Automatic process control system and method for low-code platform

Abstract

The invention discloses an automatic process control system and method for a low-code platform, and relates to the technical field of computers.The method comprises the steps that source data of a user on the low-code platform are obtained according to the preset frequency, and snapshots are generated; the source data and predefined cue words are spliced to serve as large language model input; the large language model outputs structured suggestions, and the suggestions are classified and displayed for users to confirm; after the user confirms, recording related data as a training sample and adjusting the suggestion display frequency; constructing an event chain library, calculating a matching degree between an actual operation event chain and an operation event chain in the event chain library according to an editing distance algorithm, and recommending a component type; according to the method, the local network speed and computing power data of the user are obtained, the learning rate of the large language model is weighted and adjusted after normalization, training is completed in a local background, intelligentization and high efficiency of automatic process control of the low-code platform are achieved, and user operation experience and platform applicability are improved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, specifically to an automated process control system and method for low-code platforms. Background Technology

[0002] The technological background of low-code platforms stems from the contradiction between the deepening of enterprise informatization, the surge in software demand, and the shortage of professional development resources. Traditional software development models rely on professional engineers to complete the entire process of requirements analysis, coding implementation, testing, and deployment, resulting in long development cycles, slow response times, and high communication costs, making it difficult to meet the needs of rapidly changing business requirements and personalized customization. Against this backdrop, low-code platforms emerged and have rapidly played a crucial role in enterprise digital transformation.

[0003] Despite the numerous advantages of low-code platforms, they often lack intelligent interaction with users in practical applications. These platforms are mostly based on rule engines and fixed templates, failing to understand the user's true intentions. When users encounter difficulties meeting their needs while using a low-code platform, they typically have to search within the platform's pre-set instructions. The platform does not proactively recommend components or other suggestions, lacking real-time user assistance capabilities and thus demonstrating insufficient intelligence. Summary of the Invention

[0004] The purpose of this invention is to provide a solution to the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: Step S1: Obtain the current page browsed by the user on the low-code platform, the selected component type and its attributes, the bound data fields and their data types, and the content input and selected by the user as source data according to a preset frequency, and generate a snapshot of the source data; concatenate the predefined prompt words with the source data as input data for the large language model; Furthermore, step S1 also includes: Step S1-1: Classify components according to their functions to obtain component types, and use the parameters corresponding to the components as component attributes; convert source data into text data and generate a snapshot of the source data for archiving; predefine prompt words, which are used to limit the large language model to read the source data and provide suggestions, and the prompt words are text data; trim the prompt words to meet the length limit when the large language model reads them. Step S1-2: Preload the external formula. The external formula is a program used to parse dynamic content into structured data that can be directly read by the large language model. The dynamic content is logical language that cannot be directly determined by static text or fixed rules and changes in real time according to the context, and cannot be directly read by the large language model. Add a label to the prompt word to call the external formula, and concatenate the prompt word with the label to call the external formula with the source data in sequence as the input data of the large language model.

[0006] Step S2: The large language model outputs structured suggested content based on the input data; the suggested content generated by the large language model is categorized and displayed; the user confirms the suggested content; Furthermore, step S2 also includes: The input data is transmitted to the large language model. After reading the input data, the large language model determines the user's operational intent based on the prompt words and preset keywords in the source data, and maps the natural language in the user's input and selection into structured data. The operational intent includes the page the user is about to browse, the type of component to be selected and its attributes, the data field to be bound and its data type, and the content to be input and selected. The large language model outputs structured suggestion content based on the operational intent and the mapped structured data. The suggestion content generated by the large language model is categorized according to the corresponding operational intent and displayed to the user. The process of the user confirming the suggestion content is as follows: the user chooses to accept or reject the suggestion content; if the user chooses to accept, the suggestion content is retained in a separate window, which the user can actively close; if the user chooses to reject, the display of suggestion content ends.

[0007] Step S3: After the user confirms the suggested content, adjust the frequency of displaying the suggested content by category; and record the suggested content, the source data, and the user's confirmation of the suggested content as training samples for the large language model. Furthermore, step S3 also includes: Step S3-1: Set the frequency adjustment period and suggestion detection time. The initial frequency adjustment period starts from the first time the suggestion content is shown to the user. When the suggestion content is not shown to the user within one suggestion detection time, it is considered to be the end of one frequency adjustment period. Calculate the ratio of the number of times the user accepts the suggestion content to the total number of times the suggestion content is shown to the user within one frequency adjustment period as the suggestion acceptance rate R. Record the interaction time interval from showing the suggestion content to the user to choosing to accept or reject the suggestion content. Set a baseline time interval. Calculate the ratio of the user's average interaction time interval to the baseline time interval within one frequency adjustment period as the interaction matching degree M. Set an interaction time interval threshold. If the interaction time interval exceeds the interaction time interval threshold, it is considered invalid data and is not included in the calculation of the average interaction time interval. Step S3-2: Weight the suggestion acceptance rate R and the reciprocal of the interaction matching degree M to obtain the suggestion fit degree S: ; Where 'a' represents the weight of the suggestion acceptance rate, and 'b' represents the weight of the reciprocal of the interaction matching degree. The sum of 'a' and 'b' is 1. When the suggested fit is less than the preset fit threshold, the frequency of showing suggested content to the user will be reduced by a preset amount; the user's feedback will be sent to the large language model, and the suggested content, source data and confirmation of the user's suggested content will be archived as training samples for the large language model.

