Low-code development method, device, electronic equipment, storage medium and program product

Abstract

The disclosure provides a low-code development method, device, electronic equipment, storage medium and program product, relating to the fields of artificial intelligence technologies such as generative model, large language model, low-code technology, etc. The method comprises: obtaining input interface description information of a front-end interface constituting an expected development result; inputting the interface description information into a preset first generative large language model to obtain an interface description language output by the first generative large language model; obtaining input association indication information for associating back-end data constituting the expected development result with the front-end interface; inputting the association indication information into a preset second generative large language model to obtain a database management language output by the second generative large language model; generating a corresponding front-end interface according to the interface description language, and associating corresponding database data with the front-end interface according to the database management language to obtain the expected development result. The method can significantly reduce the threshold requirement for the use of the development user.

Description

Technical Field

[0001] This disclosure relates to the field of data processing technology, specifically to the fields of artificial intelligence technology such as generative models, large language models, and low-code technology, and particularly to a low-code development method, apparatus, electronic device, computer-readable storage medium, and computer program product. Background Technology

[0002] For platform users, in today's era of rapid information development, the market demand for internal system informatization from individuals and enterprises is increasing. However, the field of system platform construction usually has high technical barriers for non-technical personnel, and they can only meet their internal informatization needs by spending a lot of money to build an internal R&D team or by purchasing standardized platform systems.

[0003] Current low-code development platforms still present a high barrier to entry for users or offer limited customization (especially in the details of the user interface), failing to truly meet users' actual needs. Summary of the Invention

[0004] This disclosure provides a low-code development method, apparatus, electronic device, computer-readable storage medium, and computer program product.

[0005] In a first aspect, embodiments of this disclosure propose a low-code development method, comprising: acquiring interface description information as input, which constitutes a front-end interface of the desired development result; inputting the interface description information as input to a preset first generative large language model to obtain an interface description language output by the first generative large language model; acquiring association instruction information as input, which associates back-end data constituting the desired development result with the front-end interface; inputting the association instruction information as input to a preset second generative large language model to obtain a database management language output by the second generative large language model; generating a corresponding front-end interface according to the interface description language, and associating the front-end interface with corresponding database data according to the database management language to obtain the desired development result.

[0006] Secondly, embodiments of this disclosure propose a low-code development apparatus, comprising: an interface description information acquisition unit configured to acquire input interface description information constituting a front-end interface of a desired development result; an interface description language generation unit configured to input the interface description information as input information into a preset first generative large language model to obtain an interface description language output by the first generative large language model; an association instruction information acquisition unit configured to acquire input association instruction information constituting a back-end data of a desired development result and the front-end interface; a database management language generation unit configured to input the association instruction information as input information into a preset second generative large language model to obtain a database management language output by the second generative large language model; and a desired development result generation unit configured to generate a corresponding front-end interface according to the interface description language and associate corresponding database data with the front-end interface according to the database management language to obtain the desired development result.

[0007] Thirdly, embodiments of this disclosure provide an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to implement the low-code development method as described in the first aspect.

[0008] Fourthly, embodiments of this disclosure provide a non-transitory computer-readable storage medium storing computer instructions that enable a computer to implement the low-code development method as described in the first aspect when executed.

[0009] Fifthly, embodiments of this disclosure provide a computer program product including a computer program that, when executed by a processor, can implement the steps of the low-code development method as described in the first aspect.

[0010] To minimize the difficulty and barrier to entry for developing platforms or systems using low-code development platforms, the low-code development solution provided in this disclosure introduces generative large language modeling technology into the low-code development platform scenario. By constructing a relatively small number of fine-tuned training samples and sorted samples, the low-code development platform can convert user-inputted development commands in natural language or image form into interface description languages ​​and database management languages ​​that are easier for computers to recognize and process. This allows for the generation of corresponding front-end interfaces based on the interface description language and the implementation of association and binding operations between front-end interface elements and back-end database data based on the database management language. This enables the desired development results to be obtained quickly and easily. Furthermore, since no specialized development commands need to be input, the usage threshold of the low-code development platform is effectively reduced.

[0011] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0012] Other features, objects, and advantages of this disclosure will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0013] Figure 1 This is an exemplary system architecture to which this disclosure can be applied;

[0014] Figure 2 A flowchart illustrating a low-code development method provided in this disclosure embodiment;

[0015] Figure 3 A flowchart illustrating a training method for a first generative large language model provided in an embodiment of this disclosure;

[0016] Figure 4 A flowchart illustrating another training method for a first generative large language model provided in an embodiment of this disclosure;

[0017] Figure 5 A flowchart illustrating a training method for a second generative large language model provided in this embodiment of the disclosure;

[0018] Figure 6 A flowchart illustrating another training method for a second generative large language model provided in this embodiment of the disclosure;

[0019] Figure 7 A flowchart illustrating a method for generating a front-end interface according to an interface description language, provided in this embodiment of the disclosure;

[0020] Figures 8-1 to 8-12 This is a specific schematic diagram provided for a particular scenario of roadway defects in the embodiments of this disclosure;

[0021] Figure 9 A structural block diagram of a low-code development apparatus provided in this disclosure embodiment;

[0022] Figure 10 This is a schematic diagram of the structure of an electronic device suitable for implementing a low-code development method, provided as an embodiment of the present disclosure. Detailed Implementation

[0023] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding; these should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description. It should be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0024] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0025] Figure 1 An exemplary system architecture 100 is shown, in which embodiments of the low-code development methods, apparatus, electronic devices, and computer-readable storage media of this disclosure can be applied.

[0026] like Figure 1 As shown, system architecture 100 may include terminal devices 101, 102, and 103, a network 104, and a server 105. Network 104 serves as the medium for providing communication links between terminal devices 101, 102, and 103 and server 105. Network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0027] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various applications for enabling information communication between the terminal devices 101, 102, and 103 and server 105 can be installed. These applications include low-code development applications, model training applications, and instant messaging applications.

[0028] Terminal devices 101, 102, and 103 and server 105 can be either hardware or software. When terminal devices 101, 102, and 103 are hardware, they can be various electronic devices with displays, including but not limited to smartphones, tablets, laptops, and desktop computers. When terminal devices 101, 102, and 103 are software, they can be installed in the aforementioned electronic devices, and can be implemented as multiple software programs or software modules, or as a single software program or software module; no specific limitation is made here. When server 105 is hardware, it can be implemented as a distributed server cluster composed of multiple servers, or as a single server. When server 105 is software, it can be implemented as multiple software programs or software modules, or as a single software program or software module; no specific limitation is made here.

[0029] Server 105 can provide various services through its built-in applications. Taking a low-code development application that provides low-code development services as an example, when running this low-code development application, server 105 can achieve the following effects: First, it receives interface description information from terminal devices 101, 102, and 103 via network 104, which constitutes the front-end interface of the desired development result. Next, it inputs this interface description information as input to a preset first generative large language model to obtain the interface description language output by the first generative large language model. Then, it receives association instruction information from terminal devices 101, 102, and 103 via network 104, which associates the back-end data constituting the desired development result with the front-end interface. Then, it inputs this association instruction information as input to a preset second generative large language model to obtain the database management language output by the second generative large language model. Finally, it generates the corresponding front-end interface according to the interface description language and associates the corresponding database data with the front-end interface according to the database management language to obtain the desired development result.

