A multitask parallel scheduling method and device

Abstract

The present disclosure relates to a multi-task parallel scheduling method and device. The method comprises: obtaining e-commerce product information and a pre-processing procedure to be executed by a main state machine, wherein the pre-processing procedure is used to represent the sequence of tasks executed by the main state machine; based on the e-commerce product information and the pre-processing procedure, obtaining a plurality of sub-state sets; based on a crowd sub-state set in the sub-state set, constructing a first preset number of text sub-tasks and a second preset number of image sub-tasks; calling a large language model to execute the text sub-tasks to obtain a first execution result, and calling an image generation model to execute the image sub-tasks to obtain a second execution result. The present disclosure is based on the pre-processing procedure, so that the main state machine executes fixed business tasks using a fixed pre-processing procedure, does not rely on a large model for splitting decision, makes the stage execution order fixed and the business logic strictly controllable, and avoids the problems of unreliable task splitting, stage omission and task redundancy caused by large model autonomous decision.

Description

Technical Field

[0001] This disclosure relates to the fields of artificial intelligence and e-commerce technology, specifically to a multi-task parallel scheduling method and apparatus. Background Technology

[0002] In current intelligent task collaboration architectures represented by general large-model programming agents, a master agent and sub-agent collaboration mode is commonly adopted. The master agent decomposes and dispatches tasks, while the sub-agents have independent contexts and return result summaries after execution.

[0003] This architecture has the following problems in e-commerce optimization scenarios: Task splitting relies on the autonomous decision-making of the large model. Since the large model is essentially a probabilistic prediction machine, it cannot guarantee that the splitting results are consistent with fixed stages such as audience analysis, comment mining, visual hooks, and comprehensive reports. At the same time, the large model cannot execute strict business logic and can only guess what to do next based on probability. Therefore, it may randomly skip stages that it considers "irrelevant", resulting in missed stages or because the large model tends to "over-generate" to cover all possibilities when dealing with uncertain tasks, generating redundant subtasks. Summary of the Invention

[0004] In view of this, this disclosure provides a multi-task parallel scheduling method and apparatus to solve the problem that the task splitting in the prior art relies on the autonomous decision-making of a large model. Since the large model is a probabilistic prediction machine, it cannot guarantee consistency with fixed business stages, and is prone to omission of stages or generation of redundant sub-tasks.

[0005] Firstly, this disclosure provides a multi-task parallel scheduling method, which includes: The process involves obtaining e-commerce product information and pre-processing procedures to be executed by the main state machine. The pre-processing procedures are used to characterize the order in which the main state machine executes tasks. Based on e-commerce product information and pre-processing procedures, multiple sub-state sets are obtained; Based on the population sub-state set in the sub-state set, construct a first preset number of text sub-tasks and a second preset number of image sub-tasks; The large language model is called to perform the text subtask, and the first execution result is obtained. The image generation model is called to perform the image subtask, and the second execution result is obtained.

[0006] In one optional implementation, based on e-commerce product information and pre-processing procedures, multiple sub-state sets are obtained, including: The pre-processing flow is executed serially to determine the current pre-processing state. Based on the previous state and e-commerce product information, we obtain the first sub-state set, the second sub-state set, and the third sub-state set. Based on the previous state, the first sub-state set, and the e-commerce product information, the fourth sub-state set is obtained; Based on the previous state, the first sub-state set, the second sub-state set, the third sub-state set, and the fourth sub-state set, the fifth sub-state set is obtained.

[0007] In one optional implementation, based on the crowd sub-state set in the sub-state set, a first preset number of text sub-tasks and a second preset number of image sub-tasks are constructed, including: Get the set of sub-states of the crowd in the sub-state set; The total number of people and the number of image variants under each people are determined based on the set of sub-states of the population. Based on the total number of people and the number of image variants, text subtasks and image subtasks are determined.

[0008] In one optional implementation, a large language model is invoked to perform a text subtask to obtain a first execution result, and an image generation model is invoked to perform an image subtask to obtain a second execution result, including: Obtain the execution summary corresponding to each of the multiple sub-state sets; Based on the execution summary and text sharing basic information, the first context information of the text subtask is determined, and the text sub-state machine for processing the text subtask is obtained. Based on the text-shared basic information, the crowd labels in the crowd sub-state set, and the prompt words when generating corresponding image variants for each crowd label, the second context information of the image sub-task is obtained, and the image sub-state machine for processing the image sub-task is obtained. Send the text sub-state machine and the image sub-state machine to the asynchronous task queue. Generate the first prompt word based on the first context information, and call the large language model based on the first prompt word to obtain the first execution result. Generate the second prompt word based on the second context information, and call the image generation model based on the second prompt word to obtain the second execution result.

[0009] In one alternative implementation, the method further includes: The first and second execution results are aggregated to extract text summary information. The second execution result is grouped into image summary groups according to the total number of people, and multiple grouped image summaries are extracted. The text summary and grouped image summaries are assembled according to the display template format to obtain the page to be displayed.

[0010] In one alternative implementation, the method further includes: If the second execution result is a preset result, obtain the location information of the image variant corresponding to the preset result; Select a preset image from the preset placeholder image library to fill in the position information.

