Search

CN-121979634-A - Task execution method of large-model intelligent agent system and related equipment

CN121979634ACN 121979634 ACN121979634 ACN 121979634ACN-121979634-A

Abstract

The invention discloses a task execution method and related equipment of a large-model intelligent body system, which are used for responding to a received task to be processed to acquire task description information of the task to be processed, dynamically generating a corresponding target workflow execution link for the task description information based on a pre-constructed workflow knowledge graph, wherein the workflow knowledge graph is used for representing an association relation among a plurality of historical workflow units, and controlling the large-model intelligent body system to execute the task to be processed according to the target workflow execution link. The invention aims to systematically organize and understand the internal association among workflows, intelligently generate an optimal execution path for a dynamically-changed complex task, avoid the problems that deep inference association is difficult to support, and an optimal execution link which accords with logic and is attached to a task target cannot be automatically and reliably generated due to workflow isolated storage calling and a fixed matching strategy, reduce the dependence on a large amount of manual intervention, and improve the efficiency and quality of a generated link.

Inventors

  • LI PENG
  • CHEN FAHAO
  • LI AO
  • YANG SHANGPENG

Assignees

  • 西安交通大学

Dates

Publication Date
20260505
Application Date
20260114

Claims (10)

  1. 1. A method of task execution for a large model agent system, the method comprising: responding to a received task to be processed, and acquiring task description information of the task to be processed; Dynamically generating a corresponding target workflow execution link for the task description information based on a pre-constructed workflow knowledge graph, wherein the workflow knowledge graph is used for representing the association relation among a plurality of historical workflow units; And controlling the large-model intelligent agent system to execute the task to be processed according to the target workflow execution link.
  2. 2. The method for executing tasks of a large model intelligent agent system according to claim 1, wherein nodes of the workflow knowledge graph are historical workflow units, and characteristic data of the nodes comprise at least one of function types, applicable scene labels and historical execution performance indexes of the historical workflow units.
  3. 3. The method for task execution of a large model intelligent agent system according to claim 2, wherein the edges of the workflow knowledge graph are used for representing the relationship between two connected historical workflow units, and the relationship comprises at least one of a time sequence connection relationship, a function complementation relationship, a scene co-occurrence relationship and a semantic similarity relationship.
  4. 4. The method for executing tasks of a large model intelligent agent system according to claim 1, wherein the dynamically generating a corresponding target workflow execution link for the task description information based on a pre-constructed workflow knowledge graph comprises: Encoding the task description information into task feature vectors; Based on the matching degree of the task feature vector and the node feature in the workflow knowledge graph, and combining the link rationality indicated by the association relation of the edges in the graph, carrying out link searching in the workflow knowledge graph; and determining the optimal node sequence obtained by searching as the target workflow execution link.
  5. 5. The method for executing tasks in a large model intelligent agent system according to claim 4, wherein the link searching in the workflow knowledge graph based on the matching degree of the task feature vector and the node feature in the workflow knowledge graph and combined with the connection rationality indicated by the association relationship of edges in the graph comprises: calculating the matching score between the task feature vector and each node feature by adopting a neural network model, and calculating the connection score of the edge-based relationship between adjacent nodes in the candidate links; and determining the optimal node sequence through a path searching algorithm according to the matching score and the connection score.
  6. 6. The method for performing tasks in a large model intelligent agent system according to claim 5, wherein controlling the large model intelligent agent system to perform the tasks to be processed according to the target workflow execution link comprises: Sequentially scheduling and executing corresponding historical workflow units according to the sequence indicated by the target workflow execution link; During execution, the output of the current workflow unit is used as the input or context information of the subsequent workflow unit.
  7. 7. The method for performing tasks in a large model intelligent agent system according to claim 1, wherein after controlling the large model intelligent agent system to perform the tasks to be processed according to the target workflow execution link, the method further comprises: evaluating the target workflow execution link according to the execution result of the task to be processed; if the evaluation result meets the preset condition, the node sequence and the relation thereof corresponding to the target workflow execution link are updated into the workflow knowledge graph as new knowledge.
  8. 8. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that the processor implements a task execution method of a large model agent system according to any of claims 1 to 7 when executing the computer program.
  9. 9. A computer readable storage medium storing a computer program, wherein the computer program when executed by a processor implements a task execution method of a large model agent system according to any one of claims 1 to 7.
  10. 10. A computer program product, which when executed by a processor implements a method of task execution for a large model agent system according to any one of claims 1 to 7.