[0030] It should be noted that the interface description information and associated indication information can be temporarily obtained from the terminal devices 101, 102, and 103 via the network 104, or they can be pre-stored locally on the server 105 through various means. Therefore, when the server 105 detects that this data is already stored locally (for example, when it starts processing previously left-behind tasks), it can choose to directly obtain this data from the local storage. In this case, the exemplary system architecture 100 may not include the terminal devices 101, 102, and 103 and the network 104.

[0031] Because providing low-code development services requires significant computing resources and power, the low-code development methods provided in the subsequent embodiments of this disclosure are generally executed by a server 105 with strong computing power and abundant computing resources. Correspondingly, the low-code development device is also generally located within the server 105. However, it should also be noted that when terminal devices 101, 102, and 103 also possess sufficient computing power and resources, they can also perform the aforementioned calculations performed by the server 105 through their installed low-code development applications, thereby outputting the same results as the server 105. Especially when multiple terminal devices with different computing capabilities exist simultaneously, but the low-code development application determines that its terminal device has strong computing power and abundant remaining computing resources, it can allow the terminal device to perform the aforementioned calculations, thereby appropriately reducing the computing pressure on the server 105. Accordingly, the low-code development device can also be located within the terminal devices 101, 102, and 103. In this case, the exemplary system architecture 100 may also exclude the server 105 and the network 104.

[0032] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0033] Please refer to Figure 2 , Figure 2 A flowchart of a low-code development method provided for embodiments of this disclosure, wherein process 200 includes the following steps:

[0034] Step 201: Obtain the input interface description information that constitutes the front-end interface of the desired development result;

[0035] This step is intended for implementation entities of low-code development methodologies (e.g., Figure 1 The server 105 shown obtains information from the development object via an input device (e.g., ...). Figure 1 The interface description information input by the terminal devices 101, 102, and 103 shown is used to form the front-end interface that constitutes the desired development result.

[0036] The expected development result refers to the development object's expectation of the development result. It can be understood as the expected development result, which can be an interactive interface, an application, a website, or even a set of functional systems or platforms for actual application.

[0037] In general, regardless of the desired development outcome, it should have a front-end interface (also known as a user interface) for interacting with users. In addition to the front-end interface, a back-end database is usually required so that the back-end database can be displayed on the front-end interface in a suitable form.

[0038] The interface description information may include: natural language descriptions of the elements constituting the front-end interface (e.g., creating a blue navigation bar containing three items: A1, A2, and A3); and image-based reference images that mimic the elements constituting the front-end interface (e.g., a screenshot of a navigation bar from another system). Specifically, the elements constituting the front-end interface may include: the way the interface is divided into areas, the controls contained within the interface, the arrangement order of different controls, and the style including colors and line distribution. Of course, in some special cases, the interface description information may exist in other variant forms besides text and images, such as voice information, but this can be converted into natural language descriptions in text form.

[0039] Step 202: Input the interface description information as input information into the preset first generative large language model to obtain the interface description language output by the first generative large language model;

[0040] Building upon step 201, this step aims to have the aforementioned executing entity input interface description information as input to a preset first generative large language model, thereby obtaining the interface description language output by the first generative large language model. That is, the first generative large language model, through training, should possess the ability to convert the input interface description information into an interface description language that is easier for computers to recognize. Specifically, the first generative large language model can be obtained by fine-tuning an existing basic generative large language model used as a training base, or it can be completely retrained.

[0041] Interface description languages ​​(IDEs) are languages ​​or markup languages ​​used to describe the layout, appearance, and interactions of user interfaces. They provide developers and designers with a structured way to define and create user interfaces for applications, websites, or software. They have the following characteristics and functions:

[0042] Layout Description: The interface description language describes the arrangement, size, and position of interface elements. This may include grid systems, relative positioning, and adaptive layouts; Style Definition: Allows defining the appearance of interface elements, including colors, fonts, borders, etc. It typically supports features similar to Cascading Style Sheets (CSS); Interaction Behavior: Describes user behavior when interacting with interface elements, such as the impact of clicks, hovers, drags, and drop operations on the interface; Components and Modularization: Supports component-based or modular development, enabling the definition and reuse of a specific set of interface elements, improving development efficiency.

[0043] Some common interface description languages ​​include:

[0044] HTML / CSS: HTML (Hypertext Markup Language) is used to describe the structure of web pages, while CSS (Cascading Style Sheets) is used to define styles and layouts; XML: Although not specifically created for interface design, XML is used to define structured data and sometimes to describe interface elements and their attributes; XAML (Extensible Application Markup Language): Primarily used in Microsoft's WPF (Windows Presentation Foundation) and Silverlight platforms to define application interfaces; JSON: Although JSON is more commonly used for data exchange, it is sometimes used to describe simple interface elements and interactions.

[0045] In addition to being easily recognized by computers, application interface description languages ​​also have the following uses and advantages:

[0046] Cross-platform compatibility: Interface description languages ​​facilitate cross-platform UI implementation, enabling applications to maintain consistency across different devices and screen sizes; Maintainability and extensibility: Using interface description languages ​​makes it easier to modify, extend, and maintain the interface; Separation of concerns: Separating interface description from application logic allows developers and designers to focus on different tasks, improving work efficiency; Standardized and shared components: Standardized interface components can be created, allowing these components to be reused in different projects, improving development efficiency.

[0047] Furthermore, in order to improve the degree of conformity between the interface description language output by the first generative large language model and the user's expectations, the aforementioned execution entity can perform semantic understanding on the interface description information, and complete and adjust the interface description information based on the semantic understanding results (such as eliminating ambiguities and loopholes in the original interface description information), so as to use the completed and adjusted interface description information as input information to the first generative large language model, in order to improve the accuracy of the output interface description language.

[0048] Step 203: Obtain the association instruction information that associates the input backend data, which constitutes the expected development result, with the frontend interface;

[0049] Building upon step 202, this step aims to have the aforementioned executing entity obtain association indication information that links the input backend data constituting the desired development result with the frontend interface. Specifically, this association indication information is used to indicate which data in the subsequent database is associated with which elements in the frontend interface, so that, when the correct associations exist, the data in the subsequent database is presented in the correct form and at the correct location on the frontend interface.

[0050] Similar to interface description information, this association instruction information may also include: natural language association instruction information describing the relationship between the backend data and the frontend interface in natural language form, and graphical association instruction information illustrating the relationship in diagrammatic form (e.g., correspondence diagrams, binding diagrams, correspondence-output diagrams, etc.). Of course, in some special cases, association instruction information may exist in other variant forms besides text and images, such as voice information, but this can be converted into natural language association instruction information in text form.

[0051] Step 204: Input the association instruction information as input information into the preset second generative large language model to obtain the database management language output by the second generative large language model;

[0052] Building upon step 203, this step aims to have the aforementioned executing entity input the association indication information as input to a preset second generative large language model, thereby obtaining the database management language output by the second generative large language model. That is, this second generative large language model, through training, should possess the ability to convert the input association indication information into a database management language that is easier for computers to recognize and process. Specifically, this second generative large language model can be obtained by fine-tuning an existing basic generative large language model used as a training base, or it can be completely retrained.

