Short video full-process customer acquisition system based on agent cooperation

Abstract

The application provides a short video full-process customer acquisition system based on agent cooperation, and relates to the technical field of artificial intelligence. The input layer of the system is used for receiving task data input by a user. The scheduling layer adopts an agent scheduling engine. The agent scheduling engine is used for determining at least one subtask and a subtask execution sequence according to the task data, and distributing the subtask based on a publish / subscribe mode and allocating computing power resources based on a weighted round robin algorithm. The function layer includes an account positioning module, a benchmark account disassembly module, an account foundation building module, a topic library building module, a video creation module and a clipping and publishing module which are independently encapsulated in the form of agents. Based on the subtask execution sequence and the computing power resources, the account positioning module, the benchmark account disassembly module, the account foundation building module, the topic library building module, the video creation module and the clipping and publishing module execute corresponding subtasks. The application can automatically create short videos to acquire customers and improve efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a short video full-process customer acquisition system based on agent cooperation. BACKGROUND

[0002] Short video creation customer acquisition refers to attracting target audience attention through short video full-process (covering account positioning, topic planning, content creation, video shooting and editing, publishing and operation, data review and optimization, etc.) to ultimately realize customer conversion. The core is to reach potential customers through high-quality short video content, and rely on data optimization to continuously improve customer acquisition efficiency. It is an important way for current enterprises and individual entrepreneurs to obtain traffic and achieve business success.

[0003] In related technologies, short video creation customer acquisition each link (topic conception, copywriting, shooting, editing, packaging, time selection and data interpretation) is dominated by manual operation, but manual operation is unstable. The efficiency of short video creation customer acquisition is low. SUMMARY

[0004] The problem solved by the present application is how to improve the efficiency of video creation customer acquisition.

[0005] To solve the above problems, the present application provides a short video full-process customer acquisition system based on agent cooperation, which comprises an input layer, a scheduling layer and a function layer.

[0006] The input layer is used to receive user input task data.

[0007] The scheduling layer adopts an agent scheduling engine, which is used to determine at least one subtask and a subtask execution order according to the task data, and distribute the subtasks based on a publish / subscribe mode and allocate computing resources based on a weighted round robin algorithm.

[0008] The function layer includes an account positioning module, a benchmark account disassembly module, an account foundation building module, a topic library building module, a video creation module and a editing and publishing module independently encapsulated in the form of an agent. Based on the subtask execution order and the computing resources, the account positioning module, the benchmark account disassembly module, the account foundation building module, the topic library building module, the video creation module and the editing and publishing module execute the corresponding subtasks.

[0009] Optionally, the agent scheduling engine includes a task distributor, a resource manager and an exception processor.

[0010] The task distributor is configured to parse the sub-tasks of the task data, determine sub-task priorities and sub-task dependency relationships of the sub-tasks by using an asynchronous pipeline scheduling algorithm, determine an execution order of the sub-tasks according to the sub-task dependency relationships, and distribute the sub-tasks to each agent in the functional layer based on the publish-subscribe mode;

[0011] The resource manager is configured to monitor resource usage, and allocate the computing resource according to the sub-task priorities based on the resource usage;

[0012] The exception processor is configured to handle task exceptions based on a three-level exception handling mechanism.

[0013] Optionally, the determination of the sub-task priorities and the sub-task dependency relationships of the sub-tasks by using the asynchronous pipeline scheduling algorithm comprises:

[0014] identifying a pre-dependence of each of the sub-tasks according to a directed acyclic graph of the task data, and determining the sub-task dependency relationships;

[0015] determining the sub-task priorities by using a priority formula, the priority formula comprising:

[0016] Priority = α×Urgency + β×(1 / ResourceCost) + γ×(1 / EstimatedTime);

[0017] wherein Priority is the sub-task priority, Urgency is an urgency degree, ResourceCost is a resource cost, EstimatedTime is an estimated time, and α, β and γ are weight coefficients.

[0018] Optionally, the handling of the task exceptions based on the three-level exception handling mechanism comprises:

[0019] when the execution of the sub-task fails, performing local retry by using a smart retry strategy matrix;

[0020] when the local retry fails, re-distributing the sub-task whose execution fails to a backup agent by using the task distributor;

[0021] when the execution of the backup agent fails, triggering a degradation scheme and pushing a human intervention notification.

[0022] Optionally, the account positioning module comprises a target audience identification agent, a content direction intelligent decision-making agent, and a positioning label generation agent.

[0023] The target audience identification agent is configured to determine an audience portrait according to the task data, wherein the audience portrait comprises basic attributes, interest labels, and priority rankings.

[0024] The content direction intelligent decision-making agent is configured to determine a short video content direction suggestion scheme according to the obtained popular data and the audience portrait.

[0025] The positioning label generation agent is configured to determine a short video positioning label according to high-frequency words in the industry in the task data.

[0026] Optionally, the video creation module comprises a script agent and a lip-sync digital human agent.

[0027] The script agent is configured to extract popular scripts of popular short videos, and generate innovative scripts of different styles based on the popular scripts.

[0028] The lip-sync digital human agent is configured to analyze script semantics of the innovative scripts, split the innovative scripts into lip-sync paragraphs, add SSML voice marks, expressions or action instructions, generate a structured script that can be parsed by a digital human platform, determine a digital human image matched with the script semantics and audience preferences, call an API of the digital human platform according to the script semantics, the structured script and the digital human image, and generate a lip-sync digital human video.

[0029] Optionally, the functional layer further comprises a data analysis and optimization module, which is configured to analyze short video key data indicators after the video is released by the clipping and publishing module, and obtain short video success reasons and failure reasons.

[0030] Optionally, the functional layer further comprises a data review agent, which is configured to generate a review report according to the short video key data indicators.

[0031] Optionally, the short video full-process customer acquisition system based on agent cooperation further comprises a support layer, wherein the support layer comprises a knowledge base and a model base, the knowledge base stores industry rules and competitor data, and the model base integrates NLP models, ASR models and digital human models.

[0032] Optionally, the short video full-process customer acquisition system based on agent cooperation further comprises a storage layer, wherein the storage layer adopts a distributed database and is configured to store various data in the input layer, the scheduling layer and the functional layer.

[0033] The short video full-process customer acquisition system based on agent cooperation has the following advantages:

[0034] The task data input by the user is received through the input layer, complete and accurate data basis is provided for subsequent task processing from the starting point of the whole process, the coherence of the initial data flow of the link is ensured, on the basis, the intelligent agent scheduling engine adopted by the scheduling layer can accurately determine at least one subtask and the subtask execution order according to the input task data, the task is published to the message queue through the publish / subscribe mode, and each intelligent agent of the function layer subscribes the task according to its own ability to realize decoupling and dynamic expansion, and the computing resource is dynamically allocated based on the weighted round robin algorithm, effectively solving the confusion and uneven resource allocation problems of traditional manual task allocation, not only ensuring the orderliness of task execution, but also improving the resource utilization, providing core scheduling support for the efficient execution of subtasks of each module in the function layer, and the account positioning module, the benchmark account disassembly module, the account foundation building module, the topic library building module, the video creation module and the editing and publishing module in the function layer are independently encapsulated in the form of intelligent agents, can execute corresponding subtasks according to the subtask execution order and the allocated computing resource, the account positioning module can accurately determine the target audience and the content direction to improve the customer acquisition accuracy, the benchmark account disassembly module can generate disassembly reports and optimization suggestions, the topic library building module can dynamically update the topic library, the video creation module can automatically generate video scripts and digital population video, and the editing and publishing module can intelligently edit and select the best publishing time, the seamless connection between modules according to the process avoids the unstable problems caused by manual work, promotes the transformation of short video customer acquisition from the traditional manual low-efficiency mode to the automatic and intelligent industrialized assembly line mode, and improves the efficiency of video creation customer acquisition. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 A structure schematic diagram of the short video whole-process customer acquisition system based on intelligent agent cooperation provided by the embodiment of the present application is shown. DETAILED DESCRIPTION

[0036] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings. Although some embodiments of the present application are shown in the drawings, it should be understood that the present application can be implemented in various forms, and should not be interpreted as being limited to the embodiments described herein, on the contrary, these embodiments are provided to make the present application more thorough and complete. It should be understood that the drawings and embodiments of the present application are only for illustrative purposes, and are not intended to limit the scope of protection of the present application.

[0037] It should be understood that each step described in the method embodiment of the present application can be executed in different order and / or in parallel. In addition, the method embodiment can include additional steps and / or omit the execution of the steps shown. The scope of the present application is not limited in this respect.

