Method for generating system architecture diagram by automatically optimizing cue word driven reasoning large model

Abstract

The invention provides a method for generating a system architecture diagram by automatically optimizing a cue word-driven reasoning large model. The method comprises the following steps of: preparing a multi-mode and reasoning large model and a sample data picture; generating content description and an initial layout requirement text; inputting the combined cue words into a large reasoning model, generating an HTML5 code, and converting the HTML5 code into a target picture; comparing the similarity between the target picture and the sample layout by the multi-modal model, and outputting a difference if the similarity does not reach a threshold value; and optimizing the layout requirement text based on the difference, and iterating until the similarity reaches the standard. Through cross-modal model collaboration, cue word closed-loop optimization, multi-role expert simulation and other technical innovations, the core pain points of strong manual dependence, insufficient flexibility and limited accuracy in traditional system architecture diagram generation are effectively solved, the technical spanning from passive auxiliary drawing to active intelligent generation is realized, and the system architecture diagram generation efficiency is improved. And the method has remarkable advantages in the aspects of automation degree, adaptability and generation quality.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically a method for generating system architecture diagrams by automatically optimizing prompt words to drive a large inference model. Background Technology

[0002] In the software development process, manually drawing software architecture diagrams is a tedious yet crucial task when writing system design documents or architecture plans. Although developers typically complete the layered and modular design of the software before drawing the architecture diagram, and have hierarchical descriptions and the names of the modules included in each layer (hereinafter referred to as "content descriptions"), the following problems currently exist:

[0003] 1. Many popular and mature large models in China, such as Deepseek v3 / r1, are non-multimodal models and cannot directly generate images. Even large models with multimodal capabilities, such as Wenxin Yiyan and Grok, cannot correctly display Chinese characters when generating images, which greatly limits the application of generating system architecture diagrams using large models.

[0004] 2. The characteristic of a system architecture diagram is that its graphical elements must be layered according to technical concepts and grouped by module relationships. The position and hierarchy of the graphics reflect the technical design philosophy. Therefore, the prompts must be very accurate and detailed to reflect these layout requirements. However, manually optimizing prompts is time-consuming and labor-intensive, requiring a large amount of tokens and resources, and sometimes still fails to achieve the desired effect. Summary of the Invention

[0005] This invention provides a method for generating system architecture diagrams by automatically optimizing prompt words to drive a large inference model, thereby solving the problems existing in the prior art and improving the efficiency and accuracy of generating system architecture diagrams.

[0006] A method for generating system architecture diagrams by automatically optimizing prompt words to drive a large inference model includes the following steps:

[0007] Step 1: Model and Data Preparation

[0008] Select a multimodal large model and an inference large model, and prepare sample data images;

[0009] Step 2: Generate the initial prompt word component

[0010] Based on the sample data images, generate content description text and initial layout requirement text;

[0011] Step 3: Generate architecture diagram code and images

[0012] The content description text and initial layout requirement text from step two are combined into prompt words, which are then input into the inference model to generate the HTML5 page code corresponding to the system architecture diagram. The HTML5 page code is then converted into the target image.

[0013] Step 4: Similarity Comparison and Difference Analysis

[0014] Using the multimodal large model selected in step one, the layout similarity of the target image generated in step three is compared with that of the sample data image. If the similarity does not reach the preset threshold, the difference analysis results are output.

[0015] Step 5: Iterative optimization loop for prompt words

[0016] Based on the difference analysis results in step four, optimize the layout requirement text in step two, generate new prompt words, and repeat steps three to four until the similarity between the target image and the sample data image reaches a preset threshold to obtain the optimal layout requirement text.

[0017] Furthermore, in step one, the multimodal large model includes gok, and the inference large model includes one of deepseek, Tongyi Qianwen (Qwen3-Instruct text-specific version), Zhipu Qingyan, Huawei Pangu NLP series models, and Tencent Hunyuan;

[0018] The sample data image is a reference architecture diagram that includes the system architecture layer structure, module hierarchy, and connection methods.

