Cross-modal case matching and solution generation method and device

Abstract

The application discloses a cross-modal case matching and scheme generation method, relates to the technical field of artificial intelligence, and in particular relates to a cross-modal case matching and scheme generation method and device, which comprises the following steps: obtaining case data in a case library and analyzing the case data to obtain case information, wherein the case information comprises case semantic labels; obtaining comprehensive feature data of a target land plot, wherein the comprehensive feature data comprises land plot basic attribute data, land plot spatial feature data, land plot land use property data, land plot real scene image data and a POI gap matrix; screening matched cases based on the comprehensive feature data, the case semantic labels and a scoring model comprising an AI inference main engine and a scoring bottom-up vice engine; generating scheme generation constraint information in combination with the comprehensive feature data and case information of the matched cases, and generating a scheme generation statement based on the scheme generation constraint information; and calling a scheme generation model to obtain a candidate design scheme. The application can improve case matching stability and reduce scheme generation homogenization risks.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for cross-modal case matching and solution generation. Background Technology

[0002] Urban micro-renewal targets, such as corner plots, vacant land, under-bridge spaces, and slope green spaces, are typically characterized by irregular areas, complex surrounding infrastructure, restricted land use attributes, and significant differences in site spatial conditions. Existing technologies include schemes that utilize basic plot information, surrounding POI data, case study materials, and large language models or multimodal models to generate urban micro-renewal design suggestions. These schemes typically first obtain the target plot's location, area, distribution of surrounding facilities, and reference cases, then concatenate this information into prompts, input them into a language model or visual language model, and generate a utilization guide, design description, or redevelopment plan text.

[0003] However, in existing solutions, the solution generation process typically relies on single-path reasoning. After completing case matching, the system often directly generates a single design solution based on a best case or a few similar cases. The output is easily influenced by the matched cases, prompt examples, and the model's prior preferences. When the prompts implicitly contain a certain functional tendency, the model may repeatedly generate similar functional positioning or spatial organization methods, resulting in strong homogenization of solutions between different plots and between different batches of generated solutions. This makes it difficult to fully cover the multiple feasible directions for the same plot in terms of functional positioning, spatial structure, and service population.

[0004] Meanwhile, even when existing technologies require models to "generate multiple solutions at once," these solutions are often merely textual rewrites or partial variations of the same design concept. For plots with rigid land use characteristics, special geometric shapes, or obvious infrastructure gaps, the models may also ignore constraints such as land use attributes, site dimensions, and shortcomings in surrounding facilities, generating solutions that do not match the site conditions or are not feasible.

[0005] Furthermore, when users provide feedback and revisions to the generated solutions in multiple rounds, existing systems typically continuously add the complete historical dialogue, the original solution, and user feedback to the context. As the number of iterations increases, the input context expands, not only increasing model invocation costs and response latency but also easily diluting early key modification intentions, causing subsequent generated results to deviate from the direction already confirmed by the user. If multiple solutions are generated simultaneously, existing streaming output mechanisms also struggle to independently distinguish and isolate errors in different solution channels; an anomaly in one channel may affect the overall output process.

[0006] Therefore, in view of the shortcomings of the existing technology, there is an urgent need to provide a method for generating intelligent urban corner land solutions that can generate multiple directions, low bias and stable iteration based on land conditions and facility gaps. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a technology that can generate intelligent solutions for urban corner land in multiple directions, with low bias and stable iteration based on land conditions and facility gaps.

[0008] The objective of this invention is achieved through the following technical solution: Therefore, according to one aspect of the present invention, a cross-modal case matching and solution generation method is provided, comprising the following steps: acquiring case data from a case library and parsing it to extract corresponding case information; wherein, the case information includes case semantic tags, and the fields of the case semantic tags include several of the following: scale classification, shape features, terrain topology, location background, and dominant gaps; acquiring comprehensive feature data of the target plot, wherein the comprehensive feature data includes basic attribute data of the plot, spatial feature data of the plot, land use nature data of the plot, real-scene image data of the plot, and POI gap matrix; based on the comprehensive feature data and case semantic tags... The system uses a pre-defined scoring model to match cases in the case library with the target plot, selecting matching cases. The scoring model includes an AI inference main engine and a scoring fallback sub-engine. Combining comprehensive feature data and case information from the matching cases, it generates scheme generation constraint information. This constraint information includes several design direction constraints, each of which satisfies the land use constraints corresponding to the plot's land use nature data. Combining comprehensive feature data, case semantic tags from the matching cases, and scheme generation constraint information, it generates a scheme generation statement. The system then calls the scheme generation model and inputs the scheme generation statement to obtain at least one candidate design scheme.

[0009] Specifically, based on a scoring model using comprehensive feature data, case semantic tags, and preset rules, cases in the case library are matched with target plots to select matching cases. This process includes the following steps: extracting structured features from the comprehensive feature data and combining them with preset rules and the case semantic tags of each case to synthesize an inference input statement; inputting the inference input statement into the AI ​​inference main engine and obtaining the output results, which include matching cases; if the AI ​​inference main engine malfunctions and cannot obtain output results, it switches to a locally configured scoring fallback engine and performs the following steps: using the scoring fallback engine, multi-dimensional tag comparisons are performed on each case to calculate the original score for each case; based on the confidence coefficient of each tag, the confidence level of each original score is adjusted to calculate the case score for each case; and matching cases are selected based on the case scores of each case.

[0010] Another specific step involves generating a scheme generation statement by combining the comprehensive feature data, the case semantic tags of the matching cases, and the scheme generation constraint information. This includes the following steps: using regular expressions to extract real-scene image sequences from the land parcel real-scene image data and / or the case information of the matching cases; and combining the comprehensive feature data, the case semantic tags of the matching cases, the scheme generation constraint information, and the real-scene image sequences to generate a scheme generation statement.

[0011] The above, combining comprehensive feature data and case information from matching cases, generates scheme generation constraint information, specifically including the following steps: extracting design logic from the case information of matching cases; inputting the design logic and comprehensive feature data into a large language model to generate and output design direction descriptions for several target plots as design direction constraints; wherein, the design direction constraints differ from each other in at least one dimension of functional positioning, spatial organization, and service objects, and the land use nature data of the plots serves as the constraint condition for generating design direction constraints.