[0038] The term "include" and variations thereof, as used in this document, mean the same as "comprise" or "comprises." The term "based on" means "based, at least in part, on." The term "one embodiment" means "at least one embodiment." The term "another embodiment" means "at least one additional embodiment." The term "some embodiments" means "at least some embodiments." The term "optional" means "optional in at least some embodiments." Related terms such as "one embodiment," "an embodiment," "some embodiments," "another embodiment," "at least one embodiment," "at least one other embodiment," "at least one further embodiment," "at least one additional embodiment," "at least one more embodiment," and the like, mean "at least one but not necessarily all embodiments of the invention." The terms "the invention" and "the present invention" refer to one or more embodiments of the invention and not necessarily all embodiments need have all features indicated in this summary. The terminology used in this document is for the purpose of describing particular embodiments only and is not intended to be limiting of other embodiments of the invention. Other embodiments of the invention will be apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. For example, the specification can contemplate embodiments and applications other than those described in this summary. It will be apparent to one skilled in the art that the present invention can be practiced under conditions and parameters other than those described in this summary. The terminology used in the description of the particular embodiments illustrated in the drawings is not intended to limit the scope of the present invention. It is possible, for example, that in other embodiments "the invention" can be practiced under conditions and parameters distinct from those described in this summary. Accordingly, the specification and figures are to be regarded in an illustrative manner and are not intended to limit the scope of the invention.

[0039] It should be noted that the terms "one", "multiple", "a number of" mentioned in the present application are illustrative rather than restrictive, and those skilled in the art should understand that "one" or "multiple" should be understood as "one or more" unless otherwise explicitly indicated in the context.

[0040] The names of the messages or information exchanged between the plurality of devices in the embodiments of the present application are only for illustrative purposes, and are not intended to limit the scope of the messages or information.

[0041] To solve the problems in the related art, the short video full-process customer acquisition system based on agent cooperation is provided.

[0042] As shown in Figure 1 The short video full-process customer acquisition system based on agent cooperation provided by the embodiments of the present application includes an input layer, a scheduling layer and a function layer.

[0043] The input layer is configured to receive task data input by a user.

[0044] Specifically, the input layer is configured to receive task data input by a user, that is, task introduction, product details, video links and account parameters, and to implement input adaptation of various formats such as text, links and files, to ensure that various data can smoothly enter the scheduling layer and the function layer.

[0045] The scheduling layer adopts an agent scheduling engine, which is configured to determine at least one subtask and a subtask execution sequence according to the task data, and distribute the subtasks based on a publish / subscribe mode and allocate computing resources based on a weighted round robin algorithm.

[0046] Specifically, the intelligent agent scheduling engine decomposes complex task requirements in the task data into at least one subtask and determines a subtask execution sequence, so as to sequentially perform the full-process links of customer acquisition, i.e., topic conception, copywriting, shooting and scene setting, editing and packaging, time selection for delivery, and data interpretation. The publish-subscribe mode refers to publishing each subtask to a message queue, and an executing intelligent agent subscribes to a corresponding task type according to its own set capability, thereby taking the subtask and executing the subtask according to the subtask execution sequence. The weighted round robin algorithm is commonly used in load balancing, task scheduling, network forwarding and other scenarios, and its core goal is to proportionally allocate tasks or requests. Through the weighted round robin algorithm, the computing power resources of the system can be reasonably allocated, thereby improving the computing efficiency of the system.

[0047] The function layer includes an account positioning module, a benchmark account disassembly module, an account foundation building module, a topic library building module, a video creation module, and a editing and publishing module independently encapsulated in the form of intelligent agents. Based on the subtask execution sequence and the computing power resources, the account positioning module, the benchmark account disassembly module, the account foundation building module, the topic library building module, the video creation module, and the editing and publishing module execute corresponding subtasks.

[0048] Specifically, each module in the functional layer is built using one or more intelligent agents. Each agent in a module can achieve its corresponding function through orchestration and debugging. The account positioning module receives task data from users, such as business introductions and product details, to clarify the account's target audience, content direction, and positioning tags. For example, for a cake shop, the account's theme could be determined as "creative cake making and sharing," targeting young people who love desserts. Meanwhile, the account positioning module can deeply analyze the creator's background, help clarify the target audience profile, achieve content differentiation positioning, and provide specific content planning suggestions; the benchmark account decomposition module is used to decompose benchmark accounts, generate decomposition reports and account optimization suggestions. The workflow covers nodes such as obtaining benchmark account information, batch obtaining video lists, extracting video scripts, generating decomposition reports, generating reference suggestions, and writing to Lark documents, and finally storing the report in Lark; the account basic construction module is used to generate multiple account basic construction schemes based on account positioning and the system's knowledge base, covering the design of personal profile pages such as avatar, nickname, and introduction, the improvement of basic information such as geographical location and contact information, and providing suggestions for pinned videos; the topic selection library construction module is used to collect competitor content, posting frequency and themes in real time, analyze interaction data, gain insights into content trends, keep abreast of industry dynamics and competitor movements, and mine content. The creation of content and the selection of topics can be achieved through processes such as streamlining workflows, setting character personas and logic, binding workflows, setting triggers, and testing releases. The topic selection module can include a hotspot monitoring agent and a topic planning agent. The hotspot monitoring agent is used to promptly acquire trending videos in the field, helping to grasp market trends, learn creation techniques, discover business opportunities, and understand competitors. Its construction process is similar to that of the benchmark account monitoring agent, which collects hot videos through a specific workflow. The topic planning agent is used to automatically generate a dynamically updated topic library by combining hotspot monitoring data and benchmark account content. The video creation module can imitate the scripts of existing popular short videos and generate new short videos. The editing and publishing module is used to edit the videos created in the video creation module and select the best release time based on data analysis.

[0049] For example, the input layer, scheduling layer and functional layer of this application can be built using the button platform or other intelligent agent platforms.

[0050] In this embodiment, the task data input by the user is received through the input layer, which provides complete and accurate data basis for subsequent task processing from the beginning of the whole process, guarantees the coherence of the initial data flow of the link, and on this basis, the agent scheduling engine adopted by the scheduling layer can accurately determine at least one subtask and the subtask execution order according to the input task data, publish the task to the message queue through the publish / subscribe mode, and let each agent of the function layer subscribe to the task according to its own ability to realize decoupling and dynamic expansion. At the same time, based on the weighted round-robin algorithm, the computing resource is dynamically allocated, effectively solving the chaos and uneven resource allocation problems of traditional manual task allocation, not only guaranteeing the orderliness of task execution, but also improving the resource utilization rate, providing core scheduling support for the efficient execution of subtasks by each module of the function layer. The account positioning module, benchmark account disassembly module, account foundation building module, topic library building module, video creation module and editing and publishing module in the function layer are independently encapsulated in the form of agents, which can execute corresponding subtasks according to the subtask execution order and allocated computing resources. The account positioning module can accurately determine the target audience and content direction to improve the customer acquisition accuracy, the benchmark account disassembly module can generate disassembly reports and optimization suggestions, the topic library building module can dynamically update the topic library, the video creation module can automatically generate video scripts and digital population video, and the editing and publishing module can intelligently edit and select the best publishing time. Seamless connection between modules according to the process avoids the problem of instability caused by manual work, promotes the transformation of short video customer acquisition from the traditional inefficient manual mode to the automated and intelligent industrial pipeline mode, and thus improves the efficiency of video creation customer acquisition.

[0051] Optionally, the agent scheduling engine comprises a task distributor, a resource manager and an exception processor.

[0052] The task distributor is configured to analyze the subtasks of the task data, determine the subtask priority and subtask dependency relationship of the subtasks by using an asynchronous pipeline scheduling algorithm, determine the subtask execution order according to the subtask dependency relationship, and distribute the subtasks to each agent in the function layer based on the publish / subscribe mode.

[0053] The resource manager is configured to monitor the resource usage, and allocate the computing resource according to the subtask priority based on the resource usage.

[0054] The exception processor processes task exceptions based on a three-level exception processing mechanism.

[0055] Specifically, the agent scheduling engine includes a task distributor, a resource manager, and an exception processor. The task distributor first decomposes complex tasks in task data into at least one subtask, determines subtask priorities and subtask dependency relationships of the subtasks by using an asynchronous pipeline scheduling algorithm, determines a subtask execution order according to the subtask dependency relationships, and allocates the subtasks that can be executed in parallel to different agents to enable the subtasks that can be executed in parallel to be processed simultaneously to improve processing efficiency. The subtasks are distributed to each agent in the functional layer based on a publish / subscribe mode. The resource manager monitors resource usage, i.e., resource usage of devices such as CPUs, memories, and GPUs, analyzes resource requirements of tasks, dynamically calculates load weights of each agent based on a weighted round-robin algorithm and resource usage historical data, and allocates computing power resources based on resource usage, subtask priorities, resource requirements, and load weights to avoid resource overload or idling and to reserve necessary computing resources for high-priority tasks to ensure fairness and efficiency of resource allocation. The exception processor monitors execution states of each agent, discovers abnormal conditions in a timely manner, and processes the abnormal conditions through a three-level exception processing mechanism to ensure normal operation of the system.