[0019] Furthermore, in step two, the content description text defines the text information that needs to be presented in the system architecture diagram, including module names, hierarchical divisions, and functional descriptions.

[0020] The initial layout requires the text to include rules for arranging graphic elements, including layer positions, hierarchical relationships, relative distances, and line connection methods. The initial layout also requires the text to be generated manually or reused from historical optimization results.

[0021] Furthermore, in step three, the combination of the prompt words includes: defining the system roles of the reasoning model as a cross-domain expert team, which includes technical illustrators, visualization experts, system architects, and front-end engineers;

[0022] The prompts also include constraints: "Page content strictly follows the content description text", "Page layout strictly follows the layout requirements text", and "Text and graphics, and graphics and graphics do not overlap".

[0023] Furthermore, in step four, the criteria for judging the layout similarity comparison include:

[0024] (1) Ignore the differences in color and the differences in the thickness of the lines on the graph;

[0025] (2) Compare whether the text on similar positions in the diagram is the same;

[0026] (3) Compare whether the overall layout is similar and whether the levels are the same;

[0027] (4) Compare whether the number of graphics in the same row is the same, and compare whether the number of graphics in the same column is the same;

[0028] (5) Compare whether elements in the same layer are similar, and compare whether elements in the same module are similar;

[0029] (6) Compare whether the division and combination relationships of the graphic elements on the diagram are the same;

[0030] (7) Compare whether the size and position of each graphic element on the graph are the same;

[0031] (8) Only when all of items 2-7 are satisfied will the two images be judged to be similar, and a similarity evaluation result will be given: yes means similar, no means dissimilar.

[0032] (9) If they are not similar, please explain in detail the differences between the second picture and the first picture in items 2-7 above;

[0033] The difference analysis results include specific difference types, including items (2)-(7) above.

[0034] Furthermore, in step five, the termination condition for the iterative optimization loop is: the similarity between the target image and the sample data image reaches a preset threshold, or the number of loops reaches the upper limit.

[0035] Furthermore, the role definition of the cross-domain expert team is achieved through system role instructions in the prompt words, which is used to constrain the large inference model to integrate visual design rules and system architecture logic when generating HTML5 page code.

[0036] Furthermore, the HTML5 page code generated in step three contains SVG graphic elements, and the attributes of the SVG graphic elements correspond one-to-one with the parameters in the layout requirement text, including position coordinates, size ratio, and connection relationship.

[0037] Furthermore, the layout similarity comparison in step four is achieved through the image feature extraction function of a multimodal large model, specifically including:

[0038] (1) Perform graphic element segmentation on the sample data image and the target image, and extract the bounding box coordinates and text labels of each element;

[0039] (2) The matching degree of the topological relationship of the element position is calculated by the cosine similarity algorithm, and the consistency of the text label is calculated by the edit distance algorithm.

[0040] Furthermore, the layout requirement text generated in steps two through five is combined with any content description text to drive the inference model to generate HTML5 page code corresponding to different system architecture diagrams.

[0041] The present invention has the following beneficial effects:

[0042] 1. It effectively solves the problem that large inference models cannot directly generate system architecture diagrams containing Chinese characters, thus expanding the application of large models in the field of software development.

[0043] 2. By automatically optimizing prompts, the workload and cost of manual prompt optimization are reduced, the efficiency of generating system architecture diagrams is improved, and the accuracy and compliance of the generated results are ensured.

[0044] 3. Innovatively, the image understanding capabilities of the multimodal large model are transferred to the inference large model, realizing the complementary advantages and collaborative work between different models, and providing new ideas and methods for using large models for graphics generation. Attached Figure Description

[0045] Figure 1 This is a flowchart of the method for generating a system architecture diagram of a large inference model driven by automatically optimized prompt words, as described in this invention.