[0053] Database management languages ​​refer to the languages ​​used to manage and manipulate data in databases. For example, SQL is the command language used to manage and manipulate relational databases. Similarly, it also includes the corresponding command languages ​​used to manage and manipulate non-relational databases. These will not be listed here.

[0054] Step 205: Generate the corresponding front-end interface according to the interface description language, and associate the front-end interface with the corresponding database data according to the database management language to obtain the desired development result.

[0055] Based on steps 202 and 204, this step aims to have the aforementioned execution entity generate the corresponding front-end interface according to the interface description language, and then associate the front-end interface with the corresponding database data according to the database management language, thereby finally obtaining the desired development result.

[0056] To minimize the difficulty and barrier to entry for developing platforms or systems using low-code development platforms, the low-code development method disclosed herein introduces generative large language modeling technology into the low-code development platform scenario. By constructing a relatively small number of fine-tuned training samples and sorted samples, the low-code development platform can convert user-inputted development commands in natural language or image form into interface description languages ​​and database management languages ​​that are easier for computers to recognize and process. This allows for the generation of corresponding front-end interfaces based on the interface description language and the implementation of association and binding operations between front-end interface elements and back-end database data based on the database management language. This facilitates the quick and easy attainment of the desired development results. Furthermore, since no specialized development commands are required, the low-code development platform's usage threshold is effectively lowered.

[0057] To better understand the process of training a first generative large language model capable of representing the correspondence between interface description information and interface description language, this embodiment also demonstrates... Figure 3 and Figure 4The training process for the first generative large language model is illustrated for two different forms of interface description information. Figure 3 One specific implementation shown includes the following steps:

[0058] Step 301: Obtain the first basic generative large language model trained based on the general same modality training samples, and obtain the fine-tuned training samples based on natural language description information and corresponding interface description language.

[0059] Among them, the general homomodal training samples used to train the first basic generative large language model are simply data whose input and expected output are of the same modality, such as text input and text expected output of the same modality as natural language description information and interface description language.

[0060] This step also demonstrates that this embodiment does not use a complete retraining to obtain the first generative large language model. Instead, it uses the first basic generative large language model as the base and fine-tunes it by constructing fine-tuning training samples to obtain the best possible training effect while reducing the required training samples and improving training efficiency.

[0061] Step 302: Use the fine-tuning training samples to perform supervised fine-tuning on the first basic generative large language model to obtain the first training large model;

[0062] Based on step 301, this step aims to have the aforementioned executing entity perform supervised fine-tuning (SFT) on the first basic generative large language model using fine-tuned training samples to obtain the first training large model.

[0063] Supervised fine-tuning involves pre-training a neural network model (the source model) on the source dataset. Then, a new neural network model (the target model) is created. The target model replicates all the model design and parameters of the source model, except for the output layer. These model parameters contain knowledge learned from the source dataset, and this knowledge is equally applicable to the target dataset. The output layer of the source model is closely related to the labels of the source dataset and is therefore not used in the target model. During fine-tuning, an output layer with an output size equal to the number of classes in the target dataset is added to the target model, and the model parameters of this layer are randomly initialized. When training the target model on the target dataset, it is trained from scratch up to the output layer, and the parameters of the remaining layers are fine-tuned based on the parameters of the source model.

[0064] Alternatively, LoRA (Low-Rank Adaptation of Large Language Models) fine-tuning can be employed for generative large language models. The basic principle of LoRA fine-tuning is to freeze the pre-trained model weights and parameters. While freezing the original model parameters, additional network layers are added to the model, and only the parameters of these newly added network layers are trained. Because the number of these new parameters is relatively small, this not only significantly reduces the cost of fine-tuning but also achieves similar results to fine-tuning with all model parameters involved.

[0065] This approach is well-suited for large language models because, as these models grow, the number of parameters increases dramatically, making fine-tuning all parameters impractical. LoRA fine-tuning, proposed by Microsoft, significantly reduces the number of trainable parameters for downstream tasks by fine-tuning only newly added parameters. The basic principle of LoRA fine-tuning is that each layer of a neural network involves matrix multiplication. The weight matrices in these layers typically have full rank. When adapted to a specific task, pre-trained language models have low "intrinsic dimensionality," allowing them to learn effectively even when randomly projected into smaller subspaces.

[0066] It should be noted that the LoRA fine-tuning method and the SFT fine-tuning method can be combined and used simultaneously.

[0067] Step 303: The sorted samples, consisting of multiple interface description languages ​​ranked by accuracy, corresponding to the natural language description information of the same front-end interface, are used to strengthen the training of the first training model using a reinforcement learning method based on human feedback, so as to obtain the second training model.

[0068] Based on step 302, this step aims to have the aforementioned executing entity use a reinforcement learning approach based on human feedback to reinforce the first training model by taking the sorted samples of interface description languages ​​consisting of multiple natural language description information corresponding to the same front-end interface, sorted by accuracy, and train them to obtain a second training model. This is to further improve the output of the obtained second training model to better meet the needs and expectations of users or human users.

[0069] Step 304: Use the reparameter method to merge the model parameters of the second training large model with the model parameters of the first basic generative large language model to obtain the first generative large language model.

[0070] Based on step 303, this step aims to have the aforementioned executing entity merge the model parameters of the second training large model with the model parameters of the first basic generative large language model using a reparameterization method, thereby obtaining the first generative large language model.

[0071] Parameter reparameterization techniques refer to methods that reparameterize a model to make it easier to optimize or converge during training. These techniques are commonly used in optimization algorithms such as gradient descent to improve optimization efficiency and stability. Below are some examples of common parameter reparameterization techniques in the field of model optimization:

[0072] 1) Batch Normalization:

[0073] Batch Normalization is a technique that normalizes the input for each batch and then scales and translates the results. This not only helps alleviate the vanishing gradient problem but also makes the model converge more easily during training. Batch Normalization introduces learnable parameters, namely scaling and translation parameters, which can be adjusted during backpropagation.

[0074] 2) Weight Normalization:

[0075] Weight Normalization is a technique for normalizing the weights of a neural network. It decomposes the weights into scaling factors of the magnitude and proportions of the orientation, making model training more stable. This reparameterization method can accelerate convergence and improve the model's generalization performance.

[0076] 3) Layer Normalization:

[0077] Similar to Batch Normalization, Layer Normalization is a method that normalizes the input to each layer. It does not rely on batch statistics but normalizes the features of each sample. This helps handle the statistical characteristics of different samples, making the model easier to optimize.

[0078] 4) Gradient Clipping:

[0079] Gradient clipping is a technique that prevents gradient explosion by scaling the magnitude of the gradient. During training, if the gradient exceeds a predetermined threshold, it is scaled. This helps prevent gradient explosion and makes the model more stable.

[0080] 5) Parameterized Activation Functions:

[0081] Some activation functions, such as Parametric Rectified Linear Unit (PReLU) and Exponential Linear Unit (ELU), change the shape of the activation function by introducing learnable parameters. This can help the model better adapt to the data and improve optimization performance.

[0082] 6) Neural Architecture Search:

[0083] In neural network architecture search, the network structure is searched and reparameterized to make the network easier to train while maintaining certain performance.

[0084] The common goal of these reparameterization techniques is to improve the convergence, stability, and generalization performance of a model, making it easier to find suitable parameter values ​​during optimization. By cleverly reparameterizing the model, the efficiency of optimization algorithms can be improved, training time reduced, and the overall performance of the model enhanced.

