Low-code platform development method and system

Abstract

The invention discloses a low-code platform development method and system, and the method comprises the steps: collecting business demand information related to a target business according to an obtained demand instruction, and converting the business demand information into first business demand data in a picture format; the method comprises the following steps: firstly, acquiring first business demand data, then analyzing the first business demand data based on a pre-trained optical character recognition large model to obtain second business demand data, then storing the second business demand data in a low-code knowledge base to obtain an updated low-code knowledge base, and if a method stream corresponding to a target business needs to be created, creating a method stream corresponding to the target business. And if the target business is updated, generating method flow configuration metadata corresponding to the target business by utilizing a pre-trained low-code field special model based on the updated low-code knowledge base, and finally, rendering the method flow configuration metadata based on a low-code platform to obtain a method flow corresponding to the target business. Through the method, the flow arrangement capability and the business demand development efficiency in the low-code development process can be effectively improved.

Description

Technical Field

[0001] This application relates to the fields of low-code development and artificial intelligence technology, and in particular to a low-code platform development method and system. Background Technology

[0002] With the accelerated development of digital transformation, low-code development platforms have become core tools for enterprises to build applications due to their ability to simplify the development process. How to improve the intelligence level of low-code development platforms is an important technical issue.

[0003] Existing low-code development platforms primarily connect to general large models through plug-in methods during the development of business requirements. In terms of method flow orchestration, they usually rely on preset templates to manually match business requirements. Specifically, it requires manually selecting the combination logic of existing data, third-party interfaces and system methods, and then manually configuring API (Application Programming Interface) call parameters.

[0004] However, existing low-code development methodologies, by directly adopting general, large models and manually matching preset templates, do not fully utilize existing data and third-party interfaces. This results in code generated for user-defined code editing scenarios that may not conform to the API specifications and development constraints of the low-code platform. Therefore, there is a need to provide a low-code platform development methodology with stronger workflow orchestration capabilities and higher efficiency in developing business requirements. Summary of the Invention

[0005] The purpose of this invention is to provide a low-code platform development method and system to solve the problems of poor process orchestration capabilities and low efficiency in developing business requirements in the existing low-code development process.

[0006] To solve the above-mentioned technical problems, the present invention is implemented as follows: In a first aspect, the present invention provides a low-code platform development method, the method comprising: Based on the obtained demand instructions, collect business demand information related to the target business, and convert the business demand information into first business demand data in image format; The first business requirement data is parsed and processed based on a pre-trained optical character recognition large model to obtain the second business requirement data. The second business requirement data is structured text information. The optical character recognition large model is a model trained based on multiple first business requirement data samples, multiple first method flow data samples, and a preset first loss function for various different businesses. The optical character recognition large model is a dual-component architecture large model composed of a visual compressor and a decoder. The second business requirement data is stored in the low-code knowledge base to obtain the updated low-code knowledge base. The low-code knowledge base is a semantic knowledge base built on the low-code platform method flow configuration metadata, historical method flow configuration metadata, historical second business requirement data, and historical method flow data. If it is necessary to create a method flow corresponding to the target business, then based on the updated low-code knowledge base, the pre-trained low-code domain-specific model is used to generate the method flow configuration metadata corresponding to the target business. The low-code domain-specific model is a model obtained by fine-tuning the base model based on the low-code knowledge base and the metadata specification of the low-code platform method flow. The method flow configuration metadata is rendered using a low-code platform to obtain the method flow corresponding to the target business.

[0007] Optionally, the pre-trained optical character recognition large model parses and processes the first business requirement data to obtain the second business requirement data, including: Input the first business requirement data into the pre-trained optical character recognition large model; The optical character recognition big model uses a visual compressor based on a window attention mechanism to perform local feature extraction processing on the first business requirement data in image format to obtain a first visual unit sequence. The optical character recognition large model performs convolutional compression on the first visual unit sequence using a visual compressor to obtain the second visual unit sequence; The optical character recognition large model performs global semantic modeling on the second visual unit sequence through a visual compressor to obtain a third visual unit sequence. Each visual unit in the third visual unit sequence includes text semantics and document logical layout semantics. The third visual unit sequence is input into the decoder to obtain the second service requirement data.

[0008] Optionally, the business requirement information includes one or more of the following: documents, pages, images, and prototype files. The step of converting the business requirement information into first business requirement data in image format includes: If the business requirement information is an image, then the business requirement information is the first business requirement data; If the business requirement information is a webpage, then a headless browser is used to render the business requirement information to obtain the first business requirement data in webpage image format. If the business requirement information is a document, then the business requirement information is virtually printed to obtain the first business requirement data in document image format; If the business requirement information is information other than images, pages, and documents, then the business requirement information is converted into first business requirement data in image format by calling the application programming interface related to the business requirement information.

[0009] Optionally, the training method for the large-scale optical character recognition model includes: Obtain multiple first business requirement data samples and multiple first method flow data samples for various different businesses; The acquired first business requirement data samples and first method flow data samples are respectively input into the visual compressor of the optical character recognition large model to obtain compressed first business requirement data samples and first method flow data samples. The compressed first business requirement data sample and the first method flow data sample are respectively input into the decoder of the optical character recognition large model to obtain the corresponding output results. Based on the output results and the preset first loss function, it is determined whether the optical character recognition large model has converged. If not, the optical character recognition large model is trained again based on the first business requirement data sample and the first method flow data sample until the optical character recognition large model converges, and the trained optical character recognition large model is obtained.

[0010] Optionally, before parsing and processing the first business requirement data based on a pre-trained optical character recognition large model to obtain the second business requirement data, the method further includes: The first business requirement data is subjected to image enhancement processing.

[0011] Optionally, the format of the second business requirement data includes any one of the following: LaTeX format, structured JSON format, and plain text format.