[0008] Step S4: Construct an event chain library and record the user's current actual operation event chain. Calculate the matching degree between the operation event chains in the event chain library and the actual operation event chain. Recommend the final selected component type in the operation event chain with the highest matching degree as predicted content to the user. The user can choose whether to display it. If the user chooses to display it, the component information in the predicted content is shown to the user, and the actual operation event chain is stored in the event chain library. If the user refuses to display it, the final selected component types in the operation event chains are recommended to the user in descending order of matching degree. Furthermore, step S4 also includes: Step S4-1: Number the different windows of the low-code platform to obtain a window sequence; classify the different windows according to the component types contained within the windows to obtain window types; construct an event chain library, in which the operation event chain consists of the window sequence opened by the user, the window type, and the finally selected component type; the finally selected component type is the component type that the user selected and did not perform an undo operation within a preset detection time; an operation event chain is the sequence from the first time a user finally selects a component type to the next time a user finally selects a component type. Step S4-2: Set the event chain prediction start threshold n. When a user continuously opens n different window sequences of the low-code platform without ultimately selecting a component type, extract the window sequence and window type opened by the user since the last component type selection, as the user's actual operation event chain; calculate the weighted edit distance between the user's actual operation event chain and the window sequence and window type of the operation event chain in the event chain library using the edit distance algorithm as the matching degree. Window sequences and window types are encoded separately, and the window sequences are converted into window sequence lists and the window types into window type lists according to the opening order. The edit distances D1 and D2 are calculated between the window sequence lists and window type lists of the user's actual operation event chain and the corresponding operation event chain lists and window sequence lists in the event chain library, where D1 represents the edit distance of the window sequence lists and D2 represents the edit distance of the window type lists. D1 is normalized to D1' based on the historical largest edit distance of the window sequence list, and D2 is normalized to D2' based on the historical largest edit distance of the window type list. D1' and D2' are weighted and summed to obtain the comprehensive distance D. The matching degree P = 1 - D. The operation event chain with the highest matching degree is extracted, and the final selected component type in that operation event chain is recommended to the user as predicted content. The user can then choose whether to directly select that component type. Step S4-3: During the recommendation process to the user, the newly generated window sequence and window type are added to the matching degree calculation process, and the new predicted content is continuously displayed to the user at the frequency mentioned above until the user actively stops the predicted content display process; when the final selected component type is generated, the actual operation event chain is stored as the operation event chain in the event chain library.

[0009] Step S5: Obtain the user's local network speed and computing power data, adjust the learning rate of the large language model based on the network speed and computing power data, and train the large language model on the user's local machine.

[0010] Furthermore, step S5 also includes: The system acquires real-time network speed and computing power data from the user's local machine, and obtains maximum network speed and computing power data according to a preset training period. The real-time network speed and computing power data are then normalized based on the maximum network speed and computing power data, and a weighted calculation is performed to adjust the learning rate for training the large language model. The training of the large language model is completed in the background on the user's local machine. ; Where t represents the number of training iterations; This represents the adjusted learning rate for the t-th iteration; This represents the initial learning rate; and All are decay coefficients, used to adjust the learning rate during the initial stages of training; C norm N represents the normalized real-time computing power data; norm α represents the normalized real-time network speed; α and β represent the computing power weight and network speed weight, respectively, and their sum is 1.

[0011] An automated process control system for a low-code platform includes an input construction module, a suggestion processing module, a sample frequency tuning module, an event chain module, and a local training module. The input construction module generates source data and concatenates prompt words as input data for a large language model. The suggestion processing module outputs structured suggestions from the large language model and displays them to the user for confirmation. The sample frequency tuning module records training samples and adjusts the display frequency of suggested content based on user feedback. The event chain module constructs an event chain library and displays predicted content based on the matching degree between actual operation event chains and the operation event chains in the library. The local training module adjusts the learning rate based on the user's local network speed and computing power data and trains the large language model locally. The output of the input construction module is connected to the input of the suggestion processing module; the output of the suggestion processing module is connected to the input of the sample frequency tuning module; the output of the sample frequency tuning module is connected to the input of the event chain module; and the output of the event chain module is connected to the input of the local training module.