[0012] More specifically, the scheme generation constraint information also includes inverse illusion strong boundary constraints and image stretching constraints; generating scheme generation statements specifically includes the following steps: for each design direction constraint, each design direction constraint is combined with the inverse illusion strong boundary constraints and image stretching constraints to generate a scheme generation statement corresponding to the design direction constraint; calling the scheme generation model and inputting the scheme generation statement to obtain at least one candidate design scheme, specifically including: inputting each scheme generation statement into the scheme generation model to obtain the candidate design scheme corresponding to each design direction constraint.

[0013] More specifically, the various scheme generation statements are input into the scheme generation model to obtain candidate design schemes corresponding to each design direction constraint. This includes the following steps: establishing corresponding scheme generation channels for each design direction constraint; each scheme generation channel has a corresponding channel identifier; and inputting the various scheme generation statements into the scheme generation model through each scheme generation channel for parallel reasoning to generate candidate design schemes corresponding to each design direction constraint.

[0014] The above also includes the following steps: For each candidate design scheme, construct a corresponding scheme iteration container; wherein, the scheme iteration container is used to store the current feedback slot, compressed historical information, and scheme version sequence; receive feedback information input by the user for any candidate design scheme, and determine the candidate design scheme corresponding to the feedback information; write the feedback information into the current feedback slot corresponding to the candidate design scheme, and obtain the design direction constraints, scheme generation statement, compressed historical information, and scheme version sequence corresponding to the candidate design scheme; generate a scheme iteration statement based on the feedback information, design direction constraints, scheme generation statement, compressed historical information, and scheme version sequence; input the scheme iteration statement into the scheme generation model to generate the iterated candidate design scheme.

[0015] According to another aspect of the present invention, a cross-modal case matching and scheme generation apparatus is provided, comprising: a case parsing module, used to acquire case data in a case library, parse it, and extract corresponding case information; wherein the case information includes case semantic tags; a case filtering module, used to acquire comprehensive feature data of a target plot, and based on the comprehensive feature data, case semantic tags, and a scoring model with preset rules, match cases in the case library with the target plot to filter out matching cases; a constraint generation module, used to combine the comprehensive feature data and the case information of the matching cases to generate scheme generation constraint information; wherein the scheme generation constraint information includes several design direction constraints, and each design direction constraint satisfies the land use constraints corresponding to the land use nature data of the plot; a statement generation module, used to combine the comprehensive feature data, the case semantic tags of the matching cases, and the scheme generation constraint information to generate a scheme generation statement; and a scheme generation module, used to call a scheme generation model and input the scheme generation statement to obtain at least one candidate design scheme.

[0016] According to another aspect of the present invention, a computing device is provided, including a memory, a processor, and computer instructions stored in the memory and executable on the processor, wherein the processor executes the instructions to implement the steps of the cross-modal case matching and solution generation method described above.

[0017] According to another aspect of the present invention, a computer-readable storage medium is provided that stores computer instructions which, when executed by a processor, implement the steps of cross-modal case matching and scheme generation as described above.

[0018] The beneficial effects of this invention are: 1. By generating multiple design directions with substantial differences, and then generating corresponding solutions based on each design direction, we can avoid the homogenization of results caused by generating solutions directly based on a single case or a single prompt word, and improve the diversity of design solutions for urban corner plots.

[0019] 2. By using land use, plot area, geometry, facility gap diagnosis results, and reference case features as constraints for direction generation and scheme generation, the matching degree between the generated scheme and the actual conditions of the target plot can be improved, and the risk of inconsistency between the generated content and site attributes can be reduced.

[0020] 3. By using anti-biased prompt word processing methods, the implicit influence of case matching results or prompt word examples on the model output direction is reduced, enabling the model to perform scheme reasoning around the characteristics of the target plot itself, thereby improving the objectivity and stability of scheme generation.

[0021] 4. By establishing independent generation channels for different design directions and returning each solution separately through streaming output with channel identifiers, multiple solutions can be generated in parallel and displayed independently, improving solution generation efficiency and avoiding the impact of a single path anomaly on the output of other solutions.

[0022] 5. By employing a finite memory window and context compression mechanism during multi-round human-computer interaction, it is possible to retain key modification intentions in continuous user feedback while controlling the context length and model call cost, thereby improving the continuity and controllability of solution iteration. Attached Figure Description

[0023] The invention can be better understood by describing exemplary embodiments disclosed herein in conjunction with the accompanying drawings, in which: Figure 1 The diagram shown is a schematic flowchart of a cross-modal case matching and solution generation method according to Embodiment 1 of the present invention; Figure 2 The diagram shown is a schematic diagram of the program modules of a cross-modal case matching and solution generation device according to Embodiment 1 of the present invention; Figure 3 The diagram shown is a hardware structure schematic of a computing device according to Embodiment 1 of the present invention. Detailed Implementation

[0024] The following describes specific embodiments of the present invention. It should be noted that, in order to provide a concise description, this specification cannot exhaustively describe all features of the actual embodiments. It should be understood that, in the actual implementation of any embodiment, just as in any engineering or design project, various specific decisions are often made to achieve the developer's specific goals and to meet system-related or business-related constraints, and this can change from one embodiment to another. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this invention, some design, manufacturing, or production modifications based on the technical content disclosed in this invention are merely conventional technical means and should not be construed as insufficient content of this invention.

[0025] Unless otherwise defined, the technical or scientific terms used in the claims and description shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in the specification and claims of this patent application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. The terms "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "comprising" or "including" and similar terms mean that the element or object preceding "comprising" or "including" encompasses the element or object listed following "comprising" or "including" and its equivalents, and do not exclude other elements or objects. The terms "connected" or "linked" and similar terms are not limited to physical or mechanical connections, nor are they limited to direct or indirect connections.

