City planning and design standard intelligent question-answering system based on large language model and GraphRAG

Abstract

The invention relates to an urban planning and design standard intelligent question-answering system based on a large language model and GraphRAG, and the system comprises a data collection and preprocessing module which is used for comprehensively collecting urban planning and design standard literatures, including national standard, industry standard, local standard and group standard multi-source documents, the standard knowledge extraction and structuring module is used for carrying out deep analysis on a planning standard text by utilizing a large language model; the graph construction and storage module constructs a city planning standard knowledge graph based on an extraction result, and the semantic retrieval and GraphRAG enhancement module constructs an enhanced question and answer mechanism based on GraphRAG in combination with the knowledge graph and text vector retrieval. The question and answer generation and result explanation module is used for inputting candidate sub-graph contents as context prompts into the large language model and generating accurate and structured answers, and the system interaction and visualization module is used for providing a visual man-machine interaction interface and supporting natural language questioning, result browsing and knowledge graph navigation.

Description

TECHNICAL FIELD

[0001] The present application relates to a city planning design standard intelligent question answering system based on a large language model and GraphRAG, belonging to the cross field of artificial intelligence and city planning, and specifically relates to a city planning design standard intelligent question answering system and method combining a large language model (LLM) and a graph retrieval-augmented generation (GraphRAG), for realizing semantic analysis, structured knowledge organization and intelligent question answering reasoning of city planning standard documents. BACKGROUND

[0002] Under the background of the increasing complexity of current city construction and management, building rules and planning standards, as important basis for ensuring city space order and safety, have complex content, numerous provisions, and are published in the form of text documents or PDFs. Designers and reviewers need to frequently consult relevant provisions in actual work, but due to the length of document content and complexity of terminology, the retrieval efficiency is low, which easily causes understanding deviation and omission.

[0003] The existing retrieval method is mostly based on keyword matching, which cannot understand the context logic, hierarchical structure and mutual reference relationship between provisions of building rules. For example, when a user queries a certain building height limit, the system can only return the fragment containing the keyword, but cannot automatically associate to related definition provisions, exception conditions or cross-chapter constraint provisions. This retrieval method lacks semantic understanding ability and is difficult to meet the accuracy and systematicness requirements of building design and review links.

[0004] In recent years, with the development of large language models and retrieval-augmented generation (RAG) technologies, some intelligent question answering systems based on semantic retrieval have appeared. This kind of method can introduce external knowledge into the model to improve the coverage and flexibility of question answering. However, in the complex building rule document scenario, the traditional RAG still has the following deficiencies: first, it lacks structured expression of semantic relationship and reference logic between provisions, making it difficult to support multi-hop reasoning; second, it lacks hierarchical organization of domain knowledge, and cannot reflect the superior-inferior relationship and constraint relationship between provisions; third, the accuracy and traceability of question answering results are weak, and users cannot verify the source of the answers. The existing retrieval-augmented technology based on GraphRAG is difficult to be directly applied to the city planning field, it is difficult to identify entities in design rules, and it is also difficult to accurately retrieve relevant laws and regulations. SUMMARY

[0005] The present application is just for the technical problems existing in the prior art, providing an urban planning design standard intelligent question and answer system based on a large language model and GraphRAG, which breaks through the limitations of traditional RAG in processing complex semantic relationships of urban planning standards. Through the weighted fusion strategy of "graph path relevance and semantic similarity", the multi-hop reasoning ability across clauses is realized. By writing prompt words and modifying entity types for the city design field, accurate retrieval of city design regulations is realized.