[0012] Optionally, the model training method for the low-code domain-specific model includes: Obtain metadata samples of multiple method flow configurations for various different businesses, data samples of multiple second business requirements for various different businesses, data samples of multiple second method flows for various different businesses, and metadata specifications of method flows for low-code platforms; The acquired method flow configuration metadata samples, multiple second business requirement data samples, multiple second method flow data samples, and the metadata specification of the low-code platform method flow are respectively input into the base model. The base model is then fine-tuned based on a preset second loss function to obtain the fine-tuned model. Finally, the fine-tuned model is optimized using reinforcement learning with human feedback to obtain the trained low-code domain-specific model.

[0013] Optionally, based on the updated low-code knowledge base, method flow configuration metadata corresponding to the target business is generated using a pre-trained low-code domain-specific model, including: Based on the second business requirement data, the method flow configuration metadata, method flow data, and business requirement data related to the target business are retrieved from the updated low-code knowledge base through the low-code domain-specific model. Input the metadata specification of the low-code platform method flow, the recalled method flow configuration metadata, method flow data, and business requirement data into the low-code domain-specific model to obtain the method flow configuration metadata corresponding to the target business.

[0014] Optionally, after rendering the method flow configuration metadata based on a low-code platform to obtain the method flow corresponding to the target business, the method further includes: Save the method flow configuration metadata; The method flow corresponding to the target business is converted into first method flow data in image format; The first method stream data is parsed and processed based on a pre-trained large-scale optical character recognition model to obtain the second method stream data; The second method stream data is stored in a low-code knowledge base.

[0015] Secondly, the present invention provides a low-code platform development system, the system comprising: The business requirement collection module is used to collect business requirement information related to the target business based on the obtained requirement instructions. The business requirement conversion module is used to convert the business requirement information into first business requirement data in image format. The business requirement parsing module is used to parse and process the first business requirement data based on a pre-trained optical character recognition large model to obtain the second business requirement data. The second business requirement data is structured text information. The optical character recognition large model is a model trained based on multiple first business requirement data samples, multiple first method flow data samples, and a preset first loss function. The optical character recognition large model is a dual-component architecture large model composed of a visual compressor and a decoder. The low-code knowledge base module is used to construct a semantic knowledge base based on the low-code platform method flow configuration metadata, historical method flow configuration metadata, historical second business requirement data, historical method flow data, and the second business requirement data and method flow data parsed and processed by the business requirement parsing module. A low-code domain-specific model is used to determine whether a method flow corresponding to the target business needs to be created based on the requirement instruction. If a method flow corresponding to the target business needs to be created, the low-code knowledge base is used to generate configuration metadata for the method flow corresponding to the target business. The low-code domain-specific model is a model obtained by fine-tuning and training the base model based on the low-code knowledge base and the metadata specification of the low-code platform method flow. The method flow generation module is used to render the method flow configuration metadata based on a low-code platform to obtain the method flow corresponding to the target business.

[0016] Optionally, the low-code platform development system further includes: A business requirement storage module is used to store the number of the first business requirements and the relationship between the first business requirement data. The method flow orchestration module is used to provide basic business logic and to perform visual orchestration of the basic business logic through interaction with the low-code intelligent agent, generating low-code platform method flow configuration metadata and method flow data. The method flow configuration conversion module is used to convert the method flow data generated by the method orchestration module into method flow data in image format. The method flow configuration storage module is used to store low-code modified method flow configuration metadata generated by the method flow orchestration module and image-formatted method flow data generated by the method flow configuration conversion module. The low-code platform agent is used to provide configuration suggestions to the method flow orchestration module and control the method flow orchestration module to perform visual orchestration based on the acquired requirement instructions.

[0017] Thirdly, the present invention provides an electronic device including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the low-code platform development method provided above.

[0018] The technical solution provided in this application may include the following beneficial effects: This specification provides a low-code platform development method. By processing business requirement information into images, it lays the foundation for subsequent parsing and processing. Furthermore, it fully utilizes third-party interfaces during the processing of business requirements in different formats, thereby improving the compatibility of the generated code with the low-code platform API specifications and thus increasing the efficiency of business requirement development. The large-scale optical character recognition model in this specification is a model trained based on multiple first business requirement data samples, multiple first method flow data samples, and a preset first loss function from various different business scenarios. This large-scale optical character recognition model is a dual-component architecture model composed of a visual compressor and a decoder. This dual-component architecture model can form a complete compression-decompression process. The final parsed second business requirement data is structured text information. The use of the visual compressor and decoder not only improves recognition accuracy but also significantly reduces computational load and allows for more efficient text representation using visual units, thereby improving business process efficiency and data quality. This is beneficial for improving workflow orchestration capabilities and development efficiency in the low-code development process. By storing the second business requirement data in a low-code knowledge base and updating the knowledge base in a timely manner, the data quality of the low-code knowledge base can be improved. This provides higher-quality foundational data for the low-code domain-specific model, thereby improving the accuracy of the method flow configuration metadata corresponding to the target business. The low-code domain-specific model in this specification is a model obtained by fine-tuning and training the base model based on the low-code knowledge base and the low-code platform's method flow metadata specifications. It can fully utilize the existing data characteristics of the low-code platform and deeply integrate with the low-code editing nodes of the low-code platform, thereby providing more accurate prompts for specific business scenarios, which is conducive to improving process orchestration capabilities and business requirement development efficiency. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating a low-code platform development method provided in an embodiment of this application; Figure 2 This is a schematic diagram of a business requirement processing flow provided in an embodiment of this application; Figure 3This is a schematic diagram of a method orchestration process provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a low-code platform development system provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of another low-code platform development system provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0022] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of this application.

[0023] To better understand this application, the embodiments of this application will be explained in detail below with reference to the accompanying drawings.

[0024] Example 1 like Figure 1 As shown in the embodiments of this specification, a low-code platform development method is provided. The execution subject of this method can be a terminal device or a server. The terminal device can be a mobile phone, tablet computer, or a computer device such as a laptop or desktop computer. The server can be a single server or a server cluster composed of multiple servers. Specifically, the method may include the following steps: In step S1, based on the obtained demand instructions, business demand information related to the target business is collected, and the business demand information is converted into first business demand data in image format.