[0011] In one optional implementation, based on the preceding state and e-commerce product information, a first sub-state set, a second sub-state set, and a third sub-state set are obtained, including: Based on the previous state, the large language model is called to analyze the product description information in the e-commerce product information, obtain the user group classification corresponding to the e-commerce product information, and generate the first sub-state set. Based on the previous state, the large language model is called to mine the competitor review information in the e-commerce product information to obtain the review status of the e-commerce product information and generate the second sub-state set. Based on the preceding state, the large language model is invoked to analyze the product characteristic information in the e-commerce product information, obtain the copy title of the e-commerce product information, and generate a third sub-state set.

[0012] In one optional implementation, a fourth sub-state set is obtained based on the previous state, the first sub-state set, and the e-commerce product information, including: Based on the preceding state, the large language model is invoked to analyze the e-commerce product information and each group of people contained in the first sub-state set, to obtain the visual main image of each group of people and generate the fourth sub-state set.

[0013] In one optional implementation, a fifth sub-state set is obtained based on the previous state, the first sub-state set, the second sub-state set, the third sub-state set, and the fourth sub-state set, including: Based on the preceding states, the large language model is invoked to analyze the output results of the first, second, third, and fourth sub-state sets to obtain test metrics and generate the fifth sub-state set.

[0014] Secondly, this disclosure provides a multi-task parallel scheduling apparatus, the apparatus comprising: The first acquisition module is used to acquire e-commerce product information and the pre-processing process to be executed by the main state machine. The pre-processing process is used to characterize the order in which the main state machine executes tasks. The first module is used to obtain multiple sub-state sets based on e-commerce product information and pre-processing procedures; The construction module is used to construct a first preset number of text subtasks and a second preset number of image subtasks based on the crowd sub-state set in the sub-state set; The second module is used to call the large language model to perform the text subtask and obtain the first execution result, and to call the image generation model to perform the image subtask and obtain the second execution result.

[0015] Thirdly, this disclosure provides a computer device, including: a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the multi-task parallel scheduling method described in the first aspect or any corresponding embodiment.

[0016] Fourthly, this disclosure provides a computer-readable storage medium storing computer instructions for causing a computer to execute the multi-task parallel scheduling method described in the first aspect or any corresponding embodiment thereof.

[0017] Fifthly, this disclosure provides a computer program product, including computer instructions for causing a computer to execute the multi-task parallel scheduling method described in the first aspect or any corresponding embodiment thereof.

[0018] In this embodiment, by acquiring e-commerce product information and the pre-processing flow to be executed by the main state machine, the main state machine executes fixed business tasks using a fixed pre-processing flow. Based on the e-commerce product information and the pre-processing flow, multiple sub-state sets are obtained. Then, based on the crowd sub-state set in the sub-state set, a first preset number of text sub-tasks and a second preset number of image sub-tasks are constructed. This achieves the splitting of crowd rules into sub-tasks without relying on a large model for splitting decisions. This ensures that the task splitting results are certain, the stage execution order is fixed, and the business logic is strictly controllable. It avoids the technical problems of unreliable task splitting, stage omissions, and task redundancy caused by the autonomous decision-making of a large model. Then, the large language model is called to execute the text sub-tasks to obtain the first execution result, and the image generation model is called to execute the image sub-tasks to obtain the second execution result. Based on this, the parallel execution of text and image tasks is achieved, which greatly improves the overall efficiency of e-commerce product detail page optimization. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the specific embodiments of this disclosure or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating a multi-task parallel scheduling method according to an embodiment of the present disclosure; Figure 2 This is a complete flowchart of a multi-task parallel scheduling method according to an embodiment of the present disclosure; Figure 3 This is a structural block diagram of a multi-task parallel scheduling device according to an embodiment of the present disclosure; Figure 4 This is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present disclosure. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0022] It is understood that before using the technical solutions disclosed in the various embodiments of this disclosure, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this disclosure in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0023] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this disclosure, "a plurality of" means two or more, unless otherwise expressly specified.

[0024] In current intelligent task collaboration architectures represented by general large-model programming agents, a master agent and sub-agent collaboration mode is commonly adopted. The master agent decomposes and dispatches tasks, while the sub-agents have independent contexts and return result summaries after execution.

[0025] This architecture has the following problems in e-commerce optimization scenarios: First, task decomposition relies on the autonomous decision-making of the large model. Since the large model is essentially a probabilistic prediction machine, it cannot guarantee that the decomposition results are consistent with fixed stages such as audience analysis, comment mining, visual hooks, and comprehensive reports. At the same time, the large model cannot execute strict business logic and can only guess what to do next based on probability. Therefore, it may randomly skip stages that it considers "irrelevant", resulting in missed stages or because the large model tends to "over-generate" to cover all possibilities when dealing with uncertain tasks, generating redundant subtasks.

[0026] Secondly, the context management adopts a complete isolation strategy, which results in duplicate storage when multiple sub-agents need the same basic product information, and partial sharing cannot be achieved when differentiated configurations are required. Third, the summary aggregation only performs simple splicing and does not include failure handling logic. When a sub-agent fails to execute, the main agent cannot automatically downgrade, causing the overall task to be interrupted.