[0012] The input construction module further includes a source data generation unit and a prompt processing unit; the source data generation unit is used to obtain in real time the current page browsed by the user on the low-code platform, the selected component type and its component attributes, the bound data fields and their data types, and the content input and selected by the user as source data; the prompt processing unit is used to predefine prompt word templates and process prompt words, and preload plug-in formulas to concatenate input data.

[0013] The suggestion processing module further includes a suggestion generation unit and a display confirmation unit; the suggestion generation unit is used by the large language model to determine the operation intent based on the input data and output structured suggestion content; the display confirmation unit is used to classify the suggestion content and display it for user confirmation; the sample frequency tuning module further includes a sample recording unit and a frequency adjustment unit; the sample recording unit is used to record the suggestion content after user confirmation, the source data, and the user's feedback as training samples for the large language model; the frequency adjustment unit is used to calculate the suggestion acceptance rate and interaction matching degree, and adjust the display frequency according to the suggestion fit.

[0014] The event chain module further includes an event chain library construction unit and a matching update unit; the event chain library construction unit is used to construct the event chain library; the matching update unit is used to calculate the matching degree between the actual operation event chain and the operation event chains in the event chain library, recommend the final selected component type in the operation event chain with the highest matching degree, and update the event chain library; the local training module further includes a parameter adjustment unit and a local training unit; the parameter adjustment unit is used to obtain the user's local network speed and computing power data to adjust the learning rate of the large language model training; the local training unit is used to train the large language model locally on the user's local machine according to the adjusted learning rate.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention acquires source data such as the user's current page, selected components and attributes, bound data fields and types on the low-code platform at a preset frequency, generates snapshots, and concatenates predefined prompt words as input to a large language model. The large language model outputs structured suggestions, and the platform records training samples and adjusts the suggestion display frequency based on user feedback. This enables the platform to understand the user's operation intentions in real time, proactively provide accurate component configuration suggestions, reduce the time users spend manually filtering and configuring, and effectively improve the intelligent assistance capabilities and development efficiency of the low-code platform.

[0016] 2. This invention constructs an event chain library, forming an operation event chain from the sequence and type of windows opened by the user and the type of the final selected component. When the user opens a preset number of windows without selecting a component, the matching degree between the actual operation event chain and the event chain in the library is calculated, and the optimal component is recommended. At the same time, the new operation is included in the event chain library, enabling the platform to predict personalized components based on the user's historical operation habits, avoid repeated operations, adapt to the usage logic of different users, and significantly enhance the process automation and personalized service capabilities of the low-code platform.

[0017] 3. This invention obtains real-time local network speed and computing power data from the user, normalizes and weights the data to adjust the learning rate of the large language model, and completes model training in the local background. This allows the model training process to adapt to the user's device resource conditions, avoiding training delays caused by network or computing power limitations. At the same time, the local training mode protects the privacy and security of user operation data, ensures that model performance is continuously optimized with use, and improves the stability and data security of the low-code platform. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an automated process control method for a low-code platform according to the present invention. Figure 2 This is a schematic diagram of an automated process control system for a low-code platform according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1: As Figure 1 As shown, the present invention provides a technical solution, an automated process control method for a low-code platform, the automated process control method comprising the following steps: Step S1: Obtain the current page browsed by the user on the low-code platform, the selected component type and its attributes, the bound data fields and their data types, and the content input and selected by the user as source data according to a preset frequency, and generate a snapshot of the source data; concatenate the predefined prompt words with the source data as input data for the large language model; Step S1 also includes: Step S1-1: Classify components according to their functions to obtain component types, and use the parameters corresponding to the components as component attributes; convert source data into text data and generate a snapshot of the source data for archiving; predefine prompt words, which are used to limit the large language model to read the source data and provide suggestions, and the prompt words are text data; trim the prompt words to meet the length limit when the large language model reads them. Step S1-2: Preload the external formula. The external formula is a program used to parse dynamic content into structured data that can be directly read by the large language model. The dynamic content is logical language that cannot be directly determined by static text or fixed rules and changes in real time according to the context, and cannot be directly read by the large language model. Add a label to the prompt word to call the external formula, and concatenate the prompt word with the label to call the external formula with the source data in sequence as the input data of the large language model.

[0021] For example: Acquire source data at a preset frequency of 5 seconds per acquisition: Current page: "Form Design Page"; Selected component type: "Single-line text box", component properties: "Width 500px, Required"; The bound data field is "Customer Name", and the data type is string. User input: "Test Form" (form name); Generate a snapshot of the source data (copy the above data for verification); The predefined prompt template, "Based on the form components and data fields currently being used by the user, recommends components and configuration rules that can be added later," is trimmed to 50 characters (to meet the length limit of the large language model) and concatenated with the source data to form the input data: "Based on the form components and data fields currently being used by the user, recommends components and configuration rules that can be added later. Current page: Form design page; Component: Single-line text box (500px wide, required); Binded field: Customer name (string); User input: Test form."