[0025] The target business-related requirement information can be one or more of the following: documents, pages, images, and prototype files. In other words, the above requirement information can be any one or a combination of different types of information such as user requirement-related documents, sample pages or web pages, requirement-related sample images, and requirement-related prototype files.

[0026] After collecting the aforementioned business requirement information, it is processed into images. In implementation, screenshots can be used to convert the requirement information into image format. Alternatively, different methods can be used depending on the type of requirement information; for example, online conversion tools can be used for documents, while screenshots can be used for web pages. The processing of the first business requirement data into image format can take many forms. The following is one optional processing method, which may include the following: S11: If the business requirement information is an image, then the business requirement information is the first business requirement data.

[0027] If the business requirement information is an image, then it can be used directly without conversion.

[0028] S12: If the business requirement information is a webpage, then use a headless browser to render the business requirement information and obtain the first business requirement data in webpage image format.

[0029] Using a headless browser to render business requirements information can present web page information more realistically and completely. The rendering process is controllable, which helps improve data accuracy. Furthermore, it has a higher degree of automation and programmability, which helps improve the efficiency of low-code platform development.

[0030] S13: If the business requirement information is a document, then perform virtual printing processing on the business requirement information to obtain the first business requirement data in document image format.

[0031] By using virtual printing to process business requirements information in document form to obtain image-formatted data, data format consistency can be ensured, data accuracy can be improved, and virtual printing is compatible with multiple document formats, making it widely applicable. Virtual printing also facilitates batch processing, which helps to improve data processing efficiency.

[0032] S14: If the business requirement information is information other than images, pages, and documents, then the business requirement information is converted into first business requirement data in image format by calling the application programming interface related to the business requirement information.

[0033] If the business requirement information is information other than images, pages, and documents, processing it as an image by calling the API can make full use of third-party interfaces. This is beneficial for the optical character recognition model to fully understand the third-party interfaces, and makes the code generated in the code editing scenario more in line with the API specifications of the low-code platform, thereby improving the efficiency of business requirement development.

[0034] In step S2, the first business requirement data is parsed and processed based on the pre-trained optical character recognition big data model to obtain the second business requirement data.

[0035] Among them, the second business requirement data is structured text information, and the optical character recognition large model is a model obtained by training multiple first business requirement data samples, multiple first method flow data samples and a preset first loss function based on multiple different business data samples. The optical character recognition large model is a dual-component architecture large model composed of a visual compressor and a decoder.

[0036] The first business requirement data sample could be the delivery method and address information of logistics business, or the order quantity and product specifications of e-commerce order business. The first method flow data sample could be the tracking information flow of logistics business, or the shipping flow of e-commerce order business. The first loss function could be a contrastive loss function, a self-supervised pre-training loss function, etc.

[0037] The processing in step S2 above can take many forms. The following is one optional processing method, which may include the following: S21: Input the first business requirement data into the pre-trained optical character recognition large model.

[0038] S22: The large-scale optical character recognition model uses a visual compressor based on a window attention mechanism to perform local feature extraction on the first business requirement data in image format to obtain the first visual unit sequence.

[0039] In implementation, a lightweight visual compressor can be used. The large OCR (Optical Character Recognition) model can utilize a window attention mechanism based on SAM-base (Segment Anything Model) to divide the first business requirement data in image format into local windows to extract local features. These local features can include fine-grained features such as the position of text regions, font, line height, and character spacing. The output first visual unit sequence is the first visual token sequence. For example, 4096 tokens are output through step S22, corresponding to an initial image resolution of 1024×1024.

[0040] S23: The large-scale optical character recognition model performs convolutional compression on the first visual unit sequence using a visual compressor to obtain the second visual unit sequence.

[0041] Continuing with the example in step S22, when performing convolutional compression on the first visual unit sequence, two 3×3 convolutions (with a stride of 2) can be used, expanding the number of channels from 256 to 1024. This step reduces the number of the first visual unit sequence from 4096 to 1024, and then further to 256, achieving a 16-fold spatial compression. This results in a final second visual unit sequence with 256 channels. This significantly reduces the computational load while preserving key spatial structures (such as table lines and paragraph layouts), which is beneficial for further improving business process efficiency and low-code development efficiency.

[0042] S24: The large optical character recognition model performs global semantic modeling on the second visual unit sequence through a visual compressor to obtain the third visual unit sequence. Each visual unit in the third visual unit sequence includes text semantics and document logical layout semantics.

[0043] Continuing with the example in step S23, when the OCR large model performs global semantic modeling (i.e., global semantic understanding) on ​​the second visual unit sequence, it can utilize CLIP-large (Contrastive Language-Image Pretraining - Large Model) to understand the overall logical structure of the document (such as the hierarchical relationship between title, body text, table, and figure caption), thereby performing global semantic modeling on the compressed second visual unit sequence. The output third visual sequence (e.g., 256 visual units) includes textual semantics and document logical layout semantics for each visual unit. This method can further efficiently represent the textual information of visual units while improving recognition accuracy, which is beneficial for improving data quality, business process efficiency, and low code development efficiency.

[0044] S25: Input the third visual unit sequence into the decoder to obtain the second business requirement data.

[0045] In implementation, the decoder employs a MoE decoder (Mixture of Experts Decoder). The third visual unit sequence is input into the MoE decoder, which performs semantic decoding to directly generate the second business requirement data in structured text form. Taking a lightweight visual compressor and a MoE decoder as an example, when processing documents, the lightweight visual compressor acts like a "rapid scanner," not requiring accuracy in recognizing every single character, but focusing on the stroke combinations, arrangement patterns, and key features of the recognized characters. The lightweight visual compressor ultimately outputs a string of visual tokens (i.e., the third visual unit sequence, similar to a compressed file). This visual token is small in size but contains the core information of the recognized text. The MoE decoder acts like a "team of expert translators," where different "experts" specialize in handling different dimensions of information in the document (e.g., some experts specialize in printed text, some in artistic fonts, and some in tables). The MoE decoder automatically assigns the visual tokens output by the lightweight visual compressor to the most suitable expert for interpretation, ultimately outputting accurate text content.

