Content generation method based on hierarchical multiple Agents

Abstract

The invention discloses a content generation method based on layered multiple Agents. The method comprises the following specific operation steps: (1) a global Agent is used for centrally managing task information, state information and execution stage information of all current content generation tasks, and giving a reply to a user in combination with global instantiation information when obtaining user input every time; (2) local Agents: solving sub-problems by the Agents by using various tools, and integrating human feedback into an Agent framework to jointly complete a task, so that the completion progress of the task is manually intervened, and the generation task is cooperatively solved; and (3) dynamically amplifying the content: generating the dynamically amplified content in two modes, namely content generation step Agent amplification and content generation mode Agent amplification. The method has the beneficial effects that the coupling between content generation modes is reduced, and the task completion accuracy and efficiency are improved; according to the invention, dynamic amplification of content generation steps and generation modes can be rapidly realized, so that different business requirements and changes of the generation modes can be met.

Description

Technical Field

[0001] The present invention relates to the technical field related to content generation, and in particular to a content generation method based on hierarchical multi-agent. Background Art

[0002] With the continuous improvement of the capabilities of Large Language Models (LLM), the use of a multi-agent framework to complete tasks of various complexities has been proven to be an efficient method. However, in current content generation tasks, multiple generation methods are often involved, such as generating videos and texts. Different generation methods often involve multiple steps such as material collection, and finally content generation. As the number of content generation methods increases, the number of corresponding agents also increases, resulting in the following problems when using Multi-Agent to schedule all agents: 1. There may be a certain coupling between the step descriptions of different content generation methods, such as generating videos and generating texts, which increases the complexity of using Multi-Agent to select the correct agent to complete the current content generation task, thereby reducing the accuracy of the response.

[0003] 2. Existing content generation methods usually lack flexibility and scalability, and it is difficult to quickly adapt to new business needs or changes in generation methods. Therefore, how to quickly and effectively expand content generation methods has become an urgent problem that needs to be solved. Summary of the invention

[0004] The present invention aims to overcome the above-mentioned deficiencies in the prior art and provides a content generation method based on hierarchical multi-agents, which reduces the coupling between content generation modes and realizes dynamic expansion.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: A content generation method based on hierarchical multi-agent uses global agent and local agent to hierarchically manage content generation status and steps, realizes orderly scheduling of multiple agents, and then realizes rapid expansion of content generation tasks. The specific operation steps are as follows: (1) Global Agent: It is used to centrally manage the task information, status information, and execution phase information of all current content generation tasks. It obtains user input each time and gives user responses based on the global instantiation information. (2) Local Agent: The agent uses various tools to solve sub-problems, and integrates human feedback into the agent framework to jointly complete tasks, realize the completion progress of human intervention tasks, and cooperate to solve the generated tasks; (3) Dynamically augmented content: There are two ways to dynamically augment content generation: Agent augmentation of content generation steps and Agent augmentation of content generation methods. Based on the above two methods, Agent calls adapted to different tasks are implemented to meet the needs of dynamically changing content production methods.

[0006] This method uses a hierarchical agent collaboration method, through global agent scheduling tasks and local agent management tasks, so as to reduce the coupling between content generation methods and improve the accuracy and efficiency of task completion. This method also provides a method for dynamically expanding content generation, which can quickly realize the dynamic expansion of content generation steps and generation methods to meet different business needs and changes in generation methods.

[0007] Preferably, in step (1), each time user input is obtained, a user reply is given in combination with the global instantiation information, and the specific information exchange is as follows: (11) Based on the user's question, GPT-4o selects the most appropriate task for the current question in combination with the current task information. If the current task status information is 1, it will not be changed; if the status information is 0, the status information will be changed to 1 until the task is completed; if there is no task response for a long time, the current task will be closed and the task status will be updated from 1 to 0; (12) After the task status information is opened, the data information in the current session is stored in Redis. When the user inputs the next round, the data information of the current task stored in Redis and the execution stage information of the current task are combined to decide whether to open a new task session or continue the existing session stage.

