Task execution methods, devices, and intelligent agents based on large models

Abstract

This invention provides a task execution method, apparatus, and intelligent agent based on a large model, relating to the field of artificial intelligence technology, particularly to the fields of large language models and AI assistants. The task execution method based on a large model includes: utilizing a large model to analyze the requirement information of the task to be processed and the execution trajectory information of historical tasks to determine the target intelligent agent from multiple candidate intelligent agents to execute the task to be processed. This reduces the probability of context breakage and incoherent decision-making caused by historical tasks and the task to be processed being scheduled to different intelligent agents due to their interrelationship. Furthermore, by utilizing a large model to analyze the type information of the task to be processed, the target execution unit is determined from multiple candidate execution units associated with the target intelligent agent. This at least solves the scheduling problem of intelligent agents with different capability boundaries, providing a reasonable execution link for the execution of the task to be processed and reducing the latency and information loss caused by redundant links.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, particularly to the fields of large language models and AI assistants, and specifically to task execution methods, devices, and intelligent agents based on large models. Background Technology

[0002] With the widespread application of large-scale models in fields such as intelligent assistance, enterprise AI (Artificial Intelligence) platforms, intelligent programming, and intelligent office, the diversity of task requirements determined based on user requests is gradually increasing. Faced with diverse task requirements, how to dynamically schedule intelligent agents with different capability boundaries to meet diverse task requirements has become a topic of great interest. Summary of the Invention

[0003] This invention provides a task execution method, apparatus, and intelligent agent based on a large model.

[0004] According to one aspect of the present invention, a task execution method based on a large model is provided, comprising: using the large model to analyze the requirement information of the task to be processed and the execution trajectory information of historical tasks, determining a target intelligent agent for executing the task to be processed from multiple candidate intelligent agents; wherein, there is an association between the task to be processed and the historical tasks; the execution trajectory information records whether there is a dependency relationship between the historical execution unit called by the historical intelligent agent during the execution of the historical task and the historical task; using the large model to analyze the type information of the task to be processed, determining a target execution unit from multiple candidate execution units associated with the target intelligent agent; and calling the target execution unit to execute the task to be processed to obtain processing result information.

[0005] According to another aspect of the present invention, a task execution device based on a large model is provided, comprising: a first determining module, a second determining module, and an execution module.

[0006] The first determination module is used to determine the target agent for executing the task from multiple candidate agents by analyzing the requirement information of the task to be processed and the execution trajectory information of historical tasks using a large model; wherein, there is a relationship between the task to be processed and the historical tasks; the execution trajectory information records whether there is a dependency relationship between the historical execution units called by the historical agents during the execution of historical tasks and the historical tasks.

[0007] The second determination module is used to determine the target execution unit from multiple candidate execution units associated with the target agent by analyzing the type information of the task to be processed using a large model.

[0008] The execution module is used to call the target execution unit to execute the task to be processed and obtain the processing result information.

[0009] According to another aspect of the present invention, an intelligent agent is provided, comprising: an input module, a processing module, and an output module.

[0010] The input module is used to receive the requirement information of the task to be processed and the execution trajectory information of the historical tasks. The execution trajectory information records whether there is a dependency relationship between the historical execution unit called by the historical agent during the execution of the historical task and the historical task.

[0011] The processing module is used to determine the target task based on the requirement information and execution trajectory information received by the input module, determine the target large model based on the target task, and obtain the processing result information by calling the target large model to execute the methods described above.

[0012] The output module is used to output the processing results obtained by the processing module.

[0013] According to another aspect of the present invention, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method described above.

[0014] According to another aspect of the present invention, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause the computer to perform the method described above.

[0015] According to another aspect of the present invention, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the method described above.

[0016] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0017] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of the invention.

[0018] Figure 1 An exemplary system architecture for applying a large-model-based task execution method and apparatus according to embodiments of the present invention is shown.

[0019] Figure 2 A flowchart of a task execution method based on a large model according to an embodiment of the present invention is shown.

[0020] Figure 3The diagram illustrates a recording structure for the execution trajectory information of historical tasks according to an embodiment of the present invention.

[0021] Figure 4 A schematic diagram illustrating the determination of a target intelligent agent based on task requirement information of the task to be processed and execution trajectory information of historical tasks according to an embodiment of the present invention is shown.

[0022] Figure 5 A schematic diagram illustrating the determination of a target agent from candidate agents according to another embodiment of the present invention is shown.

[0023] Figure 6 A schematic diagram illustrating the determination of a target execution unit based on the type information of the task to be processed, according to an embodiment of the present invention, is shown.

[0024] Figure 7 A schematic diagram illustrating the determination of a target execution unit based on the type information of the task to be processed, according to another embodiment of the present invention, is shown.

[0025] Figure 8 A block diagram of a task execution device based on a large model according to an embodiment of the present invention is shown;

[0026] Figure 9 A block diagram of an intelligent agent according to an embodiment of the present invention is shown.

[0027] Figure 10 A block diagram of an electronic device suitable for implementing a large-model-based task execution method according to an embodiment of the present invention is shown. Detailed Implementation

[0028] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0029] As multimodal large models and various toolchains are increasingly deployed in enterprise-level intelligent agents, the execution carriers at the backend of intelligent agents are showing a trend of heterogeneity: different execution frameworks or carrier units (hereinafter referred to as Harness) have significant differences in terms of the types of models that can be accessed, the sets of available tools, runtime constraints, cost and latency characteristics; within the same execution framework, functional models targeting different capability boundaries may also be deployed simultaneously, such as: plain text models, multimodal models, code models, low-cost fast models and high-precision inference models, etc.

[0030] In this context, user requests often carry heterogeneous multimodal information and its contextual history. Heterogeneous multimodal information may include, but is not limited to, text, images, documents, audio and video, structured data, etc.

[0031] In related examples, a specific execution framework or model service is usually selected during the deployment or configuration phase, making it difficult to effectively perceive the task requirements of the task to be executed. For example, when a task to be processed, including images or videos, is mistakenly scheduled to a model that only supports text, an additional OCR (Optical Character Recognition) / parsing / frame extraction toolchain is required to parse the images or videos, resulting in a lengthy task execution chain, increased latency, and information loss.

[0032] If tasks are scheduled in a coarse-grained manner using only rules or classifiers and assigned to a specific execution framework, context breaks may occur when switching between execution frameworks within the same session, leading to inconsistent decision-making. Furthermore, the scheduling strategy lacks auditable execution traces, making it difficult to review and optimize.

[0033] In view of this, embodiments of the present invention record the historical execution trajectory information of historical tasks, noting whether there are dependencies between historical execution units invoked by historical agents during the execution of historical tasks and historical tasks. By utilizing a large model to analyze the demand information of the task to be processed and the execution trajectory information of historical tasks, the target agent for executing the task to be processed is determined. This reduces the probability of context breakage and inconsistent decision-making caused by related historical tasks and the task to be processed being scheduled to different agents. Simultaneously, it provides a traceable information foundation for reviewing and optimizing the scheduling strategy after task execution. Furthermore, by dynamically perceiving the demand and type of the task to be processed, the large model dynamically determines the target execution unit for executing the task, providing a reasonable execution path for the task to be processed, reducing latency and information loss caused by redundant paths, and further improving the execution efficiency of the task to be processed.