[0027] To address the aforementioned problems, according to an embodiment of this disclosure, a multi-task parallel scheduling method embodiment is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0028] This embodiment provides a multi-task parallel scheduling method, suitable for the collaborative scheduling of text analysis and image generation tasks in product detail page optimization. It can be used with controllers / computers / processors that have data processing and computing capabilities, such as... Figure 1 As shown, the specific steps of this method are as follows: Step S101: Obtain e-commerce product information and the pre-processing flow to be executed by the main state machine. The pre-processing flow is used to characterize the order in which the main state machine executes tasks.

[0029] Optionally, in this embodiment of the disclosure, the e-commerce product information includes product name, product category, core selling points, price range, competitor review data, product white background image, etc., all of which are basic input information necessary for optimizing the product details page.

[0030] For the server, it can obtain the aforementioned e-commerce product information in real time. For example, the server can obtain information / data input by the user to describe the product, such as the user inputting the product name "Ergonomic OfficeChair", the price range "199 to 299", and the core selling point "relieves back pain from prolonged sitting". The server can obtain competitor review data provided by the user in real time, and then call a large language model to mine the competitor reviews, extracting the three most frequent pain points (insufficient lumbar support, seat cushion sagging, non-adjustable armrests) and three positive points (smooth wheels, easy installation, breathable fabric), with each conclusion accompanied by a quote from the original review.

[0031] In this embodiment, it is necessary to obtain the pre-processing flow to be executed by the main state machine. This pre-processing flow is pre-defined, fixed, and immutable, used to characterize the execution order of tasks performed by the main state machine. For example, initializing the main state machine... ,in For analyzing the status of the population, To dig out the status of comments, For visual hook state, For the trust copy status, For A / B test status, To achieve the comprehensive report status, the main state machine schedules the execution flow in the following order: audience analysis → comment mining → visual hooks → trust copywriting → A / B testing → comprehensive report. Specifically, the execution process is as follows: → → → → → .

[0032] Therefore, the embodiments of this disclosure ensure that the task execution is carried out in a fixed order through a pre-processing process, without relying on AI autonomous decision-making.

[0033] Step S102: Based on e-commerce product information and pre-processing procedures, multiple sub-state sets are obtained.

[0034] Optionally, the main state machine can execute each preceding state according to the preceding processing flow (e.g., , Each preceding state independently calls the large language model to complete a single-dimensional analysis task, without performing multi-task mixed processing.

[0035] After each state is executed, the analysis results of that state are stored in the context pool to form a corresponding set of sub-state results, and finally, multiple sets of sub-states are formed.

[0036] In this embodiment of the disclosure, the main state machine executes the first 5 pre-states, and finally obtains 5 sub-state sets, which correspond to: crowd sub-state set, comment mining sub-state set, visual hook scheme sub-state set, trust copywriting sub-state set, and A / B testing scheme sub-state set.

[0037] Step S103: Based on the crowd sub-state set in the sub-state set, construct a first preset number of text sub-tasks and a second preset number of image sub-tasks.

[0038] Optionally, by The large language model is used to analyze the sub-state set of the output audience to obtain the target audience, for example, outputting 3 target audiences. k=3, representing remote working programmers, people undergoing back rehabilitation, and e-sports players, respectively. Each group includes structured information such as demographic characteristics, pain points, and purchase motivations.

[0039] Then, based on the number of the target population, a first preset number of text subtasks and a second preset number of image subtasks can be constructed. The first preset number can be 1, and the value of the second preset number is related to the number of the target population and the number of image variants set for each target population. The second preset number = number of population k × number of population image variants m.

[0040] Therefore, the embodiments of this disclosure complete the construction of text subtasks and image subtasks only through numerical calculations, which belongs to a fixed splitting rule and can avoid the problems of stage omissions and redundant tasks caused by the probabilistic decision-making of large models.

[0041] Step S104: Call the large language model to execute the text subtask and obtain the first execution result; call the image generation model to execute the image subtask and obtain the second execution result.

[0042] Optionally, the text subtask is performed by the large language model, with the input being a dedicated context constructed by the context manager, and the output being a comprehensive text analysis result (i.e., the first execution result). The image subtask is performed by the text-generated image model, with the input being visual hook scheme prompts and product reference images, and the output being the main product image (i.e., the second execution result).

[0043] In this embodiment, the text subtask and the image subtask have no data dependency and can be executed in parallel, which greatly reduces the overall time consumption and improves the optimization efficiency of e-commerce product detail pages.

[0044] In this embodiment, by acquiring e-commerce product information and the pre-processing flow to be executed by the main state machine, the main state machine executes fixed business tasks using a fixed pre-processing flow. Based on the e-commerce product information and the pre-processing flow, multiple sub-state sets are obtained. Then, based on the crowd sub-state set in the sub-state set, a first preset number of text sub-tasks and a second preset number of image sub-tasks are constructed. This achieves the splitting of crowd rules into sub-tasks without relying on a large model for splitting decisions. This ensures that the task splitting results are certain, the stage execution order is fixed, and the business logic is strictly controllable. It avoids the technical problems of unreliable task splitting, stage omissions, and task redundancy caused by the autonomous decision-making of a large model. Then, the large language model is called to execute the text sub-tasks to obtain the first execution result, and the image generation model is called to execute the image sub-tasks to obtain the second execution result. Based on this, the parallel execution of text and image tasks is achieved, which greatly improves the overall efficiency of e-commerce product detail page optimization.

