Business function dynamic adaptation processing method based on AI interaction

Abstract

The invention relates to the technical field of system development, and particularly discloses a service function dynamic adaptation processing method based on AI interaction, which comprises the following steps: acquiring industry service data, and training a pre-selected large model; determining a preset standard function platform data requirement; according to data requirements, formulating an output format specification of the large model; identifying the components of the standard function module with expandability in the standard function platform, and performing standardized packaging and definition; obtaining a business requirement input by a user through a natural language through the large model, and generating preset format data meeting a standard function platform input requirement; and the standard function platform analyzes the preset format data, calls the packaged component based on the analyzed data, and generates a standard function module required by the target service system through pre-written generation logic. By adopting the technical scheme of the invention, the multi-modal learning process can be simplified, the system building process can be simplified, and the service system meeting the user requirements can be generated.

Description

Technical Field

[0001] The present invention relates to the technical field of system development, and particularly to a method for dynamically adapting and processing business functions based on AI interaction. Background Art

[0002] Currently, there are significant deficiencies in the technical integration and engineering implementation of AI-based system automation applications, which restricts the large-scale application of AI technology in complex business scenarios. The existing problems include:

[0003] 1. Difficulties in multi-modal learning: Firstly, the labor and time costs are high. In the current conventional mode of applying AI to system automation, it is necessary to learn three types of widely different materials simultaneously, namely industry business documents, program codes, and database structures. However, multi-modal learning faces a large amount of manual annotation work, which not only requires a large amount of manpower but also consumes a lot of time. For example, when processing natural language descriptions in business documents, accurately annotating their corresponding relationships with program codes and database structures requires a lot of effort from professionals. Secondly, the data fusion is difficult. There are significant complexities and differences in the structures, semantic expressions, etc. of different modal data, resulting in the learning effect often failing to meet the expected goals. Effectively integrating the abstract concepts in business documents, the logical structures of program codes, and the storage relationships of database structures has become a major problem in multi-modal learning. Then, there are data security risks leading to difficult data acquisition. In the process of multi-modal data processing, due to the involvement of various types of data such as business documents, program codes, and database data, their sources are extensive and storage methods are diverse. During the data collection, transmission, storage, and use processes, they are extremely vulnerable to security risks such as data leakage and tampering. For example, program codes may involve the intellectual property rights of key algorithms, and database structures are associated with a large amount of business data. Once these data are leaked, it will cause huge losses to the enterprise. Due to considerations of risks such as data security, privacy protection, and compliance, enterprises are often reluctant to submit these materials. This results in AI being unable to obtain sufficient such materials for effective learning, and thus becomes one of the important reasons why AI cannot successfully learn and develop business systems. Finally, the model generality is poor. The existing multi-modal learning models have poor generality for different industries and different business scenarios. Many models perform well in specific fields or specific datasets, but when applied to other fields or different business scenarios, they often require a large amount of re-adjustment and training and cannot quickly adapt to the new environment. This is because the term expressions, program code styles, and database structures vary greatly in different industries, and the existing multi-modal learning methods are difficult to extract universal features and patterns, restricting their wide application in diverse business scenarios.

[0004] 2. Poor code readability and difficult operation and maintenance: First, the cost of code understanding and operation and maintenance is high. The code generated by AI based on business usually requires professional programmers to read and understand, which brings great inconvenience in the subsequent operation and maintenance process of the business system. The labor cost of professional programmers is relatively high, and due to the uniqueness of the code generated by AI, additional time may be required for code interpretation and maintenance work, which undoubtedly significantly increases the application threshold of AI programming. Taking the business system of large enterprises as an example, if it frequently relies on professional programmers to interpret and maintain the code generated by AI, it will not only greatly increase the operation and maintenance cost, but also may lead to delays in system fault repair due to excessive interpretation time, which will have an adverse impact on the normal operation of the enterprise's business. Second, it is difficult for team collaboration and knowledge inheritance. Since the style and logic of the code generated by AI may be quite different from traditional code, there will also be many difficulties in team collaborative development and knowledge inheritance. After new members join the team, it is difficult to understand and master the code generated by AI, which is not conducive to the efficient collaboration of the team and the effective inheritance of knowledge. Finally, the code scalability is insufficient. Existing AI programming has deficiencies in code scalability. As the business develops and changes, the business system needs to continuously add new functions and optimize existing functions, which requires the code to have good scalability. However, the code generated by AI often lacks sufficient consideration for future business changes during design, and the code structure is not flexible enough to be effectively extended and modified. For example, when an enterprise needs to add a new business module to an existing business system, it may be necessary to rewrite and adjust a large amount of code generated by AI, which not only increases the development cost and time, but also may introduce new errors and risks.

[0005] 3. There are problems with cross-platform compatibility: With the advancement of enterprise digital transformation, business systems often run on a variety of different platforms and environments, including different operating systems (such as Windows, Linux, Mac OS, etc.), hardware architectures (such as x86, ARM, etc.), and cloud platforms (such as Alibaba Cloud, Tencent Cloud, AWS, etc.). However, the code generated by AI performs poorly in terms of cross-platform compatibility. Due to the different characteristics and requirements of different platforms, the code generated by AI may run normally on one platform, but encounter compatibility problems on other platforms, such as the code cannot be compiled, runtime errors, performance degradation, etc. For example, for a business system that needs to run on both mobile and PC sides, the code generated by AI may run well on the Windows system of the PC side, but have display abnormalities, function unavailability, etc. on the Android or iOS system of the mobile side. This not only increases the development cost and operation and maintenance difficulty of the enterprise, but also limits the application scope and flexibility of the business system.

[0006] In summary, these problems are intertwined and jointly hinder the efficient implementation and wide promotion of AI-driven automation. Therefore, there is an urgent need for a method for dynamically adapting and processing business functions based on AI interaction that can effectively solve the above problems. Summary of the Invention

[0007] The present invention provides a method for dynamically adapting and processing business functions based on AI interaction, which can simplify the multi-modal learning process, streamline the system setup process, and generate a business system that meets user requirements.

[0008] To solve the above technical problems, the present application provides the following technical solutions:

[0009] The method for dynamically adapting and processing business functions based on AI interaction includes the following:

[0010] Obtain industry business materials, train a pre-selected basic large model based on the business materials, and obtain a trained large model;

[0011] Determine the format, content, and semantic requirements of each standard function module for input data in the preset standard function platform as data requirements;

[0012] According to the data requirements of the standard function platform, formulate an output format specification for the large model;

[0013] Conduct a structural analysis of the pre-stored function components, identify the components with scalability, and perform standardized encapsulation on the components;

[0014] Obtain the business requirements input by the user through natural language by the large model, combine with the default prompt words, and generate preset format data that meets the input requirements of the standard function platform according to the preset output format specification;

[0015] The standard function platform parses the preset format data according to the predefined data format, and based on the parsed data, calls the encapsulated components, and generates the standard function modules required for the target business system through the pre-written generation logic.

[0016] Further, it specifically includes:

[0017] S1. Select a basic large model with semantic parsing ability and natural language interaction; determine whether the basic large model meets the requirements for business system development. If it meets, jump to step S4; if it does not meet, jump to step S2;

[0018] S2. Collect industry business materials, including business process documents and industry standard specifications;

[0019] S3. Train the basic large model based on the business materials to obtain a trained large model;

[0020] S4. Determine the format and content requirements of each standard function module in the preset standard function platform for the input data as data requirements.

[0021] S5. Set default prompt words according to the data requirements of the standard function platform and formulate the output format specification for the large model.

[0022] S6. Integrate the output format specification into the interaction interface between the standard function platform and the large model.

[0023] S7. Conduct a structural analysis of the pre-stored function components, identify the components with scalability, perform standardized encapsulation on these components, and place the encapsulated components in the standard function platform.

[0024] S8. Obtain the business requirements input by the user through natural language via the large model, generate a natural language description corresponding to the standard function framework, and after obtaining user confirmation, combine with the default prompt words and generate preset format data that meets the input requirements of the standard function platform according to the preset output format specification.