[0034] Figure 1 An exemplary system architecture for applying a large-model-based task execution method and apparatus according to embodiments of the present invention is shown.

[0035] It is important to note that Figure 1The examples shown are merely illustrative of system architectures applicable to embodiments of the present invention, intended to help those skilled in the art understand the technical content of the present invention. They do not imply that embodiments of the present invention cannot be used in other devices, systems, environments, or scenarios. For instance, in another embodiment, an exemplary system architecture applicable to the large-model-based task execution method and apparatus may include a terminal device. However, the terminal device can implement the large-model-based task execution method and apparatus provided by the embodiments of the present invention without interacting with a server.

[0036] like Figure 1 As shown, the system architecture 100 according to this embodiment may include: a meta-scheduling execution framework 101, a candidate agent 102, and a candidate execution unit 103.

[0037] Users can interact with the meta-scheduling execution framework 101 through a standardized message interface that includes information and attachments. The meta-scheduling execution framework 101 can completely shield users from the differences between the underlying candidate agents, providing a unified session entry point.

[0038] The meta-scheduling execution framework 101 may include a routing module 1012 and a meta-context engine 1011, which collaborate bidirectionally. The routing module 1012, based on the user's pending task requirements, extracts the execution trajectory information of historical tasks associated with the requirements from the meta-context engine. It then analyzes the requirements and execution trajectory information using a large model to determine the matching agent A1 from candidate agents 102. The routing module 1012 is responsible for scheduling the user's pending task requirements and attachments to agent A1, and after receiving a response, uniformly feeds back the reply information to the user. The meta-context engine 1011 does not participate in sending or receiving responses; it only records the corresponding content and the execution trajectory information of each round of complex interactions in the historical session after completion, providing contextual basis for the routing module 1012's scheduling decisions.

[0039] Candidate agents 102 may include Harnesses with different capabilities, such as intelligent assistant Harnesses, dedicated function Harnesses, etc. A Harness can be understood as an agent built around an execution unit (also called a model) to constrain the behavioral boundaries of the execution unit and extend the capability boundaries of the execution unit. For example... Figure 1 As shown, candidate agents 102 may include agents A1 to A2. n , where n is an integer greater than 1.

[0040] The execution unit pool associated with candidate agent 102 is configured with execution units to implement the capabilities carried by the agent. These execution units can be specific pre-trained large models, such as large vision models, code generation models, multimodal large models, etc. Figure 1As shown, the execution unit pool of agent A1 can be configured with execution units M. 11 ~Execution Unit M 1m m is an integer greater than 1. Agent A n The execution unit pool can be configured with execution units M. n1 ~Execution Unit M nt t is an integer greater than 1. There is no mathematical relationship between t and m and n. The number of execution units in the execution unit pool configured for each candidate agent can be the same or not exactly the same.

[0041] like Figure 1 As shown, the execution unit M used to execute the task is matched based on the requirement information of the task to be processed. 11 It can be done by calling execution unit M 11 The pending task is executed to generate a response message, which is then fed back to the user via the routing module 1012.

[0042] In the technical solution of this invention, the collection, storage, use, processing, transmission, provision, disclosure and application of any type of information, such as user personal information, all comply with the provisions of relevant laws and regulations, necessary confidentiality measures have been taken, and they do not violate public order and good morals.

[0043] In the technical solution of the present invention, the user's authorization or consent is obtained before acquiring or collecting the user's personal information.

[0044] Figure 2 A flowchart of a task execution method based on a large model according to an embodiment of the present invention is shown.

[0045] like Figure 2 As shown, the method 200 includes operations S210 to S230.

[0046] In operation S210, the large model is used to analyze the demand information of the task to be processed and the execution trajectory information of historical tasks to determine the target agent for performing the task to be processed from multiple candidate agents.

[0047] In operation S220, the target execution unit is determined from multiple candidate execution units associated with the target agent by analyzing the type information of the task to be processed using a large model.

[0048] In operation S230, the target execution unit is invoked to execute the task to be processed and obtain the processing result information.

[0049] The requirements for the task to be processed may include, but are not limited to, the task type, context information, and data type of the data to be processed.

[0050] Historical tasks and pending tasks are related to each other. For example, a historical task may be a preceding task generated by the same user in multiple rounds of conversations on the same or related topics, similar to a pending task. For instance, a historical task generated by a user in the first round of conversation may be a task to optimize the code at the architectural level, while a pending task generated in the third round of conversation may be a task to optimize the code optimized in the first round at the algorithm level.

[0051] The execution trajectory information records whether there are dependencies between historical execution units invoked by the historical agent during the execution of historical tasks and the historical tasks themselves. Such dependencies can include, but are not limited to, data-level dependencies, state-level dependencies, or configuration-level dependencies.

[0052] For example, data-level dependencies may include intermediate variables generated by historical execution units invoked by the historical agent during the execution of historical tasks. If these intermediate variables can be directly used to invoke subsequent tasks handled by the historical agent, it indicates that there is a data-level dependency between the historical execution unit and the historical task.

[0053] For example, configuration-level dependencies may include historical execution units invoked by the historical agent during the execution of historical tasks, which execute historical tasks based on specific prompts, toolsets, or model parameters configured for those historical execution units, indicating a configuration-level dependency between historical execution units and historical tasks.

[0054] For example, state-level dependencies may include a historical agent generating a new state during the execution of a historical task, such as: the file system pointer moving to a new location, the database transaction being in an uncommitted state, and replacing the execution unit would lead to data overwriting or write conflicts. In this case, there is a state-level dependency between the historical execution unit and the historical task.

[0055] Candidate intelligent agents can be dynamically combined based on multiple execution units to carry different capabilities. For example, intelligent assistants can carry capabilities related to intelligent office work, such as workflow orchestration, data organization, and simulate human operations to realize cross-application and cross-system processing workflows, such as sending emails and querying information.

[0056] For example, the execution trajectory information of historical tasks records that all previous historical tasks were processed by agent A1, and there is a data dependency between the historical execution units called by agent A1 and the historical tasks. That is, intermediate variables generated during the execution of historical tasks are recorded in agent A1. When the needs of subsequent tasks match the capabilities of agent A1, agent A1 can be called to process the tasks. This allows the intermediate variables described above to be directly accessed when the tasks are executed, without the need for cross-agent data migration, further improving the continuity between multiple tasks generated in the same session.

[0057] In this embodiment of the invention, the execution trajectory information of historical tasks provides reference information for matching intelligent agents to process tasks, so that the large model can determine a scheduling mode with higher contextual relevance for tasks under the perception of previous historical scheduling patterns.

[0058] In some embodiments, if it is determined that there is no dependency between the historical execution unit invoked by the historical agent during the execution of the historical task and the historical task, an agent other than the historical agent can be selected as the target agent from multiple candidate agents. In this case, since the execution trajectory information of the historical task records the scheduling mode of the preceding task, when switching agents, the target agent can obtain the scheduling mode of the preceding task as context from the execution trajectory information of the historical task, enabling the target agent to perceive the execution state of the preceding task and realizing context transfer between agents.

[0059] After identifying the target agent, a large model can be used to analyze the type information of the task to be processed and determine the target execution unit from multiple candidate execution units associated with the target agent.

