Data processing method and device, electronic equipment, medium and program product

Abstract

The application discloses a data processing method and device, electronic equipment, medium and program product; the application obtains target information; according to the target information and the processing record of the target information, a current subtask to be processed is determined from the target information, and an agent to be called corresponding to the current subtask to be processed is determined from an agent pool; the agent to be called is called from the agent pool to process the current subtask to be processed, and a task processing result of the current subtask to be processed is obtained; the task processing result of the current subtask to be processed is added in the processing record, and an updated processing record of the target information is obtained; and a processing result of the target information is obtained from the updated processing record. Therefore, the application dynamically determines the processing flow of the target information according to the current processing progress of the target information, so that the flexibility and adaptability of the target information processing process are improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, specifically to a data processing method, apparatus, electronic device, medium, and program product. Background Technology

[0002] A large-scale intelligent agent is an intelligent entity possessing the capabilities of environmental perception, autonomous understanding, decision-making, and action execution. Built upon a large model, it can simulate independent thought processes and gradually achieve predetermined goals. In practical applications, several intelligent agents can collaborate through task orchestration, enabling the large model to continuously execute tasks and ultimately achieve an information processing objective.

[0003] However, in existing technologies, the task orchestration of intelligent agents during information processing is typically pre-defined by developers. The order of task processing is fixed in the code and cannot be adjusted during the process. This approach results in a fixed intelligent agent and tasks being called each time throughout the entire information processing process, leading to poor flexibility. Especially when intelligent agents face new information processing tasks or complex, changing scenarios, this task orchestration method cannot optimally address new challenges, resulting in poor adaptability. Summary of the Invention

[0004] This application provides a data processing method, apparatus, electronic device, medium, and program product that can improve the flexibility and adaptability of the target information processing process.

[0005] This application provides a data processing method, comprising: acquiring target information; determining a current subtask to be processed from the target information and a processing record of the target information, and determining a corresponding intelligent agent to be invoked from an intelligent agent pool, wherein the processing record includes the task processing results of historically processed subtasks of the target information; invoking the intelligent agent to be invoked from the intelligent agent pool to process the current subtask to be processed, thereby obtaining the task processing result of the current subtask to be processed; adding the task processing result of the current subtask to the processing record to obtain an updated processing record of the target information; and obtaining the processing result of the target information from the updated processing record.

[0006] This application embodiment also provides a data processing apparatus, including: an information acquisition unit for acquiring target information; an agent determination unit for determining a current sub-task to be processed from the target information and a processing record of the target information, and determining an agent to be invoked from an agent pool corresponding to the current sub-task to be processed, the processing record including the task processing results of historically processed sub-tasks of the target information; an agent invocation unit for invoking the agent to be invoked from the agent pool to process the current sub-task to be processed, and obtaining the task processing result of the current sub-task to be processed; a result addition unit for adding the task processing result of the current sub-task to the processing record, and obtaining an updated processing record of the target information; and an information processing unit for obtaining the processing result of the target information from the updated processing record.

[0007] In some implementations, the agent determination unit is specifically used to: acquire agent information of the agents in the agent pool; determine the association between the current subtask to be processed of the target information and the agent information based on the target information and the processing record of the target information; and determine the agent to be invoked from the agent pool based on the association between the current subtask to be processed and the agent information.

[0008] In some implementations, determining the current subtask to be processed of the target information based on the target information and the processing record of the target information includes: determining the task information of the next current subtask to be processed of the target information based on the target information and the processing record of the target information; determining the agent to be invoked from the agent pool based on the association relationship between the current subtask to be processed and the agent information includes: determining the agent to be invoked from the agent pool based on the association relationship between the task information of the next current subtask to be processed and the agent information, wherein the agent to be invoked is used to process the next current subtask to be processed.

[0009] In some embodiments, the data processing device further includes a state determination unit, which is configured to: determine the processing completion status of the target information based on the target information and the processing record of the target information; when the processing completion status is incomplete, determine the current sub-task to be processed from the target information based on the target information and the processing record of the target information, and determine the intelligent agent to be invoked corresponding to the current sub-task to be processed from the intelligent agent pool.

[0010] In some implementations, determining the processing completion status of the target information based on the target information and the processing record of the target information includes: performing semantic understanding on the target information and the processing record of the target information to extract a first context feature, wherein the first context feature represents the semantic association between the target information and the processing record of the target information; and decoding the first context feature to obtain the processing completion status of the target information.

[0011] In some implementations, the step of performing semantic understanding on the target information and the processing records of the target information to extract the first contextual feature includes: extracting keywords corresponding to sub-tasks in the target information from the target information; and performing semantic understanding on the keywords corresponding to the sub-tasks and the processing records of the target information to extract the first contextual feature.

[0012] In some embodiments, the data processing apparatus further includes an agent training unit, which is used to: acquire multiple agents, which are trained through different subtasks; and add the identification information of the agents to the agent pool.

[0013] In some implementations, the agent invocation unit is used to: invoke the agent to be invoked from the agent pool; and use the agent to be invoked to process the current subtask based on the target information and the processing record of the target information to obtain the task processing result of the current subtask.

[0014] In some embodiments, the intelligent agent includes a task processing model and a task execution tool corresponding to the task processing model. The step of using the intelligent agent to be invoked to process the current subtask based on the target information and the processing records of the target information to obtain the task processing result of the current subtask includes: using the task processing model in the intelligent agent to be invoked to perform semantic understanding on the target information and the processing records of the target information to extract second contextual features, the second contextual features representing the semantic relationships between subtasks in the target information; decoding the second contextual features to obtain the task execution elements of the current subtask; and using the task execution tool in the intelligent agent to be invoked to execute the current subtask based on the task execution elements to obtain the task processing result of the current subtask.

[0015] In some implementations, the information processing unit is used to: perform semantic understanding on the updated processing record to extract a third context feature, the third context feature representing the semantic association between the task processing results of the historically processed subtasks in the updated processing record; and decode the third context feature to obtain the processing result of the target information.

[0016] In some embodiments, the data processing apparatus further includes a return execution unit, which is configured to: use the updated processing record as the processing record of the target information, and return the execution step to determine the agent to be invoked and subsequent steps from the agent pool based on the target information and the processing record of the target information, until the target information processing is completed.

[0017] This application also provides an electronic device, including a processor and a memory, wherein the memory stores a plurality of instructions; the processor loads instructions from the memory to execute steps in any of the data processing methods provided in this application.

[0018] This application also provides a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute steps in any of the data processing methods provided in this application.

[0019] This application also provides a computer program product, including a computer program or instructions, which, when executed by a processor, implement the steps in any of the data processing methods provided in this application.

[0020] This application embodiment can obtain target information; based on the target information and the processing record of the target information, determine the current sub-task to be processed from the target information, and determine the intelligent agent to be invoked corresponding to the current sub-task to be processed from the intelligent agent pool, the processing record including the task processing results of the historically processed sub-tasks of the target information; invoke the intelligent agent to be invoked from the intelligent agent pool to process the current sub-task to be processed, and obtain the task processing result of the current sub-task to be processed; add the task processing result of the current sub-task to the processing record to obtain the updated processing record of the target information; and obtain the processing result of the target information from the updated processing record.

[0021] In this application, based on the target information and its real-time processing progress (i.e., processing records), the sub-tasks to be processed for the target information are dynamically determined according to the current processing progress. This allows for dynamic determination of the target information processing flow based on its real-time processing progress, improving the flexibility of the target information processing process and its adaptability to different target information. Simultaneously, when facing new information processing tasks or complex, changing scenarios, by evaluating and invoking the necessary intelligent agents during the target information processing, it is ensured that each sub-task of the target information can be matched with a suitable intelligent agent for processing, thereby improving the adaptability of the intelligent agents and the accuracy of task processing. Furthermore, in this embodiment, an intelligent agent pool is used to manage available intelligent agents, thereby dynamically selecting a suitable intelligent agent from the pool to process the currently pending sub-tasks, meeting the need to dynamically adjust the task processing flow according to the actual task progress and ensuring the flexibility of the task processing process. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1a This is a schematic diagram of a scenario illustrating the data processing method provided in an embodiment of this application;

[0024] Figure 1b This is a flowchart illustrating the data processing method provided in an embodiment of this application;

[0025] Figure 1c This is a schematic diagram of the data processing flow provided in the embodiments of this application;

[0026] Figure 2a This is a flowchart illustrating a data processing method provided in another embodiment of this application;

[0027] Figure 2b This is a schematic diagram of the data processing system provided in an embodiment of this application;

[0028] Figure 2c This is a schematic diagram of a data processing flow provided in another embodiment of this application;

[0029] Figure 3 This is a schematic diagram of the structure of the data processing apparatus provided in the embodiments of this application;

[0030] Figure 4 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0031] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0032] This application provides a data processing method, apparatus, electronic device, medium, and program product.

