy using the big model, updating the initial execution result of the initial task according to the operation information carried by the interactive operation for the initial task, and obtaining a first target execution result related to the initial task.

[0117] According to an embodiment of the present disclosure, the first target execution result is used as input data to orchestrate at least one uncompleted initial task or target task in the strategy.

[0118] For example, the initial execution result might be "match result data between Team A and Team B." If the referee verifies and changes the match result after the match, the match result data will change. The target object can perform an interactive operation on the final match result after the change to modify the match score data in the initial execution result "match result data between Team A and Team B," thereby obtaining the modified final score data as the first target execution result. This final score data is then input into another initial task, "Analyze the match between Team A and Team B," which has a dependency relationship with the initial task. Thus, the large model can be used to process the "final score data" and output a match analysis text.

[0119] According to the embodiments of the present disclosure, during the execution of the task orchestration strategy, the target object can modify the initial execution result of the initial task in real time and apply the modified target execution result to the subsequent initial task or target task that has not yet been completed. In this way, during the execution of the orchestration strategy, the target object can obtain the execution result that meets the needs of the target object in a timely manner through the interactive operation of the target object, and can adjust the input data of the subsequent initial task or target task by adjusting the target execution result generated based on the operation information, so as to adjust the execution result of the orchestration strategy in accordance with the needs of the target object in a timely manner, improve the matching degree between the target execution result and the demand intention of the target object, and timely and conveniently improve the generation efficiency of feedback information, avoid the redundant computing overhead and waiting time caused by repeated generation of task orchestration strategies, and meet the actual demand intention of the target object with high quality, so as to improve the user experience.

[0120] In some implementations, in the process of executing a task scheduling strategy using a large model, updating at least one initial task in the task scheduling strategy based on the interactive operation of the target object may also include: using the large model to plan tasks based on the first target execution result and demand information, and obtaining a target task that has a dependency relationship with at least one unexecuted initial task.

[0121] In one embodiment, the first large model can be used to process the requirement information "Generate a press release for the match between Team A and Team B" and the first target execution result of the modified final score of "2:1". The first large model can be used to understand the intent of the requirement information and the first target execution result, thereby obtaining the interaction intent of "The match has been re-decided, and a press release needs to be generated based on the new score and the re-deciding process." The first large model can generate the target task "Retrieve the real-time game commentary text for the match between Team A and Team B" by processing the interaction intent. This target task can then be added to the task scheduling strategy, and the second large model can be used to execute the second target task to obtain the second target execution result of "real-time game commentary text." The second large model can then semantically fuse the second target execution result with the modified first target execution result to obtain a press release reporting on the re-decided match.

[0122] According to an embodiment of the present disclosure, using a large model to perform task planning based on the first target execution result and demand information can include using the first large model to understand the intention of the target object to perform interactive operations based on the first target execution result generated based on the operation information, as well as the demand information and other contextual content to determine the actual interactive intention of the target object performing the interactive operation. In this way, task planning can be performed based on the actual interactive intention generated by the target object through the execution of the interactive operation, and a target task that can meet the actual interactive intention can be generated, so that the modified first target execution result can be input into the target task as input data, so that the second large model can be used to execute the target task based on the first target execution result, so that the target execution result can meet the actual interactive intention of the target object and improve the matching degree between the feedback information and the actual demand intention of the target object. It is achieved based on dynamic adjustment of the task scheduling strategy to meet the needs of the target object in a timely manner, improve the data quality of the feedback information, and improve the user experience.

[0123] In some embodiments, based on the target execution result determined by executing the target task using the big model, determining the feedback information that matches the demand intention of the target object may include: based on the style attributes related to the demand information, using the big model to semantically fuse multiple target execution results to obtain feedback information.

[0124] According to embodiments of the present disclosure, style attributes can represent the language style of text or copywriting, the image style attributes of an image, the pronunciation style attributes of audio data, and so on. Style attributes can be the intent content of a requirement determined by understanding the intent of the requirement information using a large model, or they can be the information content carried in the requirement information. Embodiments of the present disclosure do not limit the specific type of style attributes or the specific method for determining style attributes.

[0125] In one embodiment, a large model is used to semantically fuse multiple target execution results, which may include prompt words determined based on style attributes, and semantically fuse multiple text segments to obtain feedback information expressed based on style attributes.