[0060] The type information of the task to be processed can include the task type and the data types of the input / output during task execution. Task types can include, but are not limited to: question-answering tasks, reasoning tasks, code generation tasks, or multimedia data creation tasks such as text / video. Input / output data types can include, but are not limited to: multimodal data such as text, images, video, and audio (any one or more combinations thereof).

[0061] The multiple candidate execution units associated with the target agent can be multiple candidate models stored in the model pool of the target agent. The multiple candidate models can include pre-trained large models, such as large language models, multimodal large models, large vision models, etc., and can also include pre-trained small models, such as object detection models, speech recognition models, etc.

[0062] The target execution unit determined from a plurality of candidate execution units associated with the target agent may include one execution unit or a combination of multiple execution units.

[0063] For example, the task to be processed is a text question answering task, the input / output data types are both text data, and the determined target execution unit can be a large language model.

[0064] In this embodiment of the invention, when the target execution unit includes multiple execution units, the large model can perceive the execution logic between historical execution units in the execution trajectory information of previous historical tasks, and determine the logical relationship between multiple target execution units based on the requirements of the task to be processed, so as to generate an execution path for executing the task to be processed. The calling order of the multiple target execution units is determined according to this logical relationship, so that the task to be processed is executed according to this calling order to obtain processing result information.

[0065] This invention records the execution trajectory information of historical tasks, identifying dependencies between historical execution units invoked by historical agents during task execution and historical tasks. A large model analyzes the requirements of the task to be processed and the execution trajectory information of historical tasks to determine the target agent for executing the task. This reduces the probability of context breakage and inconsistent decision-making caused by related historical tasks and the task to be processed being scheduled to different agents. Simultaneously, it provides a traceable information foundation for reviewing and optimizing scheduling strategies after task execution. Furthermore, by dynamically perceiving the requirements and type of the task to be processed, the large model dynamically determines the target execution unit for executing the task, providing a reasonable execution path for the task, reducing latency and information loss caused by redundant paths, and further improving the execution efficiency of the task.

[0066] Figure 3 The diagram illustrates a recording structure for the execution trajectory information of historical tasks according to an embodiment of the present invention.

[0067] like Figure 3 As shown, the record structure 300 can be stored sequentially according to the session rounds. In each session round, the record structure 300 records user information input by the user and agent response information.

[0068] For example, the record structure of the first round of the session stores user information Qe1, agent response information Re1, and execution trajectory Tr1. In the execution trajectory Tr1, in addition to the scheduling patterns and dependencies described above, it may also include, but is not limited to, agent identifiers used to process historical tasks generated in the first round of the session, call paths of historical execution units invoked by the agent during the execution of historical tasks, identifiers of the invoked historical execution units, and the processing patterns of the historical units for the data required by the historical tasks.

[0069] In this embodiment of the invention, the scheduling mode of the preceding tasks recorded in the execution trajectory information of historical tasks, as well as whether there is a dependency relationship between the historical execution units called by the historical agent during the execution of historical tasks and the historical tasks, can be used as a structured log to provide a basis for information tracing and problem investigation for the processing result information of tasks generated in each round of sessions.

[0070] like Figure 3 As shown, in the record structure of the second round of sessions, in addition to storing user information Qe2 and agent response information Re2, the execution trajectory Tr1 can be updated based on the actual execution trajectory of the task generated for the second round of sessions to obtain the execution trajectory Tr2. This ensures that the execution trajectory of each round of sessions records the historical execution trajectory of the preceding task, providing contextual information for agent scheduling of subsequent tasks and the scheduling of historical units associated with the agent, which can be referenced by the large model.

[0071] The historical execution trajectory of historical tasks preserves the complete execution trajectory of historical tasks, providing a structured basis for lossless context transfer during cross-agent switching and scheduling decisions for subsequent tasks.

[0072] According to an embodiment of the present invention, by using a large model to analyze the requirement information of the task to be processed and the execution trajectory information of historical tasks, the target intelligent agent for performing the task to be processed is determined from multiple candidate intelligent agents. This may include the following operations: in response to determining that there is a dependency relationship between historical execution units and historical tasks, the large model is used to perform correlation analysis on the requirement information and the functional description information of historical intelligent agents to obtain the correlation analysis results; based on the correlation analysis results, the target intelligent agent for performing the task to be processed is determined from multiple candidate intelligent agents.

[0073] Functional description information can describe the capabilities carried by the historical agent, such as image recognition capabilities and code generation capabilities. These capabilities are determined by at least one or any combination of multiple candidate historical execution units associated with the historical agent.

[0074] Given the established dependencies between historical execution units and historical tasks, a large model is used to perform correlation analysis on requirement information and functional descriptions of historical agents. By perceiving the requirement information of the task to be processed, the large model can determine the capability requirements of the target agent for executing the task. For example, if the task to be processed is a code optimization task, the capability requirements determined by the large model through perceiving the requirement information would at least include: code language parsing capability, algorithm complexity estimation capability, etc.

[0075] Therefore, a correlation analysis can be performed based on this capability requirement and the functional description information of historical intelligent agents to obtain correlation analysis results that indicate whether historical intelligent agents have the ability to perform the task to be processed.

[0076] Finally, if the correlation analysis results indicate that the historical agent has the ability to execute the task to be processed, the historical agent can be identified as the target agent. This method of directly using the historical agent called by the previous historical task can continue to use the intermediate variables or configuration parameters generated during the execution of the previous historical task, reducing redundant data migration operations caused by cross-agent tasks, further improving the coherence of the historical agent in executing multiple related tasks. At the same time, combining the trajectory information of historical tasks for task scheduling reduces the probability of context breakage caused by executing multiple related tasks across agents.

[0077] If the association analysis results indicate that the historical agent possesses the capability to execute the task to be processed, a target agent matching the capability requirements can be determined from the candidate agents. Since the execution trajectory information of historical tasks in this embodiment, in addition to the dependencies and scheduling modes described above, also records new state information generated during the execution of each preceding historical task, this new state information may include, but is not limited to, intermediate variables, configuration parameters, or pointer states. Therefore, even when multiple related tasks are executed across agents, when the target agent is invoked to execute the task to be processed, it can still read the context information related to the task to be processed from the execution trajectory information of the historical tasks. This ensures that the target agent can perceive the context information of preceding historical tasks when executing the task to be processed, further improving the accuracy of the processing results.

[0078] In some embodiments, in order to further improve the accuracy of the correlation analysis results of the large model for demand information and the functional description information of historical agents, the large model can be constrained to output the correlation analysis results in the form of matching degree by means of prompt words.

[0079] According to an embodiment of the present invention, determining a target agent for performing a task from multiple candidate agents based on association analysis results may include the following operations: in response to determining that the matching degree between the functional description information and the requirement information indicated by the association analysis results is greater than a predetermined threshold, determining a historical agent as a target agent; and in response to determining that the matching degree between the functional description information and the requirement information indicated by the association analysis results is less than or equal to a predetermined threshold, using a large model to analyze the requirement information and the functional description information of each of the multiple candidate agents, determining a target agent that matches the requirement information from the multiple candidate agents.