Description

Task execution method of large-model intelligent agent system and related equipment Technical Field The invention belongs to the technical field of large-model intelligent agents, and particularly relates to a task execution method and related equipment of a large-model intelligent agent system. Background With the development of large language model technology, intelligent agent systems constructed based on large language models have become an important tool for processing complex tasks. Such agents are capable of understanding user instructions, invoking various tools or modules, and completing a particular goal by planning and executing a series of steps. In this process, to improve the efficiency and consistency of task processing, the industry introduced the concept of workflow. A workflow can be regarded as a standardized flow of data collection, data analysis to report generation in sequence for the abstraction and solidification of a class of steps, logic and dependencies required to successfully complete a task, for example, the writing of a market report. By multiplexing such prefabricated workflows, the intelligent agent can avoid repeated planning of common tasks, thereby improving response speed. In practical application scenarios, especially when facing open, complex or brand new tasks, the existing workflow-based task execution method faces limitations, and restricts adaptability and efficiency of the intelligent agent. In particular, the prior art suffers from the major problem that, first, existing systems typically view different workflows as independent, static task templates for storage and invocation. Efficient associative modeling is lacking between these workflows. For example, the system knows that data collection and report generation are two independent workflows, but cannot understand the strong timing associations, functional complementarity, or overlapping of applicable scenarios that may exist between them. Such isolation results in the experience library of the agent being discrete in nature and unstructured, and failing to form a systematic domain knowledge network, making it difficult to support deeper reasoning and association. Secondly, when encountering a complex new task that cannot be directly matched by a single prefabricated workflow, most methods rely on a fixed matching strategy based on keywords or semantic similarity to find a most similar isolated workflow for the task. This approach cannot cope with scenes that require flexible combination and concatenation of multiple workflow links. For example, a task of evaluating new product market potential and composing investment advice may require multiple workflows in series, such as market data collection, bid analysis, financial predictive modeling, and risk report generation. Due to the lack of understanding and evaluation mechanisms of how to effectively link between workflows, it is difficult in the prior art to automatically and reliably generate such an optimal execution link that meets logic and task goals, often relying on a great deal of manual intervention or resulting in inefficient quality of the generated link. Disclosure of Invention Aiming at the problems existing in the prior art, the invention provides a task execution method and related equipment of a large-model intelligent agent system, which aim to systematically organize and understand the internal association among workflows and intelligently generate an optimal execution path for a dynamically-changed complex task, so that the problems that deep inference association is difficult to support, an optimal execution link which accords with logic and is attached to a task target cannot be automatically and reliably generated due to workflow isolation storage and a fixed matching strategy are avoided, dependence on a large amount of manual intervention is reduced, and the efficiency and quality of link generation are improved. In order to solve the technical problems, the invention is realized by the following technical scheme: according to a first aspect of the present invention, there is provided a task execution method of a large model agent system, applied to the large model agent system, the method comprising: responding to a received task to be processed, and acquiring task description information of the task to be processed; Dynamically generating a corresponding target workflow execution link for the task description information based on a pre-constructed workflow knowledge graph, wherein the workflow knowledge graph is used for representing the association relation among a plurality of historical workflow units; And controlling the large-model intelligent agent system to execute the task to be processed according to the target workflow execution link. In a possible implementation manner of the first aspect, the nodes of the workflow knowledge graph are historical workflow units, and the characteristic data of the nodes include at least one o