[0056] Exemplarily, the resource requirement analysis step of the resource manager includes:

[0057] The subtasks are subjected to resource type identification. The resource types include CPU-intensive tasks, memory-intensive tasks, I / O-intensive tasks, and GPU-intensive tasks. The CPU-intensive tasks include video transcoding, model inference, and data analysis. The memory-intensive tasks include large file processing and batch data loading. The I / O-intensive tasks include file downloading, database reading and writing, and network requests. The GPU-intensive tasks include deep learning model training and image processing. Then, resource consumption is quantified, historical data is counted, resource usage patterns of past tasks are analyzed, resource effect data is determined according to a resource consumption formula, resource usage fluctuations in a task execution process and a delay in resource release after a task is completed are considered, resource conflict detection is performed to identify a situation in which multiple subtasks compete for the same resource, a key resource that limits overall system performance is found, and resource utilization is determined by using a resource utilization formula. The resource consumption formula includes:

[0058] ResourceCost = BaseLoad + TaskComplexity × ScaleFactor.

[0059] wherein ResourceCost is resource consumption data, BaseLoad is a basic load, TaskComplexity is task complexity, and ScaleFactor is a scale factor.

[0060] The resource utilization formula includes:

[0061] ResourceUtilization = UsedResource / TotalResource x 100%;

[0062] wherein ResourceUtilization is resource utilization, UsedResource is used resource, and TotalResource is total resource.

[0063] Exemplarily, the task distributor can adopt an optimization scheduling strategy for scheduling optimization, and the optimization scheduling strategy includes:

[0064] An integrated priority formula is adopted to determine an integrated priority to prioritize scheduling of tasks on a critical path or a task combination with high resource utilization, and the integrated priority formula includes:

[0065] Priority1 = a x CriticalPathWeight + b x ResourceEfficiency + g x TaskUrgency;

[0066] wherein Priority1 is integrated priority, CriticalPathWeight is critical task weight, ResourceEfficiency is resource efficiency, TaskUrgency is task urgency, and a, b and g are weight coefficients.

[0067] Exemplarily, the failure analysis step of the exception processor includes:

[0068] Failure signal acquisition, i.e., abnormal code analysis: failure classification according to returned HTTP status code and error code; monitoring execution time: detecting whether task execution time exceeds a preset threshold; resource state checking: monitoring CPU, memory, network and other resource usage; dependent service state: checking third-party API, database connection and other external dependencies.

[0069] Failure root cause analysis, i.e., analyzing network failure: network delay, connection timeout, DNS resolution failure; resource shortage: memory overflow, CPU overload, insufficient disk space; service failure: third-party API flow limiting, service unavailability, authentication failure; data anomaly: input format error, data corruption, illegal parameters; system failure: process crash, system restart, insufficient permissions.

[0070] The fault influence evaluation is classified into critical (system level), high (module level), medium (task level) and low (negligible) according to the severity level, the influence range analysis includes single task influence, module influence and system influence, and the recovery time estimation is based on historical data to predict the time required for fault recovery.

[0071] Optionally, the asynchronous pipeline scheduling algorithm is used to determine the subtask priority and the subtask dependency of the subtask, which comprises:

[0072] According to the directed acyclic graph of the task data, the pre-dependence of each subtask is identified to determine the subtask dependency;

[0073] A priority formula is used to determine the subtask priority, and the priority formula comprises:

[0074] Priority = α × Urgency + β × (1 / ResourceCost) + γ × (1 / EstimatedTime);

[0075] Wherein, Priority is the subtask priority, Urgency is the urgency, ResourceCost is the resource cost, EstimatedTime is the estimated time, and α, β and γ are weight coefficients.

[0076] Specifically, the principle of the asynchronous pipeline scheduling algorithm is based on task dependency analysis of a directed acyclic graph (DAG), combined with a priority queue and resource constraints, to realize parallel scheduling and execution of tasks. The core idea of the algorithm is to convert a serial business process into a task pipeline that can be executed in parallel, maximizing system throughput to improve efficiency. The specific steps of the asynchronous pipeline scheduling algorithm are as follows: a directed acyclic graph is constructed, i.e., G=(V, E), where V is a set of task nodes, and E is a set of dependency relationship edges. Then, the in-degree and out-degree of each task node are calculated to identify a set of tasks that can be executed in parallel (nodes with an in-degree of 0), and task dependency analysis is performed to obtain subtask dependency relationships. Next, priority calculation is performed using a priority formula, in which the weight coefficients can be dynamically adjusted according to the business scenario, the urgency is the urgency of the subtask, the resource cost is the resource cost required to complete the subtask, and the estimated time is the estimated time to complete the subtask. Next, resource matching and allocation are performed. The priority queue can be scanned to select the task with the highest priority and the required resources that can be met by the current system, and the resource availability of the target agent (CPU, memory, and concurrency limit) is checked to ensure that the target agent has the ability to process the task and prevent execution failure due to resource overload. After confirming that the task is executable and the agent has resources, the actual resource allocation step is entered, in which the required resource quota is allocated to the task (such as allocating 2-core CPU and 4GB memory to the "AI editing" task), and these resources are locked (to prevent being occupied by other tasks). The system resource management table is marked with "a certain agent has occupied XX resources", and the resource usage data is synchronized in real time (such as the GPU usage rate rising from 30% to 50%), realizing the "exclusive allocation" of resources and avoiding resource contention among multiple tasks, while allowing the scheduling engine to monitor the resource status in real time to provide accurate data for the next scheduling. Next, asynchronous task scheduling is performed, in which the task is packaged as an asynchronous execution unit and submitted to the execution queue of the corresponding agent. The task execution status and resource usage are continuously monitored, and when resource imbalance is detected, dynamic adjustment of task allocation is performed. In addition, predictive scheduling can also be performed: based on historical data to predict future resource demand and perform scheduling optimization in advance. During the execution of the agent, abnormality detection is performed, and when the task execution time exceeds the expected time, resource enhancement or task splitting is started. When the system resources are insufficient, task queuing or degradation processing is started. When the current task fails, a retry mechanism or a backup plan is started.

[0077] By way of example, the steps of task dependency analysis include:

[0078] The directed acyclic graph is topologically analyzed to identify the precedence dependency of each subtask, determine which task output is the input of other tasks, find the longest dependency chain that affects the overall execution time, and ensure that the task flow has no deadlock risk, wherein the precedence dependency is divided into strong dependency, weak dependency and conditional dependency, the strong dependency means that task B must wait for task A to complete before starting (such as copywriting extraction depending on video download completion), the weak dependency means that task B can start after task A is partially completed (such as video editing can start after part of the material is downloaded), and the conditional dependency means that whether to execute task B is determined according to the result of task A (such as whether to publish according to the content audit result), thereby identifying the parallel degree, that is, tasks without mutual dependency can be executed simultaneously, and then determining the theoretical maximum parallel degree according to the theoretical maximum parallel degree formula, while ensuring that parallel tasks will not cause data conflicts, and the theoretical maximum parallel degree formula is:

[0079] Max_Parallel = Total_Tasks / Critical_Path_Length;

[0080] Wherein, Max_Parallel is the theoretical maximum parallel degree, Total_Tasks is the total number of tasks, and Critical_Path_Length is the critical path length.

[0081] Optionally, the three-level exception handling mechanism is used to handle task exceptions, including:

[0082] When the subtask execution fails, an intelligent retry strategy matrix is used for local retry;

[0083] When the local retry fails, the task distributor is used to reassign the failed subtask to a backup agent;

[0084] When the backup agent fails, a degradation scheme is triggered, and a manual intervention notification is pushed.

[0085] Specifically, when the subtask execution fails, a smart retry strategy matrix is adopted for local retry, the smart retry strategy matrix includes exponential backoff retry (suitable for network failure, temporary service unavailability), fixed interval retry (suitable for resource competition, queue full), immediate retry (suitable for transient failure, data anomaly) and delayed retry (suitable for system maintenance, timing failure), the exponential backoff retry refers to that the retry interval is 1 second, 2 seconds, 4 seconds, 8 seconds, 16 seconds, or the retry interval is determined according to a retry interval formula, the maximum number of retries is 5 times, and the applicable scenarios are network timeout, API flow limiting, temporary service failure, the fixed interval retry refers to that the retry interval is fixed at 30 seconds, the maximum number of retries is 3 times, and the applicable scenarios are resource shortage, queue full, concurrency conflict, and condition checking is performed, that is, resource availability is checked before each retry; the immediate retry refers to that the retry interval is 0 seconds, the maximum number of retries is 2 times, and the applicable scenarios are data format error, parameter anomaly, transient network jitter, and if continuous failure occurs, the downgrade scheme is immediately entered; the delayed retry refers to that the retry interval is 5 minutes, 10 minutes, 30 minutes, the maximum number of retries is 3 times, and the applicable scenarios are system maintenance window, timing task conflict, and intelligent scheduling is performed to avoid known system maintenance time. When the local retry fails, the task distributor is used to reassign the failed subtask to a backup agent, the backup agent is deployed in a function layer, and each module in the function layer has a corresponding backup agent; when the backup agent fails, a downgrade scheme is triggered, and a manual intervention notification is pushed, the downgrade scheme includes primary downgrade, secondary downgrade and tertiary downgrade, the primary downgrade is function downgrade, which can reduce output quality, such as reducing video resolution and reducing special effects, can simplify the processing flow, such as skipping non-critical steps, and can use cached data, such as using historical successful results; the secondary downgrade is service replacement, which can switch to a backup service, such as using a backup third-party service, perform local processing, such as converting cloud processing to local processing, and perform manual intervention, such as marking the task as needing manual processing; the tertiary downgrade is task suspension, which can add the failed task to a delay queue, send a failure notification to an administrator, and manually restart after the system recovers. The retry interval formula includes:

[0086] RetryInterval = BaseInterval×2^(RetryCount-1);

[0087] wherein, RetryInterval is the retry interval, BaseInterval is the base interval, and RetryCount is the number of retries.