[0080] To further reduce the illusion of large models in the correlation analysis process, the correlation analysis dimensions of the large model and the weights for each correlation analysis dimension can be constrained in the prompt words. This allows the large model to analyze the matching degree between requirement information and functional description information in a targeted manner based on each correlation analysis dimension, and the final matching degree is obtained by weighting the matching degree of each correlation analysis dimension.

[0081] In some embodiments, the correlation analysis dimensions and the weights of each correlation analysis dimension can be pre-configured according to the scenario requirements of the task to be processed, or they can be generated by analyzing the requirement information of the task to be processed using a large model.

[0082] If the correlation analysis results indicate that the matching degree between the functional description information and the requirement information is greater than a predetermined threshold, it means that the historical agent has the ability to perform the task to be processed, and the historical agent can continue to be used to perform the task to be processed.

[0083] If the correlation analysis results indicate that the matching degree between the functional description information and the requirement information is less than or equal to a predetermined threshold, it means that the historical agent does not have the ability to perform the task to be processed, and it is necessary to re-determine the target agent that matches the capability requirement from the candidate agents based on the requirement information.

[0084] For example, the requirement information and the functional descriptions of multiple candidate agents can be input into a large model. Similar to the association analysis method used by the large model to analyze the requirement information and the functional descriptions of historical agents, a matching degree is generated between the requirement information and each of the multiple candidate agents. Then, the candidate agent with the highest matching degree can be determined as the target agent. Alternatively, the matching degrees can be sorted from high to low, and the sorting results can be displayed to the user, allowing the user to determine the target agent.

[0085] By constraining the large model to output the correlation analysis results in terms of the matching degree between functional description information and requirement information, the correlation analysis results are quantified, simplifying the semantic parsing operation of the correlation analysis results in the subsequent selection of target intelligent agents. Threshold judgment can be used to determine whether to use the historical intelligent agent, reducing the computing resources required to schedule multiple related tasks to the target intelligent agent, further shortening the response latency of task execution, and simplifying the execution link in the task scheduling process.

[0086] According to an embodiment of the present invention, by using a large model to analyze the requirement information of the task to be processed and the execution trajectory information of historical tasks, a target intelligent agent for performing the task to be processed is determined from multiple candidate intelligent agents. This may include the following operation: in response to determining that there is no dependency relationship between the historical execution unit and the historical task, the target intelligent agent matching the requirement information is determined from multiple candidate intelligent agents by analyzing the requirement information and the functional description information of each of the multiple candidate intelligent agents.

[0087] For example, the absence of a dependency between historical execution units and historical tasks can characterize that there are no data dependencies, state dependencies, or configuration dependencies between historical execution units and historical tasks. Therefore, a large model can be used to perceive the capability requirements of the target intelligent agent for executing the task to be processed, and to determine the target intelligent agent that matches the capability requirements from the candidate intelligent agents.

[0088] Although the pending task will not schedule new state information generated during the execution of the historical task if there is no dependency between the historical execution unit and the historical task, the context information of the historical task may still be needed during the execution of the pending task to maintain the continuity between the tasks because the historical task and the pending task are related to each other.

[0089] Therefore, embodiments of the present invention record context information related to historical tasks in the execution trajectory information of historical tasks. Even when multiple related tasks are executed across agents, when a target agent is invoked to execute a task to be processed, the target agent can read the context information related to the task to be processed from the execution trajectory information of historical tasks. This ensures that the target agent can perceive the context information of preceding historical tasks when executing the task to be processed, further improving the coherence of multiple related tasks during execution, thereby enabling the execution decision to output processing result information with higher accuracy.

[0090] Figure 4 A schematic diagram illustrating the determination of a target intelligent agent based on task requirement information of the task to be processed and execution trajectory information of historical tasks according to an embodiment of the present invention is shown.

[0091] like Figure 4As shown, the execution trajectory information of historical tasks can include the execution units invoked during the execution of historical tasks generated in the three rounds of sessions, thus forming a call chain of multiple execution units. For example, the historical task generated in the first round of sessions is scheduled to agent A1, which then invokes execution unit M. 11 Execution. The historical tasks generated in the second round of the session are scheduled to agent A1, which then calls execution unit M. 13 Execution. The historical tasks generated in the third round of the session are scheduled to agent A1, which then calls execution unit M. 14 Execution. Therefore, the historical agent can be identified as agent A1.

[0092] First, it can be determined whether there is a dependency between the requirements of the task to be processed and the execution unit called by agent A1.

[0093] In this embodiment of the invention, when the same historical agent invokes different historical execution units in multiple historical tasks, and there is a dependency relationship between any historical unit and a historical task, the existence of a dependency can be determined. This is because when the same historical agent executes each historical task, the new state information generated by the invoked historical execution units can be stored as memory within the historical agent, so that subsequent tasks can directly invoke it.

[0094] Then, it is determined whether the function of agent A1 matches the requirements of the task to be processed. If they match, agent A1 is used directly to process the task. If they do not match, a new target agent is determined from the candidate agents.

[0095] When the same historical agent invokes different historical execution units in multiple historical tasks, and there are no dependencies between all historical units and historical tasks within these multiple historical execution units, it can be determined that there is no dependency. The target agent can then be directly determined from the candidate agents.

[0096] In some embodiments, the execution trajectory information of historical tasks may contain multiple different historical agents. In this case, the agent with dependencies can be identified from among the multiple historical agents based on the dependency state between the historical execution units invoked by the same historical agent and the historical task. The historical agent with dependencies and matching requirements is then identified as the target agent.

[0097] In some embodiments, when multiple historical agents simultaneously invoke historical execution units and historical tasks that are all dependent on each other, the historical agent to be continued can be determined through requirement matching. When the functions and requirements of the multiple historical agents are all matched, the historical agent to be continued can be determined based on the matching degree. When the matching degree between the functions and requirements of the multiple historical agents is the same, the user can determine the historical agent to be continued.

[0098] In addition to determining the target agent from the candidate agents based on the matching degree, the actual running state of each candidate agent can also be considered, thereby reducing the probability of invalid scheduling of the tasks to be processed.

[0099] According to an embodiment of the present invention, the intelligent agents matching the demand information include a first intelligent agent and a second intelligent agent; the first matching degree between the first functional description information of the first intelligent agent and the demand information is greater than the second matching degree between the second functional description information of the second intelligent agent and the demand information. Utilizing a large model to analyze the demand information of the task to be processed and the execution trajectory information of historical tasks, the target intelligent agent for performing the task to be processed is determined from multiple candidate intelligent agents. This may further include the following operations: acquiring first operating state information of the first intelligent agent and second operating state information of the second intelligent agent; and, in response to determining that the first operating state information indicates that the first intelligent agent is operating abnormally, and the second operating state information indicates that the second intelligent agent is operating normally, determining the second intelligent agent as the target intelligent agent.

[0100] Figure 5 A schematic diagram illustrating the determination of a target agent from candidate agents according to another embodiment of the present invention is shown.

[0101] like Figure 5 As shown, the large model matches the requirement information of the task to be processed with the functional description information of each candidate agent. The agents that match the requirements from the candidate agents can include agent A1 and agent A2. n .

[0102] In this embodiment of the invention, the target intelligent agent can be determined by combining the matching degree between the functional description information and the requirement information of the intelligent agent and the operating status of the intelligent agent.