[0046] Figure 2 This is a schematic diagram of sample data images prepared according to an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of the content description text of an embodiment of the present invention;

[0048] Figure 4 This is a schematic diagram of the initial layout requirements text in an embodiment of the present invention;

[0049] Figure 5 This is a schematic diagram of the target image in an embodiment of the present invention;

[0050] Figure 6 This is a schematic diagram illustrating a cross-disciplinary team composed of multiple technical experts, as described in an embodiment of the present invention.

[0051] Figure 7 This is a schematic diagram of the judgment criteria for layout similarity comparison in an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] This invention provides a method for generating system architecture diagrams using an automatically optimized prompt word-driven inference model. The method involves the large model automatically looping and iteratively optimizing multiple rounds to generate text with the optimal layout requirements. Since the layout requirements are the same for different system architecture diagrams, prompt words can be combined with various content descriptions to meet the requirements for generating different system architecture diagrams. Please refer to the overall structure diagram. Figure 1 The method includes the following steps:

[0054] Step 1: Model and Data Preparation

[0055] Choose a large model with multimodal capabilities, such as GROOK, for comparing the sample data images with the result images from step three.

[0056] Choose a large inference model, such as Deepseek, to generate HTML5 page code based on prompt words.

[0057] Prepare sample data images (i.e., a system architecture diagram, such as...) Figure 2 ).

[0058] Step 2: Generate the initial prompt word component

[0059] Prepare descriptive text corresponding to the sample data images (e.g.) Figure 3 As shown in the figure, the content description text defines the text information that needs to be presented in the system architecture diagram, including module names, hierarchical divisions and functional descriptions;

[0060] Prepare initial layout requirements text (e.g.) Figure 4 As shown, the layout requirements for all system architecture diagrams are the same, such as the hierarchical position of each module, the subordinate relationship, the relative distance, the line connection method and other key information. The layout requirements for the first use can be generated manually and can be relatively simple and rough, because it will be iteratively optimized through subsequent cycles.

[0061] Step 3: Combine and generate prompt words and input them into the large-scale reasoning model.

[0062] The content description text from step two and the initial layout requirements text are combined into a cue word combination to generate cue words. This cue word drives the inference model to generate HTML5 page code describing the architecture diagram, and then the HTML5 page code is converted into the target image (e.g., Figure 5 The first generated target image may be quite different from the sample data image, but after multiple iterations to optimize the layout requirements, it will gradually become similar to the sample data image.

[0063] This step has the following key points:

[0064] First, the combination of content description and layout requirements is not simply about creating context; it also requires defining the system role of the larger model as a cross-domain team composed of multiple technical experts (e.g., Figure 6 As shown in the figure, this allows for the understanding of multiple technologies, including illustrators, visualization experts, system architects, etc. Practice has proven that this is more effective than defining the large model as a front-end engineer or other single-skill role.

[0065] (1) Senior Technical Illustrator - Proficient in SVG technology and visual design.

[0066] (2) Visualization expert - skilled at transforming complex concepts into intuitive images.

[0067] (3) Educational content designer - focuses on the clarity and effectiveness of knowledge delivery.

[0068] (4) System Architect - Familiar with system architecture and proficient in various middleware and frameworks.

[0069] (5) Front-end development engineer - proficient in HTML, CSS and JavaScript.

[0070] These five expert roles can collaboratively activate the multi-dimensional artificial intelligence capabilities of the large model: Senior technical illustrator activates the AI ​​generation capability of graphic specifications to ensure the accuracy of SVG drawing; visualization expert activates the AI ​​analysis and intuitive transformation capability of architectural logic; educational content designer activates the AI ​​structured expression capability of knowledge transfer; system architect activates the AI ​​judgment capability of technology stack adaptation; and front-end development engineer activates the AI ​​compliance generation capability of HTML5 code. The superposition of multi-dimensional AI capabilities makes the effect far exceed the setting of a single role.