[0085] Figure 3 The steps described above provide an implementation scheme for quickly obtaining a first generative large language model that can represent the correspondence between natural language interface description information and interface description language through fine-tuning training. This enables a low-code development platform that is embedded or can call the first generative large language model to convert user-input natural language commands into a more easily understood interface description language, thereby eliminating the need for users to input more specialized code and lowering the barrier to entry.

[0086] Furthermore, the interface description language corresponding to the front-end interface that meets the expected development result of the first preset requirement can be saved as the first valid sample. Then, the model parameters of the first generative large language model can be adjusted using this first valid sample. This is equivalent to using the first valid sample as a "godden" sample and repeatedly reinforcing the training to increase the probability of the model outputting the "godden" sample.

[0087] Unlike Figure 3 The implementation method shown is as follows: Figure 4 This demonstrates another implementation scheme for training the first generative large language model, including the following steps:

[0088] Step 401: Obtain the second basic generative large language model trained based on general cross-modal training samples, and obtain fine-tuned training samples based on image description information and corresponding interface description language.

[0089] Step 402: Use fine-tuning training samples to perform supervised fine-tuning on the second basic generative large language model to obtain the third training large model;

[0090] Step 403: The sorted samples, consisting of multiple interface description languages ​​ranked by accuracy, corresponding to the image description information of the same front-end interface, are used to strengthen the training of the third training model using a reinforcement learning method based on human feedback, so as to obtain the fourth training model.

[0091] Step 404: Use the reparameter method to merge the model parameters of the fourth training large model with the model parameters of the second basic generative large language model to obtain the first generative large language model.

[0092] contrast Figure 3 As shown in steps 301-304, in the solution provided by steps 401-404 in this embodiment, since the fine-tuning training samples are composed of image description information in image form and corresponding interface description language in text form, the generative large language model used as the base should be a second generative large language model trained based on general cross-modal training samples. Correspondingly, the sorted samples should also be constructed from multiple interface description languages ​​ordered by accuracy corresponding to the image description information of the same front-end interface. This enables the low-code development platform embedded or capable of calling the first generative large language model to convert the reference interface image input by the user into a more easily understood interface description language, thereby eliminating the need for the user to input more professional code language, lowering the usage threshold, and making it easier to imitate the interface elements of the reference interface image and achieve rapid migration of interfaces of the same style.

[0093] Furthermore, the interface image corresponding to the front-end interface that meets the second preset requirement can be saved as a second valid sample. Then, the model parameters of the first generative large language model can be adjusted using this second valid sample. This is equivalent to using the second valid sample as a "godden" sample and repeatedly reinforcing the training to increase the probability of the model outputting the "godden" sample.

[0094] To better understand the process of training a second generative large language model capable of representing the correspondence between associated instruction information and database management language, this embodiment also demonstrates... Figure 5 and Figure 6 The training process for the second generative large language model is illustrated for two different forms of associated instruction information. Figure 5 One specific implementation shown includes the following steps:

[0095] Step 501: Obtain the third basic generative large language model trained based on the general same modality training samples, and obtain the fine-tuning training samples based on natural language association indication information and the corresponding database management language.

[0096] Step 502: Use the fine-tuning training samples to perform supervised fine-tuning on the third basic generative large language model to obtain the fifth training large model;

[0097] Step 503: The sorted samples consisting of multiple database management languages ​​corresponding to the natural language association indication information of the same association form, sorted by accuracy, are used to reinforce the fifth training model using a reinforcement learning method based on human feedback, so as to obtain the sixth training model.

[0098] Step 504: Use the reparameter method to merge the model parameters of the sixth training large model with the model parameters of the third basic generative large language model to obtain the second generative large language model.

[0099] contrast Figure 3 As shown in steps 301-304, in this embodiment, the solution provided by steps 501-504 also uses a basic generative large language model trained based on general homomodal training samples. However, the fine-tuning training samples are composed of natural language association indication information and corresponding database management language. Correspondingly, the sorting samples should also be constructed from multiple database management languages ​​ordered by accuracy corresponding to the natural language association indication information of the same association form. This enables the low-code development platform embedded or capable of calling the second generative large language model to convert the natural language instructions input by the user into a database management language that is easier for the computer to recognize and used to perform corresponding operations on the database data. This eliminates the need for the user to input more professional code language and lowers the barrier to entry.

[0100] Unlike Figure 5 The implementation method shown is as follows: Figure 6 This demonstrates another implementation scheme for training a second generative large language model, including the following steps:

[0101] Step 601: Obtain the fourth basic generative large language model trained based on general cross-modal training samples, and obtain fine-tuning training samples based on graphical association indication information and corresponding database management language.

[0102] Step 602: Supervised fine-tuning of the fourth basic generative large language model using fine-tuned training samples to obtain the seventh training large model;

[0103] Step 603: The sorted samples, consisting of multiple database management languages ​​ranked by accuracy and corresponding to the graphical association indication information of the same association form, are used to reinforce the seventh training model using a reinforcement learning method based on human feedback, so as to obtain the eighth training model.

[0104] Step 604: Use the reparameter method to merge the model parameters of the eighth training large model with the model parameters of the fourth basic generative large language model to obtain the second generative large language model.

[0105] contrast Figure 5 As shown in steps 501-504, in the solution provided by steps 601-604 in this embodiment, since the fine-tuning training samples are composed of graphical association indication information in image form and corresponding database management language in text form, the generative large language model used as the base should be a basic generative large language model trained on general cross-modal training samples. Correspondingly, the sorted samples should also be constructed from multiple database management languages ​​sorted by accuracy corresponding to graphical association indication information of the same association form. This enables the low-code development platform embedded or capable of calling the second generative large language model to convert the association indication graph input by the user into a database management language that is easier for the computer to recognize and used to perform corresponding operations on the database data, thereby eliminating the need for the user to input more professional code language and lowering the usage threshold.

[0106] Based on any of the above embodiments, considering that the process of generating a front-end interface based on interface description information often involves simultaneously utilizing multiple forms of interface description information to continuously adjust the desired front-end interface, one implementation method, including but not limited to, can be found in [reference needed]. Figure 7 The flowchart shown includes the following steps:

[0107] Step 701: In response to the simultaneous presence of natural language description information and image description information, determine the reception time of receiving natural language description information and image description information respectively;

[0108] Step 702: Determine the primary and secondary description information of the front-end interface in chronological order of receipt time;

[0109] Step 703: Generate an initial front-end interface according to the interface description language corresponding to the primary description information, and adjust the initial front-end interface according to the interface description language corresponding to the secondary description information to obtain the final front-end interface.

[0110] It should be understood that, under normal circumstances, different forms of interface description information should be received at different times. Based on the time sequence shown by the reception time, it is determined which is the primary description information, which is the secondary description information, the tertiary description information, etc., and then adjustments are made continuously on the interface corresponding to the interface description language of the previous description information according to the order of the description information. By switching between two different forms of interface description information, the efficiency of obtaining the final desired front-end interface can be improved and the adjustment time can be shortened.

[0111] Based on any of the above embodiments, after generating the desired development result, a target version number corresponding to the desired development result can be determined, and then the target version number can be used as the release version number of the desired development result. The desired development result can then be released according to the release version number so that it can truly play its role.