[0033] Specifically, the data processing device can be integrated into an electronic device, such as a terminal or server. The terminal can be a mobile phone, tablet, smart Bluetooth device, laptop, or personal computer (PC); the server can be a single server or a server cluster consisting of multiple servers.

[0034] In some embodiments, the data processing apparatus may also be integrated into multiple electronic devices, such as multiple servers, with the data processing method of this application being implemented by the multiple servers.

[0035] In some embodiments, the terminal can also be used as a server to implement some or all of the functions of a server.

[0036] For example, refer to Figure 1a This data processing method can be applied to a server that can obtain target information from a terminal. Based on the target information and its processing records, the server determines the current subtask to be processed from the target information and the corresponding agent to be invoked from the agent pool. The processing records include the historical processing results of subtasks processed by the target information. The server invokes the agent to be invoked from the agent pool to process the current subtask and obtains its processing result. The processing result of the current subtask is added to the processing records to obtain the updated processing records for the target information. The server obtains the processing result of the target information from the updated processing records and returns the processing result of the target information to the terminal.

[0037] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0038] The following sections provide detailed descriptions. It should be noted that the order of the following embodiments is not intended to limit the preferred order of the embodiments. It is understood that in the specific embodiments of this application, user-related data such as target information, samples, and processing results are involved. When the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0039] In this embodiment, a data processing method is provided, such as... Figure 1b As shown, the specific flow of this data processing method can be as follows:

[0040] 110. Obtain target information.

[0041] The target information refers to the information to be processed. It can be a problem, requirement, task, or a combination of these forms. The target information may include specific content or the task objective of processing it. For example, when the target information is a problem, it may include the specific question and the expected type of answer or the form of the processing result. In this embodiment, the processing result of the target information can be determined based on a knowledge base, text, database, or other information sources. These sources can be open-domain or closed-domain information sources, or a combination of both.

[0042] For example, in this embodiment, the target information may include information comprising multiple sub-tasks. These sub-tasks may be tasks of the same type or different types. For instance, the multiple sub-tasks may include one or more of the following: query task, reasoning task, generation task, classification task, translation task, etc. It should be noted that in this embodiment, the sub-tasks may be explicitly or implicitly included in the target information.

[0043] For example, the target information could be "Please help arrange a two-day travel itinerary for four adults, preferably outdoor activities, relaxing and comfortable." This information implicitly includes several query tasks (i.e., subtasks) such as "check the weather," "check hotels," "check flight tickets," "check train tickets," "check car rentals," "check attractions," and "cost assessment," etc.

[0044] For example, the target information could be "Please show with pictures and text why the universe is expanding". This information implicitly includes several sub-tasks such as "Retrieving relevant information on the expansion of the universe" (i.e., query task), "Inferring from the retrieved information" (i.e., reasoning task), and "Generating graphic content based on the reasoning results" (i.e., generation task), etc.

[0045] For example, the target information could be "to query train ticket information from A to B on a certain day, and to query the weather at each station of train XX on a certain day". This information explicitly includes the following multiple query tasks (i.e., sub-tasks): "query train tickets" (i.e. query task), "query weather" (i.e. query task), etc.

[0046] 120. Based on the target information and the processing record of the target information, determine the current subtask to be processed from the target information, and determine the intelligent agent to be invoked corresponding to the current subtask to be processed from the intelligent agent pool. The processing record includes the task processing results of the historically processed subtasks of the target information.

[0047] Among them, the currently pending subtasks refer to the subtasks that need to be processed during the processing of target information; that is, the subtasks that are currently pending processing are those that are processing target information but have not yet been processed. The historically processed subtasks refer to the subtasks that have been completed and whose processing results (i.e., task processing results) have been obtained during the processing of target information.

[0048] In this context, "processing records" refers to information used to record the task processing results of historically processed subtasks that contain target information. The task processing results of historically processed subtasks refer to the results generated by those subtasks. Different historically processed subtasks can generate different task processing results depending on their task type. For example, the task processing results of a subtask can be used to record all information during execution, including: time, agent name, execution tool name, and execution result.

[0049] In this context, the agent pool refers to a collection of agents capable of performing tasks, and this agent pool can be used to manage available agents. For example, in the embodiments of this application, the agent pool itself is a container, which can be represented by data structures such as arrays, lists, sets, or dictionaries.

[0050] In this context, an intelligent agent refers to an intelligent entity built upon a large model, possessing the capabilities of environmental perception, autonomous understanding, decision-making, and action execution. The intelligent agent to be invoked refers to the intelligent agent used to process the currently pending sub-task. For example, in this embodiment, an intelligent agent, as a single entity in an intelligent agent pool responsible for task processing, can perform specific processing or operations according to the requirements of the sub-task to cooperate with other intelligent agents to achieve the task objective of the target information. Exemplarily, the intelligent agent can be constructed from a large language model (such as a dialogue generation model (ChatGPT), a pre-trained language model (BERT), etc.) and can process sub-tasks through natural language processing (NLP) to provide real-time responses or perform corresponding tasks.

[0051] For example, based on the target information and its processing records, the sub-tasks of the target information to be processed next (i.e., the current sub-tasks to be processed) can be inferred, so as to determine the intelligent agent that can be used to process the sub-task from the intelligent pool. In the prior art, task orchestration is usually preset, and the task processing order is fixed. Compared with the prior art, the embodiments of this application are based on the target information and its real-time processing progress (i.e., processing records) to dynamically determine the subsequent processing of the target information according to the current processing progress. Therefore, task orchestration does not rely on preset, fixed task processing flows, and the task processing flow can be dynamically adjusted according to the actual progress, thereby improving the flexibility of task orchestration and its adaptability to different target information during task processing. At the same time, when facing new tasks or complex, changing scenarios, by evaluating and calling the required intelligent agents during task execution, it is ensured that each sub-task can be matched with a suitable intelligent agent for processing, thereby improving the adaptability of the intelligent agent and the accuracy of task processing.

[0052] Furthermore, in this embodiment, the agent is not fixed on a certain execution node. Instead, the available agents are managed through an agent pool. This allows for timely response to the processing needs of the current subtasks to be processed. Appropriate agents are dynamically selected from the agent pool to handle the tasks, thus meeting the need to dynamically adjust the task processing flow according to the actual progress and ensuring the flexibility of task orchestration during task processing.

[0053] In some implementations, the agent control module can be used to dynamically and in real-time orchestrate subtasks during task processing. That is, the agent control module determines the agents to be invoked from the agent pool.

[0054] In some implementations, pre-training for different types of tasks can yield multiple agents capable of handling different sub-tasks, meeting the needs of dynamic, complex, and changing scenarios. Simultaneously, by adding the agent's identification information to an agent pool, agents with corresponding capabilities can be scheduled from the pool to execute sub-tasks based on the identification information, without needing to focus too much on the specific agent details, thus reducing the complexity of invoking agents. Specifically, before determining the current sub-task to be processed from the target information and the processing records of the target information, and before determining the agent to be invoked from the agent pool corresponding to the current sub-task to be processed, the process further includes:

[0055] Multiple agents are acquired, and these agents are trained through different subtasks.

[0056] Add the agent's identification information to the agent pool.

[0057] The identification information of an intelligent agent may include, but is not limited to, names, labels, identifiers, and other information that can be used to identify the intelligent agent.

[0058] For example, training samples of different types of tasks can be used to pre-train the agents (i.e., candidate agents) in the agent pool, enabling each agent to handle the corresponding type of subtask. Thus, different subtasks require different agents to execute. The trained agents can be pooled into a list, such as agent_list = [agent1, agent2, agent3, agent4, agent5, agent6, agent7], where agent represents an agent. This avoids fixing agents to a specific execution node; instead, they are pooled into a list so that the agent control module can select a suitable agent from the pool to handle the current subtask.

[0059] For example, such as Figure 1cThe data processing flow shown initializes the agent control module after obtaining the target information (i.e., the original task) through external input (such as user-inputted task instructions). Initialization can include agent pooling, which involves determining the agent list and the LLM (Large Language Model) and tools used by the agents. For example, the initialization information for agent 1 (i.e., agent1) could be agent1 = initialize_agent(tools1, llm, agent = AgentType.ZERO_SHOT_REACT_DESCRIPTION, verbose = True). In addition, agent information such as name and description attributes can be added for subsequent calls to the agent control module. For example, agent 1's information could be agent1.name = "agent1" and agent1.description = "The agent can query weather info."