[0025] S9. The standard function platform obtains the preset format data, parses the preset format data according to the predefined data format, and based on the parsed data, calls the encapsulated components and generates the standard function modules required by the target business system through the pre-written generation logic.

[0026] Furthermore, it also includes:

[0027] S10. Conduct a comprehensive verification and testing on the automatically generated standard function modules to ensure the integrity and accuracy of their functions.

[0028] Furthermore, in step S2, it also judges whether the data included in the business materials needs to be labeled. If it needs to be labeled, the data included in the business materials is labeled.

[0029] Furthermore, in step S4, the standard function modules include a menu module, a form module, a workflow module, and a report module.

[0030] Furthermore, in step S8, the preset format data is JSON, XML, or YML.

[0031] Furthermore, in step S9, the standard function platform parses the generated preset format data, and the data format used by the standard function platform is consistent with the preset format data generated by the large model according to the output format specification.

[0032] Furthermore, in step S7, the configuration items of each standardized encapsulation module are also defined.

[0033] Furthermore, the configuration items include attributes, methods, and events.

[0034] The data unification method of this solution greatly optimizes the learning process of large models. The large model no longer needs to deal with complex multi-modal data and only needs to focus on a single data modality, significantly reducing the complexity and uncertainty in the learning process. At the same time, it greatly reduces the large amount of manual annotation work brought by multi-modal. Due to the unification of data formats, the annotation work can be carried out more efficiently and accurately, thus effectively improving the efficiency and quality of large model learning.

[0035] When the large model based on the data learning of unified business materials generates data in this solution, it can better balance business requirements and program logic. The large model can analyze the preset format data used in the business implementation of standard components in the standard function platform, and then restore the business description implemented by its standard components through natural language, making it highly readable. Users can communicate with the large model based on natural language to put forward maintenance requirements. The large model can modify the preset format data according to the maintenance requirements, and then transfer the preset format data to the standard function platform. The standard components in the standard function platform generate the business function modules required by the target business system according to the new preset format data. The unified representation of preset format data between the standard function platform and the large model also improves the operation efficiency. During the operation and maintenance process, non-professional programmers or even business personnel can relatively easily understand and modify the business system through natural language communication, reducing the dependence on professional programmers and effectively lowering the application threshold of AI programming.

[0036] On the one hand, this transformation brought by this solution greatly improves the development efficiency, effectively reducing the communication cost and understanding deviation between business personnel and developers. The long process of repeatedly communicating, confirming requirements and writing code in the past can now be quickly achieved through the large model, enabling business requirements to be transformed into actual business system implementation and application faster; on the other hand, it significantly enhances the flexibility and innovation of the business. Business personnel can adjust business logic in a timely manner according to market changes and business requirements without waiting for developers to modify and deploy code, enabling enterprises to respond to market dynamics more quickly and launch innovative business models and solutions.

[0037] In summary, this solution realizes the transformation from AI code programming to AI business programming, which is not only an innovation in programming methods, but also a vivid manifestation of a business-oriented innovative thinking mode. It deeply integrates AI large model technology with business, injecting strong impetus into the digital transformation and business development of enterprises, and helping enterprises stand out in the fierce market competition. Brief Description of the Drawings

[0038] Figure 1 It is a flowchart of the first embodiment of the method for dynamically adapting and processing business functions based on AI interaction;

[0039] Figure 2 Flowchart of Embodiment 3 of the method for dynamically adapting and processing business functions based on AI interaction;

[0040] Figure 3 Logic block diagram of Embodiment 3 of the method for dynamically adapting and processing business functions based on AI interaction. Detailed implementation manners

[0041] The following is a further detailed description through specific implementation manners:

[0042] Embodiment 1

[0043] As Figure 1 shown, the method for dynamically adapting and processing business functions based on AI interaction in this embodiment includes the following:

[0044] S1. Determine the basic large model

[0045] Select a basic large model with semantic parsing ability and capable of realizing natural language interaction. Judge whether the basic large model meets the requirements of business system development. If it meets, jump to step S4; if it does not meet, jump to step S2;

[0046] Specifically, when building a business system, conduct a comprehensive evaluation of various open-source or commercial basic large models on the market, covering key elements such as model architecture, performance, pre-training tasks, and actual application effects in various scenarios. The finally selected basic large model needs to meet the core condition of having strong semantic parsing ability and being able to achieve natural and fluent natural language interaction. For example, LLaMA, ChatGPT, DeepSeek, etc.

[0047] In this embodiment, a large model using the Transformer architecture is used as the basic model. Because it has efficient parallel computing ability, it performs well in language understanding and generation and achieves outstanding results in natural language processing tasks. In view of the uniqueness of industry business requirements, adaptively transform its architecture, including:

[0048] Add domain-specific layers: Embed multiple domain-specific attention layers in the middle layer of the large model. The common attention layers in the Transformer architecture are self-attention, cross-attention, and causal attention. The domain-specific attention layer in this solution is improved based on the multi-head self-attention layer. For example, in the application of the financial industry, adjust the weight distribution of attention calculation, introduce industry prior knowledge. For financial terms and words related to business logic, give higher weights when calculating attention scores, so that the model can accurately focus on industry semantic information and enhance the understanding ability of industry terms and business logic.

[0049] Optimize the embedding layer: According to the unique characteristics of industry data, specifically optimize the word embedding layer, retrain the word vectors, and make them more accurately reflect the semantic relationships between industry vocabulary, thereby effectively improving the model's representation ability for industry texts.

[0050] Improve the output layer: Improve the original Softmax function by introducing an adaptive weight mechanism. This mechanism can dynamically adjust the weight distribution of the output results according to different business scenarios and task requirements, thereby improving the accuracy and practicality of the generated text.

[0051] In the scenarios of other embodiments, it is also possible to choose to train a basic large model based on its own data. The goal of the training is to enable the model to have good semantic understanding ability and meet the needs of subsequent understanding of business materials. It is also not limited to a specific model type or framework.

[0052] That is, in this solution, users have flexible choices. Even if they use existing un-fine-tuned AI large models, they can still develop business systems by means of the methods in this solution. Compared with the fine-tuned large models, the business systems developed by un-fine-tuned large models tend to be more general and suitable for more common business scenarios in terms of functions and understanding of user businesses. While the large models that have been specifically fine-tuned can be deeply integrated into the specific business processes of users, have a more accurate grasp of business details, and thus the developed business systems have a higher degree of fit with their own business needs and can better meet the personalized and differentiated business needs of users.

[0053] S2. Collect and preprocess industry business materials

[0054] Collect industry business materials, including but not limited to business process documents, industry standard specifications, etc., and also include various relevant research reports, etc., without involving the provision of sensitive information.

[0055] Business process documents: Detail the operation processes of the company's various businesses, from the initiation of the business, intermediate links to the final completion, and clarify the operation steps, involved personnel, division of responsibilities, and time nodes of each link. For example, the procurement business process document will record the entire process from the procurement requirement submission, supplier screening, procurement contract signing, goods acceptance, warehousing to payment.

[0056] Industry standard specifications: Include the generally adopted technical standards, quality specifications, safety requirements, etc. in the industry to ensure that the company's products or services comply with industry guidelines. For example, in the machinery manufacturing industry, the standard specifications may involve product dimensional tolerances, material property indicators, surface roughness requirements, etc.

[0057] Various related research reports: Written by internal or external professional institutions of the company, covering industry dynamics, market trends, technological developments, competitor analysis, etc. For example, a research report on the application prospects of new technologies in the industry will analyze the advantages of new technologies, application cases, and their potential impacts on the company's business.

[0058] Clean and preprocess the data contained in the collected business materials, removing duplicate content, error information, and irrelevant data to ensure the accuracy, integrity, and consistency of the data.

[0059] Judge whether the data contained in the business materials needs to be labeled. If so, label the data so that the large model can better learn and understand business knowledge.