[0046] The format of the second business requirement data can include any one of the following: LaTeX format, structured JSON format, and plain text format. LaTeX format can accurately present professional symbols and logic, such as mathematical formulas, chemical formulas, equations from physics disciplines, and statistical probability formulas; structured JSON format is easy for machines to read and can be used not only for converting user and product information tables, but also for e-commerce order data, logistics tracking information, and equipment status monitoring data; plain text format can preserve the original paragraphs and line break logic of the text and can be used for product instructions, article summaries, press release text, meeting minutes, and legal and regulatory provisions.

[0047] In addition, the second business requirement data can also include more information types, such as: graphic location information, in addition to web design scenarios, can also be used for APP interface element layout (such as button, input box, pop-up window position), map marker position and other scenarios; user structured information can include membership level benefits, consumption habit tags and other dimensions, so as to fully meet diverse needs based on file type characteristics, which is conducive to improving the applicability of low-code platform development methods.

[0048] In practice, the training method for a large-scale optical character recognition model can be implemented using the following steps: A1: Obtain multiple first business requirement data samples and multiple first method flow data samples for various different businesses; A2: Input the acquired multiple first business requirement data samples and multiple first method flow data samples into the visual compressor of the optical character recognition large model to obtain compressed first business requirement data samples and first method flow data samples. A3: Input the compressed first business requirement data sample and the first method flow data sample into the decoder of the optical character recognition large model to obtain the corresponding output results. Based on the output results and the preset first loss function, determine whether the optical character recognition large model has converged. If not, continue to train the optical character recognition large model based on the first business requirement data sample and the first method flow data sample until the optical character recognition large model converges, and obtain the trained optical character recognition large model.

[0049] During the training process of this model, the first business requirement data sample and the first method flow data sample are used as input data, which can effectively improve the accuracy of the OCR large model training results and output high-quality second business requirement data.

[0050] Furthermore, prior to step S2, the low-code platform development method in this embodiment may further include standardizing the first business requirement data. Specifically, the process of standardizing the first business requirement data may include the following steps: S6: Adjust the image resolution of the first business requirement data to match the image resolution of the OCR large model.

[0051] By adjusting the image allocation rate of the primary business requirement data, data adaptability can be further improved, thereby enhancing the accuracy of data processing.

[0052] S7: Perform image enhancement processing on the first business requirement data.

[0053] In implementation, different image enhancement processing methods can be adopted according to the different low-quality problems of the primary business data. For example, for the low-quality problem of blurred images, the method of first sharpening the image and then adjusting the contrast can be adopted. For the low-quality problem of uneven lighting, multi-scale retina-cortex algorithms and contrast adjustment methods can be adopted. For the low-quality problem of tilted images, affine transformation and perspective transformation methods can be adopted.

[0054] By performing image enhancement processing on the primary business requirement data, the robustness of the OCR large model in recognizing the primary business requirement data can be effectively improved, thereby increasing the efficiency of business process processing.

[0055] See also Figure 1 As can be seen, in step S3, the second business requirement data is stored in the low-code knowledge base to obtain the updated low-code knowledge base.

[0056] Among them, the low-code knowledge base is a semantic knowledge base built on low-code platform method flow configuration metadata, historical method flow configuration metadata, historical second business requirement data, and historical method flow data.

[0057] This low-code knowledge base is built on the original method flow configuration metadata, historical method flow configuration metadata, high-quality historical second business requirement data output by the OCR large model during use, and historical method flow data of the low-code platform. It can provide more accurate data support for subsequent low-code domain-specific models, thereby improving the efficiency of low-code platform development.

[0058] When storing secondary business requirement data in a low-code knowledge base, different knowledge bases can be used depending on the type of the secondary business requirement data. For example, for output data in text / JSON format, a Weaviate vector database can be used for storage, while for structured data extracted from financial statements, invoices, and forms, a relational knowledge base can be used for storage.

[0059] In step S4, if it is necessary to create a method flow corresponding to the target business, the method flow configuration metadata corresponding to the target business is generated based on the updated low-code knowledge base and using a pre-trained low-code domain-specific model.

[0060] Among them, the low-code domain-specific model is a model obtained by fine-tuning and training the base large model based on the metadata specifications of the low-code knowledge base and the low-code platform method flow.

[0061] The model training method for the low-code domain-specific model in the embodiments of this specification can be implemented using the following steps: B1: Obtain metadata samples of multiple method flow configurations for various different businesses, data samples of multiple second business requirements for various different businesses, data samples of multiple second method flow for various different businesses, and metadata specifications of low-code platform method flows; B2: Input the acquired multiple method flow configuration metadata samples, multiple second business requirement data samples, multiple second method flow data samples, and the metadata specification of the low-code platform method flow into the base model. Based on the preset second loss function, perform instruction fine-tuning training on the base model to obtain the instruction-fine-tuned large model. Then, perform feedback optimization processing on the instruction-fine-tuned large model based on reinforcement learning using human feedback to obtain the trained low-code domain-specific model.

[0062] In implementation, after inputting the acquired method flow configuration metadata samples, second business requirement data samples, second method flow data samples, and low-code platform method flow metadata specifications into the base model, the base model can be further pre-trained based on a preset second loss function to obtain a further pre-trained model. Then, the further pre-fine-tuned model can be optimized using reinforcement learning with human feedback to obtain a trained low-code domain-specific model. Alternatively, instruction fine-tuning training can be performed to obtain an instruction-fine-tuned model, and the instruction-fine-tuned model can be optimized using reinforcement learning with human feedback to obtain a trained low-code domain-specific model. The base model can also be further pre-trained and instruction fine-tuned trained sequentially, and then the low-code domain-specific model can be optimized using reinforcement learning with human feedback.