[0071] Secondly, it's crucial to balance the imaginative potential of the large model with its rigor. Therefore, the prompts should include phrases like "page content strictly adheres to the content description text," "page layout strictly adheres to the layout requirements text," and "text and graphics, and graphics with each other, have no overlap." The large model cannot be allowed to freely imagine; otherwise, the results may not converge, leading to persistently different comparison results in step four until tokens are exhausted. The prompts are as follows:

[0072] (1) The page content strictly follows the content description text.

[0073] (2) The page layout strictly follows the layout requirements.

[0074] (3) The page conforms to the HTML5 standard.

[0075] (4) Use Chinese characters for the text in the picture as much as possible.

[0076] (5) There is no overlap between text and graphics, or between graphics.

[0077] (6) The connecting lines between modules should accurately reflect the subordinate or interactive relationship and avoid redundant connections.

[0078] (7) Each level of modules must maintain visual consistency (such as the same size and alignment), and the hierarchical boundaries must be clearly distinguishable.

[0079] (8) Graphic elements must be clearly labeled with module names, and the names must be exactly the same as the module names in the content description.

[0080] (9) All elements on the page should maintain reasonable spacing (it is recommended to be no less than 5px) to avoid crowding or excessive dispersion.

[0081] (10) The generated HTML5 page must support basic scaling (100%-150%), and the element layout should not be disordered after scaling.

[0082] These requirements are based on instruction tuning logic: Firstly, they anchor entity alignment by ensuring consistency between module naming and content description, and maintaining visual uniformity across levels, thus limiting the generation of irrelevant elements and preventing excessive speculation. Secondly, they guide the model to utilize structured generation capabilities by ensuring that connections reflect interactive relationships and that scaling is compatible, activating its ability to complete details within the technical framework. Simultaneously, spacing and overlap constraints set convergence boundaries for the results, preventing token waste due to unconstrained divergence, ultimately achieving a balance between directional capability activation and speculative control.

[0083] Step 4: Similarity Comparison and Difference Analysis

[0084] The target image and sample data images generated in step three are fed into a multimodal large model for layout similarity judgment. If the similarity does not reach the threshold, step five is executed.

[0085] This step has the following key points:

[0086] First, it is necessary to identify elements in the image that do not need to be included in the comparison and should be ignored during the comparison process (such as...). Figure 7As shown), non-core differences include color depth and line thickness. If a multimodal large model uses such non-core elements as the basis for comparison, it can easily lead to divergent judgment criteria and difficulty in convergence of the optimization process. These differences do not actually affect the technical logic and core information carried by the system architecture diagram. Secondly, it is necessary to determine the technical implementation path for similarity judgment (such as...). Figure 7 As shown in the figure, the specific methods can be based on the positional information of elements on the graph, such as image feature extraction, comparison of graphic element position coordinates, and similarity measurement of architectural structure, because these are the key points for conveying concepts and information in the system architecture diagram; thirdly, if the judgment result is "dissimilar", it is necessary to ask the multimodal large model to output detailed reasons for the differences (such as "missing number of modules at a certain level" or "misalignment of the vertical subordinate relationship between module C and module D"). This information plays a key supporting role in the subsequent manual debugging prompts and the optimization direction of precise positioning layout requirements.

[0087] The criteria for judging the layout similarity comparison include:

[0088] (1) Ignore the differences in color and the differences in the thickness of the lines on the graph.

[0089] (2) Compare whether the text on similar positions on the map is the same.

[0090] (3) Compare whether the overall layout is similar and whether the layers are the same.

[0091] (4) Compare whether the number of graphics in the same row is the same, and compare whether the number of graphics in the same column is the same.

[0092] (5) Compare whether the elements in the same layer are similar, and compare whether the elements in the same module are similar.

[0093] (6) Compare whether the division and combination relationships of the graphic elements on the graph are the same.

[0094] (7) Compare whether the size and position of each graphic element on the graph are the same.

[0095] (8) Two images are judged to be similar only if all conditions 2-7 are met. Please give the similarity evaluation result, where yes indicates similarity and no indicates dissimilarity.