[0026] Example 1 Please see Figure 1 This embodiment proposes a cross-modal case matching and solution generation method, including the following steps: S1: Obtain case data from the case library, parse it, and extract the corresponding case information.

[0027] The case information includes case semantic tags, and the fields of the case semantic tags include several of the following: scale classification, shape features, terrain topology, location background, and dominant gap.

[0028] In this embodiment, step S1 specifically includes the following steps S11 to S13: S11: Read the valid case directory and assemble the case data: Using asynchronous concurrent control flow (Promise.all), the underlying fs.stat probe is performed on the subdirectories of the case library to filter out the valid case directories and read the case data in the case directories.

[0029] The case data includes: case card files (.md), article files, and image files. Case information parsed from the case data includes case identifiers, case text, case summary (summary_excerpt), case image paths, and design logic. The case summary is formed by pre-setting the first few characters (e.g., 1500 characters) of the case text to eliminate the overhead of rereading files during subsequent large language model matching. The design logic characterizes the case's functional combination, spatial organization, and service target positioning.

[0030] S12: Read or generate case semantic tags: For each case, probe the persistent tag file tags.json, prioritize reading compliant structured tags; if no match is found, call the large language model to extract five-dimensional tags and write the tag results back to tags.json for subsequent matching and reuse.

[0031] In this embodiment, the large language model used to extract case semantic tags adopts a text and multimodal understanding model (e.g., doubao-seed-2-0-lite-260215).

[0032] Specifically, if tags.json exists and contains compliant fields such as scale classification, shape features, terrain topology, location background, and dominant gap, then the semantic tags for the case are deserialized. If tags.json is not found, the first preset characters of the case text are truncated and input into the large language model. The scale classification, shape features, terrain topology, location background, and dominant gap are extracted in a zero-shot manner, and the tag confidence coefficient Cconf is output. The value of Cconf ranges from 0 to 1.

[0033] During the label extraction process, scale classification is used to characterize the area level of the case; shape features are used to identify geometric shapes such as linear, regular or irregular; topography is used to distinguish site relationships such as under bridges, slopes, waterfronts, wetlands or flat land; location context is used to characterize the surrounding context such as old town, waterfront, park green space, protective green space or traffic nodes; dominant gap is used to characterize the functional shortcomings that the case focuses on addressing, such as public services, commercial services, sports and leisure, public space or transportation accessibility.

[0034] If the large language model call is abnormal, the output cannot be deserialized, or the five-dimensional fields are missing, the local lexical matching fallback processing is triggered: basic tags are extracted based on keywords or phrases such as area unit, under bridge, old town, waterfront, slope, water system, park green space and protective green space, and default tag confidence coefficients are assigned to the fallback tags.

[0035] S13: Establish the association between the case image set and the case identifier: Establish an index relationship between the case image set and the case identifier to obtain the image material index that can be passed along with the matching case, and call it when extracting the real scene image sequence in step S31.

[0036] In this embodiment, for the text image placeholders in the case, a regular expression capture group (including standard and distorted Markdown formats) is used to perform precise or degraded fuzzy matching in the image file set of the case library, replacing the placeholders with the absolute path of the local API route (such as / api / cases / image / ...) to ensure front-end rendering connectivity.

[0037] S2: Obtain comprehensive feature data of the target plot, and based on the comprehensive feature data, case semantic tags, and a scoring model with preset rules, match the cases in the case library with the target plot and filter out the matching cases.

[0038] The comprehensive feature data includes basic attribute data of the land parcel, spatial feature data of the land parcel, land use nature data of the land parcel, real-scene image data of the land parcel, and POI gap matrix; the scoring model includes an AI inference main engine and a scoring backup sub-engine.

[0039] In this embodiment, step S2 reads the comprehensive feature data of the target plot and the case semantic tags obtained in step S1, generates inference input statements, and prioritizes calling the AI ​​inference main engine to output matching cases; when the AI ​​inference main engine malfunctions, the scoring fallback sub-engine is switched in to calculate the case score and output matching cases for use in step S3. The AI ​​inference main engine calls a text and multimodal understanding model (e.g., doubao-seed-2-0-lite-260215) through a model call interface to perform inference and ranking of matching cases.

[0040] Specifically, step S2 includes the following steps S21 to S23: S21: Extract structured features from the comprehensive feature data, and combine them with preset rules and case semantic tags of each case to synthesize inference input statements.

[0041] Specifically, when extracting structured features, the target plot name, area, and address are read from the basic attribute data of the plot in the comprehensive feature data; the land use nature is read from the land use nature data of the plot in the comprehensive feature data; the site geometry, site feature keywords, and terrain inference results are read from the spatial feature data of the plot in the comprehensive feature data; and the counts, weighted cumulative values, weighted densities, and ranking results of the five functional dimensions of public services, commercial services, sports and leisure, public space, and transportation accessibility are read from the POI gap matrix of the comprehensive feature data. Notably, the plot's real-scene image data is not directly written as a structured field into the inference input statement in this step, but is used to extract the real-scene image sequence when generating the subsequent scheme generation statement.

[0042] Furthermore, keyword sniffing is performed based on the target plot's name, address, land use, and site characteristic keywords to identify site features such as wetlands, waterways, under bridges, elevated roads, mountains, slopes, parks, or protective green spaces. A five-dimensional density score, a list of top high-frequency facilities, the weakest functional dimension, and the total number of POIs are determined based on the POI gap matrix. The total number of POIs is used to help determine the locational context of the target plot; for example, when the total number of POIs is not less than 800, the locational context is inferred to be an old urban area or a high-density built-up area.

[0043] Furthermore, the five-dimensional density score is used to characterize the supply strength of the five functional dimensions, the Top High-Frequency Facility List is used to characterize the types of facilities that appear most frequently around the target site, the weakest functional dimension is used to characterize the dimension with the highest degree of supply insufficiency among the five functional dimensions, and the total POI is used to help determine the location background of the target site. For example, when the total POI is not less than 800, the location environment is inferred to be an old urban area or a high-density built-up area.