[0063] In the process of pre-training the model, fine-tuning the training with instructions, and using reinforcement learning with human feedback for feedback optimization, parameter-efficient fine-tuning methods such as LoRA (Low-Rank Adaptation), Prefix Tuning, and Adapter can be adopted. These methods enable the base model to gain a deeper understanding of the knowledge, terminology, and task specifications of the low-code domain, thereby acquiring low-code domain expert capabilities and further improving the efficiency of low-code platform development.

[0064] In addition, the acquired multiple method flow configuration metadata samples, multiple second business requirement data samples, multiple second method flow data samples, and low-code platform method flow metadata specifications are input into the base model. The base model is then fine-tuned based on a preset second loss function. After obtaining the fine-tuned model, the low-code domain-specific model can also be optimized using a direct preference optimization approach.

[0065] Furthermore, the model training method for the low-code domain-specific model in the embodiments of this specification may further include the following steps: B3: Obtain multiple intermediate models generated during the iterative training process of low-code domain-specific models; B4: Using the above intermediate models as teacher models and the currently trained low-code domain-specific model as student models, a knowledge distillation method is used to distill and train the student model through the teacher model to obtain the distilled low-code domain-specific model.

[0066] By employing knowledge distillation to train low-code domain-specific models, we can significantly reduce model size, lower computational resources, and improve training efficiency and accuracy while maintaining training accuracy.

[0067] Furthermore, the model training method for the low-code domain-specific model in the embodiments of this specification may further include the following steps: B5: Perform model pruning on the low-code domain-specific model trained by distillation to obtain the low-code domain-specific model after model pruning.

[0068] By pruning low-code domain-specific models, the model size can be further reduced, computational resource consumption can be decreased, and thus the efficiency of model training can be improved.

[0069] Furthermore, the model training method for the low-code domain-specific model in the embodiments of this specification may further include the following steps: B6: Perform model quantization on the low-code domain-specific model after model pruning to obtain the low-code domain-specific model after model quantization.

[0070] By performing model quantization on low-code domain-specific models, the model size can be further reduced, computational resource consumption can be decreased, and thus the efficiency of model training can be improved.

[0071] In practice, after executing steps B1-B2, the trained low-code domain-specific model can be pruned separately to obtain a pruned low-code domain-specific model. Alternatively, the trained low-code domain-specific model can be quantized separately to obtain a quantized low-code domain-specific model. Alternatively, steps B1-B4 can be executed to complete the training of the low-code domain-specific model; this is not a limitation.

[0072] See also Figure 1 As can be seen, during implementation, it can be determined whether a method flow corresponding to the target business needs to be created based on the obtained requirement instructions, thereby identifying the user intent. If a method flow corresponding to the target business does not need to be created, the process ends; if it needs to be created, step S4 is executed.

[0073] The processing in step S4 above can take many forms. The following is one optional processing method, which may include the following: S41: Based on the second business requirement data, retrieve the method flow configuration metadata, method flow data, and business requirement data related to the target business from the updated low-code knowledge base through the low-code domain-specific model; By retrieving existing data such as method flow configuration metadata, method flow data, and business requirement data related to the target business from the low-code knowledge base, the existing data and characteristics of the low-code platform can be fully utilized to improve the accuracy of low-code-specific models in recognizing user intent. This enables more targeted and accurate prompts to be provided for different business scenarios, thereby improving process orchestration capabilities.

[0074] S42: Input the metadata specification of the low-code platform method flow, as well as the recalled method flow configuration metadata, method flow data, and business requirement data into the low-code domain-specific model to obtain the method flow configuration metadata corresponding to the target business.

[0075] In step S5, the method flow configuration metadata is rendered based on the low-code platform to obtain the method flow corresponding to the target business.

[0076] The processing in step S5 above can take many forms. The following is one optional processing method, which may include the following: S51: Save method flow configuration metadata; S52: Convert the method flow corresponding to the target business into first method flow data in image format; S53: Based on the pre-trained large-scale optical character recognition model, the first method stream data is parsed and processed to obtain the second method stream data; S54: Store the second method flow data in a low-code knowledge base.

[0077] The processing methods and principles of steps S52-S54 above can be referred to the processing of business requirement information related to the target business in steps S1-S3, and will not be repeated here.

[0078] Through the above steps S51-S54, effective data generated during the low-code development process can be added to the low-code knowledge base in a timely manner, further increasing the richness and accuracy of the data. This is beneficial for further optimization of the low-code domain-specific model, thereby improving the efficiency of low-code platform development.

[0079] Furthermore, after step S5 in the embodiments of this specification, the following process may also be included: S8: Based on the obtained user feedback instructions, verify the method flow corresponding to the target business until the method flow corresponding to the target business is qualified.

[0080] By further verifying the method flow corresponding to the target business obtained in step S5, the accuracy of the generated method flow can be further improved.

[0081] In summary, the low-code platform development method described in this specification mainly includes two stages in its implementation: a business requirement processing flow and a method flow orchestration flow. For details, please refer to [link to documentation / reference]. Figure 2 and Figure 3 .

[0082] Depend on Figure 2 The business requirement processing flow is as follows: After a user submits a new requirement, the requirement personnel upload the business requirement information, such as the written requirement document, the designed interaction diagram, and the prototype file of the example function, to the low-code platform. The low-code platform first determines whether the uploaded business requirement information file type is an image. If it is an image, it stores the image. Then it determines whether it is a web page file. If it is, it uses technologies such as headless browsers to convert the web page into an image before storing it. Next, it determines whether it is a document file. If it is, it uses virtual printing technology to convert the document into an image before storing it. If it is another type of file, it uses the corresponding API interface to convert it into an image before storing it. The first business requirement data is stored, which can be done using distributed file storage or other image servers. The first business requirement data is then standardized. Local image features are extracted. Visual unit compression is performed. Global semantic understanding is performed. Semantic decoding is performed, and the compressed visual units are sent to the MoE decoder to directly generate the second business requirement data in structured text form. The structured text is then output in a structured manner, and the second business requirement data is stored in the low-code knowledge base.