[0112] To enhance understanding, this disclosure also uses the creation of a roadway pest and disease monitoring and analysis system as a specific example, combined with... Figures 8-1 to 8-12 A specific implementation scheme is given:

[0113] To make the interaction between developers and low-code development platforms as natural as human-to-human interaction, existing low-code development platforms need to be modified to possess the following three core capabilities:

[0114] I. Conversion from Natural Language to Interface Description Language:

[0115] Due to the flexibility and variability of natural language, coupled with the millions of possible combinations of different components, exhaustive one-to-one mapping cannot achieve the conversion from natural language to JSON-formatted interface description language. Therefore, leveraging the emergent capabilities of large-scale generative language models, a mapping relationship between natural language and JSON-formatted interface description language is established. Specifically, the steps are as follows:

[0116] 1) Select an appropriate large language model as the base. Considering the actual effect of the model and the latency requirements of the actual application scenario, a 7B large language model (with 7 billion floating-point parameters) is adopted.

[0117] 2) Create a dataset of 1000 natural language sentences converted to JSON;

[0118] 3) Supervised fine-tuning (SFT) was performed on the model using the LoRA method, with 100 training epochs.

[0119] 4) In order to ensure that the model outputs text in strict JSON format, a dataset of 1000 texts with ranking information was created, and a reinforcement learning method based on human feedback was used to strengthen the model output by SFT.

[0120] 5) Combine the parameters of the model obtained above with the original model to obtain the current model.

[0121] II. Conversion of Natural Language to SQL Commands:

[0122] Due to the flexibility and variability of natural language, a generative large language model is used to map natural language to SQL commands. Specifically, the steps are as follows:

[0123] 1) Select an appropriate SQL base model;

[0124] 2) Create a dataset of 10,000 natural language to SQL statements;

[0125] 3) Perform full fine-tuning of the SQL base model;

[0126] 4) Obtain the current model.

[0127] III. Conversion of images to interface description language:

[0128] To achieve the conversion of images to JSON format interface description language, a cross-modal model is used to output JSON format interface description language through language features and image features. Specifically, the steps are as follows:

[0129] 1) Select an appropriate cross-modal model as the basis;

[0130] 2) Create a dataset of 1000 images and JSON files;

[0131] 3) Fine-tuning training using cross-modal basis models and datasets;

[0132] 4) Obtain the current model.

[0133] In other words, through the above three core capabilities, the low-code development platform possesses the following functions:

[0134] 1) Function 1: Natural Language Interaction to Generate UI

[0135] By using large language model technology, users can input the content they want to be displayed in a certain module on the platform into the large language model through natural language interaction. Through the model's understanding of the requirements, the front-end interface of the content to be displayed in that module is automatically built, and a connection and interaction capability is established with the system's back-end database to realize automatic function development.

[0136] 2) Feature Two: Quick UI Migration

[0137] This feature utilizes cross-modal large model technology, allowing users to upload reference images of the UI style they need into the cross-modal large model. Through the model's understanding of UI style, color scheme, and layout, it automatically replaces the current interface and generates a UI similar in style to the reference image.

[0138] 3) Function Three: Continuous Content Interaction

[0139] Based on large model technology, this feature allows for multiple in-depth interactions with the data content of the current page.

[0140] The implementation of the above capabilities will be demonstrated below with the aid of several accompanying figures:

[0141] 1. Examples of natural language dialogue capabilities

[0142] For example, to generate a navigation component, normally a low-code platform would require dragging in a navigation component and then adding submenus. Now, using the low-code development platform provided in this embodiment, you only need to input, such as... Figure 8-1 Simply select "Generate Vertical Navigation Menu, including Driveway Defects, Sidewalk Defects, Signage, and Other Events" as shown, and it will automatically generate a menu like this. Figure 8-2 The interface shown.

[0143] Alternatively, you can interact with the interface again and add a second-level menu named "Pitfalls". Figure 8-3 As shown.

[0144] 2. UI design drafts can be directly used to generate interface examples.

[0145] For example, if you want to replicate the effect of an external interface, you only need to capture a screenshot like... Figure 8-4 The reference image shown is... Figure 8-4 It includes a title, an image, and some display data.

[0146] like Figure 8-5 and Figure 8-6 As shown, you can simply select the image to import in the selection bar to generate the closest UI using a cross-modal large model.

[0147] 3. Data is displayed directly using natural language.

[0148] like Figure 8-7 As shown, you can directly input the conditions you want to query, and the results will be rendered directly. Figure 8-8 The interface shown can also be like Figure 8-9 The example shown demonstrates how to fine-tune the displayed fields through a second dialogue. Figure 8-10 The new interface, after being fine-tuned as instructed, is shown.

[0149] 4. All controls can interact with the user a second time.

[0150] For example, in the table above, if a user wants to see detailed information about the first item, or the trend changes over the past few days, they can... Figure 8-11 and Figure 8-12 The diagram shows multiple interactions with the button and control.

[0151] By applying the solution provided in this embodiment, development efficiency can be improved. Using this platform to implement the same UI design, the development speed can be increased by more than 10 times. It can also combine cross-modal model self-learning visual style to adapt to complex UI pages. At the same time, it supports secondary intelligent interaction, summarization / filtering, etc. Moreover, it is also a true no-code development, which can generate corresponding results instantly by inputting instructions described in natural language, with a low operation threshold.

[0152] This means that by using large model technology, development capabilities that originally required code input can now be developed by UI designers who need to master design expertise, simply through natural language input or natural language plus reference style images.

[0153] Further reference Figure 9 As an implementation of the methods shown in the above figures, this disclosure provides an embodiment of a low-code development apparatus, which is similar to... Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0154] like Figure 9 As shown, the low-code development device 900 of this embodiment may include: an interface description information acquisition unit 901, an interface description language generation unit 902, an association instruction information acquisition unit 903, a database management language generation unit 904, and a desired development result generation unit 905. Specifically, the interface description information acquisition unit 901 is configured to acquire the interface description information input as the front-end interface constituting the desired development result; the interface description language generation unit 902 is configured to input the interface description information as input information into a preset first generative large language model to obtain the interface description language output by the first generative large language model; the association instruction information acquisition unit 903 is configured to acquire the association instruction information input as the back-end data constituting the desired development result and associate it with the front-end interface; the database management language generation unit 904 is configured to input the association instruction information as input information into a preset second generative large language model to obtain the database management language output by the second generative large language model; and the desired development result generation unit 905 is configured to generate a corresponding front-end interface according to the interface description language and associate the front-end interface with corresponding database data according to the database management language to obtain the desired development result.

[0155] In this embodiment, the specific processing and technical effects of the interface description information acquisition unit 901, interface description language generation unit 902, association instruction information acquisition unit 903, database management language generation unit 904, and expected development result generation unit 905 in the low-code development device 900 can be referred to respectively. Figure 2 The relevant descriptions of steps 201-205 in the corresponding embodiments will not be repeated here.

[0156] In some optional implementations of this embodiment, the interface description information includes: natural language description information describing the elements constituting the front-end interface in natural language form, and image description information describing the elements constituting the front-end interface in the form of an interface reference image. The elements include: the area division method within the interface, the controls contained within the interface, the arrangement order between different controls, and the style including color and line distribution.