[0096] (9) If they are not similar, please explain in detail the differences between the second picture and the first picture in items 2-7 above.

[0097] This solution reduces noise interference and improves judgment robustness by filtering non-core features (color, lines); it ensures the objectivity of the evaluation by relying on structured feature matching such as position coordinate comparison and structural similarity measurement; it provides accurate supervision signals for prompt word optimization by forcing difference attribution output; and it avoids judgment ambiguity by hard constraints of "full satisfaction of 2-7 items", effectively supporting the convergence of optimization loops, which is in line with the engineering implementation logic of AI visual semantic understanding.

[0098] Step 5: Iterative optimization loop for prompt words

[0099] Based on the difference analysis results in step four, the layout requirement text is optimized using a multimodal large model, new prompt words are reconstructed, and steps two through four are repeated until the comparison results between the target image and the sample data images are similar. The resulting layout requirement is the optimal result.

[0100] Currently, there is no better method for generating system architecture diagrams than manually drawing them. Traditional methods mainly rely on developers or designers to manually draw architecture diagrams based on the layered and modular design of the software. This approach is not only time-consuming and labor-intensive but also prone to human error. Although some drawing software tools (such as Visio and PowerPoint) can be used to assist in drawing, these tools cannot automatically generate architecture diagrams that meet technical requirements based on text descriptions, still requiring a significant amount of manual operation and adjustment. In addition, there are some template-based automated tools, but their flexibility and accuracy are limited, and they cannot adapt to the diverse needs of different projects and software architectures. They also cannot generate system architecture diagrams by intelligently optimizing prompts to drive reasoning and large-scale models, as is the case in this invention.

[0101] Compared with traditional methods, this invention has the following characteristics:

[0102] 1. Solves the problems of inefficiency and error-proneness in traditional manual drawing.

[0103] Traditional methods rely on manual drawing (such as using Visio or PowerPoint), which is time-consuming, labor-intensive, and prone to human error. This invention achieves automated generation of system architecture diagrams through cross-modal model collaboration and intelligent optimization of prompt words, significantly reducing manual operations, improving efficiency, and lowering the error rate.

[0104] 2. Overcoming the limitations of template-based tools in terms of flexibility and accuracy.

[0105] Traditional template-based automation tools have limited flexibility and accuracy, and cannot adapt to the diverse needs of different projects and software architectures. This invention, through multimodal semantic parsing, loop feedback optimization, and multi-role expert simulation, can dynamically adapt to diverse architectural requirements and generate personalized architecture diagrams that meet technical requirements.

[0106] 3. Implement intelligent optimization-driven generation logic

[0107] Traditional tools cannot generate architecture diagrams for large-scale inference models driven by prompt words; this invention innovatively transforms natural language requirements into executable generation instructions through prompt word optimization strategies (closed-loop iteration, difference-driven correction) and cross-modal model collaboration, achieving end-to-end intelligent generation of "requirements-layout-code", with significant technological advantages.

[0108] In summary, this invention effectively addresses the core pain points of traditional system architecture diagram generation, namely "strong reliance on human intervention, insufficient flexibility, and limited accuracy," through technological innovations such as cross-modal model collaboration, closed-loop optimization of prompt words, and multi-role expert simulation. It achieves a technological leap from "passive assisted drawing" to "active intelligent generation," demonstrating significant advantages in automation, adaptability, and generation quality.