[0126] In one embodiment, the style attribute can be the information content in the operation information carried by the interactive operation. The target object inputs the style attribute through the interactive operation, and the style attribute of the feedback information can be adjusted in real time based on actual needs. This allows the target object to modify the style attribute of the feedback information and the quality of the feedback information through real-time interaction with the large model, thereby improving the matching of the style attribute of the feedback information to the specific scene.

[0127] In some embodiments, the style attribute can be related to a specified demand scenario. For example, the style attribute can be "the language expression style of giving a report and explanation to middle school students". Based on the style prompt words determined by the style attribute, the large model can be prompted to perform stylized semantic fusion on the speech segments output by multiple target tasks, so that the resulting speech manuscript can meet the expression style of the demand scenario of "giving a report and explanation to middle school students", and the feedback information can match the scenario requirements of the target object.

[0128] Figure 5 The system architecture diagram of the information processing system according to the embodiment of the present disclosure is schematically shown.

[0129] like Figure 5 As shown, the information processing system 500 can apply the large model-based information processing method provided by the embodiment of the present disclosure. The information processing system 500 may include a model service layer, a security risk control layer, an intelligent scheduling layer, a message communication layer, a tool service layer, a multimodal rendering layer, a knowledge enhancement layer, a file management layer, and a memory layer.

[0130] The model service layer can deploy multiple large models for executing the large model-based information processing method provided according to the embodiment of the present disclosure. For example, the model service layer can provide a reasoning large model, a scheduling large model, an execution large model, and a reflection large model. Among them, the reasoning large model and the scheduling large model can be used as the first large model for performing task scheduling and initial task updates, and the execution large model can be used as the second large model for executing the initial task or the target task. The reflection large model can check the initial execution results or the target execution results of each output of the second large model and output the reflection results. Therefore, based on the reflection results, the second large model can be iteratively prompted to re-execute the initial task or the target task to output high-quality execution results that meet the needs of the target object.

[0131] The task orchestration strategy output by the reasoning model can represent the dependencies between multiple initial tasks using topological forms such as Chain of Thought (CoT), Tree of Thoughts (ToT), and Graph of Thoughts (GoT). The scheduling model can use a model routing mechanism to select the optimal second-largest model from a library of pre-set large models to execute the initial or target task. For example, the second-largest model can be selected based on the latency, computational cost, and accuracy metrics of the pre-set large models. The reflection model establishes a multi-dimensional error classification system and constructs an anomaly detection matrix based on quality evaluation criteria such as logical contradictions, factual errors, and task execution anomalies. This evaluates the quality of the execution results and uses the reflection results as prompts to generate the second-largest model, thereby controlling the second-largest model to output more optimal execution results. This enables a closed-loop execution process of reflection, correction, and verification for the task orchestration strategy. Furthermore, the reflection model can reflect on the execution results based on the public or private knowledge base in the knowledge enhancement layer to improve the accuracy of the reflection results.

[0132] The scheduling model can select parallel mode, serial mode, or other processing modes based on the dependencies indicated by the task orchestration strategy to schedule the second model to execute the task orchestration strategy. At the same time, the scheduling model can also understand the tool description information and provide tool resources to the execution model based on the model context protocol. The execution model can select the target tool based on multiple tool descriptions and call the target tool to execute the initial task or target task through the model context protocol to improve the accuracy of tool calling. At the same time, the execution model can also perform stylized fusion of the execution results based on style attributes, so that the style attributes of the feedback information determined based on the target execution results meet the needs of the target object.

[0133] The intelligent scheduling layer uses a multi-mode scheduling engine to support multiple scheduling modes, including parallel, asynchronous, and serial models, for scheduling large model outputs. It can also allocate computing resources across multiple computing units or devices based on a load balancing mechanism to meet the execution efficiency requirements of the task orchestration strategy. The task coordination mechanism module of the intelligent scheduling layer can also monitor and manage the status of initial and target tasks throughout their entire execution lifecycle, including creation, execution, suspension, and resumption, and supports features such as task resuming from breakpoints.

[0134] The message communication layer can build standardized communication protocols to enable message communication between different large models and between models and tools. It also categorizes and stores messages such as the operation information of target object interactions, the execution results of large model outputs, and the execution status of initial tasks or target tasks.