[0060] In some implementations, the completion status of the target information can be determined first based on the target information and processing records to ensure that the agent is only invoked to process the pending subtasks for incomplete target information. This approach, by designing a target information processing completion status detection mechanism, ensures that the task progress is checked before each agent invocation. This avoids repeatedly invoking the agent to process pending subtasks for already processed target information, thus preventing wasted computing resources. Specifically, based on the target information and its processing records, the current pending subtask is determined from the target information, and the corresponding agent to be invoked for the current pending subtask is determined from the agent pool, including:

[0061] Based on the target information and its processing records, determine the completion status of the target information processing.

[0062] When the processing completion status is incomplete, based on the target information and the processing record of the target information, the current subtask to be processed is determined from the target information, and the corresponding intelligent agent to be invoked is determined from the intelligent agent pool.

[0063] The processing completion status of target information refers to the status indicating whether the target information has been processed. For example, the processing completion status can be "completed" or "not completed". Specifically, the processing completion status can be represented as "completed" or "processing result successfully generated" (i.e., completed), or it can be represented as "not completed" or "processing".

[0064] For example, the processing records of historically processed subtasks can be read one by one from the processing record of the target information. The task status corresponding to each processing record can be analyzed step by step based on the target information. For instance, the processing results of subtasks 1 through 3 can be read sequentially from the target information processing record results ["Subtask 1 processing result", "Subtask 2 processing result", "Subtask 3 processing result"]. Each time a subtask's processing result is read, the target information is identified as having been processed when that subtask is completed. If the target information is not processed, the subtask is marked as "incomplete". If any subtask (such as subtask 3) is read and the target information is identified as processed, it is marked as "completed", thus confirming that the target information has been processed. If, after reading the processing results of all subtasks, all subtasks are marked as incomplete, the target information's processing status is determined to be incomplete. At this point, the target information and its processing record can be obtained to extract the currently pending subtasks from the target information, and based on the target information and its processing record, the agent to be invoked from the agent pool can be determined to continue processing the target information.

[0065] In some implementations, when the processing completion status is "complete," the processing result of the target information can be obtained from the target information processing record. Thus, when target information processing is detected to be complete, the processing results of all historically processed subtasks can be read from the target information processing record, and the target information processing result is obtained by combining or integrating these processing results.

[0066] In some implementations, implicit semantic relationships between target information and its processing records can be extracted using a large language model. Based on these extracted implicit semantic relationships, the processing completion status of the target information can be directly obtained with only one prediction process, improving the efficiency and accuracy of determining the processing completion status. Specifically, determining the processing completion status of the target information based on the target information and its processing records includes:

[0067] Semantic understanding is performed on the target information and the processing records of the target information to extract the first context features, which represent the semantic relationship between the target information and the processing records of the target information.

[0068] Decode the first context feature to obtain the processing completion status of the target information.

[0069] Semantic understanding, in the field of natural language processing, involves extracting semantic feature representations (i.e., contextual features) from text based on its contextual information. For example, one or more combinations of large language models such as BERT and GPT can be used to perform semantic understanding on the target information and its processing records to obtain the first contextual features.

[0070] For example, target information and its processing records can be input into a large language model (such as BERT, GPT, etc.) to encode the target information and processing records to obtain contextual features (i.e., first contextual features). Since the large language model can capture the implicit semantic relationships between the target information and processing records in the context when encoding them, these semantic relationships can characterize the relationship between the target information and the processing records; that is, they can be regarded as implicit semantic associations between the target information and the processing records.

[0071] For example, such as Figure 1c The data processing flow shown allows the intelligent agent control module to call the large language model to determine whether the target information has been executed completely based on the original target information (i.e., target information) and processing records.

[0072] A large language model can decode the target information's processing completion status from the first contextual features. During the decoding and output generation process, the language model can fully utilize the information contained in the first contextual features (i.e., the implicit semantic association between the target information and the processed record) to understand the semantic relationship between them. This allows the model to extract the association between the target information and the processed record, which represents the processing completion status of the target information. Understandably, a high degree of association between the target information and the processed record indicates that the target information has been processed successfully, while a low degree of association indicates that the target information has not been processed successfully. The large language model can be pre-trained to handle the processing completion status of the output target information. Therefore, only one model processing step is needed to determine the processing completion status of the target information.

[0073] In some implementations, the target information is further processed and used by a large language model to directly determine the processing completion status of the target information. This further processing method may include, but is not limited to, extracting key content from the target information (such as keywords corresponding to sub-tasks). For example, keywords related to the corresponding sub-tasks can be extracted from the target information, and these extracted keywords can be combined into keyword combinations corresponding to possible sub-tasks. These keyword combinations corresponding to sub-tasks are then used as input to the large language model to obtain more accurate features and improve the accuracy of determining the task processing status. Specifically, semantic understanding is performed on the target information and its processing records to extract first contextual features, including:

[0074] Extract keywords from the target information that correspond to the sub-tasks within the target information;

[0075] Semantic understanding is performed on the processing records of keywords and target information corresponding to subtasks to extract first contextual features.

[0076] For example, target information can be input into a large language model (such as BERT, GPT, etc.). The large language model performs semantic understanding on the target information and outputs keyword combinations corresponding to sub-tasks in the target information. In this way, keyword combinations related to the corresponding sub-tasks can be extracted from the target information, such as ["Subtask 1: Keyword 1, Keyword 2", "Subtask 2: Keyword 2, Keyword 3", "Subtask 3: Keyword 4, Keyword 5"]. The large language model can be pre-trained to process the target information and output keyword combinations corresponding to sub-tasks in the target information. The keyword combinations corresponding to sub-tasks in the target information and the processing records of the target information can be input into a language model (such as BERT, GPT, etc.). The large language model encodes the keyword combinations and processing records to obtain the first contextual features and decodes and outputs the predicted processing completion status of the target information.

[0077] In some implementations, based on target information and processing records, the currently pending subtask in the target information and its association with agents in the agent pool can be determined to more accurately allocate the currently pending subtask to the most suitable agent, ensuring precise subtask allocation. Specifically, based on target information and its processing records, the currently pending subtask is determined from the target information, and the agent to be invoked corresponding to the currently pending subtask is determined from the agent pool, including:

[0078] Obtain agent information from agents in the agent pool;

[0079] Based on the target information and its processing records, determine the current subtask to be processed for the target information;

[0080] Determine the association between the currently pending subtasks and the agent information;

[0081] Based on the association between the current subtask to be processed and the agent information, the agent to be invoked is determined from the agent pool.

[0082] Agent information refers to information related to an agent, which can be used to represent the agent's name, state, capabilities, etc. For example, agent information may include, but is not limited to, at least one of the following: agent identifier, agent state (e.g., performing a task, idle, etc.), agent capabilities (e.g., the types of subtasks that the agent can specifically handle), etc.

[0083] For example, one can obtain agent information for all agents in the agent pool, such as directly retrieving agent information from a list of agents in the agent pool. Combining the target information, processing records, and agent information from the agent pool, the agent most relevant to the current subtask to be processed related to the target information is determined as the agent to be invoked.

[0084] In some implementations, the currently pending subtask is considered a single subtask. Since subtasks within the target information may have dependencies—that is, the execution of one subtask depends on the result of another—for example, the subtask "inferring from the retrieved information" in the target information "Please illustrate why the universe is expanding" depends on the subtask "Retrieving information related to the expansion of the universe." If these two subtasks run concurrently, it may lead to incorrect execution order or inconsistent results. Therefore, one currently pending subtask can be processed in each round, gradually and sequentially calling the agent to process the subtasks within the target information. This makes the execution process of the subtasks within the target information more independent and clear, avoiding errors or conflicts caused by concurrent execution.

[0085] In some implementations, the current subtask to be processed can be multiple subtasks. Multiple subtasks within the target information subtask may not have dependencies on each other. For example, within the target information "organize new product launch," the subtasks "write promotional text" and "query personnel information" do not have dependencies on each other. Multiple subtasks without dependencies can be designated as the current subtask to be processed, and the agents in the agent pool that handle these multiple subtasks can be designated as agents to be invoked. This allows multiple subtasks to be processed in the current round, reducing waiting time and improving processing efficiency.

[0086] To more accurately determine the current subtask to be processed, the subtasks of the target information can be filtered according to dependencies to obtain at least one subtask without dependencies as the current subtask to be processed. For example, in some embodiments, determining the current subtask to be processed based on the target information and its processing records includes: determining candidate subtasks to be processed based on the target information and its processing records; when there are multiple candidate subtasks to be processed, determining the current subtask to be processed from among the candidate subtasks based on the dependencies between them. The current subtask to be processed includes candidate subtasks that have no dependencies on other candidate subtasks to be processed, and prerequisite tasks among candidate subtasks that have dependencies on other candidate subtasks to be processed.

[0087] For example, based on target information and processing records, at least one subtask to be processed can be identified. If there is only one subtask, it is designated as the current subtask. If there are multiple subtasks, the dependencies between them are determined, and the subtask without dependencies is designated as the current subtask. If dependencies exist between multiple subtasks, the preceding task among them is designated as the current subtask.