[0022] Preload external formulas (such as programs that parse dynamic validation rules), add the annotation "Call formula F1 to parse field correlation" to the prompt, and finally include this annotation in the input data.

[0023] Step S2: The large language model outputs structured suggested content based on the input data; the suggested content generated by the large language model is categorized and displayed; the user confirms the suggested content; Step S2 also includes: The input data is transmitted to the large language model. After reading the input data, the large language model determines the user's operational intent based on the prompt words and preset keywords in the source data, and maps the natural language in the user's input and selection into structured data. The operational intent includes the page the user is about to browse, the type of component to be selected and its attributes, the data field to be bound and its data type, and the content to be input and selected. The large language model outputs structured suggestion content based on the operational intent and the mapped structured data. The suggestion content generated by the large language model is categorized according to the corresponding operational intent and displayed to the user. The process of the user confirming the suggestion content is as follows: the user chooses to accept or reject the suggestion content; if the user chooses to accept, the suggestion content is retained in a separate window, which the user can actively close; if the user chooses to reject, the display of suggestion content ends.

[0024] For example: Large language models output structured suggestions based on input data: Recommendation 1: "It is recommended to add a 'Phone Number Input Box' component, set the attributes to 'width 500px, required', and bind the data field 'Contact Number' (data type: integer)". Recommendation 2: "It is recommended to set the validation rules for the 'Customer Name' field: the length should not exceed 20 characters."

[0025] The suggestions are displayed in categories ("Component addition suggestions" and "Validation rule suggestions"), and the user confirms "Accept suggestion 1, ignore suggestion 2".

[0026] If the large language model's suggestions include dynamic content such as time, permissions, and organizational hierarchy, the system will detect the presence of "plug-in formulas" in the prompts. These plug-in formulas are bounded by <**> and provide a description of the logical calculation and function definition within the prompt. The system preprocesses this content before calling the large language model; for example, it parses "this week" into "start_date=2024-06-24, end_date=2024-06-30" before inserting it into the large language model's input. These plug-in formulas are implemented using languages ​​such as Python and JavaScript, operate independently of the model, and can handle complex logic such as dynamic dates, organizational tree mappings, and user permission conversions.

[0027] If the user's intent involves data querying or generating conditional statements, the NL2SQL capability of the large language model is triggered. The system provides the large language model with prompts such as the target data table structure, field meanings, field types, and sample data, so that the large language model can generate standard SQL query statements or process judgment scripts based on natural language.

[0028] Step S3: After the user confirms the suggested content, adjust the frequency of displaying the suggested content by category; and record the suggested content, the source data, and the user's confirmation of the suggested content as training samples for the large language model. Step S3 also includes: Step S3-1: After the user confirms the suggested content, they choose to accept, reject, or ignore the suggestion as the user's feedback result; set the frequency adjustment period and suggestion detection time. The initial frequency adjustment period starts from the first time the suggested content is shown to the user. When the suggested content is not shown to the user within one suggestion detection time, it is considered to be the end of one frequency adjustment period; calculate the ratio of the number of times the user accepts the suggested content to the total number of times the suggested content is shown to the user within one frequency adjustment period as the suggestion acceptance rate R; record the interaction time interval from showing the suggested content to the user to the user choosing to accept or reject the suggested content, set a baseline time interval, and calculate the ratio of the user's average interaction time interval to the baseline time interval within one frequency adjustment period as the interaction matching degree M; set an interaction time interval threshold. If the interaction time interval exceeds the interaction time interval threshold, it is considered invalid data and is not included in the calculation of the average interaction time interval; Step S3-2: Weight the suggestion acceptance rate R and the reciprocal of the interaction matching degree M to obtain the suggestion fit degree S: ; Where 'a' represents the weight of the suggestion acceptance rate, and 'b' represents the weight of the reciprocal of the interaction matching degree. The sum of 'a' and 'b' is 1. When the suggested fit is less than the preset fit threshold, the frequency of showing suggested content to the user will be reduced by a preset amount; the user's feedback will be sent to the large language model, and the suggested content, source data and user confirmation of the suggested content will be archived as training samples for the large language model.