[0045] As an optional embodiment, step S102 above includes: Step S1021: Execute the pre-processing flow serially to determine the current pre-processing state; Step S1022: Based on the previous state and e-commerce product information, obtain the first sub-state set, the second sub-state set, and the third sub-state set; Step S1023: Based on the previous state, the first sub-state set, and the e-commerce product information, obtain the fourth sub-state set; Step S1024: Based on the previous state, the first sub-state set, the second sub-state set, the third sub-state set, and the fourth sub-state set, the fifth sub-state set is obtained.

[0046] Optionally, the server controls the master state machine strictly according to → → → → The process is executed sequentially, and the next pre-processing state is only entered after each step is completed and the result is output.

[0047] Then, based on the current execution state and e-commerce product information, the first sub-state set, the second sub-state set, and the third sub-state set are obtained.

[0048] Among them, the first sub-state set: the corresponding sub-state set of the population, which is obtained by segmenting product information into user segments by the large language model.

[0049] The second sub-state set: corresponds to the comment mining sub-state set, which is obtained by performing high-frequency word and sentiment analysis on competitor comments using a large language model.

[0050] The third sub-state set: corresponds to the trust copywriting sub-state set, which is generated by the large language model based on product characteristics to produce titles and selling point copywriting.

[0051] Based on the previous state, the first sub-state set, and the e-commerce product information, the fourth sub-state set is obtained.

[0052] Among them, the fourth sub-state set: corresponds to the visual hook scheme sub-state set, which depends on the first sub-state set + e-commerce product information, and generates a visual grabbing scheme for a specific group of people by a large language model.

[0053] Based on the previous state, the first sub-state set, the second sub-state set, the third sub-state set, and the fourth sub-state set, the fifth sub-state set is obtained.

[0054] The fifth sub-state set corresponds to the A / B test scheme sub-state set. It requires the combined results of the first four sub-states to determine the test variables, sample size, and judgment indicators. The A / B test is the verification stage of the overall optimization scheme and relies on all the preceding analytical conclusions.

[0055] Further, step S1022 above includes: Step a1: Based on the previous state, call the large language model to analyze the product description information in the e-commerce product information, obtain the user group classification corresponding to the e-commerce product information, and generate the first sub-state set. Step a2: Based on the previous state, call the large language model to mine the competitor review information in the e-commerce product information, obtain the review status of the e-commerce product information, and generate the second sub-state set; Step a3: Based on the previous state, call the large language model to analyze the product characteristic information in the e-commerce product information, obtain the copy title of the e-commerce product information, and generate the third sub-state set.

[0056] Optionally, if the previous state is At this point, the large language model is invoked to analyze and output the three target groups. The groups are remote working programmers, people undergoing back rehabilitation, and e-sports players. Each group includes structured information such as demographic characteristics, pain points, and purchase motivations, resulting in the first sub-state set. The first sub-state set outputs the group segmentation, pain points, and purchase motivations.

[0057] If the previous state is At this point, the large language model is invoked to mine the competitor reviews provided by the user, extracting the 3 most frequent pain points (insufficient lumbar support, seat cushion collapse, and non-adjustable armrests) and 3 positive points (smooth wheels, easy installation, and breathable fabric). Each conclusion is accompanied by a quote from the original review, resulting in a second sub-state set. The output of the second sub-state set is: high-frequency pain points, advantages, and the basis for the original review quote.

[0058] If the previous state is At this point, the large language model is invoked to transform the product characteristics into a four-dimensional copywriting matrix of "characteristics → benefits → trust → emotion", resulting in the third sub-state set. The output of the third sub-state set is: product highlights and optimized product title.

[0059] Further, step S1023 above includes: Step b1: Based on the previous state, call the large language model to analyze the e-commerce product information and each group of people contained in the first sub-state set to obtain the visual main image of each group of people and generate the fourth sub-state set.

[0060] Optionally, if the previous state is ,at this time The large language model is used to generate five main visual images (also known as visual hook schemes) for each group. Each scheme includes a description of the image composition, hook text, and validity statement, resulting in a fourth sub-state set. The fourth sub-state set is then output. The solution revolves around "visualizing lumbar and back support during prolonged sitting". The plan focuses on "lumbar protection and posture correction during the rehabilitation period". The plan revolves around "body stability support during high-intensity combat".

[0061] Further, step S1024 above includes: Step c1: Based on the previous state, the large language model is called to analyze the output results of the first sub-state set, the second sub-state set, the third sub-state set, and the fourth sub-state set to obtain test indicators and generate the fifth sub-state set.

[0062] Optionally, if the previous state is ,at this time The output results of the first to fourth sub-state sets are analyzed by calling the large language model to generate an A / B test plan. The test indicators that constitute the A / B test plan include: test variables (main image scheme), sample size (1200 exposures per variant), test period (14 days covering two weekends), and success indicators (click-through rate increase ≥8% and conversion rate not decrease). The fifth sub-state set is obtained from these test indicators. The output of the fifth sub-state set is: a set of A / B test execution plans that can be directly implemented.

[0063] The execution output results of each of the above sub-state sets are stored in the context pool in order. .