[0044] Furthermore, area magnitudes are generated based on the target plot's area. These area magnitudes are divided into five levels according to a five-level gradient: micro, small, medium, large, and mega. Specifically, micro is less than 300 square meters, small is 300 to 3000 square meters, medium is 3000 to 20,000 square meters, large is 20,000 to 80,000 square meters, and mega is greater than 80,000 square meters. The area magnitudes are then written into the inference input statement to support fault-tolerant matching between adjacent sizes.

[0045] When synthesizing the input statement for inference, the extracted structured features of the target plot, the semantic tags of each candidate case, the case summary, and the design logic are written into the same inference input object, and preset rules are written into the inference input object. The preset rules include five dimensions: design method transferability, functional gap matching, site feature fit, scale, and location and form. The preset weights of the five dimensions can be set as follows: design method transferability 30 points, functional gap matching 25 points, site feature fit 20 points, scale 15 points, and location and form 10 points. Among them, design method transferability is the highest weight dimension, which is used to encourage the AI ​​inference engine to read the design logic and case summary, rather than just performing a superficial matching based on the case semantic tags.

[0046] S22: Input the inference input statement into the AI ​​inference main engine and obtain the output results; the output results include matching cases.

[0047] Specifically, the inference input statements are fed into the AI ​​inference main engine, and the case ranking results are obtained using a low-entropy structured output method. Sampling temperature, probability sampling threshold, and random seed are set as stable output parameters, and the output format is limited to a structured JSON object.

[0048] In this embodiment, when the AI ​​inference main engine calls the text and multimodal understanding model, it sets the sampling temperature to 0.0001, the probability sampling threshold (top_p) to 0.1, the random seed to 1234, and declares the response_format as json_object to reduce the random fluctuation of the case ranking results under the same input.

[0049] The output includes case identifiers, sorting results, and matching reasons.

[0050] Specifically, the output is an ordered JSON array, with each record including at least a case identifier, sort order, and a matching reason. The matching reason can be limited to 30 characters and is used to explain the correspondence between the matched case and the target plot in terms of design techniques, functional gaps, site characteristics, scale, or location.

[0051] After performing JSON parsing validation, case identifier matching validation, and valid quantity validation on the output results, the cases ranked first are determined as matching cases.

[0052] S23: Verify the output of the main AI inference engine and switch to the scoring fallback engine when an anomaly occurs.

[0053] Specifically, the output of the AI ​​inference engine is validated by JSON parsing, case identifier matching, and valid number. If the validation passes, the matching cases output by the AI ​​inference engine are used; if the validation fails, local scoring is performed as a fallback.

[0054] More specifically, AI inference main engine anomalies include AI network blocking, JSON parsing failure, output case identifiers not matching in the local case pool, or zero valid cases.

[0055] If the output can be parsed, the output case identifier can be matched in the local candidate case pool, and the number of valid cases is not zero, then the top-ranked cases will be determined as matching cases.

[0056] If the main AI inference engine malfunctions and cannot obtain output results, it will switch to the local scoring fallback engine and execute the following steps S231 to S233.

[0057] S231: Through the scoring fallback engine, multi-dimensional label comparisons are performed on each case to calculate the original score corresponding to each case.

[0058] Specifically, a scoring-based fallback engine compares the target plot with each candidate case using multi-dimensional labels. These multi-dimensional labels include dimensions of scale, location, gaps, topography, and shape.

[0059] In this embodiment, the scale dimension is scored out of 30 points, and the score is determined based on the gap between the five-order gradient scale levels: 30 points for a gap of 0, 18 points for a gap of 1, 8 points for a gap of 2, and 0 points for a gap greater than or equal to 3. The location dimension is scored out of 20 points, and the score is determined based on the consistency of the urban environment inferred from the total number of POIs. The gap dimension is scored out of 20 points, with 20 points awarded for a precise hit of the dominant gap and 12 points awarded for a near-synonymous hit of the dominant gap in the preset affinity mapping. The topography dimension is scored out of 15 points, with 15 points awarded for a precise hit of the topography label and 10 points for a hit of the topography affinity map; the topography affinity map can set the labels of wetland, lake, and river as bidirectional near-synonymous relationships. The shape dimension is scored out of 15 points, and a Boolean judgment is performed based on the consistency of geometric shape. The scores of each dimension are added together to obtain the original score for the corresponding case.

[0060] S232: Based on the confidence coefficient of each label, perform confidence conversion on each original score to calculate the case score for each case.

[0061] Specifically, the label confidence coefficient Cconf generated in step S1 is read, and the confidence of the original score is converted according to the formula Score = RawScore × (0.6 + 0.4 × Cconf), where RawScore is the original score and Score is the case score.

[0062] When Cconf is missing or not a numeric type, Cconf is assigned a value of 0.5, making the multiplier 0.8. The constant term 0.6 in the formula is used to retain 60% of the basic feature score in the extreme case where Cconf approaches 0, avoiding the case score being directly reduced to zero due to the uncertainty of label extraction.

[0063] S233: Filter out matching cases based on the case scores of each case.

[0064] Specifically, candidate cases are sorted in descending order according to their scores, and the top-ranked cases are selected as matching cases. Matching cases include case identifier, case semantic tags, case summary, design logic, case image set, matching reason, and case score.

[0065] In one implementation, the top-3 cases in the ranking are selected as the reference case set for generating subsequent solutions.

[0066] S3: Combine comprehensive feature data and case information of matching cases to generate solution generation constraint information; and combine comprehensive feature data, case semantic tags of matching cases and solution generation constraint information to generate solution generation statements.

[0067] The scheme generation constraint information includes several design direction constraints, and each design direction constraint satisfies the land use constraints corresponding to the land use nature data of the plot.

[0068] Specifically, the constraint information for scheme generation also includes inverse illusion strong boundary constraints and image stretching constraints.

[0069] Step S3 specifically includes the following steps: S31: Using regular rules, extract real-scene image sequences from the real-scene image data of the land parcel and / or the case information of the matching cases.