[0088] Exemplarily, the abnormal processing strategy can be determined by a dynamic strategy selection algorithm, the dynamic strategy selection algorithm includes:

[0089] The fault feature extraction specifically includes: mapping specific error codes generated in the system operation process to a preset fault type system to realize preliminary classification of faults; extracting effective processing experience by searching a historical fault processing database, analyzing the execution effect of historical cases similar to the current fault characteristics under different processing strategies; and combining the real-time load state of the current system, traffic characteristics of the time period, and other context information to construct a multi-dimensional fault feature vector to provide a comprehensive basis for subsequent strategy selection. The strategy matching decision specifically includes: based on the extracted fault feature vector, using a decision tree algorithm to search a preset retry strategy library to preliminarily select candidate strategies that adapt to the current fault scenario; calculating the comprehensive score of each candidate strategy by using a preset formula "Success_Rate x 0.4 + Recovery_Time x 0.3 + Resource_Cost x 0.3", wherein Success_Rate is the historical success rate, Recovery_Time is the average recovery time, and Resource_Cost is the resource consumption cost, and selecting the strategy with the highest score as the optimal execution strategy; dynamically adjusting the weight coefficients in the above formula according to the real-time feedback data in the strategy execution process, so that the strategy selection can adapt to the system state changes. The retry execution and monitoring specifically includes: in the strategy execution process, real-time tracking of the execution state, result data and time-consuming information of each retry to form a complete retry process record; according to the correspondence between the fault type and the strategy type, statistically analyzing the actual success rate of different strategies in various fault scenarios to update the strategy effect evaluation database; based on the above statistical data, adaptively optimizing the strategy library, including eliminating low-success-rate strategies, adding combined strategies, adjusting strategy parameters, and the like, so that the strategy selection mechanism can continuously adapt to the long-term running characteristics of the system.

[0090] Exemplarily, after processing the exception, fault learning and optimization can also be performed, a fault mode library can be constructed to record common fault types and solutions, a successful case library can be constructed to save successful recovery processing strategies, failure case analysis can be performed to analyze the root cause of retry failure, and predictive fault processing can be performed, which can predict potential faults based on system indicators, can actively enable a degradation strategy before a fault occurs, and can optimize resource allocation according to fault statistics.

[0091] Optionally, the account positioning module includes a target audience identification intelligent agent, a content direction intelligent decision-making intelligent agent, and a positioning label generation intelligent agent.

[0092] The target audience identification intelligent agent is configured to determine an audience portrait according to the task data, wherein the audience portrait includes basic attributes, interest labels, and priority sorting.

[0093] The content direction intelligent decision-making agent is configured to determine a short video content direction suggestion scheme according to the obtained popular data and the audience portrait.

[0094] The positioning tag generation agent is configured to determine a short video positioning tag according to high-frequency words in the industry in the task data.

[0095] Specifically, the target audience identification agent extracts core keywords in the task description according to the task data using a TF-IDF algorithm, identifies an industry category based on a Naive Bayes classifier, analyzes product features using a Word2Vec model, constructs a task feature vector containing core keywords, an industry category, and product features, establishes a people tag system containing dimensions such as age, gender, interest, and consumption ability, and calculates a matching degree of the task feature vector and the people tag using a cosine similarity algorithm, wherein the weight of age is 0.25, the weight of interest is 0.35, the weight of consumption ability is 0.25, and the weight of behavior is 0.15. Then, the audience priority is calculated according to an audience priority formula, the core audience and the potential audience are determined, and thus the audience portrait containing basic attributes, interest tags, and priority ranking is obtained. The audience priority formula includes:

[0096] Priority_Score=Audience_Size×0.3+Conversion_Rate×0.4+Competition_Index×0.3;

[0097] wherein Priority_Score is a priority score, Audience_Size is an audience size, Conversion_Rate is a conversion probability, and Competition_Index is a competition intensity.

[0098] The content direction intelligent decision-making agent is used to determine a short video content direction suggestion scheme according to obtained popular data and audience portraits. The steps include content trend analysis, specifically including: collecting popular content data of each target platform through network crawler technology, covering multiple dimensions such as view count, like count, share count, and comment count; using the ARIMA model in time series analysis to model the collected historical data, identifying the change rule of content heat and predicting the trend in the future period; using the K-means clustering algorithm to automatically classify mass content topics and aggregate to form topic groups with similar characteristics; calculating the heat score of each topic based on the preset formula "Heat_Score=View_Count×0.4+Like_Count×0.3+Share_Count×0.2+Comment_Count×0.1", wherein Heat_Score is the heat score, View_Count is the view count, Like_Count is the like count, Share_Count is the share count, and Comment_Count is the comment count, to quantitatively evaluate the attention degree of the content. The competition analysis algorithm specifically includes: using a text similarity algorithm (such as cosine similarity) to extract and compare the features of the content published by competitors, analyze their core content strategies, topic preferences, and forms; by comparing and analyzing the business characteristics and content characteristics of competitors, identifying content gaps and differentiation opportunities in the market that have not been fully covered; calculating the value score of each differentiation opportunity based on the formula "Opportunity_Score=Market_Gap×0.5+Trend_Alignment×0.3+Resource_Match×0.2", wherein Market_Gap is the market gap, Trend_Alignment is the matching degree with the trend, and Resource_Match is the adaptation degree with the self-resource, to evaluate the feasibility and potential value of each opportunity. Content direction recommendation, specifically including: based on the business core characteristics, target audience preference labels, and market opportunities identified in the above steps, building a multi-dimensional decision tree model to preliminarily screen potential content directions; using collaborative filtering algorithm combined with historical success case data to accurately recommend and sort the screened content directions, generating a candidate direction list with the highest adaptation degree; calculating the confidence of each recommended direction by the formula "Confidence=Historical_Success_Rate×0.6+Algorithm_Score×0.4", wherein Confidence is the confidence, Historical_Success_Rate is the historical success rate of similar directions, and Algorithm_Score is the algorithm comprehensive score, finally outputting the content direction suggestion with high confidence and execution priority.

[0099] The positioning label generation agent is used to determine the short video positioning label according to the industry high-frequency words in the task data. The steps include: label candidate set generation, specifically including: performing word segmentation processing on the task description text in the input task data, and counting the frequency of each word, and extracting the industry high-frequency words as the initial label candidate words; using the Word2Vec semantic vector model to perform semantic expansion on the initial label candidate words to generate expanded words with associated meanings, and enriching the candidate label range; matching the above words with the pre-constructed industry label library, importing general labels and sub-field labels in the industry, and supplementing the candidate labels of professional dimensions; calculating the comprehensive score of each candidate label based on the preset formula "Tag_Score = Frequency × 0.4 + Relevance × 0.4 + Uniqueness × 0.2", wherein Tag_Score is the comprehensive score of the candidate label, Frequency is the frequency of the word in the business description, Relevance is the semantic relevance to the core features of the business, and Uniqueness is the uniqueness of the label (to avoid excessive repetition with general labels), to form a preliminary label candidate set. Label optimization and screening, specifically including: using the edit distance algorithm to calculate the text similarity between candidate labels, and performing deduplication processing on labels with a similarity exceeding a preset threshold, and retaining labels with more accurate semantics; calculating the weight value of each label in the business description and industry context based on the TF-IDF algorithm, and sorting the importance from high to low according to the weight; using a genetic algorithm to optimize the combination of labels, with the goal of "maximizing coverage + minimizing redundancy", and iteratively selecting a complementary label combination; according to the business scenario requirements and the label comprehensive score, finally selecting the Top-K (K is a configurable parameter) labels that best represent the core features of the business, with industry recognition and uniqueness, to form a final label set.