[0103] For example, operational status information can characterize the availability and load status of an agent. For instance, if an agent is periodically optimizing parameters based on historical feedback and is currently unavailable, this can be interpreted as the agent malfunctioning. Conversely, if an agent is available but its load has reached saturation, this can also be interpreted as the agent malfunctioning.

[0104] like Figure 5As shown, agent A1 and agent A are matched with the requirements. n All are in an available state. The black portion of the progress bar represents the agent's current load, and the white portion represents the difference between the current load and the saturation load. Therefore, it can be seen that agent A1 is in an unsaturated load state, which can be interpreted as agent A1 operating normally. Agent A n The operating state is near load saturation, which can be understood as agent A. n An error occurred.

[0105] In some embodiments, the first matching degree between the functional description information and the requirement information of agent A1 is greater than that of agent A. n The second degree of matching between the functional description information and the requirement information. Since the running state of agent A1 is unsaturated, agent A1 can be determined as the target agent.

[0106] In some embodiments, the first matching degree between the functional description information and the requirement information of agent A1 is less than or equal to that of agent A1. n The second degree of matching between the functional description information and the requirement information. At this time, although agent A... n Its capabilities are better suited for performing tasks, however, due to the limitations of agent A... n The operating state is nearing load saturation. If the tasks to be processed are scheduled to agent A... n It may be due to agent A n Overload and abnormal operation, affecting agent A n The performance of the agent is such that agent A1 can still be identified as the target agent.

[0107] By sensing the operational status information of intelligent agents in real time, tasks to be processed are scheduled to target intelligent agents that have processing capabilities and are operating normally, thereby further improving scheduling efficiency.

[0108] In this embodiment of the invention, the running status information of each candidate agent can be stored in a cache so that the large model can obtain the real-time running status of the agent from the cache when performing task scheduling.

[0109] In addition, to reduce the involvement of malfunctioning agents in demand matching, information about malfunctioning agents can be recorded in the execution trajectory information of historical tasks.

[0110] Therefore, the method of this embodiment of the invention may further include the following operation: updating the execution trajectory information based on the abnormal moment information and the predetermined time window information of the first intelligent agent to obtain the updated execution trajectory information.

[0111] For example: Agent A nIf an agent is identified as having a malfunction at time T1, then agent A can be... n The execution trajectory information is updated at the abnormal time t. Since some abnormal operating states of the agent are recoverable—for example, the load of an agent at saturation will decrease after a period of time—a predetermined time window can be pre-configured. For example, it could be T1~Tt, representing the time period from time T1 to time Tt for agent A. n It does not participate in demand matching. For example... Figure 5 As shown, agent A can be updated in the execution trajectory information of the third round. n Load saturation: T1~Tt.

[0112] The updated execution trajectory information is used to determine the agent that matches the subsequent task. The subsequent task is related to the task to be processed, and the reception time of the subsequent task is later than the reception time of the task to be processed. For example, the task to be processed may be generated in the third round of the session, and the subsequent task may be generated in the fourth round or even later rounds of the session.

[0113] The predetermined time window information provides a performance recovery period for the agent that malfunctions. After the predetermined time window has expired, the execution trajectory information can be updated based on the real-time running status of the first agent.

[0114] By updating the execution trajectory information with the abnormal moment information and the predetermined time window information of the first agent, the large model no longer considers the abnormal agent as a candidate when matching requirements for subsequent tasks. This reduces the resource consumption of invalid operations in requirement matching and further improves the matching efficiency of the agent.

[0115] For certain specific tasks, such as those involving sensitive information, a predetermined mapping relationship can be used to enforce that the task be scheduled to be executed by a specific agent. This predetermined mapping relationship indicates the mapping between candidate task types and candidate agents.

[0116] Therefore, the method of this embodiment of the invention may further include the following operation: determining a plurality of candidate agents that match the task type of the task to be processed based on a predetermined mapping relationship.

[0117] The mapping relationship between candidate task types and candidate agents indicated in the predefined mapping relationship can be determined based on task scenario requirements, data security, resource constraints, etc. The predefined mapping relationship between candidate task types and candidate agents can be a one-to-many mapping.

[0118] For example, multiple candidate agents matching the type information of the task to be processed can be determined from a pool of agents based on a predetermined mapping relationship. Then, a large model is used to determine the target agent from these multiple candidate agents by analyzing the requirement information of the task to be processed and the execution trajectory information of historical tasks.

[0119] By constraining the selection range of large models based on predetermined mapping relationships, the illusion of large models is reduced, and the stability of the output results of large models is further improved, thereby improving the efficiency of task scheduling.

[0120] In practical applications, when the requirements of a task to be processed include multimodal data such as files, images, or videos, if it is scheduled to a plain text model, the model must indirectly obtain the data required by the task by calling an external read toolchain, such as a file reading tool, an OCR (Optical Character Recognition) tool, or a video frame extraction tool. This introduces additional tool call rounds, increases end-to-end latency, and the information conversion in the intermediate steps of the toolchain inevitably causes content loss.

[0121] In view of this, embodiments of the present invention utilize a large model to determine a target execution unit from multiple candidate execution units associated with the target agent by analyzing the type information of the task to be processed. This may include the following operations: analyzing the requirement information of the task to be processed using the large model to obtain task type information and data type information to be processed by the task; and determining the target execution unit from multiple candidate execution units by analyzing the task type information and data type information using the large model.

[0122] Candidate execution units can be functional models used to implement the capabilities of the target intelligent agent. In this embodiment of the invention, multiple models with different functional orientations can be deployed simultaneously within the same intelligent agent. For example: a lightweight plain text model suitable for simple question-and-answer and fast-response tasks without attachments; a plain text reasoning model suitable for deep reasoning and long-link planning without attachments; a multimodal model that natively supports image / text / video understanding; and a code-specific model suitable for code generation, etc.

[0123] Task type information may include, but is not limited to, question-and-answer tasks, reasoning tasks, text / image / video / multimodal creation tasks, code optimization / generation tasks, etc.

[0124] The data to be processed in the pending task can be files uploaded by the user as attachments to the task. The data types to be processed in the pending task can include, but are not limited to, text, audio, images, and video, or any combination of one or more of these.

[0125] For example, plain text question-answering tasks can be scheduled to a lightweight plain text model within an intelligent body that carries text semantic understanding and reasoning capabilities. Multimodal creation tasks can be scheduled to a functional model within an intelligent body that carries multimodal information understanding and reasoning capabilities. This model, with its native multimodal understanding capabilities, can directly receive and process attachment content without any intermediate tools, resulting in significantly better response efficiency and understanding quality.

[0126] This invention employs a two-layer routing mechanism. The first layer, described above, is a routing mechanism based on the execution trajectory of historical tasks and the requirements of tasks to be processed. The second layer is a routing mechanism based on the task type of the task to be processed and the data type of the data to be processed, using candidate execution units associated with the agent. This transforms the fixed configuration mode in related examples into a two-layer routing mechanism of cross-agent dynamic scheduling and dynamic configuration of execution units within the agent. This enables dynamic resource allocation for tasks, further improving the flexibility of scheduling multiple tasks generated by interconnected multi-round sessions. It also reduces the reliance on external tools for relaying data when using a fixed model to process different modalities, resulting in lengthy execution chains, increased response times, and information loss.