[0088] In some implementations, a subtask to be processed (i.e., the next subtask to be processed) can be first extracted from the target information. Then, based on the task information of this subtask, a corresponding agent to be invoked is selected from the agent pool. This allows for the real-time extraction of the task to be processed based on the target information and related information such as processing progress, enabling flexible and dynamic allocation of the next stage's subtask and increasing the flexibility of task orchestration in the overall task processing flow. Furthermore, by first identifying the current subtask to be processed based on the target information and processing history, and then further determining the corresponding agent to be invoked, the appropriate agent can be accurately matched to the subtask's task information, ensuring more precise allocation of subtasks. Specifically, based on the target information and its processing records, the current subtask to be processed for the target information is determined, including:

[0089] Based on the target information and the processing records of the target information, determine the task information of the next currently pending subtask of the target information;

[0090] Based on the association between the currently pending subtasks and agent information, the agents to be invoked are determined from the agent pool, including:

[0091] Based on the association between the task information of the next currently pending subtask and the agent information, the agent to be invoked is determined from the agent pool. The agent to be invoked is used to process the next currently pending subtask.

[0092] The task information of a subtask refers to information related to the subtask. This task information may include the task content of the subtask, or it may also include the task objective (i.e., the form of the result of the processing), etc.

[0093] For example, agent information for all agents in the agent pool can be obtained, such as directly retrieving agent information from the agent pool's list. Based on the target information and processing records, the task information of the current sub-task to be processed by the target information is determined, and the similarity between the task information of each current sub-task and the agent information (such as the description information of agent capabilities) of the agents in the agent pool's list is calculated one by one. The agent with the highest similarity is selected as the agent to be invoked.

[0094] In some implementations, a large language model can be used to determine the current subtask to be processed from the target information, and to determine the agent to be invoked from the agent pool. For example, the target information and its processing records can be input into a language model (such as BERT, GPT, etc.) to encode the target information and its processing records into feature representations (i.e., contextual features). Because the language model can capture the implicit semantic relationships within the context when encoding the target information and processing records, these semantic relationships can characterize the context and dependencies of the subtasks in the target information, providing crucial clues for task understanding and subsequent processing. Especially for subtasks implicitly contained in the target information, these subtasks can be accurately identified based on the implicit semantic relationships.

[0095] The large language model can decode contextual features to output the predicted task information of the current subtask to be processed, and the agent identifier of the agent to be invoked based on the association between the task information of the current subtask and the agent information. During the decoding and output generation process, the large language model can fully utilize the information contained in the feature representation (i.e., the context and dependencies of the subtask in the target information) to understand the specific target information of the subtask, thereby extracting the task information of the next current subtask to be processed from the target information. Furthermore, it can encode the task information of the next current subtask and the agent information to obtain new feature representations (i.e., new contextual features). Because the large language model can capture the implicit semantic relationships (i.e., new feature representations) of these information in the context when encoding the target information and processing records, these semantic relationships can characterize the implicit semantic association between the current subtask and the agent. Therefore, the agent to be invoked most relevant to the next current subtask can be determined by decoding this new feature representation. The large language model can be pre-trained to process and output the task information of the current subtask and the agent identifier of the determined agent to be invoked. Therefore, based on the extracted implicit semantic associations, the agent to be invoked can be obtained directly with only one prediction process, improving the efficiency and accuracy of determining the agent to be invoked.

[0096] For example, such as Figure 1c The data processing flow shown is such that when the agent control module determines that the target information has not been completed, it calls the LLM (Large Language Model) to determine a suitable agent (i.e., the agent to be called) from the agent pool based on the target information and the processing record of the target information.

[0097] 130. Call the agent to be called from the agent pool to process the current subtask and obtain the task processing result of the current subtask.

[0098] For example, based on the agent information of the agent to be invoked, such as the agent identifier, the agent with the corresponding identifier can be found in the agent pool and invoked to handle the current subtask. It is understood that since an agent is an entity used to undertake the corresponding subtask, in this embodiment, invoking the agent to be invoked can achieve the task objective of the current subtask; that is, the agent to be invoked can complete the task of handling the current subtask.

[0099] In some implementations, the current subtask can be processed based on overall target information and historical experience to optimize the decision-making and processing of the agent to be invoked, thereby improving the accuracy of the processing results. Furthermore, processing the current subtask based on overall target information and historical experience helps the agent to be invoked to more comprehensively understand the needs of the current subtask, improves the coherence of task execution, and ensures consistency in the processing results of the overall target information. Specifically, invoking the agent to be invoked from the agent pool to process the current subtask and obtaining the task processing result includes:

[0100] Call the agent to be called from the agent pool;

[0101] Using the agent to be invoked, the current subtask is processed based on the target information and the processing record of the target information, and the task processing result of the current subtask is obtained.

[0102] For example, based on the agent information of the agent to be invoked, such as the agent identifier, the agent with the corresponding identifier is searched in the agent pool and invoked. The target information and the processing record of the target information are input into the agent to process the current sub-task to obtain the task processing result of the current sub-task.

[0103] Understandably, since the agent to be invoked can achieve the task objective of the current subtask to be processed, even if the overall information of the objective and the processing record are input into the agent to be invoked, the agent can identify the relevant characteristics of the current subtask to be processed based on the details of the context and the historical processing record, so as to accurately execute the current subtask to be processed and avoid omissions or misjudgments in task execution.

[0104] Furthermore, the overall target information and processing records not only include the specific requirements of the current subtask to be processed, but also provide the execution background of previous tasks, the results of previous task processing, and possible decision-making basis. Upon receiving this information, the agent can identify the dependencies and logical connections between subtasks, thereby ensuring the correct execution of the current subtask. Thus, by considering the overall target information and historical experience, the agent can obtain more contextual information, avoiding misunderstandings when processing individual subtasks. This helps the agent to more clearly grasp the true goal of each subtask, thereby providing users with more accurate processing results.

[0105] In some implementations, semantic understanding based on target information and historical experience allows for precise analysis of task requirements, breaking down sub-tasks to be executed and their key elements (e.g., objectives, steps, requirements). This ensures that the execution tool can perform targeted task processing, reducing unnecessary errors or deviations. Specifically, a large language model extracts implicit semantic relationships between sub-tasks in the target information and its processing records. Based on these extracted implicit semantic relationships, only one prediction process is needed to directly obtain the task execution elements of the current sub-task, improving the efficiency and accuracy of executing it. Specifically, the intelligent agent includes a task processing model and a corresponding task execution tool. Using the invoked intelligent agent, the current sub-task is processed based on the target information and its processing records to obtain the task processing result, including:

[0106] By utilizing the task processing model in the agent to be invoked, semantic understanding is performed on the target information and the processing records of the target information to extract second context features, which represent the semantic association between subtasks in the target information.

[0107] Decode the second context feature to obtain the task execution elements of the current subtask to be processed;

[0108] Using the task execution tool in the agent to be invoked, the current subtask to be processed is executed based on the task execution elements to obtain the task processing result of the current subtask to be processed.

[0109] The task processing model is used to determine the task execution elements of the current subtask to be processed in the target information. Task execution elements refer to the elements required by the task execution tool to complete the current subtask. These elements collectively influence the task execution tool's decision-making, thereby driving the tool to execute the current subtask. Different types of subtasks may require different task execution tools, and the required task execution elements may also differ. For example, for a query task, the task execution elements required by its corresponding task execution tool (such as an online query tool or a specific database query tool) may include at least one of the following: query keywords, query conditions, data source, and query method. For a translation task, the task execution elements of its corresponding task execution tool (such as a professional translation tool) may include at least one of the following: source language, target language type, etc. For a generation task, the task execution elements of its corresponding task execution tool (such as a text generation tool) may include at least one of the following: text description, output type, etc.

[0110] For example, the task processing model can be a large language model (such as BERT, GPT, etc.) to encode the target information and the processing records of the target information to obtain contextual features (i.e., second contextual features). Since the large language model can capture the semantic relationships implicit in the context when encoding the target information and processing records, these semantic relationships can represent the context and dependencies of subtasks in the target information, providing key clues for task understanding and subsequent processing.

[0111] Since the large language model of the agent to be invoked is a pre-trained model for processing the corresponding sub-tasks, it can decode the second contextual features and output the predicted task execution elements of the current sub-task to be processed. During the decoding and output generation process, the language model can fully utilize the information contained in the feature representation (i.e., the context and dependencies of the sub-task in the target information) to understand the specific target information of the sub-task, thereby extracting the task execution elements of the current sub-task to be processed from the target information. The large language model can be pre-trained to process and output the task execution elements of the current sub-task to be processed. Thus, based on the extracted implicit semantic relationships, the task execution elements can be directly obtained with only one prediction process, improving the efficiency and accuracy of determining the processing of the current sub-task.