[0135] The file management layer can parse and store multimodal information such as speech, text, and images through a multimodal processing module. For example, it can extract semantic features from the context using sliding windows and attention mechanisms, providing input data for the large model when updating initial tasks or performing task orchestration. The file management layer can also support information queries based on scalar or vector indexes. The hybrid storage module can store structured information for scalars, semantic information for vectors, and raw information for objects through hierarchical storage. Furthermore, data generated by the large model during execution can be recorded in log files.

[0136] The tool service layer standardizes communication between internal tools and external third-party tools through the model context protocol module, supporting protocol-level extensibility to facilitate interactive data output, large model or agent access, and other functions. The preset tool library can be configured with different internal tool and external third-party tool resource interfaces to facilitate system calls.

[0137] The tool service layer also allows for visual configuration of workflows, supporting intelligent switching between serial and parallel modes, as well as configuration of retry, circuit breaking, and downgrade exception handling strategies. It also monitors metrics such as tool call success rate, tool response latency, and resource consumption, while fully recording tool input, output, and intermediate states.

[0138] The global domain of the memory layer can store multi-source data such as user behavior, conversation context, and personalized preference attributes based on the graph nodes of the graph neural network. It supports unified storage and associated retrieval of text, files, and structured data to achieve memory storage, and can combine scalar metadata indexing and vector semantic indexing to retrieve stored data.

[0139] Knowledge base content in the knowledge enhancement layer supports parsing and storage of files in various formats, including text, images, documents, audio, and video. It supports file format normalization and multi-dimensional recall of scalar and vector quantities. Public knowledge bases integrate high-quality document library content, real-time internet data, image search, and multimedia resources. Private knowledge bases support scenario-based knowledge management, including databases, personal knowledge spaces, and enterprise network drives.

[0140] The security and risk control layer supports a multi-level defense system that conducts multiple checks on input information to intercept sensitive information. It also conducts content security reviews of output content based on keywords and risk control classification models. Furthermore, it can integrate manual review to intercept unusually sensitive information. The security and risk control layer also supports privacy data compliance, a permissions management system, and review process tracking, providing diverse functionality. The multimodal rendering layer enables real-time compilation, transcoding, and playback of various files.

[0141] Figure 6The flowchart of the information processing method based on the large model according to another embodiment of the present disclosure is schematically shown.

[0142] like Figure 6 As shown, the large model-based information processing method in this embodiment can execute steps S601 to S606 based on an information processing system.

[0143] In step S601, demand information may be received, and the first large model in the model service layer may be called to execute step S602, and a thinking process may be executed according to the demand information to determine the demand intention of the target object.

[0144] In step S603, the task scheduling strategy is planned using the first large model, or the initial task is updated according to the interactive operation of the target object.

[0145] In step S604, the scheduling strategy provided by the intelligent scheduling layer, such as the second largest model, executes the initial task and the target task to obtain the initial execution result and the target execution result. For example, the second largest model can be used to call the preset tool in the tool service layer to execute the tool call task.

[0146] In step S605, the reflection model in the model layer is used to perform self-inspection on the initial execution results and the target execution results. For example, the execution results can be checked based on the knowledge base content provided by the knowledge enhancement layer to output the reflection results. When the reflection result of step S605 indicates that the self-inspection has passed, feedback information can be obtained by fusing multiple target execution results. And step S606 is executed to render the feedback information based on the multimodal rendering layer. During the execution of the information processing method, the security risk control layer can detect the security risks of the output results and the operational information of the interactive operation in real time to avoid outputting sensitive information or non-compliant information.

[0147] Figure 7 The block diagram of the information processing device based on the large model according to the embodiment of the present disclosure is schematically shown.

[0148] like Figure 7 As shown, the large model-based information processing device 700 includes: a first obtaining module 710, a second obtaining module 720 and a feedback information determining module 730.

[0149] The first obtaining module 710 is configured to perform task scheduling on the demand information related to the target object to obtain a task scheduling strategy including a plurality of initial tasks.

[0150] The second obtaining module 720 is used to update at least one initial task in the task scheduling strategy according to the interactive operation of the target object in the process of executing the task scheduling strategy using the large model to obtain the target task.