[0109] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for generating system architecture diagrams by automatically optimizing prompt words to drive a large inference model, characterized in that, Includes the following steps: Step 1: Model and Data Preparation Select a multimodal large model and an inference large model, and prepare sample data images; Step 2: Generate the initial prompt word component Based on the sample data images, generate content description text and initial layout requirement text; Step 3: Generate architecture diagram code and images The content description text and initial layout requirement text from step two are combined into prompt words, which are then input into the inference model to generate the HTML5 page code corresponding to the system architecture diagram. The HTML5 page code is then converted into the target image. Step 4: Similarity Comparison and Difference Analysis Using the multimodal large model selected in step one, the layout similarity of the target image generated in step three is compared with that of the sample data image. If the similarity does not reach the preset threshold, the difference analysis results are output. Step 5: Iterative optimization loop for prompt words Based on the difference analysis results in step four, optimize the layout requirement text in step two, generate new prompt words, and repeat steps three to four until the similarity between the target image and the sample data image reaches a preset threshold to obtain the optimal layout requirement text.

2. The method as described in claim 1, characterized in that, In step one, the multimodal large model includes gok, and the inference large model includes one of deepseek, Tongyi Qianwen (Qwen3-Instruct text-specific version), Zhipu Qingyan, Huawei Pangu NLP series models, and Tencent Hunyuan. The sample data image is a reference architecture diagram that includes the system architecture layer structure, module hierarchy, and connection methods.

3. The method as described in claim 1, characterized in that, In step two, the content description text defines the text information that needs to be presented in the system architecture diagram, including module names, hierarchical divisions, and functional descriptions; The initial layout requires the text to include rules for arranging graphic elements, including layer positions, hierarchical relationships, relative distances, and line connection methods. The initial layout also requires the text to be generated manually or reused from historical optimization results.

4. The method as described in claim 1, characterized in that, In step three, the combination of the prompt words includes: defining the system roles of the reasoning model as a cross-domain expert team, which includes technical illustrators, visualization experts, system architects and front-end engineers; The prompt also includes constraints: "Page content strictly follows the content description text", "Page layout strictly follows the layout requirements text", and "Text and graphics, and graphics and graphics do not overlap".

5. The method as described in claim 1, characterized in that, In step four, the criteria for judging the layout similarity comparison include: (1) Ignore the differences in color and the differences in the thickness of the lines on the graph; (2) Compare whether the text on similar positions in the diagram is the same; (3) Compare whether the overall layout is similar and whether the levels are the same; (4) Compare whether the number of graphics in the same row is the same, and compare whether the number of graphics in the same column is the same; (5) Compare whether elements in the same layer are similar, and compare whether elements in the same module are similar; (6) Compare whether the division and combination relationships of the graphic elements on the diagram are the same; (7) Compare whether the size and position of each graphic element on the graph are the same; (8) Only when all of items 2-7 are satisfied will the two images be judged to be similar, and a similarity evaluation result will be given: yes means similar, no means dissimilar. (9) If they are not similar, please explain in detail the differences between the second picture and the first picture in items 2-7 above; The difference analysis results include specific difference types, including items (2)-(7) above.

6. The method as described in claim 1, characterized in that, In step five, the termination condition of the iterative optimization loop is: the similarity between the target image and the sample data image reaches a preset threshold, or the number of loops reaches the upper limit.

7. The method as described in claim 4, characterized in that, The roles of the cross-domain expert team are defined through system role instructions in the prompts, which are used to constrain the large inference model to integrate visual design rules and system architecture logic when generating HTML5 page code.

8. The method as described in claim 1, characterized in that, The HTML5 page code generated in step three contains SVG graphic elements. The attributes of the SVG graphic elements correspond one-to-one with the parameters in the layout requirement text, including position coordinates, size ratio, and connection relationship.

9. The method as described in claim 5, characterized in that, The layout similarity comparison in step four is achieved through the image feature extraction function of a multimodal large model, specifically including: (1) Perform graphic element segmentation on the sample data image and the target image, and extract the bounding box coordinates and text labels of each element; (2) The matching degree of the topological relationship of the element position is calculated by the cosine similarity algorithm, and the consistency of the text label is calculated by the edit distance algorithm.

10. The method as described in claim 1, characterized in that, The layout requirement text generated in steps two through five, combined with any content description text, is used to drive the inference model to generate HTML5 page code corresponding to different system architecture diagrams.

Images