[0029] For example: Record training samples: Recommendations: The two recommendations mentioned above; Source data: Information such as pages, components, and fields from step S1; User feedback: "Accept suggestion 1, ignore suggestion 2"; Calculate the frequency adjustment parameters (taking one frequency adjustment cycle as an example): Suggestion acceptance rate R: A total of 5 suggestions were displayed during this period, and users accepted 3 of them. R = 3 / 5 = 0.6. Interaction time interval: The average response time of users to 5 suggestions is 3 seconds, the baseline time interval is 4 seconds, the interaction matching degree M=3 / 4=0.75, and its reciprocal is 1 / 0.75≈1.33; Suggested fit S: Let a = 0.6 (acceptance rate weight), b = 0.4 (reciprocal of interaction matching degree weight), then S = 0.6 × 0.6 + 0.4 × 1.33 ≈ 0.36 + 0.53 = 0.89; The preset adaptation threshold is 0.7. Since S=0.89>0.7, the current display frequency (5 seconds / time) will be maintained.

[0030] Step S4: Construct an event chain library and record the user's current actual operation event chain. Calculate the matching degree between the operation event chains in the event chain library and the actual operation event chain. Recommend the final selected component type in the operation event chain with the highest matching degree as predicted content to the user. The user can choose whether to display it. If the user chooses to display it, the component information in the predicted content is shown to the user, and the actual operation event chain is stored in the event chain library. If the user refuses to display it, the final selected component types in the operation event chains are recommended to the user in descending order of matching degree. Step S4 also includes: Step S4-1: Number the different windows of the low-code platform to obtain a window sequence; classify the different windows according to the component types contained within the windows to obtain window types; construct an event chain library, in which the operation event chain consists of the window sequence opened by the user, the window type, and the finally selected component type; the finally selected component type is the component type that the user selected and did not perform an undo operation within a preset detection time; an operation event chain is the sequence from the first time a user finally selects a component type to the next time a user finally selects a component type. Step S4-2: Set the event chain prediction start threshold n. When a user continuously opens n different window sequences of the low-code platform without ultimately selecting a component type, extract the window sequence and window type opened by the user since the last component type selection, as the user's actual operation event chain; calculate the weighted edit distance between the user's actual operation event chain and the window sequence and window type of the operation event chain in the event chain library using the edit distance algorithm as the matching degree. Window sequences and window types are encoded separately, and the window sequences are converted into window sequence lists and the window types into window type lists according to the opening order. The edit distances D1 and D2 are calculated between the window sequence lists and window type lists of the user's actual operation event chain and the corresponding operation event chain lists and window sequence lists in the event chain library, where D1 represents the edit distance of the window sequence lists and D2 represents the edit distance of the window type lists. D1 is normalized to D1' based on the historical largest edit distance of the window sequence list, and D2 is normalized to D2' based on the historical largest edit distance of the window type list. D1' and D2' are weighted and summed to obtain the comprehensive distance D. The matching degree P = 1 - D. The operation event chain with the highest matching degree is extracted, and the final selected component type in that operation event chain is recommended to the user as predicted content. The user can then choose whether to directly select that component type. Step S4-3: During the recommendation process to the user, the newly generated window sequence and window type are added to the matching degree calculation process, and the new predicted content is continuously displayed to the user at the frequency mentioned above until the user actively stops the predicted content display process; when the final selected component type is generated, the actual operation event chain is stored as the operation event chain in the event chain library.

[0031] For example: Event chain library construction: The library already stores 10 operation event chains, for example, one of them is: Window sequence: [1 (Form Design Page) → 3 (Component Selection Page) → 2 (Attribute Configuration Page)]; Window type: [Form design class → Component library class → Configuration class]; The final component type selected is: "Mobile Number Input Box".

[0032] Current user action: Continuous window opening sequence [1 (form design page) → 3 (component selection page)] (no final selected component is generated), triggering event chain prediction (event chain prediction start threshold n=2).

[0033] Matching degree calculation: The actual operation event chain window sequence list: [1,3], window type list: [form design class, component library class]; Edit distance from the example event chain in the event chain library: D1 (window sequence) = 1 (add "2"), D2 (window type) = 1 (add "configuration class"). After normalization: D1=1 / 5=0.2 (the historical maximum window sequence edit distance is 5), D2=1 / 3≈0.33 (the historical maximum window type edit distance is 3); The overall distance D = 0.6 × 0.2 + 0.4 × 0.33 ≈ 0.12 + 0.13 = 0.25, and the matching degree P = 1 - 0.25 = 0.75 (the highest matching degree).

[0034] Recommendations and feedback: The component with the highest recommended match, the "phone number input box," is selected by the user to be displayed. Store the event chain of this actual operation (window sequence [1→3], type [form design class → component library class], final component "phone number input box") into the event chain library.

[0035] Step S5: Obtain the user's local network speed and computing power data, adjust the learning rate of the large language model based on the network speed and computing power data, and train the large language model on the user's local machine.