[0112] After obtaining the task execution elements of the current subtask to be processed, these elements can be input into the corresponding task execution tool to execute the subtask and obtain the task processing result. For example, if the current subtask is a query task, task execution elements such as query keywords and query conditions can be input into the query tool, and the corresponding result can be used as the task processing result of the current subtask. Similarly, if the current subtask is a translation task, task execution elements such as source language and target language type can be input into the translation tool, and the translated text can be used as the task processing result of the current subtask. Furthermore, if the current subtask is a generation task, task execution elements such as text description and output type can be input into the generation tool, and the generated graphic result can be used as the task processing result of the current subtask.

[0113] In some implementations, the task information of the current subtask to be processed can be added to the input of the task processing model in the agent to be invoked. For example, the target information, the task information of the current subtask to be processed, and the processing record of the target information can be input into the task processing model. This allows the task processing model to better guide the model to capture the potential semantics related to the current subtask when encoding this information, thereby improving the accuracy of determining the task execution elements of the current subtask to be processed.

[0114] In some implementations, a prompt corresponding to the current subtask can be generated based on the task information of the subtask to be processed. This prompt, along with the target information and the processing record of the target information, is then input into the task processing model of the agent to be invoked for processing. This prompt guides the model more accurately to capture the potential semantics related to the current subtask, thereby improving the accuracy of determining the task execution elements of the current subtask. For example, the prompt could be "Execute subtask XX based on the following information: target information and the processing record of the target information."

[0115] 140. Add the processing results of the current subtask to the processing record to obtain the updated processing record with the target information.

[0116] The updated processing record refers to the processing record after adding the processing results of the current subtask to be processed.

[0117] For example, the processing record of target information can store the processing results of previously processed subtasks (such as subtasks 1 to 3) in the order of processing ["Subtask 1 processing result", "Subtask 2 processing result", "Subtask 3 processing result"]. After the agent processes the current subtask to be processed (such as subtask 3), the processing result of that subtask can be added to the processing record in sequence to obtain the updated processing record ["Subtask 1 processing result", "Subtask 2 processing result", "Subtask 3 processing result", "Subtask 4 processing result"].

[0118] In some implementations, the target information's objective can be achieved through multiple rounds of sub-task processing. That is, without pre-scheduling the execution order of the target information, each round processes the currently pending task based on the target information and the current processing record. This allows for dynamic adjustment of the task processing flow based on the actual progress of the target information by repeatedly and progressively executing the currently pending sub-tasks within the target information, thereby improving the flexibility of task scheduling and adaptability to different target information. Specifically, after adding the task processing result of the currently pending sub-task to the processing record to obtain the updated processing record of the target information, and before obtaining the processing result of the target information from the updated processing record, the process further includes:

[0119] The updated processing record is used as the processing record of the target information. The execution steps are then returned. Based on the target information and the processing record of the target information, the agent to be called and subsequent steps are determined from the agent pool until the target information processing is completed.

[0120] For example, after obtaining the target information, we can first break down the current subtask to be processed in the current round (such as subtask 1) based on the target information and processing records, and determine the agent 1 to process subtask 1. We then call agent 1 to process the subtask and obtain its processing result, adding it to the target information's processing record. It should be noted that when processing the target information's current subtask for the first time (i.e., in the first round), the target information's processing record will be empty because no subtasks have been processed previously.

[0121] After processing subtask 1, based on the target information and processing records (including the processing result of subtask 1), the current unprocessed subtask to be processed in the current round (such as subtask 2) can be extracted, and the agent 2 to process subtask 2 can be determined. The processing result of subtask 2 is obtained by calling agent 2 and added to the processing record of the target information. This process continues, continuously extracting and processing the current unprocessed subtasks from the target information until the target information is fully processed. For example, the completion of target information processing can be determined based on the target information and processing records, or when a current unprocessed subtask cannot be extracted from the target information (i.e., there is no current unprocessed subtask in the target information).

[0122] For example, such as Figure 1c The data processing flow shown describes a process where, after invoking an agent to process the current subtask, all information related to the subtask's execution, such as time, agent name, execution tool name, and execution result, is written into the processing record. The agent control module completes one round of agent invocation, and an execution record is added to the target information processing record. Then, the original target information (i.e., the target data), the agent pool, and the processing record are fed into the agent control module for the next round of invocation.

[0123] In some implementations, the number of times the agent to be called can be returned (i.e., the number of loops) can be limited by setting a maximum number of loops. This limits the number of times the agent control module can determine the agent to be called, thus avoiding the situation where an infinite loop occurs and causes the agent control module to malfunction.

[0124] 150. The processing result of the target information is obtained from the updated processing record.

[0125] The processing result of the target information refers to the response content determined for the question related to the target information. For example, in the embodiments of this application, the processing result of the target information may include the processing result obtained by combining or integrating the task processing results of all sub-tasks broken down from the target information.

[0126] For example, the processing results of historically processed subtasks (such as subtasks 1 to 4) can be read from the updated processing record. These results can be directly combined to obtain a single processing result containing the results of each subtask. Alternatively, the processing results of subtasks 1 to 4 can be integrated and optimized to obtain a coherent and complete processing result.

[0127] In some implementations, implicit semantic relationships between historically processed subtasks can be extracted from the processing records of target information using a large language model. Based on these extracted implicit semantic relationships, the processing result of the target information can be directly obtained with only one prediction process, improving the efficiency and accuracy of the processing result of the determined target information. Specifically, the processing result of the target information is obtained from the updated processing records, including:

[0128] Semantic understanding is performed on the updated processing records to extract third context features, which represent the semantic association between the task processing results of historically processed subtasks in the updated processing records.

[0129] The third context feature is decoded to obtain the processing result of the target information.

[0130] For example, the processing records of target information can be input into a large language model (such as BERT, GPT, etc.) to encode the processing records of target information and obtain contextual features (i.e., third contextual features). Since the large language model can capture the semantic relationships implicit in the context when encoding the processing records, these semantic relationships can characterize the relationships between the task processing results of previously processed sub-tasks, that is, they can be regarded as implicit semantic associations between the task processing results of previously processed sub-tasks.

[0131] Large language models can decode the target information from third-context features. During the decoding and output generation process, the language model fully utilizes the information contained in the third-context features (i.e., the implicit semantic relationships between the processing results of previously processed sub-tasks) to understand the semantic relationships between them. Based on these relationships, it optimizes and integrates the processing results of previously processed sub-tasks to obtain the target information. This can be achieved by pre-training the large language model to process the target information. Thus, the target information processing result can be determined in a single model processing step.

[0132] For example, such as Figure 1cThe data processing flow shown is as follows: if the intelligent agent control module determines that the question and answer process is completed, it will extract the target information processing record and hand it over to the LLM (processing record) to summarize the task processing results of the subtasks. The processing results of multiple tasks will be summarized and the final result (i.e., the processing result of the target information) will be output.

[0133] In this embodiment, a large language model can be used for at least one processing procedure, such as determining the agent to be invoked from the agent pool, determining the processing completion status of target information, processing the current subtask to be processed, and obtaining the processing result of the target information, among one or more of these procedures. These procedures may involve inputting the same or similar information; directly using a large language model for processing could lead to confusing or erroneous results. In some implementations, when using a large language model to process each procedure, prompts corresponding to the processing procedure can be added to the input of the large language model. These prompts guide the large language model to capture the necessary implicit semantics from the input information, thereby predicting the output result corresponding to the prompts. For example, the prompt for the procedure of determining the agent to be invoked from the agent pool could be: "Based on the following target information and the target information processing record, select the most suitable agent to process the current subtask to be processed. The target information and the target information processing record are XX." The prompt for the procedure of determining the processing completion status of the target information could be: "Based on the following target information and the target information processing record, check the processing progress of the target information and determine whether the target information has been completed. The target information and the target information processing record are XX." The prompt for processing the current pending subtask could be: "Process the current pending subtask based on the following target information and target information processing record. The target information and target information processing record are XX." The prompt for obtaining the target information processing result could be: "Generate the final processing result based on the target information processing record. The target information processing record is XX."

[0134] The target information processing scheme provided in this application can be applied to various task processing scenarios. For example, taking knowledge question answering as an example, the following steps are taken: 1. Obtain target information; 2. Based on the target information and its processing record, determine the current sub-task to be processed from the target information, and determine the corresponding intelligent agent to be called from the intelligent agent pool. The processing record includes the task processing results of previously processed sub-tasks of the target information; 3. Call the intelligent agent to be called from the intelligent agent pool to process the current sub-task to be processed, obtaining the task processing result of the current sub-task; 4. Add the task processing result of the current sub-task to the processing record, obtaining the updated processing record of the target information; 5. Obtain the processing result of the target information from the updated processing record.