[0064] The embodiments disclosed herein strictly execute the pre-processing flow serially to ensure a complete closed loop of business logic, without omissions or skips, resulting in a clear data structure for the generated sub-state set and ensuring that all outputs conform to the business logic.

[0065] As an optional embodiment, step S103 above includes: Step S1031: Obtain the set of sub-states of the crowd in the sub-state set; Step S1032: Determine the total number of people and the number of image variants under each people based on the set of sub-states of the population; Step S1033: Based on the total number of people and the number of image variants, determine the text subtask and the image subtask.

[0066] Optionally, as can be seen from the above embodiments, the multiple sub-state sets include a population sub-state set, and the population sub-state set includes population tags, population characteristics, and the total number of people k.

[0067] During the main state machine transition Then, the task splitter reads... Set of sub-states of the middle group and pre-set number of image variants .because The splitter generates a list of subtasks. Text subtask There is one subtask in total: image processing. A total of 9 ( Each Bind to the corresponding audience tags and The visual hook scheme generated for this group of people is the first one. A variant of the prompt word.

[0068] It's important to note that variants are different images generated from different cue words. The cue words are the input to the variants, and the variants are the output of the cue words. Each variant corresponds to a unique visual hook cue word, which is generated from the preceding state. Each group is designed separately, with differences in composition, style, and visual elements. This means that each group has 3 different prompts, which will generate 3 different main image variations for subsequent A / B testing to select the best one.

[0069] As an optional embodiment, step S104 above includes: Step S1041: Obtain the execution digests corresponding to the multiple sub-state sets respectively; Step S1042: Determine the first context information of the text subtask based on the execution summary and text sharing basic information to obtain the text sub-state machine for processing the text subtask; Step S1043: Based on the text sharing basic information, the crowd labels in the crowd sub-state set, and the prompt words when generating the corresponding image variant for each crowd label, the second context information of the image sub-task is obtained, and the image sub-state machine for processing the image sub-task is obtained. Step S1044: Send the text sub-state machine and the image sub-state machine to the asynchronous task queue, generate a first prompt word based on the first context information, and call the large language model based on the first prompt word to obtain a first execution result. Generate a second prompt word based on the second context information, and call the image generation model based on the second prompt word to obtain a second execution result.

[0070] Optionally, construct sub-state independent contexts: First, the execution summary corresponding to each sub-state set can be obtained from the large language model, where the execution summary is a... to The output results are simplified and condensed, retaining only the core conclusions and omitting the original redundant data.

[0071] Then the context manager reads Shared basic information This includes the product name "Ergonomic Office Chair," the category "Ergonomic Office Chair," and the key selling point "Relieves lower back pain from prolonged sitting." Construct the first contextual information for the text subtask: ,in for to The execution summary output for each state includes three audience names, three core pain points, recommended visual solution numbers for each audience, suggested copy titles, and key parameters for A / B testing. Then, a text sub-state machine (one machine) is constructed to process the first context information.

[0072] Construct a second contextual information for each image subtask: ,in, Includes people Labels and The complete cue word for the j-th visual hook scheme generated for this group. Then, an image sub-state machine (9 machines) is constructed to process the second contextual information.

[0073] It should be noted that, The shared basic information is copied 10 times and injected into each sub-state context. Differentiated injection based on subtask type.

[0074] Parallel dispatch and asynchronous execution: The main state machine dispatches the aforementioned 10 sub-state machine instances to the asynchronous task queue. The text sub-state calls the large language model, in order to... The first prompt word is input into the large language model, generating a comprehensive text analysis report as the first execution result. Nine image sub-states concurrently call the image generation model, each with its corresponding... The visual hook prompt is used as the second prompt. The input image generation model is used, and the white background image of the product uploaded by the user is used as the reference image. The image generation model generates a 1024×1024 resolution main image as the second execution result.

[0075] The actual concurrency is controlled by semaphores to be 3. When a substate completes and releases its semaphore, the substates waiting in the queue begin execution. The text substate execution takes approximately 8 seconds, and the 9 raw image substates are executed in 3 batches, taking a total of approximately 45 seconds. The overall stage takes approximately 45 seconds. If a serial approach is used, first the text substate and then generate the images one by one, the total time would be 8 + 9 × 15 = 143 seconds.

[0076] In this embodiment, basic information is shared via shared text, avoiding redundant storage and ensuring no interference between tasks. By injecting specialized contexts differentiated according to subtask type, the model input becomes more accurate and the output quality is higher. Simultaneously, asynchronous parallel execution reduces overall time consumption and significantly improves efficiency.

[0077] As an optional embodiment, the method further includes: Step d1: Perform summary aggregation on the first execution result and the second execution result, extract text summary information, group the second execution result into image summary groups according to the total number of people, and extract multiple grouped image summaries; Step d2: Assemble the text summary and grouped image summary according to the display template format to obtain the page to be displayed.