[0008] Preferably, the task information is a detailed description of the current task, including the task objectives, task stages and task workflow; the status information is a status control of whether the current task has been executed. If the task has been executed, the status is 1; if the task has not been executed, the status is 0; the execution stage information is a process control of multiple stages of the current task, and each stage is divided into: unfinished, in progress, and completed.

[0009] Preferably, in step (2), a complete task processing flow is as follows: (21) Select all agents needed to complete the content generation task; (22) Manual collaborative problem solving: Waiting for user input, starting the corresponding task phase or terminating the conversation when receiving a user message; (23) After the task is started, the Agent of the corresponding task stage is selected to complete the task according to the user message; the initial status of all Agents that have completed subtasks is "unfinished". After completing a stage of the task, the corresponding status is changed to "completed"; (24) When receiving user input again, if there is already an opened task, determine whether the current user needs to solve the problem of the subtask corresponding to the task. If so, continue the task until it is completed, otherwise start a new generated task. By manually controlling the information input to enter different tasks and their corresponding subtasks, it ensures that multiple tasks are processed in parallel, greatly improving the execution efficiency of the tasks.

[0010] Preferably, in step (3), for the expansion of the content generation step Agent, specifically: a new content generation step Agent needs to be added. Since the content generation step Agent is only related to the current content generation method, a corresponding content generation step Agent can be added to the local Agent of the content generation method.

[0011] Preferably, in step (3), for the expansion of the content generation method Agent, specifically: if a new content generation method is added, it is only necessary to first hand over the initialization information to the global Agent for management, and then select several related agents from the existing Agents to form a local Agent corresponding to the task according to the specific behavior flow of completing the content generation method.

[0012] The beneficial effects of the present invention are: by scheduling tasks through a global agent and managing tasks through a local agent, the coupling between content generation methods is reduced, and the accuracy and efficiency of task completion are improved; the content generation steps and generation methods can be dynamically expanded quickly to meet different business needs and changes in generation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is the global Agent information exchange flow chart; Figure 2 This is a schematic diagram of the video generation task disassembly; Figure 3 Taking the generation of video as an example, the local Agent message processing flow chart; Figure 4 It is a schematic diagram of a method for generating dynamically augmented content. DETAILED DESCRIPTION

[0014] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0015] When using a multi-agent approach to handle content generation tasks, it is often necessary to schedule all agents to select the correct agent to reply to the user. This places high demands on the context understanding ability of the LLM, which not only wastes resources and reduces reply efficiency, but may even cause confusion in agent scheduling.

[0016] At present, content generation tasks involve a variety of methods, such as text generation, video generation, etc. The process of text generation includes providing text theme information, configuring picture material information, and then generating text content; video generation requires first providing popular video material information and storyboard material information, and then generating the video after screen configuration. It can be seen that the processes of different content generation tasks all rely on a series of orderly information provision, material integration and content generation steps, and have structural similarities. Therefore, the present invention proposes a content generation method based on hierarchical multi-agent, which uses global agents and local agents to hierarchically manage content generation status and steps, realize orderly scheduling of multiple agents, and then realize rapid expansion of content generation tasks.

[0017] (1) Design a global agent to centrally manage the task information, status information, and execution phase information of all current content generation tasks. Given the similarity in the structure of the generation tasks, the following will take video generation as an example for detailed explanation. The instantiation information of video generation is shown in Table 1.

[0018] Table 1. Initial state information table of video generation

[0019] a.Task information, namely environment_prompt, is a detailed description of the current task, including task objectives, task stages, and task workflow.

[0020] b. Status information, i.e. status, is the status control of whether the current task has been executed. If the task has been executed, the status is 1; if the task has not been executed, the status is 0.

[0021] c. Execution phase information, namely task_phases, is the process control of multiple phases of the current task. For example, the video generation task is divided into multiple task phases, and each phase is divided into: unfinished, in progress, and completed.

[0022] Each time the user input is obtained, the user response is given in combination with the global instantiation information. The specific information exchange is as follows Figure 1 shown.

[0023] Step 1: Based on the user's question and the current task information, GPT-4o selects the most suitable task for the current question. If the current task status information is 1, do not change it. If the status information is 0, change the status information to 1 until the task is completed. If there is no task response for a long time, close the current task and update the task status from 1 to 0.