[0060] Specifically, the judgment of whether to label includes:

[0061] Judgment based on the association between business materials and system functions: Since the task of the large model is to generate the standard functions of the target business system based on business materials, it is necessary to judge which content in the business materials is related to the standard functions. The standard functions include menu functions, form functions, process functions, report functions, etc., corresponding to menu, form, process, report, etc. content. For example, in a business process document, if the steps of the approval process are clearly involved, it is closely related to the process function in the target business system and needs to be labeled. Another example is in historical business data. If there is data on the performance of different departments, which is related to the report function, it also needs to be labeled.

[0062] Judgment based on the integrity of system functions: To ensure the integrity of the standard functions of the generated target business system, the content in the business materials that can supplement the system function details should be labeled. For example, in the customer feedback record, it is mentioned that it is hoped to be able to conveniently query the order status in the system, which is related to the form and menu functions and needs to be labeled for the large model to learn.

[0063] Labeling includes:

[0064] Function association labeling: For business process documents, label the standard function corresponding to each process step. For example, in the sales business process, the step of "customer placing an order" is labeled as "associated with the order management form, order generation process"; the step of "order review" is labeled as "associated with the review process, review result report". For historical business data, label the types of reports that the data can be used to generate. For example, financial historical data is labeled as "can be used to generate financial reports (monthly income and expenditure reports, annual profit reports, etc.)".

[0065] User requirement labeling: In the customer feedback record, label the user's requirements for the standard functions. For example, the customer feedback "hoping to quickly see the progress of each project in the system" is labeled as "project progress query requirement, associated with the project management menu, project progress report".

[0066] Data source annotation: In various business documents, mark the data sources used to generate standard functions. For example, in industry standard specifications, if there is data on product quality inspection standards, mark it as "Product quality inspection standard data, the data source for generating quality inspection forms and quality report statements."

[0067] S3. Train the basic large model based on business documents

[0068] Divide the preprocessed business documents into a training set, a validation set, and a test set according to the first preset ratio. In this embodiment, they are divided into a training set, a validation set, and a test set according to the ratios of 80%, 10%, and 10% respectively. Among them, 80% of the training set is used for training the basic large model to enable it to fully learn industry business knowledge; 10% of the validation set is used to evaluate the performance of the basic large model in real time during training, and then guide the adjustment of hyperparameters; the final 10% test set is used to finally evaluate the generalization ability of the trained basic large model. In other embodiments, considering the complexity of factors such as data characteristics and business requirements, the first preset ratio can be fine-tuned according to the actual situation to ensure that the dataset division can fit the actual needs of the project, and then provide more scientific and reasonable data support for subsequent model training, validation, and testing.

[0069] Use the training set to train the selected basic large model. During the training process, set hyperparameters, including the learning rate, the number of training epochs, the batch size, etc.

[0070] Specifically, the key hyperparameter setting methods are as follows:

[0071] Learning rate: The initial value is set to 0.0005, and the cosine annealing decay strategy is adopted, that is, as the number of training epochs increases, the learning rate gradually decreases in the form of a cosine function. This strategy allows the large model to quickly explore the parameter space in the early stage and finely adjust the parameters in the later stage, avoiding falling into local optimal solutions.

[0072] Number of training epochs: Initially set to 50 epochs. During the training process, closely monitor the loss value and accuracy of the validation set. If the decrease in the validation set loss value is less than 0.01 within 5 consecutive epochs and the accuracy does not increase significantly, terminate the training early to prevent overfitting.

[0073] Batch size: Initially set to 64. During training, dynamically adjust according to the GPU memory occupancy and the convergence speed of the large model. If the GPU memory is sufficient and the large model converges slowly, increase the batch size to 128; if the large model oscillates or the GPU memory is insufficient, reduce the batch size to 32.

[0074] By continuously optimizing the training process, the large model can deeply learn the industry's business knowledge, enabling its vector data to fully contain the in-depth information of the industry's business, so that it can accurately understand and process various tasks related to this industry.

[0075] After the large model is trained, its input is usually text data related to the industry, such as business process descriptions, fragments of industry standard documents, customer requirement texts, etc.

[0076] When communicating with users about business, the large model will output preset format data that can be received by the standard function platform according to the user's business needs, in order to meet the requirements of the standard function platform to generate corresponding standard functions. The preset format data needs to fit the data content requirements and clearly show the structure and composition of the data. For example, if the data requirements cover 10 items of data, the corresponding data format must accurately reflect the format characteristics of each of these 10 items of data.

[0077] In this embodiment, the preset format data is JSON data, which organizes the data in a simple and clear key-value pair form, making the format of each item of data obvious at a glance; in other embodiments, the data format can also be XML, YML, etc. XML defines the data structure by means of nested tags, which can clearly present the hierarchical relationship and format of each item of data; YML uses unique indentation and specific syntax rules to accurately present the format of each item of data.

[0078] Example of JSON data:

[0079]

[0080]

[0081]

[0082] "moduleType":"datetime"

[0083] },

[0084] {

[0085] "label":"rating",

[0086] "moduleType":"rate"

[0087] },

[0088] {

[0089] "label":"color selection",

[0090] "moduleType":"color"

[0091] },

[0092] {

[0093] "label": "Step bar",

[0094] "moduleType": "steps"

[0095] },

[0096] {

[0097] "label": "Timeline",

[0098] "moduleType": "timeline"

[0099] },

[0100] {

[0101] "label": "Transfer box",

[0102] "moduleType": "transfer"

[0103] },

[0104] {

[0105] "label": "Progress bar",

[0106] "moduleType": "progress"

[0107] S4. Determine the requirements of the standard function platform

[0108] Determine the format and content requirements of the input data for each standard function module in the standard function platform as data requirements. The standard function modules include menu modules, form modules, workflow modules, report modules, etc.;

[0109] For example, in the menu module, it is required to include the representation methods of information such as defining menu names, menu types, menu hierarchy relationships, and operations corresponding to menu items;

[0110] In the form module, it is required to include the description formats of information such as clarifying the types, labels, attributes, and validation rules of form components.

[0111] The purpose is to unify the interaction data standards between the standard function platform and the large model, that is, the data format standards generated by the large model and the data format standards parsed by the standard function platform need to be consistent.

[0112] Similarly, the standard function platform sends existing menus, forms, workflows, etc. to the large model in a consistent standard for the large model to analyze and modify.

[0113] In this solution, the data requirements of different standard function platforms may be inconsistent, and the data requirements of different standard function modules may also be inconsistent. For example, the interactive data requirements of menus and forms are different.

[0114] In this embodiment, different data requirements are aggregated into a total JSON file and transmitted to the large model, or the large model transmits the aggregated total JSON file to a certain standard function platform.

[0115] In other embodiments, it is also possible to transmit the JSON of one function based on the user's real-time needs. For example, when the user needs to modify the form structure of a certain function, the standard function platform extracts and transmits the form structure corresponding to this function to the large model in JSON format. After receiving it, the large model processes and adjusts this JSON data according to the modification requirements described in the user's natural language, and then transmits the modified JSON data back to the standard function platform. After receiving the updated JSON, the standard function platform parses the information therein again and converts it into a visual corresponding function form structure to complete the interactive closed loop of the user's needs and ensure the accurate implementation and flexible adjustment of the function.

[0116] S5. Formulate prompt words and format specifications for the large model

[0117] According to the data requirements of the standard function platform, sort out a series of detailed and clear default prompt words and formulate the output format specifications for the large model.

[0118] For example, when asking the large model to generate menu data, the prompt word is "Please output menu data in the following precise JSON format: {"menuName":"Menu name","menuType":"M / C","children":[{"menuName":"Sub-menu name","menuType":"M / C","children":[...]}]}, where menuType being M represents a directory and being C represents a menu. Note that the beginning and end of the output content should not include square brackets []".

[0119] When generating form data, the prompt word is "Please provide form component data in the following JSON array format: [{"label":"Chinese name of form component","moduleType":"Component type","properties":{"Property 1":"Value 1","Property 2":"Value 2"}}]. Common component types include input, select, textarea, checkbox, etc. The property values need to be filled in accurately according to actual needs." Through these clear prompt words and strict format specifications, the large model is guided to output content adapted to the standard function platform.