[0127] The following is combined with Figure 6 and Figure 7 The process of triaging candidate execution units associated with an agent based on the task type of the task to be processed and the data type of the data to be processed is described in detail.

[0128] According to an embodiment of the present invention, determining a target execution unit from multiple candidate execution units based on task type information and data type information may include the following operations: analyzing the task type information using a large model to generate the complexity of the task to be processed; and determining the target execution unit from multiple candidate execution units by analyzing the complexity and data type information using the large model.

[0129] Figure 6 A schematic diagram illustrating the determination of a target execution unit based on the type information of the task to be processed, according to an embodiment of the present invention, is shown.

[0130] like Figure 6 As shown, the execution unit configured inside the intelligent agent A1 may include execution unit M. 11 ~Execution Unit M 1n .

[0131] First, the large model can be used to extract information from the execution unit M based on the type of the task to be processed. 11 ~Execution Unit M 1n The execution unit M that determines the type matching is determined in the middle. 11 and execution unit M 12 For example: the type of task to be processed is a plain text question-and-answer task, and the execution unit M...11 This is a lightweight, text-based model with low inference depth, suitable for simple tasks involving quick question and answer. Execution unit M 12 This is a pure text reasoning model with high reasoning depth, suitable for complex tasks that require reasoning link planning.

[0132] Then, a large model can be used to comprehensively reason about the inference depth required for the task to be processed and the length of the context by perceiving the task type, thereby generating the complexity of the task. For example, "What's the weather like today?" is a simple information query task that does not require deep reasoning and can be handled by a lightweight plain text model. On the other hand, "Please help me parse this code and optimize the code's scheduling interface" is a complex and professional code optimization task that requires not only code parsing but also understanding the code's syntax tree and scheduling framework. Therefore, a code-specific model is needed to handle it.

[0133] like Figure 6 As shown, the execution unit M with type matching is selected based on task complexity. 11 and execution unit M 12 The target execution unit identified in the text is execution unit M. 11 .

[0134] By matching the target execution unit with the task complexity, task scheduling is achieved under the dual constraints of resource cost and response latency. This ensures that execution units with high resource utilization and long response latency are only invoked when necessary, further improving resource utilization.

[0135] In some embodiments, the execution trajectory information of historical tasks may also record user feedback information on the historical execution results output by each historical execution unit, as well as performance feedback information of each historical execution unit.

[0136] According to embodiments of the present invention, the method of determining the target execution unit from multiple candidate execution units associated with the target agent by analyzing the requirement information of the task to be processed using a large model may further include the following operations: obtaining feedback information on the historical execution results of each of the multiple candidate execution units; and determining the target execution unit from multiple candidate execution units by analyzing task type information, data type information and feedback information using a large model.

[0137] Feedback information may include explicit feedback information from the user regarding the historical execution results output by each historical execution unit, as well as performance feedback information for each historical execution unit.

[0138] For example, explicit feedback can indicate a user's satisfaction with historical execution results. This satisfaction level can be obtained by analyzing user feedback on historical execution results using a large model. Feedback information can include, but is not limited to, positive / negative evaluations of the input, the number of modification operations performed on historical execution results, etc.

[0139] For example, performance feedback information may include, but is not limited to, the response latency of historical execution units, the number of times tools were called during the execution of historical tasks, the toolchain length, whether errors occurred during the execution of historical tasks, and the number of errors.

[0140] Figure 7 A schematic diagram illustrating the determination of a target execution unit based on the type information of the task to be processed, according to another embodiment of the present invention, is shown.

[0141] like Figure 7 As shown, firstly, the large model can be used to extract information from the execution unit M based on the type of the task to be processed. 11 ~Execution Unit M 1n The execution unit M that determines the type matching is determined in the middle. 11 and execution unit M 12 .

[0142] Then, the large model can be used through the perceptual execution unit M 11 and execution unit M 12 Feedback information on the historical execution results of each entity, for example: execution unit M 12 The system reported three errors during the execution of historical tasks, and users provided negative feedback on the execution results. Execution Unit M 11 There were 0 errors during the execution of historical tasks, and the response latency was within 0.5 seconds. Therefore, the target execution unit can be identified as execution unit M. 11 .

[0143] By combining the feedback information of the historical execution results of each candidate execution unit, the target execution unit for executing the pending task is dynamically determined, and the feedback information of the historical execution results is used as reference information for dynamic scheduling, which further improves the efficiency of scheduling subsequent tasks.

[0144] It should be noted that the two-layer flow splitting mechanism provided in this embodiment of the invention is not a simple series connection, but rather a collaborative feedback loop formed through a meta-scheduling execution framework.

[0145] For example, the routing decision results at the first-layer agent level directly constrain and guide the candidate set and optimization objective of routing at the execution unit level. After the routing execution at the second-layer model level, the feedback information of the model execution results can be fed back to the meta-scheduling execution framework, affecting the scheduling preferences of the agent level in subsequent rounds. This forms a complete scheduling optimization closed loop.

[0146] For example, if a certain functional model of an agent continues to produce a specific type of error, it can trigger an agent-level routing switch to a backup agent.

[0147] In addition, the decision-making basis and execution results of the two-layer routing are uniformly written into the execution trajectory information, forming an indivisible two-layer execution record, ensuring traceability and data integrity for subsequent routing optimization.

[0148] Figure 8 A block diagram of a task execution device based on a large model according to an embodiment of the present invention is shown.

[0149] like Figure 8 As shown, the task execution device 800 based on a large model may include a first determining module 810, a second determining module 820, and an execution module 830.

[0150] The first determining module 810 is used to determine the target intelligent agent for executing the task to be processed from multiple candidate intelligent agents by analyzing the requirement information of the task to be processed and the execution trajectory information of the historical task through a large model; wherein, there is a relationship between the task to be processed and the historical task; the execution trajectory information records whether there is a dependency relationship between the historical execution unit called by the historical intelligent agent during the execution of the historical task and the historical task.

[0151] The second determination module 820 is used to determine the target execution unit from multiple candidate execution units associated with the target agent by analyzing the type information of the task to be processed using a large model.

[0152] The execution module 830 is used to call the target execution unit to execute the task to be processed and obtain the processing result information.

[0153] According to an embodiment of the present invention, the first determining module 810 may include a first analysis submodule and a first determining submodule.

[0154] The first analysis submodule is used to respond to the determination that there is a dependency relationship between historical execution units and historical tasks. It uses a large model to perform correlation analysis on the requirement information and the functional description information of historical intelligent agents to obtain the correlation analysis results.

[0155] The first determination submodule is used to determine the target agent for performing the task from multiple candidate agents based on the association analysis results.

[0156] According to an embodiment of the present invention, the first determining submodule may include a first determining unit and a second determining unit.

[0157] The first determining unit is used to determine the historical agent as the target agent in response to the determination that the matching degree between the function description information and the demand information indicated by the correlation analysis results is greater than a predetermined threshold.

[0158] The second determining unit is used to determine the target agent that matches the requirement information by analyzing the functional description information and requirement information of multiple candidate agents using a large model, in response to the determination that the matching degree between the functional description information and requirement information of multiple candidate agents is less than or equal to a predetermined threshold.