[0083] As shown in section 3, the method flow orchestration process is as follows: the user interacts with the low-code agent to develop business functions. The user needs to provide relevant business requirement information, and the provided business requirement information is processed according to... Figure 2 The process is as follows: The low-code domain-specific model identifies the user's intent and determines whether the user wants to create a business flow. If a business flow needs to be created, the low-code domain-specific model retrieves method flow configuration metadata, method flow data, and business requirement data (i.e., business requirement knowledge, similar method flow configuration metadata, method flow graphs, etc.) related to the target business from the low-code knowledge base to improve the accuracy of the model in identifying the user's intent. The low-code domain-specific model combines historical similar business data retrieved from the low-code knowledge base with the function currently being processed by the user to generate the method flow configuration metadata for the user's current business function. The low-code platform renders and presents the method flow configuration metadata for the current business function. The user checks the correctness of the method flow and confirms saving. The low-code platform saves the method flow configuration metadata. After the low-code platform converts the method flow into an image, it executes... Figure 2 The process shown is the same as the process for processing the first business requirement data.

[0084] Example 2 The above describes the low-code platform development method provided in the embodiments of this specification. Based on the same idea, the embodiments of this specification also provide a low-code platform development system, such as... Figure 4 As shown.

[0085] The low-code platform development system includes: a business requirement collection module 210, a business requirement transformation module 220, a business requirement parsing module 230, a low-code knowledge base module 240, a low-code domain-specific model 250, and a method flow generation module. Among them: The business requirement collection module 210 is used to collect business requirement information related to the target business based on the obtained requirement instructions. The business requirement conversion module 220 is used to convert business requirement information into first business requirement data in image format. The business requirement parsing module 230 is used to parse and process the first business requirement data based on the pre-trained optical character recognition large model to obtain the second business requirement data. The second business requirement data is structured text information. The optical character recognition large model is a model trained based on multiple first business requirement data samples, multiple first method flow data samples and a preset first loss function. The optical character recognition large model is a dual-component architecture large model based on a visual compressor and a decoder. The low-code knowledge base module 240 is used to build a semantic knowledge base based on the low-code platform method flow configuration metadata, historical method flow configuration metadata, historical second business requirement data, historical method flow data, and second business requirement data and method flow data parsed and processed by the business requirement parsing module. The low-code domain-specific model 250 is used to determine whether a method flow corresponding to the target business needs to be created based on the requirement instructions. If a method flow corresponding to the target business needs to be created, the low-code knowledge base is used to generate the method flow configuration metadata corresponding to the target business. The low-code domain-specific model is a model obtained by fine-tuning and training the base model based on the low-code knowledge base and the low-code platform method flow metadata specification. The method flow generation module 260 is used to render and process the method flow configuration metadata based on the low-code platform to obtain the method flow corresponding to the target business.

[0086] In the embodiments described in this specification, the business requirement parsing module 230 includes two parts: a visual compressor and a decoder.

[0087] The visual compressor includes: a visual perception unit, used to acquire first business requirement data and perform local feature extraction processing on the first business requirement data in image format based on a window attention mechanism to obtain a first visual unit sequence; a convolutional compression unit, used to perform convolutional compression processing on the first visual unit sequence to obtain a second visual unit sequence; and a visual knowledge unit, used to perform global semantic modeling processing on the second visual unit sequence to obtain a third visual unit sequence, wherein each visual unit in the third visual unit sequence includes text semantics and document logical layout semantics. The decoder is used to acquire the third visual unit sequence and decode it to obtain the second business requirement data.

[0088] In this embodiment of the specification, the business requirement information collected by the business requirement conversion module 220 includes one or more of the following: documents, pages, images, and prototype files.

[0089] Business requirement conversion module 220 includes: The first conversion unit, if the business requirement information is an image, then the business requirement information is the first business requirement data; The second conversion unit, if the business requirement information is a page, uses a headless browser to render the business requirement information and obtain the first business requirement data in webpage image format. The third conversion unit performs virtual printing processing on the business requirement information if it is a document, and obtains the first business requirement data in document image format. The fourth conversion unit converts the business requirement information into first business requirement data in image format by calling the application programming interface related to the business requirement information if the business requirement information is information other than images, pages, and documents.

[0090] In this embodiment of the specification, the training method of the optical character recognition large model in the business requirement parsing module 230 includes: acquiring multiple first business requirement data samples and multiple first method flow data samples of various different businesses; inputting the acquired multiple first business requirement data samples and multiple first method flow data samples into the visual compressor of the optical character recognition large model to obtain compressed first business requirement data samples and first method flow data samples; inputting the compressed first business requirement data samples and first method flow data samples into the decoder of the optical character recognition large model to obtain corresponding output results; and judging whether the optical character recognition large model has converged based on the output results and a preset first loss function. If not, the optical character recognition large model is trained again based on the first business requirement data samples and the first method flow data samples until the optical character recognition large model converges, thus obtaining the trained optical character recognition large model.

[0091] In the embodiments described in this specification, the low-code platform development system further includes: The image processing module is used to perform image enhancement processing on the data required for the first business needs.

[0092] In the embodiments of this specification, the format of the second business requirement data includes any one of the following: LaTeX format, structured JSON format, and plain text format.