[0120] S6. Integrate the Prompt Words and Format Specifications of the Large Model

[0121] During the construction of the interaction system between the standard function platform and the large model, the default prompt words and output format specifications formulated will be integrated into the API (Application Programming Interface) specifically designed for the interaction between the two.

[0122] This API has been defined during the program development stage. It runs at the system's bottom layer and is not directly accessible to users. It is mainly responsible for transmitting data between the standard function platform and the large model and controlling the interaction logic, building a standardized and normalized interaction bridge for their communication.

[0123] S7. Analyze and Package Standard Function Modules

[0124] Conduct a structural analysis of the pre-stored function components, identify the components with scalability, and perform standardized packaging on the above components. Define the configuration items of each component in detail. The configuration items include attributes, methods, and events.

[0125] Scalability includes generality (the component can adapt to multiple scenarios) and reusability (the component can be reused across projects) to achieve the decoupling and flexible combination of the component and the specific business.

[0126] For example, after the function is packaged, with the configuration items and simple script extensions, flexible customization can be achieved in multiple dimensions such as the user interface, business logic, and data processing. Whether it is adjusting the menu style, form layout, changing the business process direction, or data storage rules, it can be easily handled. This high degree of flexibility can meet the general requirements of various business scenarios and effectively satisfy the personalized demands of special projects.

[0127] Attributes, methods, and events are all configuration items of the standard function. Specifically,

[0128] Attribute: It is an intuitive description of the appearance and characteristics of various components. Taking the button component as an example, the attributes cover the selection of font styles, the setting of font sizes, and the adjustment of colors. The configuration of these attributes directly determines the visual presentation effect of the button on the user interface and affects the user's interaction experience.

[0129] Method: As an important part of the component's behavioral logic, it defines the scripts to be executed in advance during specific operations. For example, when the form is opened, by configuring the method, scripts such as data loading and permission verification can be executed in advance to ensure that the form can run normally and efficiently after being opened, providing an accurate and error-free interaction environment for users.

[0130] Event: It is a key factor triggering the interaction behavior between components. For example, when the value of an input box component changes, this state change constitutes an event. Based on this event, an operation to trigger the hiding of another component can be configured, thereby realizing the dynamic association and interaction between components and meeting the diverse needs under complex business logics.

[0131] For example, encapsulate the components with menu functions into independent units that can be expanded and collapsed, can respond to click events, and can navigate to corresponding pages;

[0132] Encapsulate the components with form functions into functional units with functions such as data input, verification, and submission.

[0133] In the encapsulation process of this embodiment, the components are made scalable so as to be able to adapt to different business scenarios and user requirements.

[0134] S8. Obtain the business requirements of the user

[0135] Obtain the business requirements input by the user in the conversation through the large model, combine with the default prompt words, and generate preset format data that meets the input requirements of the standard function platform according to the preset output format specification. In this embodiment, it is JSON data.

[0136] Specifically, when the user communicates with the large model about the business in the form of natural language conversation, the large model understands the user's business requirements based on the conversation content and tries to generate natural language descriptions of corresponding standard function frameworks such as menus, forms, processes, and reports. In this process, relying on the business knowledge and language understanding ability learned by itself, the large model initially constructs a standard function architecture that meets the user's requirements. The user can evaluate the generated content to confirm whether the menu and form content is consistent with the business requirements, and for inconsistent requirements, communicate with the large model through natural language to inform the large model to make modifications;

[0137] Once the user confirms that the natural language description of the standard function framework by the large model is consistent with the business requirements, issue an instruction to generate relevant standard functions in the standard function platform to the large model through natural language. The request processing mechanism inside the standard function platform integrates the corresponding default prompt words and output format specification with the user's generation instruction according to the rules preset by the API and transmits it to the large model as complete input information.

[0138] After receiving the request containing complete input information, the large model first parses the default prompt words to further clarify the user's intention and the specific requirements of the standard function platform for the data, and then strictly generates JSON data that meets the input requirements of the standard function platform according to the output format specification.

[0139] In this embodiment, the standard function platform has a modification backup mechanism. When a user modifies a certain function module (such as a form) generated by the standard function platform, the platform will automatically back up the JSON data of the current form structure. After the large model modifies the JSON data, it will pass the generated new JSON data to the platform, and the platform will re-layout the form structure according to the new JSON data. During this process, the user has flexible operation permissions. If the user has a need, they can restore the form structure to the state corresponding to the previously backed-up JSON data either by leveraging the large model or through manual operations, thus easily achieving the restoration of the form structure layout.

[0140] In this embodiment, the standard function platform is also used to count the number of times a user converses with the large model and modifies the standard function framework, and determine whether the number of modifications exceeds a first threshold. If it exceeds the first threshold, the standard function platform inputs additional prompt words to the large model;

[0141] In this embodiment, the first threshold is determined based on the average number of modifications of different users in history (referring to the average number of modifications before a user issues an instruction to generate relevant standard functions in the standard function platform through natural language) combined with a set value; for example, the average number of modifications + 10; the additional prompt words are to generate at least three versions of the standard function framework, including the standard function framework of the user's original business requirements; the standard function framework of the user's original business requirements and the current business common requirements; the standard function framework of the user's original business requirements and random functions; the standard function framework of the current business common requirements is pre-statistically obtained by the standard function platform according to the usage history of all users. For example, the standard functions in the final business system of the e-commerce business are statistically analyzed and ranked to obtain the standard functions of common requirements (in this embodiment, the top 5); if some of the top 5 standard function frameworks are missing from the user's original business requirements, the missing standard function frameworks are added; in this embodiment, the random functions are randomly selected from the standard functions ranked sixth and below in the current business, and at least 3 are selected; to help the user quickly determine the standard function framework and improve the development speed.

[0142] The standard function platform also obtains the user's selection of the standard function frameworks of three versions. If the selected standard function framework includes the user's original business requirements, the modification times are reset to zero and the calculation starts again; if the selected standard function framework includes the original business requirements and the standard functions commonly used in the current business, the current modification times are reduced by 1 / 2 and the modification times continue to be counted; if the selected standard function framework includes the original business requirements and random functions, the standard function platform inputs the first defined prompt word into the large model; through different selections of the user, users with clear requirements (selecting the standard function framework of the user's original business requirements), users with clear requirements but insufficient understanding of the business system functions (selecting the standard function framework of the original business requirements and the standard functions commonly used in the current business), and users with unclear requirements (selecting the standard function framework of the original business requirements and random functions) are screened out;

[0143] The large model introduces a preset number of random functions and the standard functions of the unselected common requirements through natural language for the user to select and add to the standard function framework through natural speech;

[0144] The large model also receives the user's refresh operation, introduces new random functions through natural language, while reducing the number of introductions of new random functions and reducing the number of introductions of the standard functions of the unselected common requirements; for example, 10 random functions are introduced, and there are 3 standard functions of the unselected common requirements. After refreshing, 8 new random functions are introduced, and there are 2 remaining standard functions of the unselected common requirements. Through the continuous reduction of recommendations after refreshing, the user is guided to think about the requirements first, make a careful choice, and then input, reducing the waste of the computing power of the large model.

[0145] Record the user's refresh times. If the user's refresh times reach the set refresh times (for example, 3 times), determine whether the current load of the standard function platform reaches the set load; in this embodiment, 80% of the maximum throughput is used as the load, and in other embodiments, it can be set according to the actual situation.

[0146] If the set load is not reached, set the refresh time interval to a fixed time interval; until the user issues an instruction to generate relevant standard functions in the standard function platform through natural language. It can increase the user's thinking time, so as to select the required standard functions, reduce invalid operations, and reduce the waste of the computing power of the large model.

[0147] If the set load is reached, reset the user's refresh times to zero and start calculating again. The current load of the standard function platform is relatively high, reducing the intervention in the user and avoiding generating a large number of data parsing requirements at the same time.

[0148] Through the above settings, in actual use, users with unclear requirements can be screened out. Without affecting the normal use of the user, it can assist the user in determining the requirements and at the same time reduce the waste of system resources.