[0036] Step S5 also includes: The system acquires real-time network speed and computing power data from the user's local machine, and obtains maximum network speed and computing power data according to a preset training period. The real-time network speed and computing power data are then normalized based on the maximum network speed and computing power data, and a weighted calculation is performed to adjust the learning rate for training the large language model. The training of the large language model is completed in the background on the user's local machine. ; Where t represents the number of training iterations; This represents the adjusted learning rate for the t-th iteration; This represents the initial learning rate; and All are decay coefficients, used to adjust the learning rate during the initial stages of training; C norm N represents the normalized real-time computing power data; norm α represents the normalized real-time network speed; α and β represent the computing power weight and network speed weight, respectively, and their sum is 1.

[0037] For example: Retrieve user local data: Real-time network speed: 80Mbps, maximum network speed (within 24 hours): 100Mbps, normalized Nnorm = 80 / 100 = 0.8; Real-time computing power: 2.4GHz, maximum computing power (within 24 hours): 3.0GHz, normalized Cnorm = 2.4 / 3.0 = 0.8; The computing power weight α=0.6, and the network speed weight β=0.4 (α+β=1).

[0038] Adjusting the learning rate of the large language model (initial learning rate) =0.01, attenuation coefficient =0.9、 =0.95, current iteration number t=10): Learning rate calculation formula:

[0039] Substituting the data: η(10) = 0.01 × 0.9¹ 0 ×0.95×(0.6×0.8 + 0.4×0.8)≈0.01×0.348×0.95×0.8≈0.0026 A large language model is trained on the user's local background with an adjusted learning rate of 0.0026.

[0040] Example 2: Figure 2 As shown, this invention provides an automated process control system for a low-code platform. The system includes an input construction module, a suggestion processing module, a sample frequency tuning module, an event chain module, and a local training module. The input construction module generates source data and concatenates prompt words as input data for a large language model. The suggestion processing module outputs structured suggestions from the large language model and displays them to the user for confirmation. The sample frequency tuning module records training samples and adjusts the display frequency of suggestion content based on user feedback. The event chain module constructs an event chain library and displays predicted content based on the matching degree between actual operation event chains and the operation event chains in the event chain library. The local training module adjusts the learning rate based on the user's local network speed and computing power data and trains the large language model locally. The output of the input construction module is connected to the input of the suggestion processing module; the output of the suggestion processing module is connected to the input of the sample frequency tuning module; the output of the sample frequency tuning module is connected to the input of the event chain module; and the output of the event chain module is connected to the input of the local training module.

[0041] The input construction module further includes a source data generation unit and a prompt processing unit; the source data generation unit is used to obtain in real time the current page browsed by the user on the low-code platform, the selected component type and its component attributes, the bound data fields and their data types, and the content input and selected by the user as source data; the prompt processing unit is used to predefine prompt word templates and process prompt words, and preload plug-in formulas to concatenate input data.

[0042] The suggestion processing module further includes a suggestion generation unit and a display confirmation unit; the suggestion generation unit is used by the large language model to determine the operation intent based on the input data and output structured suggestion content; the display confirmation unit is used to classify the suggestion content and display it for user confirmation; the sample frequency tuning module further includes a sample recording unit and a frequency adjustment unit; the sample recording unit is used to record the suggestion content after user confirmation, the source data, and the user's feedback as training samples for the large language model; the frequency adjustment unit is used to calculate the suggestion acceptance rate and interaction matching degree, and adjust the display frequency according to the suggestion fit.

[0043] The event chain module further includes an event chain library construction unit and a matching update unit; the event chain library construction unit is used to construct the event chain library; the matching update unit is used to calculate the matching degree between the actual operation event chain and the operation event chains in the event chain library, recommend the final selected component type in the operation event chain with the highest matching degree, and update the event chain library; the local training module further includes a parameter adjustment unit and a local training unit; the parameter adjustment unit is used to obtain the user's local network speed and computing power data to adjust the learning rate of the large language model training; the local training unit is used to train the large language model locally on the user's local machine according to the adjusted learning rate.