[0157] In some optional implementations of this embodiment, the low-code development device 900 may further include:

[0158] The semantic understanding and correction unit is configured to perform semantic understanding on the interface description information before inputting it into the first generative large language model, and to complete and adjust the interface description information based on the semantic understanding results, so as to use the completed and adjusted interface description information as input information into the first generative large language model.

[0159] In some optional implementations of this embodiment, the interface description language is JSON-formatted interface description code.

[0160] In some optional implementations of this embodiment, the association indication information includes: natural language association indication information that describes the association between backend data and frontend interface in natural language form, and graphical association indication information that illustrates the association in the form of a diagram.

[0161] In some optional implementations of this embodiment, in response to the backend data being stored in a relational database, the database management language is the SQL command language used to manage and operate the relational database.

[0162] In some optional implementations of this embodiment, the low-code development device 900 may further include: a first generative large language model training unit, which is further configured to:

[0163] Obtain the first basic generative large language model trained based on general isomodal training samples, and obtain fine-tuned training samples based on natural language description information and corresponding interface description language.

[0164] The first basic generative large language model was subjected to supervised fine-tuning using fine-tuning training samples to obtain the first large model in training.

[0165] The sorted samples, consisting of multiple interface description languages ​​ranked by accuracy corresponding to the natural language description information of the same front-end interface, are used to strengthen the training of the first training model using a reinforcement learning method based on human feedback, thus obtaining the second training model.

[0166] The parameters of the second large training model and the first basic generative large language model are merged using the reparameterization method to obtain the first generative large language model.

[0167] In some optional implementations of this embodiment, the low-code development device 900 may further include:

[0168] The first valid sample determination unit is configured to save the interface description language corresponding to the front-end interface that meets the expected development result of the first preset requirement as the first valid sample.

[0169] The first model parameter adjustment unit is configured to adjust the model parameters of the first generative large language model using the first valid sample.

[0170] In some optional implementations of this embodiment, the low-code development device 900 may further include: a first generative large language model training unit, which is further configured to:

[0171] Obtain the second basic generative large language model trained based on general cross-modal training samples, and obtain fine-tuned training samples based on image description information and corresponding interface description language.

[0172] The second basic generative large language model was subjected to supervised fine-tuning using fine-tuned training samples to obtain the third training large model.

[0173] The sorted samples, consisting of multiple interface description languages ​​ranked by accuracy corresponding to the image description information of the same front-end interface, are used to strengthen the training of the third training model using a reinforcement learning method based on human feedback, thus obtaining the fourth training model.

[0174] The parameters of the large model in the fourth training session are merged with those of the second basic generative large language model using a parameter re-parameter approach to obtain the first generative large language model.

[0175] In some optional implementations of this embodiment, the low-code development device 900 may further include:

[0176] The second valid sample determination unit is configured to save the interface image corresponding to the front-end interface that meets the expected development result of the second preset requirement as the second valid sample.

[0177] The second model parameter adjustment unit is configured to adjust the model parameters of the first generative large language model using the second valid samples.

[0178] In some optional implementations of this embodiment, the low-code development device 900 may further include: a second generative large language model training unit, which is further configured to:

[0179] Obtain the third basic generative large language model trained based on general homomodal training samples, and obtain fine-tuned training samples based on natural language association indication information and corresponding database management language.

[0180] The third basic generative large language model was subjected to supervised fine-tuning using fine-tuned training samples to obtain the fifth training large model.

[0181] The sorted samples, consisting of multiple database management languages ​​ranked by accuracy corresponding to the natural language association indication information of the same association form, are used to reinforce the fifth training model with a reinforcement learning method based on human feedback, thus obtaining the sixth training model.

[0182] By using a parameter re-parameter approach, the model parameters of the sixth training large model are merged with the model parameters of the third basic generative large language model to obtain the second generative large language model.

[0183] In some optional implementations of this embodiment, the low-code development device 900 may further include: a second generative large language model training unit, which is further configured to:

[0184] Obtain the fourth basic generative large language model trained based on general cross-modal training samples, and obtain fine-tuning training samples based on graphical association indication information and corresponding database management language.

[0185] The fourth basic generative large language model was subjected to supervised fine-tuning using fine-tuned training samples to obtain the seventh training large model.

[0186] The sorted samples, consisting of multiple database management languages ​​ranked by accuracy and corresponding to the graphical association indication information of the same association form, are used to reinforce the seventh training model with a reinforcement learning method based on human feedback, resulting in the eighth training model.

[0187] By using a parameter re-parameter approach, the model parameters of the eighth training large model and the model parameters of the fourth basic generative large language model are merged to obtain the second generative large language model.

[0188] In some optional implementations of this embodiment, the expected development result generation unit 905 may include a front-end interface generation unit configured to generate a corresponding front-end interface according to the interface description language. The front-end interface generation unit may be further configured to:

[0189] In response to the simultaneous presence of natural language description information and image description information, the reception time of receiving natural language description information and image description information respectively is determined;

[0190] Determine the primary and secondary description information of the front-end interface according to the order of receipt time;

[0191] The initial front-end interface is generated according to the interface description language corresponding to the primary description information, and then the initial front-end interface is adjusted according to the interface description language corresponding to the secondary description information to obtain the final front-end interface.

[0192] In some optional implementations of this embodiment, the low-code development device 900 may further include:

[0193] The version number determination unit is configured to determine the target version number corresponding to the expected development result;

[0194] The release unit is configured to synthesize the target version number as the release version number of the expected development result, and release the expected development result according to the release version number.

[0195] This embodiment exists as a device embodiment corresponding to the above method embodiment.

[0196] To minimize the difficulty and barrier to entry for developing platforms or systems using low-code development platforms, the low-code development device provided in this embodiment introduces generative large language modeling technology into the low-code development platform scenario. By constructing a relatively small number of fine-tuned training samples and sorted samples, the low-code development platform can convert user-inputted development commands in natural language or image form into interface description languages ​​and database management languages ​​that are easier for computers to recognize and process. This allows the platform to generate corresponding front-end interfaces based on the interface description language and to perform association and binding operations between front-end interface elements and back-end database data based on the database management language. This enables the platform to obtain the desired development results quickly and easily. Furthermore, since no specialized development commands need to be input, the usage threshold of the low-code development platform is effectively reduced.

[0197] According to embodiments of this disclosure, this disclosure also provides an electronic device, the electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to implement the low-code development method described in any of the above embodiments.

[0198] According to embodiments of this disclosure, this disclosure also provides a readable storage medium storing computer instructions that enable a computer to implement the low-code development method described in any of the above embodiments when executed.

[0199] According to embodiments of this disclosure, this disclosure also provides a computer program product that, when executed by a processor, can implement the steps of the low-code development method described in any of the above embodiments.

[0200] Figure 10 A schematic block diagram of an example electronic device 1000 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0201] like Figure 10 As shown, device 1000 includes a computing unit 1001, which can perform various appropriate actions and processes according to a computer program stored in read-only memory (ROM) 1002 or a computer program loaded from storage unit 1008 into random access memory (RAM) 1003. The RAM 1003 may also store various programs and data required for the operation of device 1000. The computing unit 1001, ROM 1002, and RAM 1003 are interconnected via bus 1004. Input / output (I / O) interface 1005 is also connected to bus 1004.