[0070] Image paths, image filenames, image access addresses, and case image sets are extracted from the real-world image data of the land parcels and the case information of the matching cases. When the case information of a matching case contains image placeholders referenced by the case text, a regular expression capture group is used to identify standard or malformed Markdown image placeholders, and an exact match is performed in the case image set. If the exact match fails, a downgraded fuzzy match is performed based on filename similarity, number similarity, or path fragment to convert the image placeholders into accessible image paths.

[0071] Furthermore, for the real-scene image data of the land parcels, the regular expression rule / edge_land(\d+)\.(png|jpg|jpeg) / i is used to identify the sequence number in the image file name, and the order of the real-scene images is determined based on the sequence number. If the file name does not contain a sequence number, the sequence order is determined according to the upload order, storage time, or sorting field in the image access address.

[0072] For real-world images that require input to generate the model, the image processing engine performs safe dimensionality reduction and encoding: under the condition of prohibiting image scaling (withoutEnlargement=true), the image's long side is proportionally scaled to no more than 1024 pixels, JPEG compression is performed according to quality parameter 80, and the compressed image is converted into Base64 encoded data. This processing controls the input size and avoids the creation of new spatial facts due to image stretching.

[0073] S32: Extract the design logic from the case information of the matching cases, and input the design logic and comprehensive feature data into the large language model to generate and output the design direction descriptions of several target plots as design direction constraints.

[0074] Specifically, before executing the scheme reasoning, a directional divergence exploration is initiated. The target plot's name, area, land use, site geometry, POI gap matrix, five-dimensional density score, weakest functional dimension, case semantic tags of matching cases, case summaries, and design logic are collectively constructed as the input information for the directional exploration.

[0075] The directional exploration input information is fed into the large language model, and the sampling temperature is set to 0.7 to stimulate the divergence of design directions. The large language model generates no fewer than three target plot design direction descriptions with essential differences; each design direction description includes a direction name, core design concept, functional combination strategy, and reason for differentiation.

[0076] In this embodiment, the large language model used to generate the design direction description adopts a text and multimodal understanding model (e.g., doubao-seed-2-0-lite-260215).

[0077] When generating design direction descriptions, the land use constraints corresponding to the land use nature data of the plots are set as the highest priority constraints, so that each design direction constraint conforms to the land use constraints corresponding to the land use nature data of the plots.

[0078] Among them, the design direction constraints differ from each other in at least one dimension of functional positioning, spatial organization and service objects, and the land use data of the plots serve as the constraint conditions for generating the design direction constraints.

[0079] S33: For each design direction constraint, combine each design direction constraint with the inverse illusion strong boundary constraint and the image stretching constraint to generate the scheme generation constraint information corresponding to the design direction constraint; The constraint information for scheme generation is used as the constraint segment in the scheme generation statement to limit the design direction, factual boundaries, and image usage boundaries of the corresponding candidate design scheme.

[0080] Specifically, the strong boundary constraints of the reverse illusion include: candidate design schemes are mainly based on local data of the target site, with case studies used as supplementary. In this embodiment, the strong boundary constraints of the reverse illusion limit the proportion of local data to no less than 70% and the proportion of case studies to no more than 30%; at the same time, it forces a distinction between visually observed facts and spatially logical inferences.

[0081] Furthermore, the strong boundary constraints of the reverse illusion also include: it is not allowed to fabricate plot ratio, property boundaries, underground pipelines, invisible buildings, invisible elevation differences, hidden structures or other quantitative indicators that cannot be verified by input data; for content that cannot be supported by basic attribute data of the plot, spatial characteristic data of the plot, land use nature data of the plot, real scene image data of the plot or POI gap matrix, it is only allowed to be described as suggestions or matters to be verified.

[0082] Image stretching constraints include: scaling and compressing the input real-world image proportionally only, without changing the image aspect ratio, without mistakenly moving distant features to the foreground, without generating new spatial facts based on the image compression process, and declaring in the scheme generation statement that image information is only used to identify visible spatial relationships and the current environment, and is not used as a basis for inferring invisible areas.

[0083] Furthermore, anti-biased neutralization is performed when generating statements for the generated scheme to reduce output homogenization caused by prompt wording.

[0084] The anti-biased neutralization treatment includes at least one of the following treatments: 1) In the scheme generation statement, the expression involving functional positioning is neutralized, specific examples of specific land use types are removed, and only the format specification requirements and evaluation dimension framework are retained.

[0085] 2) In the case reference information of the solution generation statement, only the case identifier, case semantic tag, matching reason, case summary and design logic of the matching case should be written. No biased examples, priority prompts or additional weighted descriptions for a certain dominant gap or a certain functional type should be attached, so as to avoid the solution generation model being guided to a single functional direction by the case reference information.

[0086] S34: Combine comprehensive feature data, case semantic labels of matching cases, solution generation constraint information, and real-world image sequences to generate solution generation statements.

[0087] Specifically, the comprehensive feature data, POI gap matrix, case semantic labels of matching cases, matching reasons, design logic, design direction constraints, inverse illusion strong boundary constraints, image stretching constraints, and real-scene image sequences are combined to form a solution generation statement for the solution generation model.

[0088] In this embodiment, the scheme generation statement includes basic prompt information, basic land parcel data, POI gap diagnosis information, case reference information, design direction constraints, image input information, output format constraints, and a quality checklist. For different design direction constraints, corresponding scheme generation statements are generated. Each scheme generation statement has an independent direction constraint paragraph (directionConstraint) appended to the end, specifying the direction name, core concept, and functional combination strategy that the scheme should follow.

[0089] S4: Call the scheme generation model and enter the scheme generation statement to obtain at least one candidate design scheme.

[0090] In this embodiment, the scheme generation model adopts a text and multimodal understanding model (e.g., doubao-seed-2-0-lite-260215) to generate corresponding candidate design schemes based on each scheme generation statement. By inputting each scheme generation statement into the scheme generation model, candidate design schemes corresponding to each design direction constraint are obtained, specifically including the following steps S41 to S42: S41: Establish corresponding scheme generation channels for each design direction constraint.

[0091] Each solution generation channel has a corresponding channel identifier.