[0078] Optionally, when the large language model is invoked to obtain the first execution result, a preset prompt word template is used to tell the large language model "you should answer in this format," allowing it to automatically distribute the content to different locations during generation. Specifically... The input is a long text, which the large language model reads. Then, following the structure specified in the prompts, it writes as it thinks: Distinguishing by "Title": The prompts will require: "Finally, please write a summary starting with '## Execution Summary'." The large language model then outputs a concise summary. Distinguishing by "Field Name": The prompts will specify: "At the beginning of a line, explicitly write 'Execution Status: Success'." or "Give a score in the format 'Optimization Potential Score: 8.5 / 10'." In this way, the large language model outputs the "Status Flags" and "Numerical Estimates" to the designated locations. Distinguishing by "Format": For the suggestion list, the prompts will require: "Output key suggestions in the list format of 'Suggestion Content'." The large language model will then generate a series of action instructions to obtain the key suggestion list.

[0079] Therefore, as described above, after the text sub-state machine completes its execution, it returns a summary, including the execution status "success" and the execution summary text "The core optimization direction of this product is to enhance the visualization of the lumbar support." Option 3 (lumbar support angle annotation diagram) is expected to have the highest click-through rate, with the following key suggestions: (use lumbar support annotation in the main image, add a lumbar spine anatomy comparison diagram to the A+ page, and keep the current pricing). Optimization potential score: 8.5.

[0080] Because this embodiment creates a separate image sub-state machine for each image variant of each population, each independently calls the raw image API. Each sub-state machine is bound to only one population label and one visual hook cue word. After the call, a fixed-format summary is generated based on the API return result. If successful, it includes the image URL and cue word summary; if unsuccessful, it includes the reason for failure. There is no data dependency between different sub-state machines.

[0081] Therefore, in the 9 image sub-state machines, All three returned success and the image URL. All three returned success and the image URL. All three returned failure statuses and the reason for failure: "Raw image model service timed out". Each successfully returned image sub-state summary also includes the first 120 characters of the prompt words used in its generation.

[0082] Taking the first variant of the "programmer" user group as an example: The image sub-state machine receives the bound prompt phrase "a product image of an ergonomic chair, with the lumbar support area marked by a red dotted line and labeled 'adaptive lumbar support'," and uses this as input to call the image generation model. After the image generation model returns successfully, the image sub-state machine only retains the first 120 characters of the prompt phrase as a summary field, which is packaged together with the generation status "successful" and the image URL and returned. The final summary does not contain the complete prompt phrase; only the first 120 characters are extracted for later tracing which visual solution corresponds to this variant.

[0083] The summary aggregator aggregates the summaries of the first and second execution results, collects 10 sub-state summaries, extracts text summary information, and then groups the image summaries according to the total number of people, extracting multiple grouped image summaries, that is, grouping the image summaries according to the people's labels: The group contains 3 successful results and their corresponding URLs. The group contains 3 successful results and their corresponding URLs. The group contains 3 failure results, and the reasons for failure include: model timeout, service exception, network error, parameter error, etc.

[0084] In this embodiment of the disclosure, regarding the second execution result When the group feedback result is the preset result, i.e., a failure result, then... If the number of successful group formations is 0, then... Mark as failed image generation state, and select a preset image URL labeled "gaming chair scene illustration" from the preset placeholder image library and assign it to... The three image variant positions of the group ensure the integrity of the report structure, preventing blank spaces or display anomalies. This method of obtaining failure locations and filling placeholder images is used to degrade failed subtasks, preventing the overall task from being interrupted because the main state machine cannot automatically degrade when some subtasks fail.

[0085] The text summary information and grouped image summaries are then assembled according to a preset display template format. The final comprehensive report includes: the execution summary section displaying the text summary information returned by the text sub-state, and the main image scheme sections for each group of people. , , Displayed in three sections and Each block displays three generated main images and corresponding keyword summaries. The block displays three placeholder images with the label "Failed to generate image - using default image".

[0086] Finally, press to The output includes a 30-day execution plan, covering week 1 for main image replacement, week 2 for initiating A / B testing, and weeks 3-4 for iteration based on test data; the A / B testing suggestions are directly quoted. The output test plan. The entire comprehensive report is returned to the caller in structured JOSN format.

[0087] In this embodiment of the disclosure, the display interface is clear and intuitive by showing text summaries and grouped image summaries of the crowd separately.

[0088] like Figure 2 As shown, Figure 2 This is a complete flowchart of a multi-task parallel scheduling method according to an embodiment of the present disclosure. The specific process is as follows: Main state machine phase: The process sequentially executes S1 audience analysis, S2 comment mining, S3 visual hooks, S4 trust copywriting, and S5 A / B testing, and stores the execution results of S1 to S5 into the context pool. Then proceed to the S6 comprehensive report stage.

[0089] Parallel sub-state machine phase: Task splitter Split( Read G,m, and construct a text sub-state machine and k·m image sub-state machines, where m is the number of crowd image variants and G is the crowd set; The text sub-state machine returns a text summary, and the image sub-state machine returns an image summary. The summary aggregator is invoked for processing, and a comprehensive report is output.

[0090] This embodiment also provides a multi-task parallel scheduling device for implementing the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0091] This embodiment provides a multi-task parallel scheduling device, such as... Figure 3 As shown, it includes: The first acquisition module 301 is used to acquire e-commerce product information and the pre-processing process to be executed by the main state machine. The pre-processing process is used to characterize the order in which the main state machine executes tasks. The first module 302 is used to obtain multiple sub-state sets based on e-commerce product information and pre-processing procedures; Module 303 is used to construct a first preset number of text subtasks and a second preset number of image subtasks based on the crowd sub-state set in the sub-state set; The second module 304 is used to call the large language model to perform the text subtask and obtain the first execution result, and to call the image generation model to perform the image subtask and obtain the second execution result.