[0202] Multiple components in device 1000 are connected to I / O interface 1005, including: input unit 1006, such as keyboard, mouse, etc.; output unit 1007, such as various types of monitors, speakers, etc.; storage unit 1008, such as disk, optical disk, etc.; and communication unit 1009, such as network card, modem, wireless transceiver, etc. Communication unit 1009 allows device 1000 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0203] The computing unit 1001 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1001 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1001 performs the various methods and processes described above, such as low-code development methods. For example, in some embodiments, the low-code development method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 1008. In some embodiments, part or all of the computer program may be loaded and / or installed on device 1000 via ROM 1002 and / or communication unit 1009. When the computer program is loaded into RAM 1003 and executed by the computing unit 1001, one or more steps of the low-code development method described above may be performed. Alternatively, in other embodiments, computing unit 1001 may be configured to perform low-code development methods by any other suitable means (e.g., by means of firmware).

[0204] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0205] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0206] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0207] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0208] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with embodiments of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0209] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and Virtual Private Server (VPS) services, such as high management difficulty and weak business scalability.

[0210] According to the technical solution of this disclosure, by introducing the generative large language model technology into the low-code development platform scenario, and by constructing a relatively small number of fine-tuned training samples and sorted samples, the low-code development platform can convert user-inputted development commands in natural language or image form into interface description languages ​​and database management languages ​​that are easier for computers to recognize and process by calling the trained generative large language model. This enables the platform to generate corresponding front-end interfaces based on the interface description language and realize the association and binding operations between front-end interface elements and back-end database data based on the database management language, thereby conveniently and quickly obtaining the desired development results. Furthermore, since no specialized development commands need to be input, the usage threshold of the low-code development platform is effectively reduced.

[0211] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0212] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A low-code development method, comprising: obtaining input interface description information of a front-end interface constituting an expected development result, the interface description information comprising: natural language description information in natural language form describing elements constituting the front-end interface, and interface reference images in image form as image description information imitating the description of the elements constituting the front-end interface; obtaining a basic generative large language model and a fine-tuning training sample constituted by the natural language description information and a corresponding interface description language; performing supervised fine-tuning on the basic generative large language model using the fine-tuning training sample to obtain a training large model; using a reinforcement learning method based on human feedback to perform reinforcement training on the training large model using a plurality of ordered interface description languages corresponding to the natural language description information of the same front-end interface and corresponding to the same front-end interface, to obtain a reinforced training large model; and performing parameter merging on the model parameters of the reinforced training large model and the model parameters of the basic generative large language model using a reparameterization method to obtain a preset first generative large language model; inputting the interface description information as input information into the preset first generative large language model to obtain interface description language output by the first generative large language model; obtaining input association indication information for associating the front-end interface with back-end data constituting the expected development result; inputting the association indication information as input information into a preset second generative large language model to obtain database management language output by the second generative large language model; generating a corresponding front-end interface according to the interface description language, and associating the front-end interface with corresponding database data according to the database management language to obtain the expected development result.

2. The method of claim 1, wherein, The elements include: a region division manner within the interface, controls contained in the interface, arrangement orders between different controls, style styles including colors and line distributions.

3. The method of claim 2, wherein, Before inputting the interface description information into the first generative large language model, further comprising: performing semantic understanding on the interface description information, and completing and adjusting the interface description information according to the semantic understanding result, so as to input the completed and adjusted interface description information into the first generative large language model as input information.

4. The method of claim 1, wherein, The interface description language is interface description code in JSON format.

5. The method of claim 1, wherein, The association indication information comprises: natural language association indication information in natural language form describing the association form between the back-end data and the front-end interface, and graphical association indication information in the form of a schematic diagram embodying the association form.

6. The method of claim 1, wherein, In response to the back-end data being stored in a relational database, the database management language is a SQL instruction language for managing and operating the relational database.

7. The method of claim 2, wherein, The training process of the first generative large language model comprises: obtaining a first basic generative large language model trained based on general same-modal training samples, and obtaining fine-tuning training samples constituted by the natural language description information and corresponding interface description languages; performing supervised fine-tuning on the first basic generative large language model based on the fine-tuning training samples to obtain a first trained large model; performing reinforcement training on the first trained large model based on a reinforcement learning manner based on human feedback, to obtain a second trained large model, by using, as the reinforcement training sample, a sorting sample constituted by multiple items of interface description languages corresponding to the natural language description information of the same front-end interface and sorted according to accuracy; performing parameter merging on the model parameters of the second trained large model and the model parameters of the first basic generative large language model in a re-parameterization manner, to obtain the first generative large language model.

8. The method of claim 7, further comprising: saving, as a first valid sample, an interface description language corresponding to a front-end interface with a desired development result meeting a first preset requirement; adjusting the model parameters of the first generative large language model based on the first valid sample.

9. The method of claim 2, wherein, The training process of the first generative large language model comprises: obtaining a second basic generative large language model trained based on general cross-modal training samples, and obtaining fine-tuning training samples constituted by the image description information and corresponding interface description languages; performing supervised fine-tuning on the second basic generative large language model based on the fine-tuning training samples to obtain a third trained large model; performing reinforcement training on the third trained large model based on a reinforcement learning manner based on human feedback, to obtain a fourth trained large model, by using, as the reinforcement training sample, a sorting sample constituted by multiple items of interface description languages corresponding to the image description information of the same front-end interface and sorted according to accuracy; performing parameter merging on the model parameters of the fourth trained large model and the model parameters of the second basic generative large language model in a re-parameterization manner, to obtain the first generative large language model.

10. The method of claim 9, further comprising: saving, as a second valid sample, an interface image corresponding to a front-end interface with a desired development result meeting a second preset requirement; adjusting the model parameters of the first generative large language model based on the second valid sample.

11. The method of claim 5, wherein, The training process of the second generative large language model comprises: obtaining a third basic generative large language model trained based on general same-modal training samples, and obtaining fine-tuning training samples constituted by the natural language association indication information and corresponding database management languages; performing supervised fine-tuning on the third basic generative large language model based on the fine-tuning training samples to obtain a fifth trained large model; performing reinforcement training on the fifth trained large model based on a reinforcement learning manner based on human feedback, to obtain a sixth trained large model, by using, as the reinforcement training sample, a sorting sample constituted by multiple items of database management languages corresponding to the natural language association indication information of the same association form and sorted according to accuracy; The model parameters of the sixth large model in training are merged with the model parameters of the third basic generative large language model in a reparameterization manner to obtain the second generative large language model.

12. The method of claim 5, wherein, The training process of the second generative large language model comprises: A fourth basic generative large language model trained based on general cross-modal training samples is obtained, and a fine-tuning training sample composed of the graphically associated indication information and the corresponding database management language is obtained. The fourth basic generative large language model is fine-tuned in a supervised manner using the fine-tuning training sample to obtain a seventh large model in training; The sorting sample composed of multiple database management languages corresponding to the graphically associated indication information of the same association form and sorted according to accuracy is used to reinforce the training of the seventh large model in training in a reinforcement learning manner based on human feedback to obtain an eighth large model in training; The model parameters of the eighth large model in training are merged with the model parameters of the fourth basic generative large language model in a reparameterization manner to obtain the second generative large language model.