[0044] In this solution, a component refers to a reusable element that encapsulates specific business logic, interface elements, or interactive functions. It serves as the core foundation for users to build applications through visual operations on a low-code platform. It will be apparent to those skilled in the art that this invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered exemplary and non-limiting in all respects. The scope of the invention is defined by the appended claims rather than the foregoing description, and all changes falling within the meaning and scope of equivalents of the claims are intended to be included within this invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An automated process control method for a low-code platform, characterized in that: The method includes the following steps: Step S1: Obtain the current page browsed by the user on the low-code platform, the selected component type and its attributes, the bound data fields and their data types, and the content input and selected by the user as source data according to a preset frequency, and generate a snapshot of the source data; concatenate the predefined prompt words with the source data as input data for the large language model; Step S2: The large language model outputs structured suggested content based on the input data; the suggested content generated by the large language model is categorized and displayed; the user confirms the suggested content; Step S3: After the user confirms the suggested content, adjust the frequency of displaying the suggested content by category; and record the suggested content, the source data, and the user's confirmation of the suggested content as training samples for the large language model. Step S4: Construct an event chain library and record the user's current actual operation event chain. Calculate the matching degree between the operation event chains in the event chain library and the actual operation event chain. Recommend the final selected component type in the operation event chain with the highest matching degree as predicted content to the user. The user can choose whether to display it. If the user chooses to display it, the component information in the predicted content is shown to the user, and the actual operation event chain is stored in the event chain library. If the user refuses to display it, the final selected component types in the operation event chains are recommended to the user in descending order of matching degree. Step S5: Obtain the user's local network speed and computing power data, adjust the learning rate of the large language model based on the network speed and computing power data, and train the large language model on the user's local machine.

2. The automated process control method for a low-code platform according to claim 1, characterized in that: Step S1 includes: Step S1-1: Classify components according to their functions to obtain component types, and use the parameters corresponding to the components as component attributes; convert source data into text data and generate a snapshot of the source data for archiving; predefine prompt words, which are used to limit the large language model to read the source data and provide suggestions, and the prompt words are text data; trim the prompt words to meet the length limit when the large language model reads them. Step S1-2: Preload the external formula. The external formula is a program used to parse dynamic content into structured data that can be directly read by the large language model. The dynamic content is logical language that cannot be directly determined by static text or fixed rules and changes in real time according to the context, and cannot be directly read by the large language model. Add a label to the prompt word to call the external formula, and concatenate the prompt word with the label to call the external formula with the source data in sequence as the input data of the large language model.

3. The automated process control method for a low-code platform according to claim 1, characterized in that: Step S2 includes: The input data is transmitted to the large language model. After reading the input data, the large language model determines the user's operational intent based on the prompt words and preset keywords in the source data, and maps the natural language in the user's input and selection into structured data. The operational intent includes the page the user is about to browse, the type of component to be selected and its attributes, the data field to be bound and its data type, and the content to be input and selected. The large language model outputs structured suggestion content based on the operational intent and the mapped structured data. The suggestion content generated by the large language model is categorized according to the corresponding operational intent and displayed to the user. The process of the user confirming the suggestion content is as follows: the user chooses to accept or reject the suggestion content; if the user chooses to accept, the suggestion content is retained in a separate window, which the user can actively close; if the user chooses to reject, the display of suggestion content ends.

4. The automated process control method for a low-code platform according to claim 1, characterized in that: Step S3 includes: Step S3-1: Set the frequency adjustment period and suggestion detection time. The initial frequency adjustment period starts from the first time the suggestion content is shown to the user. When the suggestion content is not shown to the user within one suggestion detection time, it is considered to be the end of one frequency adjustment period. Calculate the ratio of the number of times the user accepts the suggestion content to the total number of times the suggestion content is shown to the user within one frequency adjustment period as the suggestion acceptance rate R. Record the interaction time interval from showing the suggestion content to the user to choosing to accept or reject the suggestion content. Set a baseline time interval. Calculate the ratio of the user's average interaction time interval to the baseline time interval within one frequency adjustment period as the interaction matching degree M. Set an interaction time interval threshold. If the interaction time interval exceeds the interaction time interval threshold, it is considered invalid data and is not included in the calculation of the average interaction time interval. Step S3-2: Weight the suggestion acceptance rate R and the reciprocal of the interaction matching degree M to obtain the suggestion fit degree S: ； Where 'a' represents the weight of the suggestion acceptance rate, and 'b' represents the weight of the reciprocal of the interaction matching degree. The sum of 'a' and 'b' is 1. When the suggested fit is less than the preset fit threshold, the frequency of showing suggested content to the user will be reduced by a preset amount; the user's feedback will be sent to the large language model, and the suggested content, source data and confirmation of the user's suggested content will be archived as training samples for the large language model.