[0151] The feedback information determination module 730 is used to determine feedback information that matches the demand intention of the target object based on the target execution result determined by executing the target task using the large model.

[0152] According to an embodiment of the present disclosure, the orchestration strategy includes multiple task topologies having dependencies with a specified initial task, and the interactive operation is related to a first task topology among the multiple task topologies; wherein the second obtaining module includes a first updating unit.

[0153] The first updating unit is used to update the first initial task related to the operation information in the first task topology by using the large model to understand the intention of the operation information of the interactive operation, wherein the feedback information is determined based on the initial execution result and the target execution result of the second initial task in the second task topology.

[0154] According to an embodiment of the present disclosure, the first updating unit includes: a semantic understanding subunit and an updating subunit.

[0155] The semantic understanding subunit is used to obtain a semantic understanding result by using a large model to perform semantic understanding on the operation information and the demand information. The semantic understanding result indicates the degree of semantic relevance between the demand information and the operation information.

[0156] The updating subunit is used to update the multiple first initial tasks in the first task topology based on the semantic understanding result.

[0157] According to an embodiment of the present disclosure, the first updating unit includes a tool calling task obtaining subunit.

[0158] The tool call task acquisition subunit is used to understand the tool call intention of the operation information based on the tool description prompt word using the first large model, and obtain the tool call task in the first task topology. The tool description prompt word represents the tool description information related to the preset tool. The tool call task instructs the second large model to call the target tool in the preset tool by executing the tool call task, execute at least one specified task in the first task topology, and obtain the target tool execution result used to determine the feedback information.

[0159] According to an embodiment of the present disclosure, the first updating unit includes a timed calling task obtaining subunit.

[0160] The scheduled task acquisition sub-unit is used to understand the execution time intention of the operation information based on the tool description prompt word using the first large model to obtain the scheduled task, wherein the tool description prompt word represents the tool description information related to the preset tool, and the scheduled task is used to instruct the second large model to call the target tool in the preset tool to perform the specified task at a specified time or specified time period that matches the execution time intention represented by the operation information.

[0161] According to an embodiment of the present disclosure, the information processing apparatus based on the large model further includes a stop execution module.

[0162] The execution stop module is used to stop executing at least one initial task in the task scheduling strategy by using the computing unit of the deployed large model based on the operation information of the interactive operation.

[0163] According to an embodiment of the present disclosure, the orchestration strategy includes multiple task topologies having dependencies with a specified initial task, the stopped state topology in the multiple task topologies includes the initial task whose execution is stopped, wherein the information processing device based on the large model also includes an execution module.

[0164] An execution module is used to use the computing unit to execute the initial task in the associated topology in the process of stopping the execution of at least one initial task in the task scheduling strategy using the computing unit of the deployed large model. The associated topology is other task topologies in multiple task topologies except the stopped state topology.

[0165] According to an embodiment of the present disclosure, the second obtaining module includes a second updating unit.

[0166] The second updating unit is used to update the initial execution result of the initial task according to the operation information carried by the interactive operation on the initial task during the process of executing the task scheduling strategy using the large model, and obtain a first target execution result related to the initial task, wherein the first target execution result is used as input data for at least one unexecuted initial task or target task in the scheduling strategy.

[0167] According to an embodiment of the present disclosure, the second obtaining module further includes a target task obtaining unit.

[0168] The target task obtaining unit is used to use the large model to perform task planning according to the first target execution result and demand information, and obtain a target task that has a dependency relationship with at least one uncompleted initial task.

[0169] According to an embodiment of the present disclosure, the first obtaining module includes: a demand intention obtaining unit and a task scheduling strategy obtaining unit.

[0170] The demand intention acquisition unit is used to use the large model to semantically understand the demand information and obtain the demand intention.

[0171] The task scheduling strategy obtaining unit is used to perform task scheduling based on demand intention and obtain a task scheduling strategy.

[0172] According to an embodiment of the present disclosure, the demand intention includes an expected time intention that characterizes the target object's expectation of obtaining feedback information; and the task scheduling strategy obtaining unit includes a task scheduling strategy obtaining sub-unit.