[0092] In this embodiment, by acquiring e-commerce product information and the pre-processing flow to be executed by the main state machine, the main state machine executes fixed business tasks using a fixed pre-processing flow. Based on the e-commerce product information and the pre-processing flow, multiple sub-state sets are obtained. Then, based on the crowd sub-state set in the sub-state set, a first preset number of text sub-tasks and a second preset number of image sub-tasks are constructed. This achieves the splitting of crowd rules into sub-tasks without relying on a large model for splitting decisions. This ensures that the task splitting results are certain, the stage execution order is fixed, and the business logic is strictly controllable. It avoids the technical problems of unreliable task splitting, stage omissions, and task redundancy caused by the autonomous decision-making of a large model. Then, the large language model is called to execute the text sub-tasks to obtain the first execution result, and the image generation model is called to execute the image sub-tasks to obtain the second execution result. Based on this, the parallel execution of text and image tasks is achieved, which greatly improves the overall efficiency of e-commerce product detail page optimization.

[0093] In some optional implementations, the first obtaining module 302 is used to serially execute the pre-processing process to determine the currently executed pre-processing state; based on the pre-processing state and e-commerce product information, to obtain a first sub-state set, a second sub-state set, and a third sub-state set; based on the pre-processing state, the first sub-state set, and the e-commerce product information, to obtain a fourth sub-state set; and based on the pre-processing state, the first sub-state set, the second sub-state set, the third sub-state set, and the fourth sub-state set, to obtain a fifth sub-state set.

[0094] In some optional implementations, the construction module 303 is used to obtain the population sub-state set in the sub-state set; determine the total number of people and the number of image variants under each population based on the population sub-state set; and determine the text sub-task and image sub-task based on the total number of people and the number of image variants.

[0095] In some optional implementations, the second obtaining module 304 is used to obtain execution summaries corresponding to multiple sub-state sets respectively; determine the first context information of the text sub-task based on the execution summaries and text-sharing basic information, and obtain the text sub-state machine for processing the text sub-task; obtain the second context information of the image sub-task based on the text-sharing basic information, the crowd labels in the crowd sub-state set, and the prompt words when generating corresponding image variants for each crowd label, and obtain the image sub-state machine for processing the image sub-task; send the text sub-state machine and the image sub-state machine to the asynchronous task queue, generate the first prompt word based on the first context information, and call the large language model based on the first prompt word to obtain the first execution result; generate the second prompt word based on the second context information, and call the image generation model based on the second prompt word to obtain the second execution result.

[0096] In some alternative embodiments, the device further includes: The extraction module is used to aggregate the first and second execution results, extract text summary information, group the second execution results into image summary groups according to the total number of people, and extract multiple grouped image summaries; The assembly module is used to assemble the text summary and grouped image summary according to the display template format to obtain the page to be displayed.

[0097] In some alternative embodiments, the device further includes: The second acquisition module is used to acquire the location information of the image variant corresponding to the preset result when the second execution result is a preset result; The filling module is used to select a preset image from the preset placeholder image library to fill in the position information.

[0098] In some optional implementations, the first obtaining module 302 is further configured to, based on the prior state, call a large language model to analyze the product description information in the e-commerce product information to obtain the user group classification corresponding to the e-commerce product information and generate a first sub-state set; based on the prior state, call a large language model to mine the competitor review information in the e-commerce product information to obtain the review status of the e-commerce product information and generate a second sub-state set; based on the prior state, call a large language model to analyze the product characteristic information in the e-commerce product information to obtain the copy title of the e-commerce product information and generate a third sub-state set.

[0099] In some optional implementations, the first obtaining module 302 is further configured to, based on the preceding state, call a large language model to analyze the e-commerce product information and each group of people contained in the first sub-state set, obtain the visual main image of each group of people, and generate a fourth sub-state set.

[0100] In some optional implementations, the first obtaining module 302 is further configured to, based on the preceding state, call the large language model to analyze the output results of the first sub-state set, the second sub-state set, the third sub-state set, and the fourth sub-state set, obtain test indicators, and generate a fifth sub-state set.

[0101] Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0102] In this embodiment, the learning community subgroup identification and vulnerability assessment device is presented in the form of a functional unit. Here, a unit refers to an ASIC (Application Specific Integrated Circuit), a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0103] This disclosure also provides a computer device having the above-described features. Figure 3 The multi-task parallel scheduling device shown.

[0104] Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of this disclosure, such as... Figure 4 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 4 Take a processor 10 as an example.

[0105] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0106] The memory 20 stores instructions executable by at least one processor 10 to cause at least one processor 10 to perform the method shown in the above embodiments.