[0092] Specifically, a corresponding number of scheme generation channels are established according to the number of design direction constraints. Each scheme generation channel is bound to a design direction constraint, a scheme generation statement, a channel identifier, and a set of independent output states. The channel identifier can be a numeric number, a string number, or a scheme tab identifier.

[0093] In this embodiment, the channel identifiers can be channel:0, channel:1, and channel:2 to represent the three-way parallel scheme generation channels.

[0094] Furthermore, each scheme generation channel shares the same scheme generation model, but maintains independent scheme generation statements, directional constraint paragraphs, and output buffers, so that candidate design schemes under different design directional constraints will not overlap with each other during the streaming output process.

[0095] S42: Input the scheme generation statements into the scheme generation model through each scheme generation channel for parallel reasoning, and generate candidate design schemes corresponding to each design direction constraint.

[0096] In this embodiment, multiple scheme generation requests are initiated simultaneously using the asynchronous parallel control flow Promise.all, and the generated fragments of each scheme generation channel are returned via the SSE protocol. Each SSE data frame carries a channel identifier so that the outputs of different channels can fill the corresponding scheme display area.

[0097] During streaming output, intermediate buffering can be disabled in request or response processing, for example, by setting X-Accel-Buffering to no to bypass Nginx buffering. If the solution generation model returns thought chain fragments and body fragments, the thought chain fragments are encapsulated as reasoning types, and the solution body fragments are encapsulated as content types, until the streaming ends and is marked DONE.

[0098] Candidate design schemes include functional positioning, spatial layout, facility configuration, target audience, implementation suggestions, reference relationships with matching cases, and correspondence with design direction constraints. If any scheme generation channel encounters an anomaly, an error frame is returned to that channel, while other scheme generation channels continue to generate schemes, avoiding termination of the overall multi-scheme output due to a single-channel failure.

[0099] S5: Iterate on candidate design solutions based on user feedback; specifically including the following steps S51 to S55: S51: For each candidate design scheme, construct the corresponding scheme iteration container.

[0100] The solution iteration container includes the current feedback slot, compressed historical information, solution version sequence, current solution text, corresponding design direction constraints, and corresponding initial solution generation statements. The solution iteration containers for different candidate design solutions are isolated from each other, enabling multiple candidate design solutions to receive feedback and revisions separately.

[0101] In this embodiment, the scheme iteration container adopts the SchemeChat structure. The current feedback slot includes a slots array of fixed capacity. Each slot includes feedback information and a corresponding revised scheme snapshot. The compressed history information is called compressedHistory. The scheme version sequence is called versions, and the scheme versions of each round are saved in chronological order starting from the initial candidate design scheme.

[0102] S52: Receive feedback information from the user regarding any candidate design scheme, and determine the candidate design scheme corresponding to the feedback information. This specifically includes the following steps: When receiving feedback information, the channel identifier or scheme identifier of the candidate design scheme is received simultaneously, and the corresponding scheme iteration container is found according to the channel identifier or scheme identifier, thereby determining the candidate design scheme corresponding to the feedback information.

[0103] The feedback information includes natural language descriptions such as adding facilities, deleting facilities, adjusting functional proportions, changing spatial organization, strengthening certain types of service targets, retaining existing design elements, or avoiding certain types of spatial elements.

[0104] S53: Write the feedback information to the current feedback slot corresponding to the candidate design scheme, and obtain the design direction constraints, scheme generation statements, compressed historical information, and scheme version sequence corresponding to the candidate design scheme. Specifically, this includes the following steps: Assemble the feedback information with the current candidate design snapshot into a new slot and push it into the current feedback slot.

[0105] In this embodiment, the current feedback slot adopts a three-slot sliding window structure with a fixed capacity of three slots. When the number of slots in the current feedback slot reaches three, the context compression process is triggered. The previously compressed summary, the revision schemes in each round within the window, the feedback information in each round, and the core content of the current scheme are input into the large language model and reduced to a structured intent summary of 500 to 800 characters, which serves as the new compressed historical information.

[0106] In this embodiment, the large language model used to generate structured intent summaries employs a text and multimodal understanding model (e.g., doubao-seed-2-0-lite-260215).

[0107] The compressed historical information retains the modification intent chain formed in chronological order, the core features of the current design, functional positioning, spatial layout, material selection, design decisions explicitly rejected by the user, and design decisions explicitly insisted upon by the user. At the same time, redundant data, detailed area allocations, and repetitive explanations that do not affect subsequent iterations are removed. After compression, the current feedback slot is cleared, and the compressed historical information is used as the prefix context for the next round of feedback slots.

[0108] By combining a three-slot sliding window with compressed historical information, the context length of subsequent inputs to the large language model remains within the range supported by the model, regardless of the number of iterations.

[0109] S54: Generate solution iteration statements based on feedback information, design direction constraints, solution generation statements, compressed historical information, and solution version sequences.

[0110] Specifically, the basic iteration constraint information, site basic data, feedback information, design direction constraints, scheme generation statement, compressed historical information, scheme version sequence, and current candidate design scheme are assembled into a scheme iteration statement. To save context capacity, the scheme iteration statement may no longer carry the Base64 data of the real scene image, but retains the site basic data, direction constraints, and necessary image fact summaries.

[0111] The scheme iteration statement is used to constrain the scheme generation model to make local or overall adjustments based on feedback information while maintaining the design direction constraints, output format and main spatial logic of the original candidate design scheme. For content not covered by the feedback information, the scheme iteration statement restricts large-scale rewriting to maintain the continuity of scheme evolution.

[0112] S55: Input the scheme iteration statement into the scheme generation model to generate candidate design schemes after iteration.

[0113] Specifically, the iterative statement generated in step S54 is input into the scheme generation model as the iterative statement to obtain the candidate design schemes after iteration. During the generation process, revisions can continue to be returned in a streaming manner using SSE.

[0114] After each iteration, the candidate design scheme is written into the corresponding scheme version sequence, and the current candidate design scheme is updated. If feedback is received again, the next iteration is performed based on the scheme iteration container.