[0173] The task scheduling strategy acquisition subunit is used to perform task scheduling on the demand content of the demand intention based on the delay prompt information determined by the expected time intention using the first large model to obtain the task scheduling strategy, wherein the second large model executes the task scheduling strategy to obtain the strategy execution time of the feedback information that matches the expected time intention.

[0174] According to an embodiment of the present disclosure, the large model-based information processing apparatus further includes a large model determination module.

[0175] The large model determination module is used to determine the second large model for performing the initial task from the multiple candidate large models based on the task attribute information of the initial task and the model description information of each candidate large model using the first large model to perform semantic understanding. The model description information represents the model performance attributes of the candidate large models.

[0176] According to an embodiment of the present disclosure, the demand intention obtaining unit includes a first processing subunit.

[0177] The first processing sub-unit is used to perform semantic understanding of demand information based on the object attribute knowledge graph using a large model, wherein the object attribute nodes in the object attribute knowledge graph represent the historical interaction information or object preference attributes of the target object.

[0178] According to an embodiment of the present disclosure, the feedback information determination module includes:

[0179] The feedback information obtaining unit is used to obtain feedback information by semantically fusing multiple target execution results using a large model based on style attributes related to the demand information.

[0180] According to an embodiment of the present disclosure, the information processing device based on the large model further includes: a push module.

[0181] The push module is configured to push at least one of the task scheduling strategy and the initial execution result of the initial task to the target object.

[0182] Figure 8 The structural block diagram of an artificial intelligence agent according to an embodiment of the present disclosure is schematically shown.

[0183] In the embodiments of the present disclosure, Figure 8 As shown, the AI ​​agent 800 may include an input module 810 , a processing module 820 and an output module 830 .

[0184] Input module 810, for receiving input information;

[0185] A processing module 820 is configured to determine a target task based on the input information received by the input module, determine a large model based on the target task, and obtain output information by calling the large model to execute the large model-based information processing method provided according to an embodiment of the present disclosure;

[0186] The output module 830 is used to output the output information obtained by the processing module.

[0187] According to an embodiment of the present disclosure, the input module 810 is responsible for receiving or perceiving information such as queries, requests, instructions, signals, or data from the outside world (e.g., a user or the external environment) and converting it into a format that can be understood and processed by the AI ​​agent 800. The input module 810 is the primary link for the AI ​​agent 800 to interact with the outside world. It enables the AI ​​agent 800 to efficiently and accurately obtain the necessary "sensory" information from the outside world and respond to this information.

[0188] In an example, the input module 810 can input the demand information described above, operation information of the interactive operation, etc.

[0189] In the example, the processing module 820 is the core support for the AI ​​agent 800 to handle complex tasks. The processing module 820 can execute the large model-based information processing method described above.

[0190] In this example, the performance of processing module 820 may be closely related to the large model underlying AI agent 800. To fully leverage the capabilities of the large model, the internal structure of processing module 820 may be designed to be highly configurable and extensible to handle a variety of different types of tasks and requirements in real-world scenarios.

[0191] In this example, after AI agent 800 receives a request voice message, processing module 820 uses the large model to process the request information and derive a task scheduling strategy. The large model then executes the task scheduling strategy and updates the initial task based on the interaction, ultimately obtaining the target task. The large model then processes the target execution result to generate feedback information, which is then passed to output module 830.

[0192] Understandably, while large language models possess excellent language understanding and generation capabilities, like humans, they are limited in the tasks they can perform without tools. However, when AI Agent 800 is empowered with tool-based capabilities, it can perform tasks such as mathematical calculations using a calculator, data analysis using Python, and weather forecasting using search engines.

[0193] In an example, the output module 830 may output the feedback information described above.

[0194] The AI ​​agent 800 according to the embodiment of the present disclosure can simply and effectively improve the level of intelligence, and enhance flexibility and versatility.

[0195] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0196] According to an embodiment of the present disclosure, an electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method described above.

[0197] According to an embodiment of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause a computer to execute the method described above.

[0198] According to an embodiment of the present disclosure, a computer program product includes a computer program, and when the computer program is executed by a processor, the computer program implements the method described above.