5. The automated process control method for a low-code platform according to claim 1, characterized in that: Step S4 includes: Step S4-1: Number the different windows of the low-code platform to obtain a window sequence; classify the different windows according to the component types contained within the windows to obtain window types; construct an event chain library, in which the operation event chain consists of the window sequence opened by the user, the window type, and the finally selected component type; the finally selected component type is the component type that the user selected and did not perform an undo operation within a preset detection time; an operation event chain is the sequence from the first time a user finally selects a component type to the next time a user finally selects a component type. Step S4-2: Set the event chain prediction start threshold n. When a user continuously opens n different window sequences of the low-code platform without ultimately selecting a component type, extract the window sequence and window type opened by the user since the last component type selection, as the user's actual operation event chain; calculate the weighted edit distance between the user's actual operation event chain and the window sequence and window type of the operation event chain in the event chain library using the edit distance algorithm as the matching degree. Window sequences and window types are encoded separately, and the window sequences are converted into window sequence lists and the window types into window type lists according to the opening order. The edit distances D1 and D2 are calculated between the window sequence lists and window type lists of the user's actual operation event chain and the corresponding operation event chain lists and window sequence lists in the event chain library, where D1 represents the edit distance of the window sequence lists and D2 represents the edit distance of the window type lists. D1 is normalized to D1' based on the historical largest edit distance of the window sequence list, and D2 is normalized to D2' based on the historical largest edit distance of the window type list. D1' and D2' are weighted and summed to obtain the comprehensive distance D. The matching degree P = 1 - D. The operation event chain with the highest matching degree is extracted, and the final selected component type in that operation event chain is recommended to the user as predicted content. The user can then choose whether to directly select that component type. Step S4-3: During the recommendation process to the user, the newly generated window sequence and window type are added to the matching degree calculation process, and the new predicted content is continuously displayed to the user at the frequency mentioned above until the user actively stops the predicted content display process; when the final selected component type is generated, the actual operation event chain is stored as the operation event chain in the event chain library.

6. The automated process control method for a low-code platform according to claim 1, characterized in that: Step S5 includes: The system acquires real-time network speed and computing power data from the user's local machine, and obtains maximum network speed and computing power data according to a preset training period. The real-time network speed and computing power data are then normalized based on the maximum network speed and computing power data, and a weighted calculation is performed to adjust the learning rate for training the large language model. The training of the large language model is completed in the background on the user's local machine. ； Where t represents the number of training iterations; This represents the adjusted learning rate for the t-th iteration; This represents the initial learning rate; and All are decay coefficients, used to adjust the learning rate during the initial stages of training; C norm N represents the normalized real-time computing power data; norm α represents the normalized real-time network speed; α and β represent the computing power weight and network speed weight, respectively, and their sum is 1.

7. An automated process control system for a low-code platform, which is applied to the automated process control method for a low-code platform as described in any one of claims 1-6, characterized in that: The system includes an input construction module, a suggestion processing module, a sample frequency tuning module, an event chain module, and a local training module. The input construction module generates source data and concatenates prompt words as input data for the large language model. The suggestion processing module outputs structured suggestions from the large language model and displays them to the user for confirmation. The sample frequency tuning module records training samples and adjusts the display frequency of suggested content based on user feedback. The event chain module constructs an event chain library and displays predicted content based on the matching degree between actual operation event chains and the operation event chains in the library. The local training module adjusts the learning rate based on the user's local network speed and computing power data and trains the large language model locally. The output of the input construction module is connected to the input of the suggestion processing module; the output of the suggestion processing module is connected to the input of the sample frequency tuning module; the output of the sample frequency tuning module is connected to the input of the event chain module; and the output of the event chain module is connected to the input of the local training module.

8. An automated process control system for a low-code platform according to claim 7, characterized in that: The input construction module also includes a source data generation unit and a prompt processing unit; The source data generation unit is used to obtain in real time the current page browsed by the user on the low-code platform, the selected component type and its component attributes, the bound data fields and their data types, and the content input and selected by the user as source data; The prompt processing unit is used to predefine prompt word templates and process prompt words, and preload external formulas to concatenate input data.

9. An automated process control system for a low-code platform according to claim 7, characterized in that: The suggestion processing module further includes a suggestion generation unit and a display confirmation unit; the suggestion generation unit is used by the large language model to determine the operation intent based on the input data and output structured suggestion content; the display confirmation unit is used to classify the suggestion content and display it for user confirmation; the sample frequency tuning module further includes a sample recording unit and a frequency adjustment unit; the sample recording unit is used to record the suggestion content after user confirmation, the source data, and the user's feedback as training samples for the large language model; the frequency adjustment unit is used to calculate the suggestion acceptance rate and interaction matching degree, and adjust the display frequency according to the suggestion fit.

10. An automated process control system for a low-code platform according to claim 7, characterized in that: The event chain module further includes an event chain library construction unit and a matching update unit; the event chain library construction unit is used to construct the event chain library; the matching update unit is used to calculate the matching degree between the actual operation event chain and the operation event chain in the event chain library, recommend the final selected component type in the operation event chain with the highest matching degree, and update the event chain library; the local training module further includes a parameter adjustment unit and a local training unit; The parameter adjustment unit is used to obtain the user's local network speed and computing power data to adjust the learning rate of the large language model training; the local training unit is used to train the large language model on the user's local machine according to the adjusted learning rate.

Images