[0159] According to an embodiment of the present invention, the first determining module 810 may further include a second determining submodule, which is used to determine a target intelligent agent that matches the requirement information from multiple candidate intelligent agents by analyzing the functional description information and requirement information of each of the multiple candidate intelligent agents in response to determining that there is no dependency relationship between the historical execution unit and the historical task, using a large model.

[0160] According to an embodiment of the present invention, the intelligent agent matching the demand information includes a first intelligent agent and a second intelligent agent, and the second determining module 820 may include a first acquiring submodule and a second determining submodule.

[0161] The first acquisition submodule is used to acquire the first operating state information of the first intelligent agent and the second operating state information of the second intelligent agent.

[0162] The second determination submodule is used to determine the second intelligent agent as the target intelligent agent in response to the determination that the first intelligent agent is operating abnormally and the second intelligent agent is operating normally.

[0163] According to an embodiment of the present invention, the first determining module 810 may further include an updating submodule, which is used to update the execution trajectory information based on the abnormal moment information of the first intelligent agent and the predetermined time window information to obtain the updated execution trajectory information; wherein, the updated execution trajectory information is used to determine the intelligent agent that matches the subsequent task; the subsequent task is associated with the task to be processed, and the receiving time of the subsequent task is later than the receiving time of the task to be processed.

[0164] According to an embodiment of the present invention, the task execution device 800 based on a large model may further include a third determining module, used to determine a plurality of candidate agents that match the type information of the task to be processed based on a predetermined mapping relationship; wherein the predetermined mapping relationship indicates the mapping relationship between the candidate task type and the candidate agent.

[0165] According to an embodiment of the present invention, the second determining module 820 may include a second analysis submodule and a unit determining submodule.

[0166] The second analysis submodule is used to analyze the requirement information of the task to be processed using the large model, and obtain the task type information and the data type information that the task to be processed needs to process.

[0167] The unit determination submodule is used to determine the target execution unit from multiple candidate execution units by analyzing task type information and data type information using a large model.

[0168] According to an embodiment of the present invention, the unit determination submodule may include a complexity analysis unit and a first determination unit.

[0169] The complexity analysis unit is used to analyze task type information using a large model and generate the complexity of the task to be processed.

[0170] The first determining unit is used to determine the target execution unit from multiple candidate execution units by analyzing complexity and data type information using a large model.

[0171] According to an embodiment of the present invention, the unit determination submodule may further include: an acquisition unit and a second determination unit.

[0172] The acquisition unit is used to obtain feedback information on the historical execution results of multiple candidate execution units.

[0173] The second determination unit is used to determine the target execution unit from multiple candidate execution units by analyzing task type information, data type information, and feedback information using a large model.

[0174] According to embodiments of the present invention, the present invention also provides an intelligent agent, an electronic device, a readable storage medium, and a computer program product.

[0175] According to an embodiment of the present invention, an intelligent agent includes: an input module, a processing module, and an output module.

[0176] The input module is used to receive the requirement information of the task to be processed and the execution trajectory information of the historical tasks. The execution trajectory information records whether there is a dependency relationship between the historical execution unit called by the historical agent during the execution of the historical task and the historical task.

[0177] The processing module is used to determine the target task based on the requirement information and execution trajectory information received by the input module, determine the target large model based on the target task, and obtain the processing result information by calling the target large model to execute the methods described above.

[0178] The output module is used to output the processing results obtained by the processing module.

[0179] According to an embodiment of the present invention, an electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method described above.

[0180] According to an embodiment of the present invention, a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to perform the methods described above.

[0181] According to an embodiment of the present invention, a computer program product includes a computer program that, when executed by a processor, implements the method described above.

[0182] Figure 9 A block diagram of an intelligent agent according to an embodiment of the present invention is shown.

[0183] like Figure 9 As shown, in embodiments of the present invention, inspired by the von Neumann architecture in modern computer theory, such as... Figure 9 As shown, the AI ​​agent 900 may include three core modules: an input module 910, an output module 920, and a processing module 930. The processing module 930 may include a control unit 931, a storage unit 932, and a computing unit 933.

[0184] The input module 910 is responsible for receiving or sensing information such as queries, requests, instructions, signals, or data from the outside world (e.g., users or the external environment), and converting it into a format that the AI ​​agent 900 can understand and process. The input module 910 is the primary link for the AI ​​agent 900 to interact with the outside world. It enables the AI ​​agent 900 to efficiently and accurately obtain the necessary "sensory" information from the outside world and respond to this information.

[0185] In the example, the input information received by the input module 910 can be the requirement information of the task to be processed and the execution trajectory information of the historical task, as described above.

[0186] In the example, the processing module 930 is the core support for the AI ​​agent 900's ability to handle complex tasks. The processing module 930 can determine the target task based on the input information received by the input module 910, determine the large model based on the target task, execute the task execution method based on the large model described above by calling the large model, and output the processing result information.

[0187] In the example, the control unit 931 in the processing module 930 will continuously interact with the storage unit 932, the arithmetic unit 933, and / or the output module 920 during operation. However, it should be noted that in the embodiments of the present invention, the control unit 931 initiates communication with the storage unit 932, the arithmetic unit 933, and / or the output module 920 as a single initiator, and there is no communication coupling between the storage unit 932, the arithmetic unit 933, and the output module 920.

[0188] In the example, the performance of the control unit 931 is closely related to the large model on which the AI ​​agent 900 is based. To fully leverage the capabilities of the large language model, the internal structure of the control unit 931 can be designed to be highly configurable and scalable to handle various types of tasks and requirements in real-world scenarios.

[0189] Storage unit 932 can be responsible for remembering information such as historical dialogues and event streams. Configuration information, target text, and data resources generated in each round can be included in storage unit 932.

[0190] In the example, after receiving a configuration generation request, the AI ​​agent 900 can determine the configuration intent from the initial text using an intent recognition model. The configuration intent can be stored in storage unit 932. The AI ​​agent 900 can retrieve relevant data resources from storage unit 932 and feed them back to control unit 931. Then, control unit 931 can use the returned data resources to obtain configuration data corresponding to the initial text. It can also retrieve relevant text data from storage unit 932 and feed it back to control unit 931. Then, control unit 931 can use the returned text data to obtain the target text and pass the target text and configuration data to output module 920.

[0191] The arithmetic unit 933 can be viewed as a predefined tool library. Renderers and display controls, as mentioned earlier, can be included in the arithmetic unit 933.

[0192] In the example, when the AI ​​agent 900 needs to render multiple output data, it can call the relevant renderer and display controls from the computing unit 933 and feed them back to the control unit 931. Then, the control unit 931 can use the fed-back renderer and display controls to render the first search result and pass it to the output module 920. It's understandable that although large language models have excellent language understanding and generation capabilities, like humans, the tasks they can solve without any tools are very limited. When the AI ​​agent 900 is given the ability to call tools, it can perform tasks such as using a calculator to complete mathematical calculations, using Python to perform data analysis, and using a search engine to complete prediction tasks.

[0193] In the example, output module 920 can output the processing result information described above.

[0194] The AI ​​agent 900 according to embodiments of the present invention can simply and effectively improve the level of intelligence, as well as enhance flexibility and versatility.