[0100] Exemplarily, the account positioning module further includes a creator background depth analysis algorithm and a large model integrated optimization scheme. The steps of the creator background depth analysis algorithm include: background information structuring, specifically including: using a named entity recognition (NER) algorithm to extract key information from the personal background description text provided by the creator, and extracting core elements including work experience, skill expertise, industry field, and educational background; sorting the extracted experience information in chronological order, and constructing a complete creator career development timeline to clearly present the growth trajectory; converting unstructured skill descriptions into standardized labels based on a preset skill system to achieve unified representation of skill information; based on the above structured data, constructing a creator feature vector:

[0101] Creator_Vector=[Experience_Years,Skill_Level,Industry_Background,Education_Level], where Creator_Vector is the creator feature vector, Experience_Years is the years of experience, Skill_Level is the skill level, Industry_Background is the industry background classification value, and Education_Level is the education level quantification value, completing the digital conversion of background information. Advantage capability identification, specifically including: calculating "Skill_Match=Σ(Skill_i×Industry_Relevance_i)" to obtain the matching degree of the creator's skills and the target industry, where Skill_Match is the matching degree, Skill_i is the proficiency score of the i-th skill, and Industry_Relevance_i is the relevance weight of the skill in the target industry; evaluating the experience value based on the formula "Experience_Value=Years×Depth_Factor×Relevance_Factor", where Experience_Value is the experience value, Years is the years of experience in the corresponding field, Depth_Factor is the experience depth coefficient, and Relevance_Factor is the relevance coefficient to the current business; comparing and analyzing the creator's skill and experience data with the industry average level and competitor data to identify their unique competitive advantages; calculating the comprehensive capability score through "Capability_Score=Technical_Skill×0.4+Experience×0.3+Creativity×0.3", where Capability_Score is the comprehensive capability score, Technical_Skill is the technical skill score, Experience is the experience score, and Creativity is the innovation capability score, quantifying the core capability level of the creator.The personalized positioning strategy specifically includes: based on the matching calculation of the creator feature vector and the preset character setting type model, screening out the character setting positioning type most suitable for the creator background and ability; combining the language style, skill characteristics and target audience preference of the creator, recommending the adapted content performance style (such as professional and rigorous type, relaxed and humorous type, etc.); generating a differentiated strategy quantitative evaluation result through the formula "Differentiation_Strategy=Unique_Advantage x Market_Gap x Execution_Ability", wherein Differentiation_Strategy is the evaluation result, Unique_Advantage is a unique advantage coefficient, Market_Gap is market blankness, and Execution_Ability is an execution ability score, and finally outputting a personalized positioning scheme including character setting positioning, content style and competition strategy.

[0102] The large model integration optimization scheme includes a three-layer structure: the first layer performs basic preprocessing, feature extraction, and structured conversion on the input data through traditional algorithms (such as statistical analysis, rule matching, etc.), to ensure the standardization and integrity of the data; the second layer introduces a large model (such as a natural language processing model, a multi-modal understanding model, etc.) to perform deep semantic understanding, context association analysis, and creative content generation on the preliminarily processed data, to improve the system's processing capability for complex information; the third layer performs compliance checking, logical consistency verification, and optimization adjustment on the results output by the large model through a rule engine, to ensure that the output results meet the requirements of the business scenario and industry standards, forming a three-layer collaborative processing mechanism of "basic processing - deep understanding - result optimization". The large model application strategy specifically includes: using pre-trained language models such as BERT and GPT to perform semantic enhancement understanding on unstructured text (such as business descriptions, user comments, etc.), to improve the analysis accuracy of ambiguous text and implicit intent; using the generation capability of the large model, combined with user business characteristics and target audience preferences, to automatically generate personalized account positioning suggestions, content creative schemes, and other non-standardized outputs; constructing a multi-modal analysis framework to correlate text information with multimedia data such as images and videos, to realize cross-modal content feature extraction and trend prediction; establishing a model iteration mechanism based on user feedback, to continuously fine-tune the large model parameters by collecting actual application effect data of the output results, to improve the model's adaptation capability to specific business scenarios. The advantages of the hybrid algorithm architecture are specifically reflected in: ensuring the accuracy of basic data processing and feature extraction through traditional algorithms, while improving the depth of understanding of complex semantics and implicit associations through large models, to realize a synergistic improvement in overall processing accuracy; adopting a resource allocation strategy of "traditional algorithm processing of routine tasks + large model focusing on key links" to reduce the overall computational resource consumption and response delay while ensuring the effectiveness of core functions, to achieve cost-effective control; using traditional algorithms to provide clear decision logic and feature weights to solve the "black box" problem of large models, to provide an interpretable decision path for the output results; processing high-concurrency and low-complexity routine tasks through traditional algorithms to ensure the system's real-time response capability, while focusing on processing low-frequency and high-complexity deep analysis tasks through large models, to achieve a balance between efficiency and effectiveness.

[0103] Optionally, the video creation module includes a script intelligent agent and a lip-sync digital human intelligent agent.

[0104] The script intelligent agent is configured to extract hot scripts of hot short videos, and generate innovative scripts of different styles based on the hot scripts.

[0105] The oral broadcast digital human intelligent agent is configured to analyze the copy semantics of the innovative copy, split the innovative copy into oral broadcast paragraphs, add SSML voice marks, expression or action instructions, generate a structured script that can be parsed by a digital human platform, determine a digital human image matched with the copy semantics and audience preferences, call a digital human platform API according to the copy semantics, the structured script and the digital human image, and generate an oral broadcast digital human video.

[0106] Specifically, the video creation module includes a script intelligent agent and a lip-sync digital human intelligent agent. The script intelligent agent is composed of a video script extraction intelligent agent, a hit video disassembly intelligent agent, a marketing intelligent agent, and a script imitation intelligent agent, to extract hot scripts of popular short videos, and generate innovative scripts of different styles based on the hot scripts. The construction process of the video script extraction intelligent agent is as follows: the intelligent agent takes a video link as input, integrates speech recognition technology and natural language processing algorithms to realize automatic transcription of voice content in the video and extraction of text scripts, and is built based on the LangGraph framework, specifically including definition of core state nodes such as downloading video, script extraction, and script output. Through the state conversion logic between the preset nodes (such as automatically triggering script extraction after video download is completed, and triggering output process after script extraction is completed), a complete intelligent agent workflow is formed, and the output structured script can be directly used as basic material for subsequent content creation. The construction process of the hit video disassembly intelligent agent is as follows: the intelligent agent focuses on the analysis of script features of hit short videos, aiming to extract reusable creation rules and skills, and is built based on the LangGraph framework, with nodes such as downloading video, script extraction, model analysis, and result output. Through the graph state management mechanism, the nodes are cooperatively linked (such as automatically activating the model analysis node after script extraction is completed), to complete the whole-process logic configuration of the intelligent agent. The construction process of the marketing intelligent agent is as follows: the intelligent agent uses the DeepSeek model to generate prompt words suitable for specific marketing scenarios, and builds an intelligent agent workflow based on the LangGraph framework. Through the configuration of style definition nodes, prompt word generation nodes, and script creation nodes, the intelligent agent realizes parameterized setting of target styles (such as "Uncle Jinguang" style). The style definition node receives user input of style feature description, the prompt word generation node generates accurate model input instructions based on the style features, and the script creation node calls a large language model to generate marketing scripts that meet the style requirements. The construction process of the script imitation intelligent agent is as follows: the intelligent agent takes the historical scripts of a target account as input, extracts language features (such as sentence structure, word preference, and emotional tendency) through a text style analysis algorithm, and combs a closed-loop workflow of "analysis-imitation-optimization" based on the LangGraph framework. The intelligent agent is set with style analysis nodes, script generation nodes, and optimization iteration nodes, and through the configuration of triggering conditions between the nodes (such as automatically entering the optimization loop after initial imitation is completed), the intelligent agent realizes autonomous iteration, and finally outputs new scripts that meet the style of the target account.