[0199] Figure 9 A schematic block diagram of an example electronic device that can be used to implement a large model-based information processing method of an embodiment of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0200] like Figure 9 As shown, device 900 includes a computing unit 901, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 902 or a computer program loaded from a storage unit 908 into a random access memory (RAM) 903. RAM 903 may also store various programs and data required for the operation of device 900. Computing unit 901, ROM 902, and RAM 903 are connected to each other via a bus 904. An input / output (I / O) interface 905 is also connected to bus 904.

[0201] Various components in the device 900 are connected to the I / O interface 905, including an input unit 906, such as a keyboard, a mouse, etc.; an output unit 907, such as various types of displays, speakers, etc.; a storage unit 908, such as a magnetic disk, an optical disk, etc.; and a communication unit 909, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 909 allows the device 900 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0202] The computing unit 901 can be any general-purpose and / or specialized processing component with processing and computing capabilities. Some examples of the computing unit 901 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various specialized artificial intelligence (AI) computing chips, various computing units that run machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 901 performs the various methods and processes described above, such as the large-model-based information processing method. For example, in some embodiments, the large-model-based information processing method can be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit 908. In some embodiments, part or all of the computer program can be loaded and / or installed onto the device 900 via the ROM 902 and / or the communication unit 909. When the computer program is loaded into the RAM 903 and executed by the computing unit 901, one or more steps of the large-model-based information processing method described above can be performed. Alternatively, in other embodiments, the computing unit 901 can be configured to perform the large-model-based information processing method through any other suitable means (e.g., via firmware).

[0203] Various embodiments of the systems and techniques described above can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), system-on-chip systems (SOCs), programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system that includes at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0204] The program code for implementing the method of the present disclosure can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0205] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of machine-readable storage media may include an electrical connection based on one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), optical fibers, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0206] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0207] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0208] A computer system may include a client and a server. The client and server are generally remote from each other and typically interact through a communication network. The client-server relationship arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The server may be a cloud server, a server in a distributed system, or a server integrated with a blockchain.

[0209] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved. This is not a limitation herein.

[0210] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the scope of protection of this disclosure.

Claims

1. A large model-based information processing method, comprising: Performing task scheduling on the demand information related to the target object to obtain a task scheduling strategy including multiple initial tasks; In the process of executing the task scheduling strategy using the large model, updating at least one of the initial tasks in the task scheduling strategy according to the interactive operation of the target object to obtain a target task; as well as Based on the target execution result determined by executing the target task using the large model, feedback information matching the demand intention of the target object is determined.

2. The method according to claim 1, wherein The orchestration strategy includes a plurality of task topologies having a dependency relationship with a specified initial task, and the interaction operation is associated with a first task topology among the plurality of task topologies; Wherein, updating at least one of the initial tasks in the task scheduling strategy according to the interactive operation of the target object includes: The large model is used to understand the intention of the operation information of the interactive operation, and to update the first initial task related to the operation information in the first task topology, wherein the feedback information is determined based on the initial execution result of the second initial task in the second task topology and the target execution result.

3. The method according to claim 2, wherein: The updating of the first initial task related to the operation information in the first task topology by using the large model to understand the intent of the operation information of the interactive operation includes: Using the large model to perform semantic understanding on the operation information and the demand information to obtain a semantic understanding result, wherein the semantic understanding result indicates a degree of semantic relevance between the demand information and the operation information; A plurality of first initial tasks in the first task topology is updated based on the semantic understanding result.

4. The method according to claim 2, wherein: The updating of the first initial task related to the operation information in the first task topology by using the large model to understand the intent of the operation information of the interactive operation includes: Based on the tool description prompt words, the first large model is used to understand the tool call intention of the operation information, and the tool call task in the first task topology is obtained. The tool description prompt words represent the tool description information related to the preset tool. The tool call task instructs the second large model to call the target tool in the preset tool by executing the tool call task, execute at least one specified task in the first task topology, and obtain the target tool execution result used to determine the feedback information.

5. The method according to claim 2, wherein: The updating of the first initial task related to the operation information in the first task topology by using the large model to understand the intent of the operation information of the interactive operation includes: Based on the tool description prompt words, the first large model is used to understand the execution time intention of the operation information to obtain a timed call task, wherein the tool description prompt words represent the tool description information related to the preset tool, and the timed call task is used to instruct the second large model to call the target tool in the preset tool to perform the specified task at a specified time or a specified time period that matches the execution time intention represented by the operation information.