[0093] In the embodiments of this specification, the model training method for the low-code domain-specific model 250 includes: acquiring multiple method flow configuration metadata samples, multiple second business requirement data samples, multiple second method flow data samples, and the metadata specification of the low-code platform method flow for various different businesses; inputting the acquired multiple method flow configuration metadata samples, multiple second business requirement data samples, multiple second method flow data samples, and the metadata specification of the low-code platform method flow into the base large model respectively; performing instruction fine-tuning training on the base large model based on a preset second loss function to obtain the instruction-fine-tuned large model; and performing feedback optimization processing on the instruction-fine-tuned large model based on reinforcement learning using human feedback to obtain the trained low-code domain-specific model.

[0094] In the embodiments described in this specification, the low-code platform development system further includes: The business requirement storage module is used to store the first business requirement data and the relationships between the first business requirement data; The method flow orchestration module provides basic business logic and visualizes the orchestration of the basic business logic through interaction with the low-code agent, generating low-code platform method flow configuration metadata and method flow data; The method flow configuration conversion module is used to convert the method flow data generated by the method orchestration module into image-formatted method flow data. The method flow configuration storage module is used to store low-code modified method flow configuration metadata generated by the method flow orchestration module and image-formatted method flow data generated by the method flow configuration conversion module. The low-code platform agent is used to provide configuration suggestions for the method flow orchestration module and control the method flow orchestration module to perform visual orchestration based on the acquired requirement instructions.

[0095] In the embodiments of this specification, the nodes corresponding to the basic business logic in the method flow orchestration module may include: data query nodes, parameter setting nodes, method call nodes, script call nodes, etc. For example, for a specific user query function, the parameter setting node can set the query parameters, the parameter setting node can query the database according to the set parameters, and the script call node can call a custom script to perform custom processing on the query results (such as dictionary conversion).

[0096] In the embodiments described in this specification, the low-code platform agent is an application layer module, which can call the base model, the low-code domain-specific model, the method flow orchestration module, etc. at its underlying level.

[0097] In this embodiment of the specification, the low-code knowledge base module is also used to store method flow configuration metadata; the business requirement conversion module 220 is also used to convert the method flow corresponding to the target business into first method flow data in image format; the business requirement parsing module 230 is also used to parse and process the first method flow data based on a pre-trained optical character recognition large model to obtain second method flow data; the low-code knowledge base module is also used to store the second method flow data.

[0098] In practical applications, low-code platform development systems can adopt methods such as... Figure 5 The structure shown.

[0099] Example 3 The above describes the low-code platform development system provided in the embodiments of this specification. Based on the same idea, the embodiments of this specification also provide an electronic device, such as... Figure 6 As shown. Electronic devices can vary considerably due to differences in configuration or performance, and may include one or more processors 301 and memory 302. Memory 302 may store one or more application programs or data. Memory 302 may be temporary or persistent storage. The application programs stored in memory 302 may include one or more modules (not shown), each module may include a series of computer-executable instructions for the electronic device. Furthermore, processor 301 may be configured to communicate with memory 302 and execute the series of computer-executable instructions in memory 302 on the electronic device. The electronic device may also include one or more power supplies 303, one or more wired or wireless network interfaces 304, one or more input / output interfaces 305, and one or more keyboards 306.

[0100] Specifically, in this embodiment, the electronic device includes a memory and one or more programs, wherein one or more programs are stored in the memory, and one or more programs may include one or more modules, and each module may include a series of computer-executable instructions for use in the electronic device, and is configured to be executed by one or more processors. The one or more programs include computer-executable instructions for performing the following: Based on the obtained demand instructions, collect business demand information related to the target business, and convert the business demand information into first business demand data in image format; The first business requirement data is parsed and processed based on the pre-trained optical character recognition large model to obtain the second business requirement data. The second business requirement data is structured text information. The optical character recognition large model is a model trained based on multiple first business requirement data samples, multiple first method flow data samples and a preset first loss function. The optical character recognition large model is a dual-component architecture large model based on a visual compressor and a decoder. The second business requirement data is stored in the low-code knowledge base to obtain the updated low-code knowledge base. The low-code knowledge base is a semantic knowledge base built on the low-code platform method flow configuration metadata, historical method flow configuration metadata, historical second business requirement data, and historical method flow data. If it is necessary to create a method flow corresponding to the target business, the method flow configuration metadata corresponding to the target business is generated based on the updated low-code knowledge base and using a pre-trained low-code domain-specific model. The low-code domain-specific model is a model obtained by fine-tuning and training the base model based on the low-code knowledge base and the metadata specification of the low-code platform method flow. The method flow configuration metadata is rendered and processed based on a low-code platform to obtain the method flow corresponding to the target business.

[0101] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A low-code platform development method, characterized in that, The method includes: Based on the obtained demand instructions, collect business demand information related to the target business, and convert the business demand information into first business demand data in image format; The first business requirement data is parsed and processed based on a pre-trained optical character recognition large model to obtain the second business requirement data. The second business requirement data is structured text information. The optical character recognition large model is a model trained based on multiple first business requirement data samples, multiple first method flow data samples, and a preset first loss function for various different businesses. The optical character recognition large model is a dual-component architecture large model composed of a visual compressor and a decoder. The second business requirement data is stored in the low-code knowledge base to obtain the updated low-code knowledge base. The low-code knowledge base is a semantic knowledge base built on the low-code platform method flow configuration metadata, historical method flow configuration metadata, historical second business requirement data, and historical method flow data. If it is necessary to create a method flow corresponding to the target business, then based on the updated low-code knowledge base, the pre-trained low-code domain-specific model is used to generate the method flow configuration metadata corresponding to the target business. The low-code domain-specific model is a model obtained by fine-tuning and training the base model based on the low-code knowledge base and the metadata specification of the low-code platform method flow. The method flow configuration metadata is rendered using a low-code platform to obtain the method flow corresponding to the target business.