[0149] S9. Analyze the data generated by the large model and generate standard function modules;

[0150] The standard function platform obtains the preset format data generated by the large model according to the output format specification. In this embodiment, it is JSON data. In the JSON data, it includes the composition data of several standard functions, such as menu composition data (which functions are included in the menu), form composition data, etc. The standard function platform parses the generated preset format data according to the predefined data format.

[0151] In this embodiment, the standard function platform parses the generated preset format data. The data format JSON used by the standard function platform is the same as the data format JSON generated by the large model according to the output format specification. Therefore, the parsing process is to generate different common standard functions according to the JSON type, including menus, forms, processes, reports, etc.

[0152] Taking the menu composition data as an example, the standard function platform extracts key information such as menu name, menu type, and sub-menu hierarchy relationship; for the form composition data, it identifies the type, label, attributes, and validation rules of each form module, etc.

[0153] Based on the parsed preset format data, the standard function platform automatically calls the encapsulated components and generates the standard function modules required by the business system in batches by writing automated generation logic. The generation logic of the standard function platform is to place one or more modular components in a certain standard function in sequence according to JSON. For example, the menu is data, and the menu program displays the corresponding menu according to the menu data.

[0154] Taking the form component as an example, the form components of the standard function include dropdown boxes, input boxes, radio boxes, timelines, pie charts, line charts, etc. In the actual application scenario, the form components will strictly arrange the above various components in the form in the order of the data set in the JSON information. At the same time, the specific attributes, methods, and triggerable events of each component will make flexible and accurate adjustments with the dynamic change of the JSON data. For example, whether the component needs to be verified, whether there is a linkage relationship between the component and other components in the form, etc. are all determined by the JSON data and change according to it.

[0155] Standard function implementation:

[0156] In system development, there is a category of functions defined as standard functions, such as menu generation, form structure construction, data addition, deletion, modification, and query operations, data import and export functions, process engine operation, report generation, page printing, and mobile interface generation. These functions can either be pre-set or reasonably anticipated during the development process. To improve development efficiency and convenience, this solution encapsulates these standard functions into components. The encapsulated components take JSON data as parameter input and can automatically implement a series of standard functions such as menu generation, form building, data addition, deletion, modification, and query, and data import and export according to the specific content within the JSON. And this JSON data used as parameters is all generated by the large model, making full use of the powerful capabilities of the large model to make the implementation of standard functions more intelligent and efficient.

[0157] In this embodiment, during the target program generation process, strictly following the design specifications and style requirements of the standard function platform, format verification and style adjustment are performed on the generated menu and form to ensure that they can meet the requirements of the business system both in appearance and function. When adjusting, the JSON format specification is verified. If the verification fails, this JSON is not run and an error is prompted.

[0158] For example, the standard function platform requires that the form data generated by the large model be presented in a specific JSON format, stipulating the name, data type, and arrangement order of each form field. The large model follows these specifications for data generation and collation.

[0159] Finally, the large model transmits the generated JSON data that conforms to the specifications to the standard function platform. The standard function platform generates corresponding standard function modules based on this data to implement the business functions required by the user. Through this interaction process, seamless docking and efficient collaboration are achieved between the standard function platform and the large model, significantly improving the accuracy and efficiency of processing user business requirements and providing a more convenient and intelligent development experience for users.

[0160] S10. Verify and optimize the automatically generated standard function modules

[0161] Conduct a comprehensive verification and test on the automatically generated standard function modules to ensure the integrity and accuracy of their functions.

[0162] For example, check whether the navigation function of the menu is normal and whether the click events of menu items can correctly trigger corresponding operations; verify whether the data input, verification, and submission functions of the form meet the expectations and whether the styles of form components can be correctly displayed on different devices and browsers.

[0163] According to the verification and test results, conduct targeted optimization and adjustment on the generated standard function modules, repair possible vulnerabilities and problems, and continuously improve the quality and stability of the generated content.

[0164] For example, when the standard function platform generates a business system through the JSON input by the large model, it automatically attempts to execute and verify the integrity and accuracy of the function. If the execution fails, it notifies the large model and returns an error message. After receiving the error message, the large model generates a new JSON and outputs it to the standard function platform until the business system is successful or the user interrupts.

[0165] The following uses a specific scenario to illustrate the application of the solution of this embodiment:

[0166] Requirement submission

[0167] User: I need an asset management system.

[0168] Large model: Okay, may I ask what industry the asset management system is for?

[0169] User: For the manufacturing industry, the automotive industry.

[0170] Initial menu generation and adjustment

[0171] Large model: Based on the understanding of the automotive manufacturing industry and the learned industry business materials, the menus used in the "asset management system" include asset procurement, asset receipt, asset inventory, asset maintenance, asset scrapping, etc. Do you think these menus meet your needs?

[0172] User: We don't need the asset maintenance menu. Also, why isn't there an asset transfer option?

[0173] Large model: The asset maintenance menu has been removed. At the same time, because of the deletion of this menu, the relevant functions have been adjusted, which may better meet your business needs. In addition, I have added the asset transfer option you want and associated the relevant data flow and operation permission settings according to the business logic. For example, asset transfer involves the transfer of assets between different warehouses, and the corresponding inventory data will be updated in real time, and only personnel with specific permissions can perform asset transfer operations.

[0174] User: Okay, that's fine.

[0175] Menu generation and data item determination

[0176] Large model: Okay, now I'm about to generate the menu (page refreshed, menu generated).

[0177] User: Generate the data items included in this menu. (The user clicks on the first menu and communicates with the large model)

[0178] Large model: The data items for the asset procurement menu have been generated, including: purchase order number, supplier name, purchase date, asset name, specification model, quantity, unit price, total price, etc. Do you think this is appropriate?

[0179] User: Appropriate, just like this.

[0180] Large model: Okay, I will immediately generate the basic functions of adding, deleting, modifying, querying, importing, and exporting for you (the page will be refreshed to generate the corresponding functions).

[0181] Through the above method, the user has created all the menu data items.

[0182] Complex business logic settings

[0183] User: The asset transfer menu is an operation involving asset transfer. It needs to use the table of the asset inventory menu as a sub-table, and the fields of the sub-table need to be totaled. The totaled number cannot be greater than the allocable asset quantity in the asset transfer menu form.

[0184] Large model: The data has been associated and set. Through data-based judgment and calculation logic, your requirements are ensured to be met. Do you think it is appropriate? Specifically, at the data level, by establishing the association relationship between data, when the asset transfer operation occurs, the system will automatically obtain the relevant data in the asset inventory menu for total calculation and compare it with the allocable asset quantity in the asset transfer menu form. If it exceeds, the operation will be prompted to fail.

[0185] User: I also need to send an email notification to role A about the asset transfer situation after the new data is saved in this menu.

[0186] Large model: Okay, the adjustment has been completed through data-based event triggering and email sending logic to ensure the timely and accurate transmission of information. When the asset transfer data is successfully saved, the system will automatically send an email containing the detailed asset transfer information to the email of role A according to the preset email sending logic. The email content includes asset name, specification model, quantity, warehouse from which transferred, warehouse to which transferred, etc.

[0187] Further function expansion and optimization

[0188] User: In the asset inventory menu, I hope to be able to classify and count the asset inventory data according to different workshops.

[0189] Large model: The setting has been completed. When entering the asset inventory data, it will be required to select the corresponding workshop, and the system will store the data classified according to the workshop. When counting, you can generate the asset inventory reports for each workshop with one click. The reports include information such as asset quantity and value.

[0190] User: In the asset procurement menu, for suppliers with whom we often cooperate, we hope to set up a quick procurement option.

[0191] Large model: This function has been added. On the asset procurement page, for suppliers marked as regular partners, a quick procurement entry will be displayed. After clicking, it can quickly fill in supplier-related information, simplifying the procurement process.

[0192] In this scenario, the large model only needs to learn one type of data, i.e., business materials, and can quickly understand the user's requirements regarding the asset management system in the automotive manufacturing industry (i.e., the target business system). Through natural language interaction with the user, it continuously adjusts and generates the business implementation of standard functions that conform to business logic. By using the standard function platform to execute program data, the entire process from menu generation, data item setting to complex business logic association and function expansion is realized, fully demonstrating the efficiency and convenience of the present invention in practical applications.