[0107] The mouth broadcast digital human intelligent agent successfully builds a full-automatic generation link from "text script input" to "digital human mouth broadcast video output" through deep integration of the intelligent agent system and the digital human anchor creation platform. The core operation logic is as follows: first, the innovative script is analyzed semantically, disassembled into paragraphs suitable for mouth broadcasting, and SSML voice markers, expressions, and action instructions are added to generate a structured script that can be parsed by the digital human platform; second, according to the script semantics and audience preferences, the digital human image is accurately matched; finally, combined with the script semantics, structured script, and selected digital human image, the digital human platform API is called to generate the mouth broadcast video. The mouth broadcast digital human intelligent agent is built through three core steps: first, filter interface plugins suitable for different digital human platforms to complete the preparation of script materials and digital human image resources; second, based on the LangGraph framework, build the intelligent agent workflow, set up key state nodes such as script analysis, image matching, voice synthesis, and video generation, and define the triggering logic and data flow rules between nodes; third, after the intelligent agent is created, it can automatically receive text script input, complete digital human matching, voice synthesis, and video rendering, and finally output a mouth broadcast video that can be directly published, effectively solving the core pain points of creators' low willingness to appear on screen and high shooting costs. The full-automatic generation link includes script intelligent analysis and preprocessing, digital human image intelligent matching, voice intelligent matching and synthesis, workflow automation arrangement, and platform integration and API calling. The steps of script intelligent analysis and preprocessing are as follows:

[0108] A large language model is used to perform semantic deep analysis on the input script, identify the emotional tendency (such as positive, neutral, and negative), tone style (such as serious, lively, and friendly), and professional level (such as academic and popular) of the script; a text classification algorithm is used to determine the type of the script, distinguishing between teaching, marketing, entertainment, and professional commentary scenarios; a keyword extraction algorithm (such as TF-IDF) is used to extract industry-specific words, emotional feature words, and style identifier words from the script; natural language processing techniques are used to analyze the language characteristics of the script to determine the appropriate speed range, inter-sentence pause rhythm, and emphasis position, providing basic parameters for subsequent voice synthesis; based on semantic logic, the long script is segmented into independent paragraphs suitable for mouth broadcasting, with each paragraph length adapted to the digital human's single sentence expression needs; SSML (Speech Synthesis Markup Language) markers are added to the segmented script, including voice control instructions such as speed, tone, and pause; based on the script word count and preset speed parameters, the final video duration is calculated and estimated; based on the script sentiment analysis results, digital human expression control instructions (such as smiling, serious, and surprised) are generated and associated with the corresponding text paragraphs to achieve synchronization of emotions and expressions.

[0109] The steps of digital human image intelligent matching are as follows:

[0110] The digital human resources are classified and labeled in multiple dimensions according to age (young, middle-aged, and old), gender, professional attributes (such as teachers, sales, and experts), and style characteristics (such as formal, casual, and cartoon); each digital human is associated with a suitable scene tag to clarify its adapted content type (such as teaching videos and product promotion) and industry field (such as education, finance, and retail); a digital human image scoring system is established, and a comprehensive score is calculated through the formula "Appearance_Score = Professionalism x 0.4 + Affinity x 0.3 + Authority x 0.3", where Professionalism is the professionalism score, Affinity is the affinity score, and Authority is the authority score, forming a quantitative image evaluation standard; the content matching degree is calculated through "Content_Match = Σ(Keyword_i x Avatar_Relevance_i)", where Keyword_i is the weight value of the text keyword, and Avatar_Relevance_i is the relevance of the digital human to the keyword; the vector dot product algorithm is used to calculate "Style_Consistency = Text_Style_Vector • Avatar_Style_Vector" to evaluate the consistency of the text style vector and the digital human style vector; based on the formula "Audience_Fit = Target_Age_Match x 0.4 + Gender_Preference x 0.3 + Industry_Match x 0.3", the audience adaptation degree is calculated, where Target_Age_Match is the matching degree with the target audience age, Gender_Preference is the gender preference fit degree, and Industry_Match is the industry matching degree; the comprehensive matching score is calculated through "Final_Score = Content_Match x 0.4 + Style_Consistency x 0.35 + Audience_Fit x 0.25", and the digital human with the highest score is selected as the optimal choice; the mapping relationship between the text sentiment and the digital human expression is established, and the corresponding expression is dynamically triggered according to the emotional tendency of the text paragraph (such as triggering a smile expression for positive content); based on the key words and logical nodes in the text, adaptive gesture actions are automatically generated (such as triggering a gesture indication action when emphasizing data); through the eye control algorithm, the real anchor's eye movement is simulated to realize the eye contact effect with virtual audiences, enhancing the interaction of the video.

[0111] The steps of sound intelligent matching and synthesis are:

[0112] The voice library is classified by timbre, divided into categories such as sweet, magnetic, professional, friendly, and authoritative; the appropriate speech rate range (120-180 words per minute) is set according to the complexity of the content (such as the density of professional terms and the logical hierarchy); the pitch variation is controlled by the formula "Pitch_Range = Base_Pitch ± Emotion_Modifier", where Base_Pitch is the base pitch and Emotion_Modifier is the emotion-based pitch adjustment value; the quality of the synthesized voice is evaluated from three dimensions: clarity, naturalness, and emotional expression; high-quality voice resources are selected based on content type matching rules (such as teaching content matching clear and professional timbre, marketing content matching friendly and sweet timbre); the voice is matched with the digital human appearance age and temperament characteristics through feature comparison (such as young digital human matching young voice); the voice style is selected according to the brand positioning of the account (such as high-end, people-oriented, and technological); the voice matching comprehensive score is calculated by "Voice_Score = Content_Fit × 0.4 + Avatar_Harmony × 0.3 + Brand_Alignment × 0.3", where Content_Fit is the content adaptation degree, Avatar_Harmony is the coordination with the digital human image, and Brand_Alignment is the brand tone matching degree; according to the results of the text sentiment analysis, the corresponding emotional rendering parameters are added in the voice synthesis process (such as raising the pitch and lowering the speech rate when happy); appropriate pauses are added before and after the key information (such as the core selling point and the conclusion) to enhance the information transmission effect; important words and key data are highlighted through stress processing algorithms; post-processing of the synthesized voice is performed, including noise reduction, volume equalization, and sound quality enhancement, to improve the naturalness and audibility of the voice.

[0113] The steps of workflow automation orchestration are:

[0114] The text input node is configured to receive user input text and verify its format and content integrity. The analysis processing node is configured to perform semantic analysis, feature extraction, and structured processing tasks in parallel. The matching decision node is configured to evaluate digital human images and voices based on a pre-set algorithm and select the optimal combination. The synthesis generation node is configured to render and generate videos by calling the digital human platform API. The quality detection node is configured to automatically detect video quality indicators such as picture clarity and voice synchronization. When the digital human or voice matching score is below a pre-set threshold, the system automatically triggers a backup solution recommendation mechanism to provide sub-optimal matching results for user selection. In the event of API call failure or video generation timeout, an exponential backoff retry strategy is adopted, and a degradation plan is initiated (such as switching to a basic digital human model) after multiple failed retries. When the quality detection finds obvious defects in the video (such as audio-visual asynchronization), the system automatically triggers a re-generation process or pushes a manual intervention notification when multiple repairs are ineffective. The platform data interface is connected to collect video play, interaction rate, and completion rate indicators. Based on the effectiveness data, a feedback model is built to dynamically adjust the weight parameters in the matching algorithm (such as increasing the matching priority of high interaction rate combinations). By recording user preferences for digital humans and voices, an individualized recommendation model is trained to make subsequent matching results more in line with user needs.

[0115] The steps of platform integration and API calling are as follows:

[0116] A unified API interface layer is designed to encapsulate the underlying calling logic of different digital human platforms, standardizing the interface for upper-layer applications. A parameter conversion mechanism is established to automatically convert standardized parameters (such as text features and image requirements) within the system into the private parameter formats required by each platform. Through interface calling optimization strategies, load balancing (avoiding single-platform request overload) and concurrency control (adjusting request frequency according to platform limitations) are achieved. A platform state monitoring module is deployed to monitor the availability and service quality of each digital human platform in real-time through heartbeat detection and response time statistics. Based on the current load, historical response speed, and calling cost of the platform, an intelligent scheduling algorithm is used to select the optimal service platform. A failover mechanism is established to automatically switch to a backup platform within 10 seconds when the primary platform experiences service anomalies, ensuring the continuity of the video generation process.

[0117] Optionally, as shown in Figure 1 , the function layer further includes a data analysis and optimization module for analyzing short video key data indicators after the video is released by the editing and publishing module, obtaining short video success and failure reasons.

[0118] Specifically, the data analysis and optimization module comprehensively collects key data indicators of short videos published by the editing and publishing module in the aspects of dissemination, interaction, conversion, etc., conducts deep mining and intelligent analysis on the data, identifies successful experiences and shortcomings of content operation, generates visual reports to intuitively present the analysis results, and outputs targeted content iteration suggestions and optimization strategies to provide data-driven operation decision support for creators. The index collection step includes:

[0119] Connect short video platform open API, user behavior tracking system, third party statistical tools, etc. to realize the comprehensive collection of video playback data, user interaction data, conversion behavior data and account operation data, establish a standardized data source processing mechanism, unify the format, field definition and time granularity of different source data, eliminate data islands, and form a complete data source system; build playback dimension indicators, including play count (cumulative play count), play completion rate (percentage of users who watched the video completely), replay rate (percentage of users who watched the same video multiple times), and sharing rate (percentage of users who shared the video); build interaction dimension indicators, including like rate (percentage of users who liked the video), comment rate (percentage of users who commented), forward rate (percentage of users who forwarded the video), and collection rate (percentage of users who collected the video); build conversion dimension indicators, including attention conversion rate (percentage of users who followed the account after watching the video), private message conversion rate (percentage of users who sent private messages after watching the video), and lead conversion rate (percentage of users who jumped to external links through the video); build content dimension indicators, including video duration, publishing time, topic label, content type, etc.; build user dimension indicators, including fan growth rate, user portrait features (age, gender, region, etc.), user activity (average daily interaction times), and user retention rate (percentage of users who interacted continuously for multiple days); deploy Kafka message queue in the data collection layer of the data analysis and optimization module to realize the streaming reception and temporary storage of real-time user behavior data and video playback data, supporting high-concurrency writing of tens of thousands of data per second; use ClickHouse time series database in the data storage layer of the data analysis and optimization module to efficiently store and compress massive time series indicator data, meeting the long-term retention needs of historical data; use SparkStreaming distributed computing framework in the data processing layer of the data analysis and optimization module to perform real-time cleaning (de-duplication, missing value completion), format conversion and indicator calculation (such as real-time play count accumulation and interaction rate instantaneous calculation) on the collected data; use Redis cache hot data (such as the play count of the top 10 videos of the day and real-time fan count) in the data service layer of the data analysis and optimization module to provide millisecond-level query response for upper-layer applications, improving the interactive experience of the report system.