6. The method according to any one of claims 1 to 4, wherein The method further comprises: Based on the operation information of the interactive operation, the computing unit where the large model is deployed stops executing at least one initial task in the task scheduling strategy.

7. The method according to claim 6, wherein: The orchestration strategy includes a plurality of task topologies having a dependency relationship with a specified initial task, and a stopped state topology among the plurality of task topologies includes the initial task whose execution is stopped; The method further comprises: In the process of stopping the execution of at least one initial task in the task orchestration strategy using the computing unit on which the large model is deployed, the computing unit is used to execute the initial task in an associated topology, where the associated topology is other task topologies among the multiple task topologies except the stopped state topology.

8. The method according to claim 1, wherein In the process of executing the task scheduling strategy using the large model, updating at least one of the initial tasks in the task scheduling strategy according to the interactive operation of the target object includes: In the process of executing the task orchestration strategy using the large model, the initial execution result of the initial task is updated according to the operation information carried by the interactive operation on the initial task, and a first target execution result related to the initial task is obtained, wherein the first target execution result is used as input data for at least one unexecuted initial task or target task in the orchestration strategy.

9. The method according to claim 8, wherein In the process of executing the task scheduling strategy using the large model, updating at least one of the initial tasks in the task scheduling strategy according to the interactive operation of the target object further includes: The large model is used to perform task planning according to the first target execution result and the demand information to obtain a target task that has a dependency relationship with at least one of the uncompleted initial tasks.

10. The method according to claim 1, wherein The task scheduling is performed on the demand information related to the target object to obtain a task scheduling strategy containing multiple initial tasks, including: Using the large model to perform semantic understanding on the demand information to obtain demand intent; and Perform task scheduling based on the demand intention to obtain the task scheduling strategy.

11. The method according to claim 10, wherein: The demand intention includes an expected time intention representing that the target object expects to obtain the feedback information; The step of performing task scheduling based on the demand intention to obtain the task scheduling strategy includes: Based on the delay prompt information determined by the expected time intention, the first large model is used to perform task scheduling on the demand content of the demand intention to obtain the task scheduling strategy, wherein the second large model executes the task scheduling strategy to obtain the strategy execution duration of the feedback information that matches the expected time intention.

12. The method according to claim 10 or 11, wherein: The method further comprises: Based on the task attribute information of the initial task, the first large model is used to perform semantic understanding according to the model description information of each of multiple candidate large models, and a second large model for performing the initial task is determined from the multiple candidate large models, where the model description information represents the model performance attributes of the candidate large models.

13. The method according to claim 10, wherein: The using the large model to perform semantic understanding on the demand information includes: Based on the object attribute knowledge graph, the large model is used to perform semantic understanding on the demand information, wherein the object attribute nodes in the object attribute knowledge graph represent the historical interaction information or object preference attributes of the target object.

14. The method according to claim 1, wherein Determining feedback information that matches the demand intention of the target object based on the target execution result determined by executing the target task using the large model includes: According to the style attributes related to the demand information, the large model is used to perform semantic fusion on the plurality of target execution results to obtain the feedback information.

15. The method according to claim 1, wherein The method further comprises: At least one of the task scheduling strategy and the initial execution result of the initial task is pushed to the target object.

16. An information processing device based on a large model, comprising: The first acquisition module is used to perform task scheduling on the demand information related to the target object and obtain a task scheduling strategy including multiple initial tasks; A second obtaining module is configured to update at least one of the initial tasks in the task scheduling strategy according to the interactive operation of the target object in the process of executing the task scheduling strategy using the large model to obtain a target task; as well as The feedback information determination module is used to determine feedback information that matches the demand intention of the target object based on the target execution result determined by executing the target task using the large model.

17. An artificial intelligence agent, comprising: An input module, used for receiving input information; a processing module, configured to determine a target task based on the input information received by the input module, determine a large model based on the target task, and execute the method according to any one of claims 1 to 15 by calling the large model to obtain output information; An output module is used to output the output information obtained by the processing module.

18. An electronic device comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 15.

19. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1 to 15.

20. A computer program product comprising a computer program which, when executed by a processor, implements the method according to any one of claims 1 to 15.

Images