[0024] Step 2: After opening the task status information, store the data information in the current session in Redis. When the user inputs the next round, combine the data information of the current task stored in Redis and the execution phase information of the current task to decide whether to open a new task session or continue the existing session phase.

[0025] By dynamically managing the session state through step1 and step2, it is possible to dynamically instantiate different generation task information during the content generation task management process and process multiple generation methods in parallel.

[0026] (2) Local Agent: In the content generation task, each method requires multiple steps to implement. Agents use various tools to solve sub-problems. At the same time, human feedback is integrated into the Agent framework to complete the task together, realize the completion progress of human intervention tasks, and cooperate to solve the generation task. A complete task processing flow is as follows: Step 1: Select all the agents needed to complete the content generation task. Taking the video generation task as an example, first, the video generation task is disassembled into imitation video disassembly, material analysis, and video generation. Figure 2 As shown. Then, select the Agent required to complete the video generation task. Select UserProxyAgent for accepting user input, Video Disassembly Agent for video disassembly, Material Parsing Agent for material parsing, and Video Generation Agent for generating videos. The specific behavior of each Agent is specified in the system_message of the generation task, as shown in Table 1.

[0027] Step 2: Human collaboration to solve the problem. Figure 3 As shown, UserProxyAgent is used to wait for user input. When a user message is received, the corresponding task phase is started or the conversation is terminated.

[0028] Step 3: After starting the task, UserProxyAgent is familiar with the subtasks and subtask steps of all other Agents, and is responsible for overall task planning, guiding other Agents and tracking task progress. According to user messages, select the Agent of the corresponding task stage to complete the task. Currently, the initial status of all Agents that have completed subtasks is "unfinished". After completing a stage of the task, the corresponding status is changed to "completed".

[0029] Step 4: When receiving user input again, if there is an opened task, determine whether the current user needs to solve the problem of the corresponding subtask of the task. If so, continue the task until it is completed, otherwise start a new generated task. By manually controlling the input of information into different tasks and their corresponding subtasks, multiple tasks can be processed in parallel, which greatly improves the execution efficiency of tasks.

[0030] By only activating the Agents of the current task scene each time, the appropriate Agent for the task stage is selected according to the user message to improve the response efficiency. In addition, through human feedback at each step, the transition of subtasks in the scene is guaranteed, and errors that occur are reflected on and corrected by humans in a timely manner to improve the accuracy of task completion.

[0031] (3) Dynamic augmentation of content generation methods: In the content generation task, in order to meet business needs, there are two ways to dynamically augment content generation: content generation step agent augmentation and content generation method agent augmentation, such as Figure 4 shown.

[0032] (31) Content generation step Agent expansion The content generation method is implemented by multiple sub-steps, such as Figure 4 As shown in video generation method 1, video disassembly, material analysis and video generation are currently required. In order to increase content marketing revenue, the generated video needs to be forwarded to public network platforms such as Douyin for release. For this purpose, a new agent for video release needs to be added. Since this agent is only related to the current video content generation method, the corresponding video release agent can be added to the local agent of the content generation method, such as Figure 4 As shown in video generation method 2, this method not only avoids the impact on other content generation methods, but also achieves rapid local agent expansion.

[0033] (32) Content generation method Agent expansion If a new video generation method is added, such as only requiring users to upload video materials for video generation, it is only necessary to first hand over the initialization information such as task objectives, task stages, and task workflow to the global agent for management, and then select the material analysis agent and video generation agent from the existing agents to form the local agent corresponding to the task according to the specific behavior flow of completing the generation method, such as Figure 4 As shown in video generation method 3, dynamic expansion of content generation methods can be quickly achieved.

[0034] Based on the above two methods, Agent calls adapted to different tasks can be quickly implemented to meet the needs of dynamically changing content production methods.