13. The method of claim 2, wherein, The generation of the corresponding front-end interface according to the interface description language comprises: In response to the simultaneous presence of the natural language description information and the image description information, the receiving time at which the natural language description information and the image description information are respectively received is determined; According to the order of the receiving time, the primary description information and the secondary description information of the front-end interface are determined; An initial front-end interface is generated according to the interface description language corresponding to the primary description information, and the initial front-end interface is adjusted according to the interface description language corresponding to the secondary description information to obtain the final front-end interface.

14. The method of any one of claims 1-13, further comprising: determining a target version number corresponding to the desired development result; taking the target version number as a release version number of the desired development result, and releasing the desired development result according to the release version number.

15. A low-code development device, comprising: an interface description information acquisition unit configured to acquire input interface description information for a front-end interface constituting a desired development result, the interface description information comprising natural language description information describing elements constituting the front-end interface in a natural language form, and image description information in the form of an interface reference image as a mimetic description of the elements constituting the front-end interface. The first generative large language model training unit is configured to: obtain a basic generative large language model and a fine-tuning training sample constituted by the natural language description information and corresponding interface description language; perform supervised fine-tuning on the basic generative large language model by using the fine-tuning training sample to obtain a training large model; perform reinforcement training on the training large model by using a reinforcement learning method based on human feedback, by using a plurality of ranking samples constituted by interface description language corresponding to the natural language description information of the same front-end interface and ranked according to accuracy, to obtain a reinforced training large model; and perform parameter merging on model parameters of the reinforced training large model and model parameters of the basic generative large language model by using a reparameterization method, to obtain a preset first generative large language model; The interface description language generation unit is configured to input the interface description information as input information into the preset first generative large language model, to obtain interface description language output by the first generative large language model; The association indication information acquisition unit is configured to obtain input association indication information for associating back-end data constituting the expected development result with the front-end interface; The database management language generation unit is configured to input the association indication information as input information into a preset second generative large language model, to obtain database management language output by the second generative large language model; The expected development result generation unit is configured to generate a corresponding front-end interface according to the interface description language, and associate corresponding database data with the front-end interface according to the database management language, to obtain the expected development result.

16. The apparatus of claim 15, wherein, The elements include: a region division method within an interface, controls contained in an interface, arrangement orders between different controls, style styles including colors and line distributions.

17. The apparatus of claim 16, further comprising: The semantic understanding and correction unit is configured to perform semantic understanding on the interface description information before inputting the interface description information into the first generative large language model, and to perform completion and adjustment on the interface description information according to the semantic understanding result, so as to input the completed and adjusted interface description information as input information into the first generative large language model.

18. The apparatus of claim 15, wherein, The interface description language is interface description code in a JSON format.

19. The apparatus of claim 15, wherein, The association indication information includes: natural language association indication information for describing an association form between the back-end data and the front-end interface in a natural language form, and graphical association indication information for embodying the association form in a form of a schematic diagram.

20. The apparatus of claim 15, wherein, In response to the back-end data being stored in a relational database, the database management language is a SQL instruction language for managing and operating the relational database.

21. The apparatus of claim 16, wherein, The first generative large language model training unit is further configured to: obtain a first basic generative large language model trained based on a general same-modal training sample, and obtain a fine-tuning training sample constituted by the natural language description information and corresponding interface description language; The first basic generative large language model is subjected to supervised fine-tuning using the fine-tuned training samples to obtain the first training large model. The sorted samples, consisting of multiple interface description languages ​​ranked by accuracy corresponding to the natural language description information of the same front-end interface, are used to reinforce the first training model with a reinforcement learning method based on human feedback to obtain the second training model. The model parameters of the second large training model and the model parameters of the first basic generative large language model are merged using a parameter re-parameter method to obtain the first generative large language model.

22. The apparatus of claim 21, further comprising: The first valid sample determination unit is configured to save the interface description language corresponding to the front-end interface that meets the expected development result of the first preset requirement as the first valid sample. The first model parameter adjustment unit is configured to adjust the model parameters of the first generative large language model using the first valid samples.

23. The apparatus of claim 16, wherein, The first generative large language model training unit is further configured as follows: Obtain the second basic generative large language model trained based on general cross-modal training samples, and obtain fine-tuned training samples based on the image description information and the corresponding interface description language. The second basic generative large language model is subjected to supervised fine-tuning using the fine-tuned training samples to obtain the third training large model. The sorted samples, consisting of multiple interface description languages ​​ordered by accuracy corresponding to the image description information of the same front-end interface, are used to reinforce the third training model with a reinforcement learning method based on human feedback to obtain the fourth training model. The model parameters of the fourth training large model and the model parameters of the second basic generative large language model are merged using the reparameterization method to obtain the first generative large language model.

24. The apparatus of claim 23, further comprising: The second valid sample determination unit is configured to save the interface image corresponding to the front-end interface that meets the expected development result of the second preset requirement as the second valid sample. The second model parameter adjustment unit is configured to adjust the model parameters of the first generative large language model using the second effective sample.

25. The apparatus of claim 19, further comprising: The second generative large language model training unit is further configured as follows: Obtain the third basic generative large language model trained based on the general same modality training samples, and obtain the fine-tuning training samples based on the natural language association indication information and the corresponding database management language. The third basic generative large language model is subjected to supervised fine-tuning using the fine-tuned training samples to obtain the fifth training large model. The sorted samples, consisting of multiple database management languages ​​ranked by accuracy corresponding to the natural language association indication information of the same association form, are reinforced and trained on the fifth training model using a reinforcement learning method based on human feedback to obtain the sixth training model. The model parameters of the sixth training large model and the model parameters of the third basic generative large language model are merged using the reparameterization method to obtain the second generative large language model.

26. The apparatus of claim 19, further comprising: The second generative large language model training unit is further configured as follows: Obtain the fourth basic generative large language model trained based on general cross-modal training samples, and obtain fine-tuning training samples based on the graphical association indication information and the corresponding database management language. The fourth basic generative large language model is subjected to supervised fine-tuning using the fine-tuned training samples to obtain the seventh training large model. The sorted samples, consisting of multiple database management languages ​​ordered by accuracy corresponding to the graphical association indication information of the same association form, are reinforced and trained on the seventh training model using a reinforcement learning method based on human feedback, to obtain the eighth training model. The model parameters of the eighth training large model and the model parameters of the fourth basic generative large language model are merged using the reparameterization method to obtain the second generative large language model.

27. The apparatus of claim 16, wherein, The expected development result generation unit includes a front-end interface generation unit configured to generate a corresponding front-end interface according to the interface description language, wherein the front-end interface generation unit is further configured to: In response to the simultaneous presence of the natural language description information and the image description information, the reception time of receiving the natural language description information and the image description information respectively is determined; Based on the order of the received times, determine the primary and secondary description information of the front-end interface; An initial front-end interface is generated according to the interface description language corresponding to the primary description information, and the initial front-end interface is adjusted according to the interface description language corresponding to the secondary description information to obtain the final front-end interface.

28. The apparatus according to any one of claims 15-27, further comprising: The version number determination unit is configured to determine the target version number corresponding to the expected development result; The release unit is configured to synthesize the target version number as the release version number of the desired development result, and release the desired development result according to the release version number.

29. An electronic device comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the low-code development method according to any one of claims 1-14.

30. A non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the low-code development method according to any one of claims 1-14.

31. A computer program product comprising a computer program that, when executed by a processor, implements the steps of the low-code development method according to any one of claims 1-14.

Images