[0135] As can be seen from the above, the embodiments of this application dynamically determine the subsequent processing of the target information based on the target information and its real-time processing progress (i.e., processing records). Therefore, task orchestration does not rely on preset, fixed task processing flows; it can dynamically adjust the task processing flow according to actual progress, thereby improving the flexibility of task orchestration and adaptability to different target information during task processing. Simultaneously, when facing new tasks or complex, changing scenarios, by evaluating and calling the necessary intelligent agents during task execution, it ensures that each subtask can be matched with a suitable intelligent agent for processing, thereby improving the adaptability of the intelligent agent and the accuracy of task processing. Furthermore, the embodiments of this application do not fix the intelligent agent on a specific execution node, but manage available intelligent agents through an intelligent agent pool. This allows for timely response to the processing needs of currently pending subtasks, dynamically selecting a suitable intelligent agent from the pool to meet the need for dynamically adjusting the task processing flow according to actual progress, ensuring the flexibility of task orchestration during task processing.

[0136] The method described in the above embodiments will be further described in detail below.

[0137] In this embodiment, the method of this application embodiment will be described in detail using an information query scenario as an example.

[0138] like Figure 2a As shown, the specific process of a data processing method is as follows:

[0139] 210. Obtain target information.

[0140] For example, embodiments of this application can be applied to processing query tasks (hereinafter referred to as original tasks) of a user input information retrieval platform. For example, the target information can be the task information of a question-and-answer task, and the task information of the original task input by the user can be obtained, such as the question of the question-and-answer task, "Please help arrange a two-day travel itinerary for four adults, preferably outdoor activities, leisure and comfort."

[0141] The data processing method in this application embodiment can be implemented through a task processing system deployed in the background of an information retrieval platform. For example... Figure 2bThe data processing system shown may include an agent control module (AgentController module) and an agent pool. The agent control module is used to dynamically and in real-time orchestrate subtasks during task processing. The agent pool (agent pool) can be represented as an initialized and pooled list of agents (e.g., gent_list). This list can store agent identifiers for agents used to handle subtasks such as "check weather," "check hotels," "check flight tickets," "check train tickets," "check car rentals," "check attractions," and "cost assessment." The corresponding agent can be invoked through these agent identifiers.

[0142] It should be noted that after the user information retrieval platform inputs the task instructions (i.e., task information) for the original task, it can send the task information to the background agent control module for further processing. Before the agent control module processes the task information, it can be initialized, for example, by using the command "Initialize Agent_Controller agent_controller = initialize_agent_controller(agent_list,llm,memory)".

[0143] 220. Based on the target information and the target information processing record, determine the completion status of the target information processing.

[0144] For example, such as Figure 2b The data processing system shown allows the intelligent agent control module to call specified LLM models (i.e., large language models) such as OpenAI, LLAMA, and Hunyuan, and determine whether the original task has been completed (i.e., determining the processing completion status) based on task information and historical execution information (i.e., processing records). Initially, the historical execution information is empty, so the determination is negative. In this embodiment, there are no restrictions on the language used in the large language model (LLM); the prompt language can be any type of language.

[0145] In this application embodiment, the processing record of target information can be in the form of a log service, such as langchain's simpleememory().

[0146] 230. When the processing completion status is incomplete, based on the target information and the processing record of the target information, determine the current subtask to be processed from the target information, and determine the intelligent agent to be invoked corresponding to the current subtask to be processed from the intelligent agent pool.

[0147] For example, when determining whether the processing is complete, the agent control module calls the LLM again to determine, based on the task information and historical execution information, the subtask to be processed by the target information is "query attractions," and selects a suitable agent from the agent pool. For instance, it first calls the agent used to process the "query attractions" subtask. Historical execution information (i.e., processing records) includes the task processing results of previously processed subtasks of the target information.

[0148] 240. Call the agent to be called from the agent pool to process the current subtask to be processed, and obtain the task processing result of the current subtask to be processed.

[0149] 250. Add the processing results of the current subtask to the processing record to obtain the updated processing record with the target information.

[0150] For example, after selecting an agent from the agent pool to handle the "search for attractions" subtask, the agent executes the "search for attractions" subtask and writes the execution information (i.e., task processing results) such as the execution time, agent name, tool used (e.g., AttractionsTool), and execution result into memory (i.e., write to historical execution records).

[0151] 260. Use the updated processing record as the processing record of the target information, and return to the execution steps. Based on the target information and the processing record of the target information, determine the current subtask to be processed from the target information, and determine the intelligent agent to be called and subsequent steps corresponding to the current subtask to be processed from the intelligent agent pool, until the target information processing is completed.

[0152] For example, the process returns to the agent control module for the next round of subtask processing. At this point, the historical execution record contains the agent's execution information for the "search for tourist attractions" subtask. The agent control module then uses the original task's information and historical execution information to determine whether the original task has been completed. This process continues through multiple rounds of original task completion status checks and agent invocations until the original task is determined to be complete.

[0153] For example, such as Figure 2c The data processing flow shown involves multiple rounds of processing, sequentially calling the corresponding intelligent agents to handle the sub-tasks in the original task: Intelligent Agent Control Module → "Query Attractions" → Intelligent Agent Control Module → "Query Weather" → Intelligent Agent Control Module → "Query Flights" → Intelligent Agent Control Module → "Query Hotels". Clearly, in this task processing flow, the execution nodes and order of the original task are not fixed. Instead, before each round of sub-task execution, the intelligent agent control module judges and adjusts based on the execution results of the previous round to determine the sub-task to be processed in the current round and calls the corresponding intelligent agent for processing.

[0154] During the process of invoking an agent to handle a corresponding subtask, the final result can be obtained through the task execution tools within the agent. For example, the scenic spot search tool in agent 1 corresponding to the "search for attractions" subtask can be used to search for and determine specific attractions; the weather search tool in agent 2 corresponding to the "search for weather" subtask can be used to search for and determine specific weather conditions; the flight search tool in agent 3 corresponding to the "search for flights" subtask can be used to search for and determine specific flights; and the hotel search tool in agent 4 corresponding to the "search for hotels" subtask can be used to search for and determine specific hotels. In practical applications, these tools can be queried through a local database or via the internet. These tools can be stored in a tool pool so that agents can invoke the corresponding task execution tools from the tool pool.

[0155] 270. The processing result of the target information is obtained from the updated processing record.

[0156] For example, the execution information from historical execution records can be compiled and summarized into a single result, which is the query result (i.e., processing result) of the original task. The information retrieval platform backend can then return and display this processing result to the user. For example, ... Figure 2c The data processing flow shown allows the intelligent agent control module to call the LLM model to ultimately integrate the task processing results of all subtasks and obtain the query results of the original task: "XX attractions: XX, weather conditions: XX, flight recommendations: XX, hotel recommendations: XX".

[0157] As can be seen from the above, in this embodiment, the agents are no longer fixed in a certain execution order during task orchestration, but are instead processed as a pool of candidate agents. In each round, the agent control module evaluates whether the original task is complete based on the original task, the agent pool, and the historical execution history. If the original task is not complete, the next agent to be invoked is selected based on historical execution information and the available agent pool. If the original task is complete, the process ends. Thus, without pre-arranging the task execution order or pre-setting the execution content, by analyzing the task objective and providing a list of supporting agents, the large model can select agents to execute based on the task objective. Furthermore, all historical processing information is used as global information to determine the node for each transition, achieving agent task orchestration without manual pre-setting or intervention.

[0158] Therefore, by implementing steps such as agent pooling, task judgment, agent selection, and execution recording, a non-fixed large-scale agent task orchestration is achieved, resulting in the following advantages: 1. Compared to existing large-scale task orchestration methods, it no longer relies on developers to determine the execution order and content, reducing dependence on personnel and minimizing human error; 2. Compared to existing large-scale task orchestration methods, the task execution order is no longer fixed and can be flexibly changed based on task execution status. For example, "search for hotels" can be performed before "search for flights," instead of the fixed order of "search for flights" followed by "search for hotels"; 3. Compared to existing large-scale task orchestration methods, the task execution content is no longer fixed and can be flexibly determined based on task execution status. For example, "search for train tickets" can be omitted, and only "search for flights" can be performed; 4. It makes task orchestration more versatile. When the agent pool contains agents with different tasks, different task schemes can be orchestrated. For instance, existing large-scale task orchestration methods can only orchestrate tasks of a specific task type, such as travel advice tasks and medical consultation tasks, requiring two separate task orchestrations. The embodiments of this application allow for the orchestration of many different tasks. It is only necessary to put the agents used for travel advice tasks and medical problem consultation tasks into the agent pool. The agent control module can find suitable agents for task orchestration on its own, without the need to create two separate task orchestrations.