[0120] The steps of AI large model deep mining and intelligent analysis of data include:

[0121] ARIMA time series analysis model is used to fit the trend and periodicity of core indicators (such as play count, fan count), identify long-term growth trends and periodic fluctuations; Pearson correlation coefficient is used to calculate the correlation strength between indicators (such as analyzing the positive correlation between play completion rate and like rate), and to explore the linkage behind the indicators; K-means clustering algorithm is used to automatically classify content (by play count, interaction rate, etc.) and users (by behavior preference, interaction frequency, etc.), identify high-value content types and core user groups; Isolation Forest algorithm is used to build an anomaly detection model to quickly identify abnormal fluctuations (such as sudden drop in play count, sudden rise in conversion rate) and potential opportunities (such as the interaction rate of a certain niche content is abnormally higher than the average value). Train the Random Forest model, based on the basic attributes of the video (duration, release time, tags) and the performance of similar historical content, predict the play count interval and potential spread range of newly released videos; use LSTM long short-term memory neural network to combine user historical interaction behavior sequences to predict the interaction probability (such as like, comment probability) of users to different types of content; use Prophet prediction model to analyze the hot trend of content field and predict the outbreak time of potential blockbuster topics; build XGBoost gradient boosting model based on content features, user portrait and historical conversion data to predict the attention conversion and lead conversion effect of different content. Through feature importance analysis (such as Gini coefficient of random forest), identify the key factors that affect the success of the video (such as the decisive role of "3-second play rate at the beginning" on the overall play count); compare and analyze low-performing content (such as videos with play count lower than the average value) and high-performing content from the perspectives of content form, release time, user matching degree, etc. to diagnose the root cause of failure; based on data gap analysis (such as insufficient content supply for a certain user group), explore new content opportunities and market gaps; combine historical optimal case data and industry benchmark values to generate quantifiable optimization suggestions (such as "controlling video length within 60 seconds can increase play completion rate by 20%").

[0122] The steps of generating visual reports include:

[0123] Dynamic dashboard design, including: developing real-time monitoring panels to display the trend and current value of core indicators (playback, interaction rate, and fan count) in real time through line charts, digital cards, etc., supporting automatic refresh; designing multi-dimensional comparative analysis charts, including index comparison column charts and radar charts for different time periods (such as this week vs. last week), different content types (such as tutorial vs. review), and different accounts (such as this account vs. benchmark account); generating user behavior heat maps (showing user interaction density at different times of the video) and content performance heat maps (showing content performance distribution by label, duration, etc.); building conversion funnel analysis charts to visually present the complete user path from "exposure to playback to interaction to conversion" and the drop-off rate at each stage. Support data drilling function, users can drill from the overall index (such as total account playback) to the specific video detail data (such as single video playback source, user portrait); provide multi-dimensional filtering function, can flexibly filter data according to time range (such as the past 7 days), content type (such as marketing), topic label (such as # food tutorial) and other conditions; develop custom report module, allow users to independently select analysis dimensions (such as by release time + content type), index combination (such as playback + attention conversion rate) and chart type, generate personalized analysis report; support report export function, can export analysis results to PDF, Excel and other formats, convenient for offline viewing and secondary processing. Responsive design is adopted to make the report system adaptive to mobile devices of different screen sizes (such as mobile phones and tablets), ensuring complete chart display and smooth interaction; the mobile interface prioritizes the display of core indicator cards (such as daily playback and fan growth), highlighting key information; set up data push mechanism, when the core indicators change significantly (such as a 50% increase in playback) or abnormally (such as a 30% drop in conversion rate), real-time push reminders are sent through the mobile terminal to ensure that users can keep abreast of account dynamics in a timely manner.

[0124] By way of example, the data analysis and optimization module also includes AI large model intelligent analysis, the steps including:

[0125] The BERT pre-training language model is used for semantic analysis of video scripts, titles, comments and other text content, to identify the emotional tendency, core theme and potential intent of the text; computer vision technology is used to analyze video image elements, to extract objects, scenes and visual styles in the image; audio feature extraction algorithms are used to analyze the style of the video background music and the emotional color of the voice tone, to evaluate its impact on user experience; a multi-modal fusion analysis model is constructed to integrate text semantics, visual elements and audio features for cross-modal correlation analysis. The common attributes of popular videos with high play counts and high interaction rates are extracted and summarized; a content preference model of target users is constructed based on user interaction data and content features; a replicable creation methodology is extracted from historical successful cases; the unique success factors that distinguish the account from competitors are identified by comparing the content features of the account and those of the competitor accounts. Based on the correlation analysis of content features and performance data, specific content improvement directions are provided; the best release time, release frequency and platform combination strategy are recommended based on the distribution of user active time and historical release data; specific strategies for fan interaction and community operation are provided; based on user portraits and behavior data, commercial monetization suggestions are provided.

[0126] By way of example, the present embodiment provides three specific application examples to assist in describing the functions of the data analysis and optimization module:

[0127] Example 1: Data analysis and optimization for a food blogger, specifically including: for the "Xiaowang's Kitchen" account (500,000 fans, average play count of 80,000 in the last 30 days, interaction rate of 3.2%, focusing on home cooking tutorial), through data analysis, it is found that: videos published between 7-9 pm have a 40% higher play count than other time periods; videos containing the "10-minute quick recipe" tag have a 60% higher play count; 25-35 year-old working women account for 70% of the fans and prefer simple and easy-to-learn recipes; among similar bloggers, videos with cost annotations have a higher interaction rate. Based on the above findings, the data analysis and optimization module outputs optimization suggestions: adjust the release time to 19:30 every day; increase the "quick recipe for office workers" series, and annotate the preparation time and cost for each dish; add keywords such as "10 minutes" and "lazy person's must-have" to the title; actively respond to users' questions in the comments section. The optimization effect is significant: the average play count is increased to 120,000 (up 50%), the interaction rate is increased to 4.8% (up 50%), the fan growth rate is increased by 35%, and commercial cooperation inquiries are increased by 80%.

[0128] Example 2: Knowledge payment account conversion optimization, including: for the "Career Growth Academy" account (30 million fans, course conversion rate 1.2%, single price 299 yuan, focusing on career skill sharing), data analysis shows that: the jump rate from video to course page is only 15%; the conversion rate of case-based videos is 3 times higher than that of theoretical videos; the average user stay time on the course introduction page is only 30 seconds; the conversion rate of 199 yuan courses is 80% higher than that of 299 yuan courses. The data analysis and optimization module accordingly suggests: increase real case sharing, reduce pure theory explanation; add "free trial" link in the video to improve jump rate; redesign course introduction page to highlight student success stories; launch a 199 yuan introductory course as a lead generation product. After optimization, the course conversion rate increased to 2.8% (increased by 133%), the jump rate increased to 35% (increased by 133%), monthly income increased from 180,000 to 420,000 (increased by 133%), and user satisfaction score increased from 4.2 to 4.7.

[0129] Example 3: E-commerce account ROI optimization, including: for the "Fashion Dressing Assistant" account (80 million fans, ROI 1:3.5, monthly sales 150 million, mainly female clothing and accessories), data analysis found that: the conversion rate of products priced between 200-500 yuan is the highest; the sales of videos published on weekends is 60% higher than that on weekdays; users in second and third-tier cities account for 65%, with higher price sensitivity; real person try-on videos have a 5 times higher conversion rate than picture displays. The data analysis and optimization module suggests: focus on promoting 200-500 yuan price range bestsellers; schedule main content for bringing goods on Friday night and weekends; increase the proportion of real person try-on and matching tutorials; negotiate exclusive coupons with brand partners to improve price competitiveness. After optimization, the ROI of bringing goods increased to 1:5.2 (up 49%), monthly sales increased to 220 million (up 47%), user repurchase rate increased by 40%, and brand cooperation single price increased by 30%.