[0035] In summary, the present invention uses a hierarchical multi-agent management method. The global agent schedules different tasks by managing the task information, state information and execution stage information of all tasks. The local agent manages the execution steps of tasks through human feedback, thereby reducing the coupling between different tasks and improving the accuracy and efficiency of task completion. In the present invention, the local agent includes one or more steps, each of which is processed by an agent. The local agent capability can be expanded by adding a corresponding step agent to the local agent, or a new local agent can be formed by using the existing step agent to achieve a new content generation method; through the expansion of the local agent, the generation steps can be added or modified separately, which is convenient for management and maintenance, and can quickly adapt to different tasks and new generation method requirements without large-scale modification of the entire system; that is, a method of dynamically expanding content generation is provided, which can quickly realize the dynamic expansion of content generation steps and generation methods to meet different business needs and changes in generation methods.

Claims

1. A content generation method based on hierarchical multi-agent, characterized in that: Use global agents and local agents to hierarchically manage content generation status and steps, and schedule multiple agents in an orderly manner, thereby achieving rapid expansion of content generation tasks. The specific steps are as follows: (1) Global Agent: It is used to centrally manage the task information, status information, and execution phase information of all current content generation tasks. It obtains user input each time and gives user responses based on the global instantiation information. (2) Local Agent: The agent uses various tools to solve sub-problems, and integrates human feedback into the agent framework to jointly complete tasks, realize the completion progress of human intervention tasks, and cooperate to solve the generated tasks; (3) Dynamically augmented content: There are two ways to dynamically augment content generation: Agent augmentation of content generation steps and Agent augmentation of content generation methods. Based on the above two methods, Agent calls adapted to different tasks are implemented to meet the needs of dynamically changing content production methods.

2. The method for generating content based on hierarchical multi-agent according to claim 1, characterized in that: In step (1), each time the user input is obtained, the user response is given in combination with the global instantiation information. The specific information exchange is as follows: (11) Based on the user's question, GPT-4o selects the most appropriate task for the current question in combination with the current task information. If the current task status information is 1, it will not be changed; if the status information is 0, the status information will be changed to 1 until the task is completed; if there is no task response for a long time, the current task will be closed and the task status will be updated from 1 to 0; (12) After the task status information is opened, the data information in the current session is stored in Redis. When the user inputs the next round, the data information of the current task stored in Redis and the execution stage information of the current task are combined to decide whether to open a new task session or continue the existing session stage.

3. A method for generating content based on hierarchical multi-agent according to claim 1 or 2, characterized in that: The task information is a detailed description of the current task, including the task objectives, task stages and task workflow; the status information is a status control of whether the current task has been executed. If the task has been executed, the status is 1; if the task has not been executed, the status is 0; the execution stage information is a process control of multiple stages of the current task, and each stage is divided into: unfinished, in progress, and completed.

4. The method for generating content based on hierarchical multi-agent according to claim 1, characterized in that: In step (2), a complete task processing flow is as follows: (21) Select all agents needed to complete the content generation task; (22) Manual collaborative problem solving: Waiting for user input, starting the corresponding task phase or terminating the conversation when receiving a user message; (23) After the task is started, the agent of the corresponding task stage is selected to complete the task according to the user message; Currently, the initial status of all Agents that have completed subtasks is "unfinished". After completing a stage of tasks, the corresponding status is changed to "completed"; (24) When receiving user input again, if there is already an opened task, determine whether the current user needs to solve the problem of the subtask corresponding to the task. If so, continue the task until it is completed, otherwise start a new generated task. By manually controlling the information input to enter different tasks and their corresponding subtasks, it ensures that multiple tasks are processed in parallel, greatly improving the execution efficiency of the tasks.

5. The method for generating content based on hierarchical multi-agent according to claim 1, characterized in that In step (3), for the expansion of the content generation step Agent, specifically: a new content generation step Agent needs to be added. Since the content generation step Agent is only related to the current content generation method, a corresponding content generation step Agent can be added to the local Agent of the content generation method.

6. The method for generating content based on hierarchical multi-agent according to claim 1, characterized in that In step (3), for the expansion of the content generation method agent, specifically: if a new content generation method is added, it is only necessary to first hand over the initialization information to the global agent for management, and then select several related agents from the existing agents to form the local agent corresponding to the task according to the specific behavior process of completing the content generation method.

Images