[0159] To better implement the above methods, this application also provides a data processing device, which can be integrated into an electronic device, such as a terminal or server. The terminal can be a mobile phone, tablet computer, smart Bluetooth device, laptop computer, personal computer, etc.; the server can be a single server or a server cluster composed of multiple servers.

[0160] For example, in this embodiment, the method of this application embodiment will be described in detail by taking the data processing device specifically integrated into the server as an example.

[0161] For example, such as Figure 3 As shown, the data processing device may include an information acquisition unit 310, an agent determination unit 320, an agent invocation unit 330, a result addition unit 340, and an information processing unit 350, as follows:

[0162] (I) Information Acquisition Unit 310

[0163] Used to obtain target information.

[0164] (II) Agent Determination Unit 320

[0165] This is used to determine the current subtask to be processed from the target information and the processing record of the target information, and to determine the intelligent agent to be invoked from the intelligent agent pool corresponding to the current subtask to be processed. The processing record includes the task processing results of the historical processed subtasks of the target information.

[0166] In some implementations, the agent determination unit is specifically used to: acquire agent information of agents in the agent pool; determine the association between the current subtask to be processed of the target information and the agent information based on the target information and the processing record of the target information; and determine the agent to be invoked from the agent pool based on the association between the current subtask to be processed and the agent information.

[0167] In some implementations, determining the current subtask to be processed for the target information based on the target information and the processing record of the target information includes: determining the task information of the next current subtask to be processed for the target information based on the target information and the processing record of the target information; determining the intelligent agent to be invoked from the intelligent agent pool based on the association relationship between the current subtask to be processed and the intelligent agent information includes: determining the intelligent agent to be invoked from the intelligent agent pool based on the association relationship between the task information of the next current subtask to be processed and the intelligent agent information, wherein the intelligent agent to be invoked is used to process the next current subtask to be processed.

[0168] In some embodiments, the data processing apparatus further includes a state determination unit, which is used to: determine the processing completion status of the target information based on the target information and the processing record of the target information; when the processing completion status is incomplete, determine the current sub-task to be processed from the target information based on the target information and the processing record of the target information, and determine the intelligent agent to be invoked corresponding to the current sub-task to be processed from the intelligent agent pool.

[0169] In some implementations, determining the processing completion status of the target information based on the target information and the processing record of the target information includes: performing semantic understanding on the target information and the processing record of the target information to extract a first context feature, the first context feature representing the semantic association between the target information and the processing record of the target information; and decoding the first context feature to obtain the processing completion status of the target information.

[0170] In some implementations, semantic understanding is performed on the target information and the processing records of the target information to extract first contextual features, including: extracting keywords corresponding to sub-tasks in the target information from the target information; and performing semantic understanding on the keywords corresponding to the sub-tasks and the processing records of the target information to extract first contextual features.

[0171] In some embodiments, the data processing apparatus further includes an agent training unit, which is used to: acquire multiple agents, which are trained through different subtasks; and add the identification information of the agents to an agent pool.

[0172] (III) Intelligent Agent Calling Unit 330

[0173] Used to call upon an agent from the agent pool to process the current subtask and obtain the task processing result of the current subtask.

[0174] In some implementations, the agent invocation unit is used to: invoke an agent to be invoked from the agent pool; and use the agent to be invoked to process the current subtask based on the target information and the processing record of the target information, so as to obtain the task processing result of the current subtask.

[0175] In some implementations, the intelligent agent includes a task processing model and a corresponding task execution tool. The agent, to be invoked, processes the current subtask based on target information and its processing records to obtain the task processing result. This includes: using the task processing model in the agent to be invoked to perform semantic understanding on the target information and its processing records to extract second contextual features, which represent the semantic relationships between subtasks in the target information; decoding the second contextual features to obtain the task execution elements of the current subtask; and using the task execution tool in the agent to be invoked to execute the current subtask based on the task execution elements to obtain the task processing result.

[0176] (iv) Result Addition Unit 340

[0177] This is used to add the processing results of the current subtask to the processing record, resulting in an updated processing record with target information.

[0178] (V) Information Processing Unit 350

[0179] This is used to obtain the processing results of the target information from the updated processing records.

[0180] In some implementations, the information processing unit is used to: perform semantic understanding on the updated processing record to extract a third context feature, the third context feature representing the semantic association between the task processing results of historically processed subtasks in the updated processing record; and decode the third context feature to obtain the processing result of the target information.

[0181] In some embodiments, the data processing apparatus further includes a return execution unit, which is used to: use the updated processing record as the processing record of the target information, and return the execution steps to determine the intelligent agent to be invoked and subsequent steps from the intelligent agent pool based on the target information and the processing record of the target information, until the target information processing is completed.

[0182] In practice, each of the above units can be implemented as an independent entity or can be arbitrarily combined to be implemented as the same or several entities. For the specific implementation of each of the above units, please refer to the previous method embodiments, which will not be repeated here.

[0183] As can be seen from the above, the data processing device of this embodiment includes an information acquisition unit, an agent determination unit, an agent invocation unit, a result addition unit, and an information processing unit. Specifically, the information acquisition unit acquires target information; the agent determination unit determines the current subtask to be processed from the target information and the processing record of the target information, and determines the agent to be invoked corresponding to the current subtask from the agent pool; the processing record includes the task processing results of historically processed subtasks of the target information; the agent invocation unit invokes the agent to be invoked from the agent pool to process the current subtask to be processed, and obtains the task processing result of the current subtask to be processed; the result addition unit adds the task processing result of the current subtask to the processing record, obtaining an updated processing record of the target information; and the information processing unit obtains the processing result of the target information from the updated processing record.

[0184] Therefore, this application embodiment, based on target information and its real-time processing progress (i.e., processing records), dynamically determines the subsequent processing procedure of the target information according to its current processing progress. This eliminates reliance on preset, fixed task processing flows for task orchestration, allowing for dynamic adjustment of the task processing flow based on actual progress, thereby improving the flexibility of task orchestration and adaptability to different target information. Simultaneously, when facing new tasks or complex, changing scenarios, by evaluating and invoking the necessary intelligent agents during task execution, it ensures that each subtask can be matched with a suitable intelligent agent for processing, improving agent adaptability and task processing accuracy. Furthermore, this application embodiment does not fix intelligent agents to a specific execution node, but manages available intelligent agents through an agent pool. This allows for timely response to the processing needs of currently pending subtasks, dynamically selecting suitable intelligent agents from the pool to meet the need for dynamic adjustment of the task processing flow based on actual progress, ensuring the flexibility of task orchestration during task processing.

[0185] This application also provides an electronic device, which can be a terminal, a server, or other similar device. The terminal can be a mobile phone, tablet computer, smart Bluetooth device, laptop computer, personal computer, etc.; the server can be a single server or a server cluster composed of multiple servers, etc.

[0186] In some embodiments, the data processing apparatus may also be integrated into multiple electronic devices, such as multiple servers, with the data processing method of this application being implemented by the multiple servers.

[0187] In this embodiment, a server will be used as an example for detailed description. For example, ... Figure 4 As shown, it illustrates a schematic diagram of the server structure involved in an embodiment of this application. Specifically:

[0188] The server may include components such as a processor 410 with one or more processing cores, a memory 420 with one or more computer-readable storage media, a power supply 430, an input module 440, and a communication module 450. Those skilled in the art will understand that... Figure 4 The server architecture shown does not constitute a limitation on the server and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements. Wherein:

[0189] Processor 410 is the control center of the server, connecting various parts of the server via various interfaces and lines. It performs various server functions and processes data by running or executing software programs and / or modules stored in memory 420, and by calling data stored in memory 420. In some embodiments, processor 410 may include one or more processing cores; in some embodiments, processor 410 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into processor 410.

[0190] The memory 420 can be used to store software programs and modules. The processor 410 executes various functional applications and data processing by running the software programs and modules stored in the memory 420. The memory 420 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the server, etc. In addition, the memory 420 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 420 may also include a memory controller to provide the processor 410 with access to the memory 420.

[0191] The server also includes a power supply 430 that supplies power to the various components. In some embodiments, the power supply 430 can be logically connected to the processor 410 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 430 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0192] The server may also include an input module 440, which can be used to receive input numeric or character information, and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0193] The server may also include a communication module 450. In some embodiments, the communication module 450 may include a wireless module, through which the server can perform short-range wireless transmission, thereby providing users with wireless broadband internet access. For example, the communication module 450 can be used to help users send and receive emails, browse web pages, and access streaming media.