[0193] The data unification method in this solution greatly optimizes the learning process of the large model. The large model no longer needs to deal with complex multi-modal data and only needs to focus on learning business materials, significantly reducing the complexity and uncertainty in the learning process. At the same time, it greatly reduces the large amount of manual annotation work brought by multi-modal data. Due to the unified data format, the annotation work can be carried out more efficiently and accurately, thus effectively improving the learning efficiency and quality of the large model.

[0194] Taking deep learning in business systems as an example, to enable the large model to understand business systems and infer new business systems, traditional multi-modal learning requires separate annotation and processing of business documents, program codes, data structures, and business data. However, after adopting the method of using only one type of data, i.e., business materials, the large model can learn and train more directly, thus converging to a better solution faster and improving the performance and accuracy of the model.

[0195] In the field of AI programming, this method enables the large model to master business and program logic faster and generate data that better meets actual needs. Since the data learned by the large model has integrated business logic, program logic, and business data, the large model can quickly extract key information from business materials, understand the relationship between business requirements and program implementation, and thus generate data more accurately. In addition, the unified business materials also facilitate the knowledge transfer of the large model and can adapt and learn faster when facing new business scenarios.

[0196] When the large model based on the unified business data learning generates data in this solution, it can better balance business requirements and program logic. Although the generated preset format data such as JSON, XML, and YML does not have high readability, the large model can analyze this data and then restore the business implemented by its components through natural language to make it highly readable. Users can communicate with the large model in natural language to propose maintenance requirements. The large model can modify the data according to the maintenance requirements and then transfer the data to the standard function platform, which generates the standard function modules required by the target business system. The unified data representation between the standard function platform and the large model also improves the operation efficiency. During the operation and maintenance process, non-professional programmers or even business personnel can relatively easily understand and modify the business system through natural language communication, reducing the dependence on professional programmers and effectively lowering the application threshold of AI programming.

[0197] For example, in a small e-commerce business system, business personnel can, according to the description of the business system functions by the large model in natural language, input requirements to the large model through natural language dialogue and adjust some simple business logics by themselves. For example, when there is a new promotion trend in the market, business personnel can immediately adjust the promotion strategy, modify the product promotion rules, and generate corresponding data through the large model and quickly update it to the e-commerce system to attract more consumers without waiting for the intervention of professional programmers.

[0198] Traditional AI programming mainly focuses on writing specific code logics to achieve specific functions, and these codes can only be put into use after being read and modified by professional programmers, which greatly limits the wide application of AI in business systems. By means of the core strategy of data standardization proposed in this solution, a major leap from AI code programming to AI business programming can be achieved. In this solution, business personnel can directly participate in the programming process. Since the data has integrated business logic and program logic, business personnel do not need to be proficient in complex code writing skills. By simply describing business requirements in natural language, the large model can learn and understand based on the unified data and then generate data that meets the requirements. The standard function platform generates the function modules required by the business system according to these data. By docking with the enterprise's business requirements, the large model does not need to access sensitive enterprise core data or privacy information, thus avoiding the risk of data security and privacy leakage. This greatly improves the flexibility and efficiency of system operation and maintenance and enhances the adaptability of AI programming to business systems. This advantage enables enterprises to respond more quickly to business changes, adjust system functions in a timely manner, and enhance business competitiveness when applying AI programming.

[0199] In summary, this solution realizes the transformation from AI code programming to AI business programming. It is not only an innovation in programming methods but also a vivid manifestation of a business-oriented innovative thinking mode. It deeply integrates AI technology with business, injecting strong impetus into the digital transformation and business development of enterprises, and helping enterprises stand out in the fierce market competition.

[0200] Example Two

[0201] The difference between this example and Example One is that in this example, it further includes:

[0202] S0. Construct the overall architecture of the standard function platform

[0203] Construct the overall architecture of the standard function platform, which includes a front-end interaction layer, an intermediate business logic processing layer, and a back-end data persistence layer;

[0204] The front-end interaction layer is used to provide users with an intuitive and convenient visual operation environment, enabling users to complete the preliminary construction of application programs through simple operations such as dragging and configuration;

[0205] The intermediate business logic processing layer is used to convert the operations of users in the front-end interaction layer into executable code logic. In this example, the program is interpreted and executed through the solution in the previously applied patent "A Program Interpretation and Execution Method".

[0206] The back-end data persistence layer is used to interface with a variety of preset databases, store the application program data created by users (including menu structures, form contents, business logic rules, etc.), and also provide fast retrieval and management functions.

[0207] This example maintains openness in technology selection and is not limited to specific programming languages or technology frameworks. The front-end interaction layer can select popular JavaScript frameworks such as Vue.js, React, or Angular to achieve rich user interaction functions; other technical solutions suitable for building visual interfaces can also be adopted.

[0208] The intermediate business logic processing layer can select languages including Java, C#, Python, etc. for development according to the team's technical reserves and project requirements, and build a stable and reliable service in combination with corresponding back-end frameworks (such as Spring,.NET Core, Django, etc.).

[0209] The back-end data persistence layer supports integration with various relational databases such as MySQL, PostgreSQL, SQL Server, as well as non-relational databases such as MongoDB and Redis to meet data storage requirements in different scenarios.

[0210] Example 3

[0211] The difference between this example and Example 1 is that in the solution of Example 1, the user interacts with the large model through natural language to generate standard programs (standard functions), such as menus, forms, processes, reports, etc.

[0212] As Figure 2 and Figure 3 shown, in this example, through the cooperation of the large model and the non-standard function platform, non-standard backend programs are generated, such as application scenarios where the business system needs to call AGV, logistics doors, weighbridges, indicator lights, loudspeakers, etc., or application scenarios such as sending emails and making network requests.

[0213] This example further includes the following steps:

[0214] S11. Collect and organize program data

[0215] Comprehensively collect the program data generated by the non-standard function platform. In this example, the non-standard function platform adopts the solution proposed in the inventor's prior patent application CN117389569A, "A Method for Interpreting and Executing Programs".

[0216] For traditional intermediate languages, after generation, the intermediate language needs to be converted into code, and this conversion process requires compilation or the generated program code needs to be compiled. In the program data of this example, the program is the data and the data is the program; it is not an intermediate language and can directly interpret and execute the program operation through preset logic without the process of code conversion.

[0217] For example, a judgment is a piece of data, the true in the judgment is a piece of data, the false is a piece of data, executing an SQL is a piece of data, set merging is a piece of data, a loop is a piece of data, debug is a piece of data, tcp call is a piece of data, email sending is a piece of data, return is a piece of data, etc. These data are all "program data components", and the program data is composed of these program data components according to business requirements. The preset logic interpretation execution program will parse and execute these data.

[0218] These program data components can also be extended according to the upgrade of the non-standard function platform, such as adding trycatch as a piece of data, executing a new communication protocol as a piece of data, etc.

[0219] The program data can be in various storage formats, which can be stored in files such as JSON, XML, YML, etc. or directly stored in the database.

[0220] Among them, each program data component has corresponding business annotations to explain the business or program flow, facilitating personnel to read the business or program ideas. These business annotations are also used for large model learning, that is, when the large model learns this data, it not only learns the program itself but also the business ideas attached to the program itself.

[0221] In the actual application of other embodiments, it also includes other supporting data, such as program dependency data, connection data between nodes, data flow rule data, etc.

[0222] Preprocess all the collected program data, including unified conversion of data formats, reasonable filling of missing values, and effective handling of outliers, etc., to ensure the integrity and consistency of the data.

[0223] S12. Fine-tune the large model based on the program data

[0224] Divide the preprocessed program data into a training set, a validation set, and a test set according to a preset ratio for targeted fine-tuning training of the large model.

[0225] The specific training method can refer to step S4.