[0130] Optionally, as shown in Figure 1 the function layer further includes a data review intelligent agent, configured to generate a review report according to the short video key data indicators.

[0131] Specifically, the review intelligent agent aims to solve the core pain points of creators "difficulty in data interpretation and ambiguity in optimization direction", by building a standardized data review workflow, integrating multi-dimensional operation data and AI big model analysis capabilities, realizing quantitative evaluation of content quality, deep mining of audience preferences, and outputting practical creation strategy optimization solutions; after the intelligent agent is deployed, it can be widely applied to single video special review, periodic account operation review and new content strategy optimization test scenarios, forming a "data collection-analysis-review-optimization" closed loop support.

[0132] Optionally, as shown inFigure 1 As shown, the short video full-process customer acquisition system based on agent cooperation further includes a support layer, the support layer includes a knowledge base and a model base, the knowledge base stores industry rules and competitor data, and the model base integrates NLP models, ASR models and digital human models.

[0133] Specifically, the support layer provides knowledge and algorithm support for each functional module of the system, including a structured knowledge base and a model base. The knowledge base is used to systematically store various types of knowledge information related to short video full-process customer acquisition, including industry rules (such as platform content publishing specifications, short video creation methodologies), competitor data (such as benchmark account operation data, industry blockbuster content features), business-specific knowledge (such as audience preference tags for different industries, content direction templates), etc., forming a reusable knowledge resource pool. The model base is used to centrally integrate various AI models that support system functions, including natural language processing (NLP) models (such as BERT, GPT series models for copywriting semantic analysis and content generation), automatic speech recognition (ASR) models (for video speech transcription and copywriting extraction), digital human driving models (for voice-over digital human image control and action generation), computer vision models (for video image feature analysis), etc., providing algorithm capability support for agent scheduling and content creation. It can carry out knowledge graph construction work, through entity extraction, relationship definition, attribute annotation, etc. Non-structured and semi-structured knowledge in the knowledge base is converted into structured knowledge graph, realizing the association storage and efficient retrieval of knowledge, and improving the accuracy of knowledge calling; at the same time, it can promote the lightweight deployment of models, through model compression (such as pruning, quantization), model distillation, etc. Technical means to reduce the computational resource consumption and running delay of AI models, combined with model service encapsulation (such as providing model calling capability through API interface), to ensure that each functional module can quickly call the required model, ultimately realizing the dual improvement of system running efficiency and knowledge utilization level, providing bottom-layer technical support for the stable and efficient operation of the short video full-process customer acquisition system.

[0134] Optionally, as shown in Figure 1 The short video full-process customer acquisition system based on agent cooperation further includes a storage layer, the storage layer uses a distributed database to store various data in the input layer, the scheduling layer and the functional layer.

[0135] Specifically, the storage layer can realize unified management of various types of data generated during system operation, specifically including centralized storage of task data input by users (such as enterprise or personal business description, product parameters, target audience information, etc.), intermediate result data generated during execution of each function module (such as sub-task decomposition results, content feature extraction data, temporary analysis reports, etc.), and structured achievement data output finally (such as account positioning reports, finished video metadata, post-mortem analysis reports, etc.), while supporting export of the above data to common formats such as PDF, Excel, CSV, etc. according to user needs, meeting the needs of offline viewing, secondary processing and external system docking; the technical implementation points are to design using a distributed database architecture, to ensure data security and reliability in high-concurrency access scenarios through data sharding storage and multi-copy backup mechanisms, to realize cross-node data sharing through a distributed file system, to ensure that each intelligent agent module (such as an account positioning intelligent agent and a data analysis intelligent agent) in the system can efficiently call the required data, and to improve data read / write efficiency through database index optimization and query caching mechanisms, ultimately meeting the data usage needs of multiple scenarios such as "user data management, intermediate process tracking, and achievement report archiving" in the short video full-process customer acquisition system, and providing data storage support for the overall stable operation of the system.

[0136] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the above-mentioned embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), etc. In this application, the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., they can be located in one place or distributed on multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiment of the present application according to actual needs. In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0137] Although the present application is disclosed as above, the protection scope of the present application is not limited to this. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present application, and these changes and modifications will fall within the protection scope of the present application.

Claims

1. A short video end-to-end customer acquisition system based on intelligent agent collaboration, characterized in that, It includes an input layer, a scheduling layer, and a functional layer; The input layer is used to receive task data input by the user; The scheduling layer employs an intelligent agent scheduling engine, which is used to determine at least one subtask and the execution order of the subtasks based on the task data, distribute the subtasks based on a publish-subscribe model, and allocate computing resources based on a weighted round-robin algorithm. The functional layer includes an account positioning module, a benchmark account decomposition module, an account basic construction module, a topic selection library construction module, a video creation module, and an editing and publishing module, all independently encapsulated in the form of intelligent agents. Based on the execution order of the subtasks and the computing resources, the account positioning module, the benchmark account decomposition module, the account basic construction module, the topic selection library construction module, the video creation module, and the editing and publishing module execute the corresponding subtasks. The intelligent agent scheduling engine includes a task dispatcher, a resource manager, and an exception handler; The task dispatcher is used to parse the subtasks of the task data, and uses an asynchronous pipeline scheduling algorithm to determine the subtask priority and subtask dependency of the subtask. Based on the subtask dependency, it determines the execution order of the subtasks, and distributes the subtasks to each agent in the functional layer based on the publish-subscribe pattern. The resource manager is used to monitor resource usage and, based on the resource usage, allocate computing resources according to the subtask priority. The exception handler is based on a three-level exception handling mechanism to handle task exceptions; The asynchronous pipeline scheduling algorithm is used to determine the subtask priority and subtask dependency relationship of the subtask, including: Based on the directed acyclic graph of the task data, identify the prerequisite dependencies of each subtask and determine the subtask dependencies; The priority of the subtask is determined using a priority formula, which includes: Priority = α×Urgency + β×(1 / ResourceCost) + γ×(1 / EstimatedTime); Where Priority is the priority of the subtask, Urgency is the urgency of the subtask, ResourceCost is the resource cost required to complete the subtask, EstimatedTime is the estimated time to complete the subtask, and α, β and γ are all weighting coefficients.

2. The short video end-to-end customer acquisition system based on intelligent agent collaboration according to claim 1, characterized in that, The three-level exception handling mechanism for handling task exceptions includes: When the subtask fails to execute, a smart retry strategy matrix is ​​used for local retry; When the local retry fails, the failed subtask is reassigned to a backup agent via the task dispatcher. When the backup agent fails to execute, a degradation scheme is triggered, and a notification requiring manual intervention is pushed.

3. The short video end-to-end customer acquisition system based on intelligent agent collaboration according to claim 1, characterized in that, The account positioning module includes a target audience identification intelligent agent, a content direction intelligent decision-making intelligent agent, and a positioning tag generation intelligent agent; The target audience identification agent is used to determine the audience profile based on the task data, wherein the audience profile includes basic attributes, interest tags, and priority ranking. The intelligent decision-making agent for content direction is used to determine a short video content direction suggestion scheme based on the acquired popular data and the audience profile; The location tag generating agent is used to determine short video location tags based on high-frequency industry terms in the task data.

4. The short video end-to-end customer acquisition system based on intelligent agent collaboration according to claim 1, characterized in that, The video creation module includes a text-based intelligent agent and a voice-over digital human intelligent agent. The copywriting AI is used to extract popular copywriting from popular short videos and generate innovative copywriting in different styles based on the popular copywriting. The spoken digital human agent is used to analyze the semantics of the innovative copy, break down the innovative copy into spoken segments, add SSML voice tags, facial expressions or action commands, generate a structured script that can be parsed by the digital human platform, determine a digital human image that matches the semantics of the copy and audience preferences, and, based on the semantics of the copy, the structured script and the digital human image, call the digital human platform API to generate a spoken digital human video.

5. The short video end-to-end customer acquisition system based on intelligent agent collaboration according to claim 1, characterized in that, The functional layer also includes a data analysis and optimization module, which is used to analyze the key data indicators of the short video after the video is published by the editing and publishing module, and to obtain the reasons for the success and failure of the short video.

6. The short video end-to-end customer acquisition system based on intelligent agent collaboration according to claim 5, characterized in that, The functional layer also includes a data review agent, which is used to generate a review report based on the key data indicators of the short video.

7. The short video end-to-end customer acquisition system based on intelligent agent collaboration according to claim 1, characterized in that, It also includes a support layer, which includes a knowledge base and a model library. The knowledge base stores industry rules and competitor data, and the model library integrates NLP models, ASR models, and digital human models.

8. The short video end-to-end customer acquisition system based on intelligent agent collaboration according to claim 1, characterized in that, It also includes a storage layer, which uses a distributed database to store various data from the input layer, the scheduling layer, and the functional layer.

Images