2. The method according to claim 1, characterized in that, The pre-trained optical character recognition big data model parses and processes the first business requirement data to obtain the second business requirement data, including: Input the first business requirement data into the pre-trained optical character recognition large model; The optical character recognition large model uses a visual compressor based on a window attention mechanism to perform local feature extraction processing on the first business requirement data in image format to obtain a first visual unit sequence. The optical character recognition large model performs convolutional compression on the first visual unit sequence using a visual compressor to obtain the second visual unit sequence; The optical character recognition large model performs global semantic modeling on the second visual unit sequence through a visual compressor to obtain a third visual unit sequence. Each visual unit in the third visual unit sequence includes text semantics and document logical layout semantics. The third visual unit sequence is input into the decoder to obtain the second service requirement data.

3. The method according to claim 1, characterized in that, The business requirement information includes one or more of the following: documents, pages, images, and prototype files. The step of converting the business requirement information into first business requirement data in image format includes: If the business requirement information is an image, then the business requirement information is the first business requirement data; If the business requirement information is a webpage, then a headless browser is used to render the business requirement information to obtain the first business requirement data in webpage image format. If the business requirement information is a document, then the business requirement information is virtually printed to obtain the first business requirement data in document image format; If the business requirement information is information other than images, pages, and documents, then the business requirement information is converted into first business requirement data in image format by calling the application programming interface related to the business requirement information.

4. The method according to claim 1, characterized in that, The training method for the large-scale optical character recognition model includes: Obtain multiple first business requirement data samples and multiple first method flow data samples for various different businesses; The acquired first business requirement data samples and first method flow data samples are respectively input into the visual compressor of the optical character recognition large model to obtain compressed first business requirement data samples and first method flow data samples. The compressed first business requirement data sample and the first method flow data sample are respectively input into the decoder of the optical character recognition large model to obtain the corresponding output results. Based on the output results and the preset first loss function, it is determined whether the optical character recognition large model has converged. If not, the optical character recognition large model is trained again based on the first business requirement data sample and the first method flow data sample until the optical character recognition large model converges, and the trained optical character recognition large model is obtained.

5. The method according to claim 1, characterized in that, Before parsing and processing the first business requirement data based on a pre-trained optical character recognition large model to obtain the second business requirement data, the method further includes: The first business requirement data is subjected to image enhancement processing.

6. The method according to claim 1, characterized in that, The format of the second business requirement data includes any one of the following: LaTeX format, structured JSON format, and plain text format.

7. The method according to claim 1, characterized in that, The model training method for the low-code domain-specific model includes: Obtain metadata samples of multiple method flow configurations for various different businesses, data samples of multiple second business requirements for various different businesses, data samples of multiple second method flows for various different businesses, and metadata specifications of method flows for low-code platforms; The acquired method flow configuration metadata samples, multiple second business requirement data samples, multiple second method flow data samples, and the metadata specification of the low-code platform method flow are respectively input into the base model. The base model is then fine-tuned based on a preset second loss function to obtain the fine-tuned model. Finally, the fine-tuned model is optimized using reinforcement learning with human feedback to obtain the trained low-code domain-specific model.

8. The method according to claim 1, characterized in that, Based on the updated low-code knowledge base, a pre-trained low-code domain-specific model is used to generate method flow configuration metadata corresponding to the target business, including: Based on the second business requirement data, the method flow configuration metadata, method flow data, and business requirement data related to the target business are retrieved from the updated low-code knowledge base through the low-code domain-specific model. Input the metadata specification of the low-code platform method flow, the recalled method flow configuration metadata, method flow data, and business requirement data into the low-code domain-specific model to obtain the method flow configuration metadata corresponding to the target business.

9. The method according to claim 1, characterized in that, After rendering the method flow configuration metadata based on a low-code platform and obtaining the method flow corresponding to the target business, the method further includes: Save the method flow configuration metadata; The method flow corresponding to the target business is converted into first method flow data in image format; The first method stream data is parsed and processed based on a pre-trained large-scale optical character recognition model to obtain the second method stream data; The second method stream data is stored in a low-code knowledge base.

10. A low-code platform development system, characterized in that, The system includes: The business requirement collection module is used to collect business requirement information related to the target business based on the obtained requirement instructions. The business requirement conversion module is used to convert the business requirement information into first business requirement data in image format. The business requirement parsing module is used to parse and process the first business requirement data based on a pre-trained optical character recognition large model to obtain the second business requirement data. The second business requirement data is structured text information. The optical character recognition large model is a model trained based on multiple first business requirement data samples, multiple first method flow data samples, and a preset first loss function. The optical character recognition large model is a dual-component architecture large model composed of a visual compressor and a decoder. The low-code knowledge base module is used to construct a semantic knowledge base based on the low-code platform method flow configuration metadata, historical method flow configuration metadata, historical second business requirement data, historical method flow data, and the second business requirement data and method flow data parsed and processed by the business requirement parsing module. A low-code domain-specific model is used to determine whether a method flow corresponding to the target business needs to be created based on the requirement instruction. If a method flow corresponding to the target business needs to be created, the low-code knowledge base is used to generate configuration metadata for the method flow corresponding to the target business. The low-code domain-specific model is a model obtained by fine-tuning and training the base model based on the low-code knowledge base and the metadata specification of the low-code platform method flow. The method flow generation module is used to render the method flow configuration metadata based on a low-code platform to obtain the method flow corresponding to the target business.

11. The low-code platform development system according to claim 10, characterized in that, The low-code platform development system also includes: A business requirement storage module is used to store the first business requirement data and the relationship between the first business requirement data; The method flow orchestration module is used to provide basic business logic and to perform visual orchestration of the basic business logic through interaction with the low-code intelligent agent, generating low-code platform method flow configuration metadata and method flow data. The method flow configuration conversion module is used to convert the method flow data generated by the method orchestration module into method flow data in image format. The method flow configuration storage module is used to store low-code modified method flow configuration metadata generated by the method flow orchestration module and image-formatted method flow data generated by the method flow configuration conversion module. The low-code platform agent is used to provide configuration suggestions to the method flow orchestration module and control the method flow orchestration module to perform visual orchestration based on the acquired requirement instructions.

12. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the low-code platform development method as described in any one of claims 1 to 9.

Images