[0195] Figure 10 A block diagram of an electronic device suitable for implementing a large-model-based task execution method according to an embodiment of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0196] like Figure 10 As shown, the electronic device 1000 includes a computing unit 1001, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 1002 or a computer program loaded from a storage unit 1008 into a random access memory (RAM) 1003. The RAM 1003 may also store various programs and data required for the operation of the electronic device 1000. The computing unit 1001, ROM 1002, and RAM 1003 are interconnected via a bus 1004. An input / output (I / O) interface 1005 is also connected to the bus 1004.

[0197] Multiple components in electronic device 1000 are connected to I / O interface 1005, including: input unit 1006, such as keyboard, mouse, etc.; output unit 1007, such as various types of displays, speakers, etc.; storage unit 1008, such as disk, optical disk, etc.; and communication unit 1009, such as network card, modem, wireless transceiver, etc. Communication unit 1009 allows electronic device 1000 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0198] The computing unit 1001 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1001 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1001 performs the various methods and processes described above, such as the large model-based task execution method. For example, in some embodiments, the large model-based task execution method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 1008. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 1000 via ROM 1002 and / or communication unit 1009. When the computer program is loaded into RAM 1003 and executed by the computing unit 1001, one or more steps of the large model-based task execution method described above can be performed. Alternatively, in other embodiments, computing unit 1001 may be configured to perform a large model-based task execution method by any other suitable means (e.g., by means of firmware).

[0199] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0200] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0201] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0202] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0203] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0204] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, distributed system servers, or servers incorporating blockchain technology.

[0205] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0206] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A task execution method based on a large model, characterized in that, The method includes: By using a large model to analyze the requirement information of the task to be processed and the execution trajectory information of historical tasks, a target intelligent agent for executing the task to be processed is determined from multiple candidate intelligent agents; wherein, there is a correlation between the task to be processed and the historical tasks; the execution trajectory information records whether there is a dependency relationship between the historical execution units called by the historical intelligent agent during the execution of the historical tasks and the historical tasks. By utilizing the large model and analyzing the type information of the task to be processed, a target execution unit is determined from multiple candidate execution units associated with the target agent; and The target execution unit is invoked to execute the task to be processed and obtain the processing result information.

2. The method according to claim 1, characterized in that, The method of using a large model to analyze the requirements of the task to be processed and the execution trajectory information of historical tasks to determine the target agent for performing the task from multiple candidate agents includes: In response to determining that a dependency exists between the historical execution unit and the historical task, the large model is used to perform a correlation analysis on the requirement information and the functional description information of the historical agent to obtain the correlation analysis result; and Based on the correlation analysis results, a target agent for performing the task to be processed is determined from the plurality of candidate agents.

3. The method according to claim 2, characterized in that, The step of determining the target agent for performing the task from the plurality of candidate agents based on the association analysis results includes: In response to determining that the correlation analysis result indicates a matching degree between the functional description information and the requirement information greater than a predetermined threshold, the historical agent is identified as the target agent; and In response to determining that the correlation analysis result indicates that the matching degree between the functional description information and the requirement information is less than or equal to the predetermined threshold, the target agent that matches the requirement information is determined from the multiple candidate agents by analyzing the requirement information and the functional description information of each of the multiple candidate agents using the large model.

4. The method according to claim 1, characterized in that, The method of using a large model to analyze the requirements of the task to be processed and the execution trajectory information of historical tasks to determine the target agent for performing the task from multiple candidate agents includes: In response to determining that there is no dependency between the historical execution unit and the historical task, the target agent that matches the requirement information is determined from the multiple candidate agents by analyzing the requirement information and the functional description information of each of the multiple candidate agents using the large model.

5. The method according to claim 1, characterized in that, The intelligent agents that match the demand information include a first intelligent agent and a second intelligent agent; The method of using a large model to analyze the demand information of the task to be processed and the execution trajectory information of historical tasks to determine the target agent for performing the task to be processed from multiple candidate agents also includes: Obtain the first operating state information of the first intelligent agent and the second operating state information of the second intelligent agent; and In response to determining that the first operating status information indicates that the first intelligent agent is operating abnormally, and the second operating status information indicates that the second intelligent agent is operating normally, the second intelligent agent is determined to be the target intelligent agent.

6. The method according to claim 5, characterized in that, The method further includes: The execution trajectory information is updated based on the abnormal moment information and the predetermined time window information of the first intelligent agent to obtain the updated execution trajectory information; The updated execution trajectory information is used to determine the agent that matches the subsequent task; the subsequent task is associated with the task to be processed, and the receiving time of the subsequent task is later than the receiving time of the task to be processed.

7. The method according to any one of claims 1-6, characterized in that, The method further includes: Based on a predetermined mapping relationship, the plurality of candidate agents that match the type information of the task to be processed are determined; The predetermined mapping relationship indicates the mapping relationship between candidate task types and candidate agents.

8. The method according to any one of claims 1-6, characterized in that, The step of using the large model to analyze the type information of the task to be processed and determining the target execution unit from multiple candidate execution units associated with the target agent includes: The large model is used to analyze the requirements information of the task to be processed, thereby obtaining task type information and the data type information that the task needs to process; and The target execution unit is determined from the plurality of candidate execution units by analyzing the task type information and the data type information using the large model.

9. The method according to claim 8, characterized in that, The step of determining the target execution unit from the plurality of candidate execution units based on the task type information and the data type information includes: The large model is used to analyze the task type information to generate the complexity of the task to be processed; and By using the large model to analyze the complexity and data type information, the target execution unit is determined from the plurality of candidate execution units.

10. The method according to claim 8, wherein, The step of using the large model to analyze the type information of the task to be processed and determining the target execution unit from multiple candidate execution units associated with the target agent further includes: Obtain feedback information on the historical execution results of each of the multiple candidate execution units; and The target execution unit is determined from the plurality of candidate execution units by analyzing the task type information, the data type information, and the feedback information using the large model.

11. A task execution device based on a large model, characterized in that, include: The first determination module is used to determine the target agent for executing the task to be processed from multiple candidate agents by analyzing the requirement information of the task to be processed and the execution trajectory information of historical tasks using a large model; wherein, there is a correlation between the task to be processed and the historical tasks; the execution trajectory information records whether there is a dependency relationship between the historical execution units called by the historical agent during the execution of the historical task and the historical task. The second determining module is used to determine the target execution unit from multiple candidate execution units associated with the target agent by analyzing the type information of the task to be processed using the large model; and The execution module is used to call the target execution unit to execute the task to be processed and obtain the processing result information.

12. An intelligent agent, characterized in that, include: The input module is used to receive the requirement information of the task to be processed and the execution trajectory information of the historical task; wherein, the execution trajectory information records whether there is a dependency relationship between the historical execution unit called by the historical agent during the execution of the historical task and the historical task; The processing module is configured to determine a target task based on the requirement information and execution trajectory information received by the input module, determine a target large model based on the target task, and execute the method described in any one of claims 1-10 by calling the target large model to obtain processing result information; and The output module is used to output the processing result information obtained by the processing module.

13. An electronic device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method of any one of claims 1-10.

14. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-10.

15. A computer program product, comprising a computer program, characterized in that, The computer program, when executed by a processor, implements the method according to any one of claims 1-10.

Images