[0115] The working principle of this embodiment is as follows: Case data in the case library is parsed into case semantic tags with confidence levels, and case text, design logic, and image materials are transformed into reusable case information; then, matching cases are selected based on the comprehensive feature data of the target plot, the POI gap matrix, and the dual-engine scoring model; then, multiple design direction constraints that are controlled by land use constraints and have substantial differences from each other are generated; then, the design direction constraints, inverse illusion strong boundary constraints, image stretching constraints, and real-scene image sequences are combined into a scheme generation statement; then, multiple candidate design schemes are generated through parallel streaming inference with channel identifiers; finally, feedback information is continuously absorbed using an independent scheme iteration container, a three-slot sliding window, and compressed historical information, so that candidate design schemes can be continuously iterated under context-controlled conditions.

[0116] Please continue reading. Figure 2This invention illustrates a cross-modal case matching and solution generation apparatus. In this embodiment, the apparatus may include or be divided into one or more program modules. One or more program modules are stored in a storage medium and executed by one or more processors to complete the invention and implement the aforementioned cross-modal case matching and solution generation method. The program module referred to in this invention is a series of computer program instruction segments capable of performing a specific function, and is more suitable than the program itself for describing the execution process of a cross-modal case matching and solution generation apparatus in a storage medium. The following description will specifically introduce the functions of each program module in this embodiment: The case analysis module is used to obtain case data from the case library, parse it, and extract the corresponding case information, including case semantic tags.

[0117] The case filtering module is used to obtain comprehensive feature data of the target plot, and based on the comprehensive feature data, case semantic tags and a scoring model with preset rules, it matches the cases in the case library with the target plot and filters out the matching cases.

[0118] The constraint generation module is used to generate scheme generation constraint information by combining comprehensive feature data and case information of matching cases. The scheme generation constraint information includes several design direction constraints, and each design direction constraint satisfies the land use constraints corresponding to the land use nature data of the plot.

[0119] The statement generation module is used to generate a solution generation statement by combining comprehensive feature data, case semantic tags of matching cases, and solution generation constraint information.

[0120] The scheme generation module is used to call the scheme generation model and input the scheme generation statement to obtain at least one candidate design scheme.

[0121] The scheme iteration module is used to construct corresponding scheme iteration containers for each candidate design scheme, receive and determine the candidate design scheme corresponding to the feedback information, write the feedback information into the corresponding current feedback slot, generate a scheme iteration statement based on the feedback information, design direction constraints, scheme generation statement, compressed historical information and scheme version sequence, and input the scheme iteration statement into the scheme generation model to generate the iterated candidate design scheme.

[0122] This embodiment also provides a computing device, such as a smartphone, tablet computer, laptop computer, desktop computer, rack server, blade server, tower server, or cabinet server (including standalone servers or server clusters composed of multiple servers) capable of executing programs. The computing device 20 in this embodiment includes, but is not limited to, a memory 21 and a processor 22 that can communicate with each other via a system bus. Figure 3As shown. It should be noted that, Figure 3 Only a computing device 20 with components 21-22 is shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0123] In this embodiment, the memory 21 (i.e., the readable storage medium) includes flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory 21 may be an internal storage unit of the computing device 20, such as the hard disk or memory of the computing device 20. In other embodiments, the memory 21 may also be an external storage device of the computing device 20, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computing device 20. Of course, the memory 21 may include both the internal storage unit and the external storage device of the computing device 20. In this embodiment, the memory 21 is typically used to store the operating system and various application software installed on the computing device 20, such as the program code of a cross-modal case matching and solution generation device in Embodiment 1. In addition, the memory 21 can also be used to temporarily store various types of data that have been output or will be output.

[0124] In some embodiments, processor 22 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip. This processor 22 is typically used to control the overall operation of computing device 20. In this embodiment, processor 22 is used to run program code stored in memory 21 or process data, for example, to run a cross-modal case matching and solution generation device to implement a cross-modal case matching and solution generation method according to Embodiment 1.

[0125] This embodiment also provides a computer-readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, server, App application store, etc., which stores a computer program. When the program is executed by a processor, it implements the corresponding function. The computer-readable storage medium of this embodiment is used to store a cross-modal case matching and solution generation device, which, when executed by a processor, implements a cross-modal case matching and solution generation method according to Embodiment 1.

[0126] In summary, according to the exemplary embodiments, the cross-modal case matching and solution generation method and apparatus of the present invention have the following beneficial effects: 1. By using persistent caching of tags.json, zero-sample five-dimensional tag extraction, and local lexical fallback, we can reduce repetitive reasoning and improve the usability of the case information parsing process.

[0127] 2. Through the low-entropy JSON rearrangement of the AI ​​inference main engine and the confidence conversion of the scoring backup sub-engine, stable matching cases can still be obtained even when the AI ​​network is abnormal, JSON parsing fails, or case identification is mismatched.

[0128] 3. By using five-dimensional scoring weights, five-order area scales, terrain affinity maps, and score conversion formulas, case matching can be expanded from simple label matching to a computable and fault-tolerant multi-dimensional matching process.

[0129] 4. By first generating multiple design direction constraints that are controlled by land use and are essentially different, and then generating schemes based on the direction constraints, the risk of homogenization of results caused by generating a single path scheme can be reduced.

[0130] 5. By employing anti-biased neutralization, inverse illusion strong boundary constraints, and image stretching constraints, we can reduce the inappropriate influence of prompts, case studies, and image compression on candidate design schemes.

[0131] 6. By generating channels, channel identifiers, and SSE multiplexing frames, multiple candidate design schemes can be generated in parallel, and other channels can continue to output even if a single channel fails.

[0132] 7. By using a solution iteration container, a three-slot sliding window, and 500 to 800-word compressed historical information and solution version sequence, the key intent chain in continuous user feedback can be preserved while controlling the context length, thereby improving the continuity of solution iteration.

[0133] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0134] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0135] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable medium, and when executed, it includes one or a combination of the steps of the method embodiments.

[0136] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0137] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method.