[0194] Although not shown, the server may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 410 in the server loads the executable files corresponding to the processes of one or more applications into the memory 420 according to the following instructions, and the processor 410 runs the applications stored in the memory 420 to realize various functions, as follows:

[0195] Obtain target information; based on the target information and its processing records, determine the current subtask to be processed from the target information, and determine the corresponding agent to be invoked from the agent pool. The processing records include the historical processing results of subtasks processed by the target information; invoke the agent to be invoked from the agent pool to process the current subtask to be processed, and obtain the processing result of the current subtask to be processed; add the processing result of the current subtask to the processing records to obtain the updated processing records of the target information; obtain the processing result of the target information from the updated processing records.

[0196] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0197] As can be seen from the above, the embodiments of this application dynamically determine the subsequent processing of the target information based on the target information and its real-time processing progress (i.e., processing records). Therefore, task orchestration does not rely on preset, fixed task processing flows; it can dynamically adjust the task processing flow according to actual progress, thereby improving the flexibility of task orchestration and adaptability to different target information during task processing. Simultaneously, when facing new tasks or complex, changing scenarios, by evaluating and calling the necessary intelligent agents during task execution, it ensures that each subtask can be matched with a suitable intelligent agent for processing, thereby improving the adaptability of the intelligent agent and the accuracy of task processing. Furthermore, the embodiments of this application do not fix the intelligent agent on a specific execution node, but manage available intelligent agents through an intelligent agent pool. This allows for timely response to the processing needs of currently pending subtasks, dynamically selecting a suitable intelligent agent from the pool to meet the need for dynamically adjusting the task processing flow according to actual progress, ensuring the flexibility of task orchestration during task processing.

[0198] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0199] Therefore, embodiments of this application provide a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to execute steps in any of the data processing methods provided in embodiments of this application. For example, the instructions can execute the following steps:

[0200] Obtain target information; based on the target information and its processing records, determine the current subtask to be processed from the target information, and determine the corresponding agent to be invoked from the agent pool. The processing records include the historical processing results of subtasks processed by the target information; invoke the agent to be invoked from the agent pool to process the current subtask to be processed, and obtain the processing result of the current subtask to be processed; add the processing result of the current subtask to the processing records to obtain the updated processing records of the target information; obtain the processing result of the target information from the updated processing records.

[0201] The storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0202] According to one aspect of this application, a computer program product or computer program is provided, comprising a computer program or instructions that, when executed by a processor, implement the steps of the methods provided in the various optional implementations of the above embodiments. The computer program / instructions are stored in a computer-readable storage medium. A processor of an electronic device reads the computer program / instructions from the computer-readable storage medium and executes the computer program / instructions, causing the electronic device to perform the methods provided in the various optional implementations of the above embodiments.

[0203] Since the instructions stored in the storage medium can execute the steps of any of the data processing methods provided in the embodiments of this application, the beneficial effects that any of the data processing methods provided in the embodiments of this application can achieve can be realized. For details, please refer to the previous embodiments, which will not be repeated here.

[0204] The foregoing has provided a detailed description of a data processing method, apparatus, electronic device, medium, and program product provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A data processing method, characterized in that, include: Obtain target information; Based on the target information and the processing record of the target information, the current sub-task to be processed is determined from the target information, and the intelligent agent to be invoked corresponding to the current sub-task to be processed is determined from the intelligent agent pool. The processing record includes the task processing results of the historical processed sub-tasks of the target information. Call the agent to be called from the agent pool to process the current subtask to be processed, and obtain the task processing result of the current subtask to be processed; Add the task processing result of the current subtask to be processed to the processing record to obtain the updated processing record of the target information; The processing result of the target information is obtained from the updated processing record.

2. The data processing method as described in claim 1, characterized in that, After adding the task processing result of the current subtask to the processing record to obtain the updated processing record of the target information, and before obtaining the processing result of the target information from the updated processing record, the method further includes: The updated processing record is used as the processing record of the target information. The execution steps are returned. Based on the target information and the processing record of the target information, the current sub-task to be processed is determined from the target information, and the agent to be invoked and subsequent steps corresponding to the current sub-task to be processed are determined from the agent pool, until the target information is processed.

3. The data processing method as described in claim 1, characterized in that, The step of determining the current subtask to be processed from the target information and the processing record of the target information, and determining the agent to be invoked corresponding to the current subtask to be processed from the agent pool, includes: Based on the target information and the processing record of the target information, determine the processing completion status of the target information; When the processing completion status is incomplete, based on the target information and the processing record of the target information, the current subtask to be processed is determined from the target information, and the corresponding intelligent agent to be invoked is determined from the intelligent agent pool.

4. The data processing method as described in claim 3, characterized in that, The step of determining the processing completion status of the target information based on the target information and the processing record of the target information includes: Semantic understanding is performed on the target information and the processing record of the target information to extract a first context feature, wherein the first context feature represents the semantic association between the target information and the processing record of the target information; The first context feature is decoded to obtain the processing completion status of the target information.

5. The data processing method as described in claim 4, characterized in that, The step of performing semantic understanding on the target information and the processing record of the target information to extract first contextual features includes: Extract keywords corresponding to subtasks in the target information from the target information; Semantic understanding is performed on the keywords corresponding to the sub-tasks and the processing records of the target information to extract the first contextual features.

6. The data processing method as described in claim 1, characterized in that, The step of determining the current subtask to be processed from the target information and the processing record of the target information, and determining the agent to be invoked corresponding to the current subtask to be processed from the agent pool, includes: Obtain the agent information of the agents in the agent pool; Based on the target information and the processing record of the target information, the current subtask to be processed for the target information is determined; Determine the association between the currently pending subtask and the agent information; Based on the association between the current subtask to be processed and the agent information, the agent to be invoked is determined from the agent pool.

7. The data processing method as described in claim 6, characterized in that, The step of determining the current subtask to be processed for the target information based on the target information and the processing record of the target information includes: Based on the target information and the processing record of the target information, determine the task information of the next current subtask to be processed for the target information; The step of determining the agent to be invoked from the agent pool based on the association between the currently pending subtask and the agent information includes: Based on the association between the task information of the next current subtask to be processed and the agent information, the agent to be invoked is determined from the agent pool, and the agent to be invoked is used to process the next current subtask to be processed.

8. The data processing method as described in claim 1, characterized in that, The step of calling the agent to be called from the agent pool to process the current subtask and obtaining the task processing result of the current subtask includes: The agent to be invoked is invoked from the pool of agents; Using the agent to be invoked, the current subtask to be processed is processed based on the target information and the processing record of the target information, so as to obtain the task processing result of the current subtask to be processed.

9. The data processing method as described in claim 8, characterized in that, The intelligent agent includes a task processing model and a task execution tool corresponding to the task processing model. The process of using the intelligent agent to be invoked to process the current sub-task based on the target information and the processing record of the target information, and obtaining the task processing result of the current sub-task, includes: Using the task processing model in the agent to be invoked, semantic understanding is performed on the target information and the processing record of the target information to extract a second context feature, which represents the semantic association between subtasks in the target information; Decoding the second context feature yields the task execution elements of the current subtask to be processed; Using the task execution tool in the agent to be invoked, the current subtask to be processed is executed based on the task execution elements to obtain the task processing result of the current subtask to be processed.

10. The data processing method as described in claim 1, characterized in that, The processing result of obtaining the target information from the updated processing record includes: Semantic understanding is performed on the updated processing record to extract a third context feature, which represents the semantic association between the task processing results of the historically processed subtasks in the updated processing record; The third context feature is decoded to obtain the processing result of the target information.

11. The data processing method according to any one of claims 1 to 10, characterized in that, Before determining the agent to be invoked from the agent pool based on the target information and the processing record of the target information, the method further includes: Multiple intelligent agents are acquired, and these agents are trained through different subtasks. Add the identification information of the intelligent agent to the intelligent agent pool.

12. A data processing apparatus, characterized in that, include: The information acquisition unit is used to acquire target information; The agent determination unit is configured to determine the current sub-task to be processed from the target information and the processing record of the target information, and to determine the agent to be invoked corresponding to the current sub-task to be processed from the agent pool. The agent to be invoked is used to process the current sub-task to be processed of the target information. The processing record includes the task processing results of the historically processed sub-tasks of the target information. The agent invocation unit is used to invoke the agent to be invoked from the agent pool to process the current subtask to be processed, and to obtain the task processing result of the current subtask to be processed; The result addition unit is used to add the task processing result of the current subtask to be processed to the processing record, so as to obtain the updated processing record of the target information. An information processing unit is used to obtain the processing result of the target information from the updated processing record.

13. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing multiple instructions; the processor loads instructions from the memory to perform the steps of the data processing method as described in any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to perform the steps of the data processing method according to any one of claims 1 to 11.

15. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by the processor, they implement the steps of the data processing method according to any one of claims 1 to 11.

Images