[0107] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0108] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0109] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0110] This disclosure also provides a computer-readable storage medium in which the methods described in this disclosure can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium after being downloaded over a network. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium may be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium may also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code that, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0111] A portion of this disclosure can be applied to computer program products, such as computer program instructions, which, when executed by a computer, can invoke or provide methods and / or technical solutions according to this disclosure through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, and installation package files. Accordingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions; the computer compiling the instructions and then executing the corresponding compiled program; the computer reading and executing the instructions; or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0112] Although embodiments of the present disclosure have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the present disclosure, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A multi-task parallel scheduling method, characterized in that, The method includes: The process involves acquiring e-commerce product information and pre-processing procedures to be executed by the main state machine, wherein the pre-processing procedures are used to characterize the order in which the main state machine executes tasks. Based on the e-commerce product information and the pre-processing flow, multiple sub-state sets are obtained; Based on the crowd sub-state set in the sub-state set, a first preset number of text sub-tasks and a second preset number of image sub-tasks are constructed; The large language model is invoked to execute the text subtask, resulting in a first execution result. The image generation model is then invoked to execute the image subtask, resulting in a second execution result.

2. The method according to claim 1, characterized in that, Based on the e-commerce product information and the pre-processing flow, multiple sub-state sets are obtained, including: The pre-processing flow is executed serially to determine the current pre-processing state. Based on the aforementioned pre-state and the e-commerce product information, a first sub-state set, a second sub-state set, and a third sub-state set are obtained. Based on the previous state, the first sub-state set, and the e-commerce product information, a fourth sub-state set is obtained; Based on the preceding state, the first sub-state set, the second sub-state set, the third sub-state set, and the fourth sub-state set, a fifth sub-state set is obtained.

3. The method according to claim 1, characterized in that, The construction of a first preset number of text subtasks and a second preset number of image subtasks based on the crowd sub-state set in the sub-state set includes: Obtain the set of population substates from the substate set; The total number of people and the number of image variants under each people are determined based on the set of sub-states of the population. The text subtask and the image subtask are determined based on the total number of people and the number of image variants.

4. The method according to claim 1, characterized in that, The process of calling the large language model to execute the text subtask and obtaining a first execution result, and calling the image generation model to execute the image subtask and obtaining a second execution result, includes: Obtain the execution digests corresponding to the multiple sub-state sets respectively; Based on the execution summary and text sharing basic information, the first context information of the text subtask is determined, and the text sub-state machine for processing the text subtask is obtained; Based on the text sharing basic information, the crowd labels in the crowd sub-state set, and the prompt words when generating corresponding image variants for each crowd label, the second context information of the image sub-task is obtained, and the image sub-state machine for processing the image sub-task is obtained. The text sub-state machine and the image sub-state machine are sent to the asynchronous task queue. A first prompt word is generated based on the first context information, and a large language model is called based on the first prompt word to obtain the first execution result. A second prompt word is generated based on the second context information, and an image generation model is called based on the second prompt word to obtain the second execution result.

5. The method according to claim 1, characterized in that, The method further includes: The first execution result and the second execution result are aggregated to extract text summary information. The second execution result is grouped into image summary groups according to the total number of people, and multiple grouped image summaries are extracted. The text summary and the grouped image summary are assembled according to the display template format to obtain the page to be displayed.

6. The method according to claim 1 or 4, characterized in that, The method further includes: If the second execution result is a preset result, obtain the location information of the image variant corresponding to the preset result; Select a preset image from the preset placeholder image library to fill in the position information.

7. The method according to claim 2, characterized in that, The process of obtaining a first sub-state set, a second sub-state set, and a third sub-state set based on the preceding state and the e-commerce product information includes: Based on the aforementioned pre-state, the large language model is invoked to analyze the product description information in the e-commerce product information, thereby obtaining the user group classification corresponding to the e-commerce product information and generating the first sub-state set. Based on the aforementioned pre-state, a large language model is invoked to mine competitor review information in the e-commerce product information to obtain the review status of the e-commerce product information and generate the second sub-state set. Based on the aforementioned pre-state, the large language model is invoked to analyze the product characteristic information in the e-commerce product information, thereby obtaining the text title of the e-commerce product information and generating the third sub-state set.

8. The method according to claim 2, characterized in that, Based on the previous state, the first sub-state set, and the e-commerce product information, a fourth sub-state set is obtained, including: Based on the aforementioned pre-state, a large language model is invoked to analyze the e-commerce product information and each group of people contained in the first sub-state set, thereby obtaining the visual main image of each group of people and generating the fourth sub-state set.

9. The method according to claim 2, characterized in that, The fifth sub-state set is obtained based on the previous state, the first sub-state set, the second sub-state set, the third sub-state set, and the fourth sub-state set, including: Based on the aforementioned pre-state, the large language model is invoked to analyze the output results of the first sub-state set, the second sub-state set, the third sub-state set, and the fourth sub-state set to obtain test metrics and generate the fifth sub-state set.

10. A multi-task parallel scheduling device, characterized in that, The device includes: The first acquisition module is used to acquire e-commerce product information and the pre-processing process to be executed by the main state machine, wherein the pre-processing process is used to characterize the order in which the main state machine executes tasks. The first obtaining module is used to obtain multiple sub-state sets based on the e-commerce product information and the pre-processing process; The construction module is used to construct a first preset number of text subtasks and a second preset number of image subtasks based on the crowd sub-state set in the sub-state set; The second obtaining module is used to call the large language model to execute the text subtask and obtain the first execution result, and to call the image generation model to execute the image subtask and obtain the second execution result.

Images