[0226] Through continuous fine-tuning training, the large model can deeply learn the data implementation methods, structural features, and business logic relationships contained in the graphical data, so that the vector data of the large model can accurately capture the core essence of the graphical data. By learning the above example program data and other supporting data, the large model can understand the program logic represented by different node_codes and the business meaning behind the text description. Furthermore, when dealing with tasks related to graphical programs, it can more accurately generate and modify program data that meets business requirements.

[0227] Among them, node_code is a type of node identifier in the program data. In a graphical program, each "node" represents a specific program logic or operation, such as a judgment, a loop, an email operation, a TCP operation, etc. node_code is used to uniquely identify these nodes, and it is like the "ID card" of each node. For example, in a graphical program, different running steps can be regarded as different nodes, and each node has a corresponding node_code. Through these identifiers, the program can identify, call, and process the functions of each node. When the large model learns node_code, it is understanding the program logic behind these nodes. For example, a certain node_code may represent a data calculation function, and after the large model learns it, it can know that encountering this node means to perform related calculation operations.

[0228] text: Refers to the textual description of various elements or operations in a graphical program. It is used to more intuitively and elaborately explain the business implications involved in the graphical program. For example, in a graphical program for e-commerce order processing, a certain node may have a text description like "Verify whether the inventory of the order items is sufficient". This text is an explanation of the business meaning of that node. When the large model learns the text description, it is understanding the meaning of the graphical program at the business level, so that when processing related tasks, it can generate or modify program data that meets the actual business needs based on these business implications. For example, when generating an order processing flow, it can ensure that the inventory verification link is included and set according to the correct business logic.

[0229] S13. Read the default program data and display it

[0230] Receive the modification operation initiated by the user for the current graphical program in the non-standard function platform, and determine whether there is already a default graphical program for the current graphical program (such as common operations like adding, deleting, modifying, querying, importing, exporting, etc.);

[0231] If there is already a default graphical program, the non-standard function platform first reads the default program data and displays it in an intuitive visual way to present it to the user.

[0232] Meanwhile, the non-standard function platform provides a simple and clear guiding interface to guide the user to input a detailed business description, such as "After completing the addition of new data, send the new data content to person A via email". At this time, the large model will insert a program data component for email sending before the return component at the end of the program data and configure it to complete the business requirements.

[0233] If there is no default graphical program, jump to step S16;

[0234] S14. Analyze the user's modification requirements and back up the data

[0235] The large model uses natural language processing technology to analyze the user's business description and clarify the specific modification intention and operation content.

[0236] Before making any modifications in this embodiment, the non-standard function platform also automatically makes a complete backup of the original program data to ensure the security and traceability of the data.

[0237] After the backup is completed, modify the program data according to the parsed modification intention, such as adding new program data components, adjusting component attributes, reconstructing the logical relationship between components, etc.

[0238] S15. Test the modified graphical program and optimize it

[0239] After the modification is completed, the non-standard function platform conducts a running test on the modified program data.

[0240] If an error message appears during the test, the non-standard function platform feeds back the current error program data to the large model, along with the error message, the execution result that does not meet the expectation, and the description of the user's original business requirements. The large model attempts to modify based on the feedback, and after the modification, it passes the new program data to the preset logic interpretation and execution program to attempt execution again until the running expectation is reached or the user interrupts.

[0241] S16. Guide the user to describe the requirements and parse them

[0242] When the user modifies the graphical program but there is no default graphical program for the current program;

[0243] The non-standard function platform gradually guides the user to describe the business requirements in detail through the guidance interface, such as "Create a graphical program for production process monitoring, including key links such as raw material procurement, processing and manufacturing, quality inspection, and product out-of-warehouse, and achieve dynamic data interaction between each link."

[0244] The large model parses the user's business description, extracts key information such as process nodes, data flow directions, and interaction logics.

[0245] S17. Build new program data and test it

[0246] Based on the parsed information, the large model attempts to build new program data. Building new program data includes reordering the program data components according to the business requirements, correcting the attribute values of the program data components, and the parameter passing relationships between the components. After the construction is completed, the generated program data is sent to the non-standard function platform for the first running test.

[0247] If an error occurs during the running process, the non-standard function platform sends the error message to the large model, and the large model analyzes the error reasons, including incorrect attribute settings of the program data components, unclear definition of logical relationships, etc.

[0248] Based on the analysis results, the large model automatically makes intelligent adjustments to the program data, such as correcting the attribute values of the program data components, the parameter passing relationships between the components, reconnecting the program data components, etc.

[0249] Because the large model has learned the program data generated by the preset logic interpretation and execution program, therefore, the large model will continuously attempt to combine these program data components according to the user's business requirements and pass the combined program data to the non-standard function platform for execution until the execution result meets the expectation.

[0250] The user can also view the structures and running processes of these "program data" through the graphical program

[0251] Note: The graphical program is just a visual display of these "program data" for easy human reading. In fact, the program data is essentially executable data.

[0252] S18. Iteratively optimize the newly constructed graphical program

[0253] Conduct a running test again and repeat this process continuously until the graphical program can run smoothly on the non-standard function platform and meet the user's business requirements.

[0254] During the iterative optimization process, the non-standard function platform records the content of each modification and the test results, providing a reference basis for subsequent optimization. At the same time, it also facilitates the user to understand the creation process and optimization history of the graphical program.

[0255] S19. Display the graphical program logic and results for user evaluation

[0256] Display the detailed business logic of the graphical program on the user interface of the non-standard function platform. For example, the detailed business logic of the graphical program can be clearly and intuitively displayed on the user interface through visualization tools such as flowcharts and state machine diagrams; through visualization tools such as flowcharts and state machine diagrams, the user can clearly understand the running process and data interaction relationship of the graphical program at a glance.

[0257] At the same time, display the running results of the graphical program in real time, including dynamically generated graphic images, real-time updated data displays, etc. The user can carefully view and evaluate the displayed business logic and running results to determine whether they fully meet their business expectations.

[0258] S20. Save the graphical program and related records

[0259] Receive the user's save operation when it meets the expectations (for example, the user clicks the save button). The non-standard function platform records the detailed modification or creation information of the graphical program, including key metadata such as the operation time, operator, and modification content.

[0260] Save the program data according to the system's storage specifications, complete the modification or creation operation of the graphical program, and provide an efficient and reliable graphical programming solution for the user.

[0261] In other embodiments, the system can also generate corresponding version records, facilitating the user to perform version backtracking and comparison when needed later, ensuring the traceability and maintainability of the entire graphical program development process.

[0262] During the saving process, version management will also be carried out for the program data used in the previous training of the large model, including example data and other supporting data, so that relevant data can be accurately reused during subsequent model optimization or system upgrade, further improving the stability and scalability of the system.

[0263] This solution generates a target program based on the program data generated by the large model. The generated target program can run without compilation, achieving what you see is what you get. In this embodiment, the solution proposed in the inventor's prior patent application CN117389569A "A Method for Interpreting and Executing Programs" (logical interpretation and execution program) is used to generate the target program. The logical interpretation and execution program is pre-written by professionals and executes program data through logical interpretation. In other embodiments, existing systems can also be used to generate the target program.

[0264] The non-standard functions in this solution refer to the functional requirements that are generated due to the dynamic changes in the business and are difficult to pre-plan and foresee. Taking the vehicle weighing business as an example, when a vehicle drives onto the weighbridge, the system needs to synchronously execute operations such as taking pictures, raising the barrier, and simultaneously issuing a voice prompt indicating that the vehicle can enter the weighbridge area at this time; then activate the laser device to check the vehicle's posture, and guide the driver to get out of the vehicle through voice; then start the weighbridge weighing process and accurately write the weighing data into the database; afterwards, prompt the driver to get back in the vehicle, raise the barrier again, and issue a voice prompt for the vehicle to leave. This series of closely related functional requirements specific to this business scenario belong to the category of non-standard functions.

[0265] The solution in this embodiment has powerful intelligent capabilities and can directly generate a target program that can be immediately put into use based on the program data generated by the large model. It is worth mentioning that this target program can run without a cumbersome compilation process, truly realizing the efficient operation experience of "what you see is what you get".