[0138] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for cross-modal case matching and solution generation, characterized in that, Includes the following steps: The system retrieves case data from the case library, parses it, and extracts the corresponding case information. The case information includes case semantic tags, and the fields of the case semantic tags include several of the following: scale classification, shape features, terrain topology, location background, and dominant gap. Acquire comprehensive feature data of the target land parcel, wherein the comprehensive feature data includes basic attribute data of the land parcel, spatial feature data of the land parcel, land use nature data of the land parcel, real-scene image data of the land parcel, and POI gap matrix; Based on the comprehensive feature data, case semantic tags, and a scoring model with preset rules, the cases in the case library are matched with the target plots to filter out matching cases; wherein, the scoring model includes an AI reasoning main engine and a scoring fallback sub-engine; Combining the comprehensive feature data and the case information of the matching cases, scheme generation constraint information is generated; wherein, the scheme generation constraint information includes several design direction constraints, and each of the design direction constraints satisfies the land use constraints corresponding to the land use nature data of the plot; the scheme generation constraint information also includes inverse illusion strong boundary constraints and image stretching constraints; By combining the comprehensive feature data, the case semantic tags of the matched cases, and the solution generation constraint information, a solution generation statement is generated; Call the scheme generation model and input the scheme generation statement to obtain at least one candidate design scheme; The generation scheme for generating statements specifically includes the following steps: For each of the design direction constraints, each of the design direction constraints is combined with the inverse illusion strong boundary constraint and the image stretching constraint to generate a scheme generation statement corresponding to the design direction constraint. The process of calling the scheme generation model and inputting the scheme generation statement to obtain at least one candidate design scheme specifically includes: inputting each of the scheme generation statements into the scheme generation model to obtain candidate design schemes corresponding to each of the design direction constraints.

2. The cross-modal case matching and solution generation method according to claim 1, characterized in that, The scoring model based on the comprehensive feature data, the case semantic tags, and preset rules matches cases in the case library with target plots and filters out matching cases, specifically including the following steps: Extract the structured features from the comprehensive feature data, and combine them with the preset rules and the case semantic tags of each case to synthesize the reasoning input statement; The inference input statement is input into the AI ​​inference main engine, and the output result is obtained; wherein, the output result includes the matching case; If the main AI inference engine malfunctions and cannot obtain the output results, it will switch to the locally configured scoring fallback engine and perform the following steps: Through the scoring backup engine, multi-dimensional label comparisons are performed on each case to calculate the original score for each case. Based on the confidence coefficient of each label, the confidence of each original score is converted to calculate the case score of each case. Matching cases are selected based on the case scores of each case.

3. The cross-modal case matching and solution generation method according to claim 1, characterized in that, The step of generating a solution generation statement by combining the comprehensive feature data, the case semantic tags of the matched cases, and the solution generation constraint information specifically includes the following steps: Using regular expressions, extract real-scene image sequences from the real-scene image data of the land parcel and / or the case information of the matching cases; The solution generation statement is generated by combining the comprehensive feature data, the case semantic tags of the matched cases, the solution generation constraint information, and the real-scene image sequence.

4. A cross-modal case matching and solution generation method according to any one of claims 1 to 3, characterized in that, The step of combining the comprehensive feature data and the case information of the matched cases to generate solution generation constraint information specifically includes the following steps: Extract the design logic from the case information of the matched cases; The design logic and the comprehensive feature data are input into the large language model to generate and output design direction descriptions for several target plots, which serve as design direction constraints. The design direction constraints differ from each other in at least one dimension of functional positioning, spatial organization, and service objects, and the land use data of the plots serve as constraints for generating the design direction constraints.

5. The cross-modal case matching and solution generation method according to claim 1, characterized in that, The step of inputting each of the scheme generation statements into the scheme generation model to obtain candidate design schemes corresponding to each of the design direction constraints specifically includes the following steps: A corresponding scheme generation channel is established for each of the aforementioned design direction constraints; wherein each of the aforementioned scheme generation channels is provided with a corresponding channel identifier; Each of the aforementioned scheme generation statements is input into the scheme generation model through each of the scheme generation channels for parallel reasoning, thereby generating candidate design schemes corresponding to each of the aforementioned design direction constraints.

6. A cross-modal case matching and solution generation method according to claim 1 or 5, characterized in that, It also includes the following steps: For each of the candidate design schemes, a corresponding scheme iteration container is constructed; wherein, the scheme iteration container is used to store the current feedback slot, compressed historical information, and scheme version sequence; Receive feedback information from the user for any of the candidate design schemes, and determine the candidate design scheme corresponding to the feedback information; Write the feedback information into the current feedback slot corresponding to the candidate design scheme, and obtain the design direction constraints, scheme generation statements, compressed historical information and scheme version sequence corresponding to the candidate design scheme; Based on the feedback information, design direction constraints, scheme generation statements, compressed historical information, and scheme version sequence, generate scheme iteration statements; The iterative statement is input into the scheme generation model to generate candidate design schemes after iteration.

7. A cross-modal case matching and solution generation apparatus, employing the cross-modal case matching and solution generation method according to any one of claims 1 to 6, characterized in that, include: The case analysis module is used to acquire case data from the case library, parse it, and extract the corresponding case information; wherein, the case information includes case semantic tags; The case filtering module is used to acquire comprehensive feature data of the target plot, and based on the comprehensive feature data, case semantic tags and a scoring model with preset rules, to match cases in the case library with the target plot and filter out matching cases. The constraint generation module is used to generate scheme generation constraint information by combining the comprehensive feature data and the case information of the matching cases; wherein, the scheme generation constraint information includes several design direction constraints, and each of the design direction constraints satisfies the land use constraints corresponding to the land use nature data of the land parcel; The statement generation module is used to combine the comprehensive feature data, the case semantic tags of the matching cases, and the solution generation constraint information to generate a solution generation statement; The scheme generation module is used to call the scheme generation model and input the scheme generation statement to obtain at least one candidate design scheme.

8. A computing device, comprising a memory, a processor, and computer instructions stored in the memory and executable on the processor, characterized in that, When the processor executes the instructions, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium storing computer instructions, characterized in that, When executed by a processor, this instruction implements the steps of the method according to any one of claims 1 to 6.

Images