[0266] For the development of non-standard function programs, this implementation uses the innovative solution in the inventor's prior patent (Patent No.: CN117389569A "A Method for Interpreting and Executing Programs"). This solution deeply analyzes and makes full use of the unique characteristics of program data when the business and the program are integrated. Relying on this technical advantage, the AI large model can cleverly bypass the complex multi-modal learning process. In the field of multi-modal learning, problems such as high data annotation costs and difficulties in data fusion have always been obstacles to the development of technology, and this patent solution has successfully circumvented these challenges. Finally, with the help of this solution, the AI large model can efficiently develop non-standard business functions, not only significantly reducing the development difficulty and cost, but also greatly improving the development efficiency and quality, strongly promoting the innovative development of the business.

[0267] By combining the solutions of Embodiment 1 and this embodiment, the standard functions and non-standard functions can be fully developed to adapt to business systems with different requirements.

[0268] For example, in the scenario where the user needs an asset management system in Embodiment 1,

[0269] Large model: The data has been associated and set. Through data-based judgment and calculation logic, it is ensured to meet your requirements. Do you think it is appropriate? Specifically, at the data level, by establishing the association relationship between data, when an asset transfer operation occurs, the system will automatically obtain the relevant data in the asset inventory menu for total calculation and compare it with the allocable asset quantity in the asset transfer menu form. If it exceeds, an operation failure will be prompted.

[0270] Combined with the solution of Embodiment 2, the following non-standard function operations can also be performed:

[0271] User: I also need to send an email notification about the asset transfer situation to Role A after the new data is saved in this menu.

[0272] Large model: Okay, the adjustment has been completed through data-based event triggering and email sending logic to ensure the timely and accurate transmission of information. When the asset transfer data is successfully saved, the system will automatically send an email containing the detailed asset transfer information to the email of Role A according to the preset email sending logic. The email content includes asset name, specification model, quantity, transfer-out warehouse, transfer-in warehouse, etc.

[0273] The product of this solution does not require a complex development environment to be installed. Users only need to download it and then start it with one key, completely eliminating the complexity of development configuration. Users can automatically generate and operate and maintain the business system by communicating with the large model in natural language, truly realizing zero programming. The large model can understand the user's needs, generate the business system, and greatly reduce the user's usage threshold.

[0274] The solution of this embodiment is based on the previously applied "A Method for Program Interpretation and Execution", and innovatively uses program data to replace traditional program code to implement program execution. Program data covers two aspects: program description and business description. In this way, the cost input of the large model in the multi-modal learning process can be effectively reduced. At the same time, this solution greatly simplifies the traditional development process, endows the large model with the ability to automatically generate business systems, enables it to adapt to a variety of complex business scenarios, helps enterprises efficiently complete the construction of business systems, and comprehensively improves the efficiency and quality in the process of enterprise digital transformation.

[0275] This solution has a great cost advantage compared with traditional development in the development mode. With the help of AI large model technology, users can create business systems without writing code, which can reduce the traditional development cost by 90%. Through the automated support of the large model, the development cycle is significantly shortened, and the efficient progress of the project can bring a higher return on investment.

[0276] This solution does not require complex installation and configuration and can be used in any development environment. Users only need to download the product and start it with one click to start using it, and they can operate and adjust it on the full-platform interface. Compared with existing products, this product has a more intuitive operation interface and is suitable for a wide range of user groups. In particular, non-technical personnel can also easily get started.

[0277] This product has a simple and intuitive operation interface, full-platform support, and a user-friendly operation process. The program data generated by it has high visualization effects and powerful cross-platform capabilities, which can bring a more convenient and efficient business system development experience to enterprises.

[0278] In summary, through a highly simplified user experience, this solution eliminates the technical threshold of existing solutions, enabling users to generate business systems through simple operations, significantly reducing development costs, while ensuring data security and privacy. Compared with traditional solutions, it not only improves the efficiency of system development but also provides a more secure, convenient operation and maintenance, and adjustment method, providing an enterprise with a more efficient and flexible solution.

[0279] The above are only embodiments of the present invention. The invention is not limited to the fields involved in this embodiment. Common knowledge such as the specific structures and characteristics known in the solution is not described in detail here. Those of ordinary skill in the art know all the common technical knowledge in the technical field to which the invention belongs before the application date or the priority date, can know all the existing technologies in this field, and have the ability to apply the conventional experimental means before this date. Those of ordinary skill in the art can, under the inspiration given in this application, combine their own abilities to improve and implement this solution. Some typical known structures or known methods should not become an obstacle for those of ordinary skill in the art to implement this application. It should be noted that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can still be made, and these should also be regarded as the protection scope of the present invention, and these will not affect the implementation effect of the present invention and the practicality of the patent. The protection scope required by this application should be based on the content of its claims, and the specific implementation manners described in the specification can be used to explain the content of the claims.

Claims

1. A method for dynamically adapting business functions based on AI interaction, characterized in that: It includes the following: Obtain industry business data, train a pre-selected basic large model based on the business data, and obtain a trained large model; Determine the format and content requirements of input data for each standard function module in the preset standard function platform as data requirements; Formulate output format specifications for large models based on data requirements of standard functional platforms; Perform structural analysis on pre-existing functional components, identify components with scalability, and perform standardized packaging on the components; The big model obtains the business requirements of users inputted in natural language, and generates data in a preset format that meets the input requirements of the standard functional platform according to the preset output format specifications; The standard function platform parses the preset format data according to the pre-defined data format, calls the encapsulated components based on the parsed data, and generates the standard function modules required by the target business system through the pre-written generation logic.

2. The method for dynamic adaptation of business functions based on AI interaction according to claim 1, characterized in that: Specifically include: S1. Select a basic large model with semantic parsing capabilities and natural language interaction; Determine whether the basic large model meets the business system development requirements. If yes, jump to step S4. If not, jump to step S2. S2. Collect industry business information, including business process documents and industry standards and specifications; S3. Train the basic big model based on the business data to obtain a trained big model; S4. Determine the format and content requirements of input data for each standard function module in the preset standard function platform as data requirements; S5. According to the data requirements of the standard functional platform, set the default prompt words and formulate the output format specifications of the large model; S6. Integrate the output format specification into the interaction interface between the standard functional platform and the large model; S7, structurally analyzing the pre-existing functional components, identifying components with scalability, performing standardized packaging on the components, and placing the packaged components on a standard functional platform; S8. The big model obtains the business requirements input by the user in natural language, generates a natural language description of the corresponding standard functional framework, and after obtaining the user's confirmation, the big model generates preset format data that meets the input requirements of the standard functional platform according to the preset output format specification in combination with the default prompt words; S9. The standard function platform obtains preset format data, parses the preset format data according to a predefined data format, calls the encapsulated components based on the parsed data, and generates the standard function modules required by the target business system through pre-written generation logic.

3. The method for dynamic adaptation of business functions based on AI interaction according to claim 2 is characterized in that: Also includes: S10. Comprehensively verify and test the automatically generated standard function modules to ensure their functional integrity and accuracy.

4. The method for dynamic adaptation of business functions based on AI interaction according to claim 3 is characterized in that: In the step S2, it is also determined whether the data included in the business data needs to be marked. If it needs to be marked, the data included in the business data is marked.

5. The method for dynamic adaptation of business functions based on AI interaction according to claim 4 is characterized in that: In step S4, the standard function modules include a menu module, a form module, a workflow module, and a report module.

6. The method for dynamic adaptation of business functions based on AI interaction according to claim 5 is characterized in that: In the step S8, the preset format data is JSON, XML or YML.

7. The method for dynamic adaptation of business functions based on AI interaction according to claim 6, characterized in that: In step S9, the standard function platform parses the generated preset format data, and the data format used by the standard function platform is consistent with the preset format data generated by the large model according to the output format specification.

8. The method for dynamic adaptation of business functions based on AI interaction according to claim 7, characterized in that: In step S7, configuration items of each standardized encapsulation module are also defined.

9. The method for dynamic adaptation of business functions based on AI interaction according to claim 8, characterized in that: The configuration items include properties, methods and events.

Images