[0006] In order to achieve the above purpose, the technical scheme of the present application is as follows: an urban planning design standard intelligent question and answer system based on a large language model and GraphRAG, the system comprises

[0007] The data acquisition and preprocessing module is used for comprehensive acquisition of urban planning design standard literature, including national standards, industry standards, local standards and group standards of multiple sources. The collected documents usually have problems such as inconsistent format, redundant symbols, and differences in typesetting, so text cleaning, segmentation, noise removal and coding into format are needed. Through preprocessing, it is ensured that the subsequent knowledge extraction link can be carried out on the basis of structured and standardized text, improving the stability and accuracy of the whole system;

[0008] The standard knowledge extraction and structuring module uses a large language model to deeply analyze the planning standard text, automatically identifies entities (such as building height, volume rate, sunshine distance, fire safety specifications, etc.) and the constraint relationship, reference relationship and logical dependence between clauses and rules, and the extraction result is converted into a structured knowledge unit, including entities, relationships, attributes and the index information of the source clauses. This module not only supports single clause analysis, but also can handle cross-chapter semantic association, providing high-quality input for subsequent knowledge graph construction.

[0009] The graph construction and storage module constructs an urban planning standard knowledge graph based on the extraction results, summarizes entities, relationships and context semantic units into nodes and edges, and the knowledge graph can reflect the hierarchical structure, constraint logic and reference chain between clauses, and support community division and theme aggregation. The constructed graph is stored in a high-performance graph database, supporting fast query, update and expansion, ensuring that the system can still maintain high efficiency in response to large-scale standard literature;

[0010] Semantic Retrieval and GraphRAG Enhancement Module: This module combines knowledge graph and text vector retrieval to build an enhanced question answering mechanism based on GraphRAG. Specifically, the system first converts the user's question into a semantic vector and locates the relevant nodes in the knowledge graph, then performs neighborhood expansion, and finally generates candidate knowledge subgraphs through a weighted fusion strategy of "semantic similarity and graph path relevance". This mechanism not only retrieves relevant clauses related to the question, but also supports multi-hop reasoning across clauses to solve complex constraint relationship judgment problems that traditional retrieval cannot handle.

[0011] Question Answering Generation and Result Explanation Module: This module inputs the candidate subgraph content into a large language model as a context prompt to generate accurate and structured answers. The answers not only contain conclusions but also include referenced clauses, evidence texts, and reasoning chains to ensure traceability and explainability. For complex questions, the system can output multi-dimensional analysis results, such as conflicts between different clauses, exceptional conditions, or applicable scope, helping users fully understand the planning standards.

[0012] System Interaction and Visualization Module: This module provides an intuitive human-computer interaction interface that supports natural language questioning, result browsing, and knowledge graph navigation. Users can view the source clauses, reasoning paths, and related entity relationships of the answers through the interface and explore the structure of the knowledge graph in a graphical manner. The interactive interface also supports multi-dimensional filtering and visualization analysis, such as navigation by standard type, clause theme, or constraint relationship, improving user experience and retrieval efficiency.

[0013] Among them, the data collection and preprocessing module extracts the main content from standardized databases or PDF documents and removes headers, footers, annotations, blank lines, and symbols. Based on regular and semantic rules, it identifies clause numbers (such as GB 50180-2018) and chapter levels. It performs segmentation and semantic chunking to provide basic data for subsequent structuring and retrieval.

[0014] Among them, the standard knowledge extraction and structuring module mainly extracts three types of entities, including meta-information such as standard number, standard name, issuing unit, issuing time, and standard type, classification structure including rule domain, major category, and medium category, and technical rules including specific rules, numerical constraints, building categories, and building levels.

[0015] The standard knowledge extraction and structuring module combines the context engineering to completely write the prompt words used by the large language model (LLM) during knowledge extraction, guides the large language model to output the expected entities and relationships according to the urban design regulations, checks the consistency of the extraction results, and outputs in a structured JSON format. The mixed verification of artificial rules and large language model generation results improves the accuracy and stability of extraction.

[0016] Among them, the atlas construction and storage module uses a graph database (such as Neo4j) for storage, performs embedding calculation on nodes and edges, generates semantic vectors to support retrieval, and realizes the association reasoning between standard provisions through the dual embedding of structure level and semantic similarity; The graph structure constructed in the atlas construction and storage module has the following content structure:

[0017] 1) Node: contains four attributes of standard, provision, term, and index;

[0018] 2) Edge: has reference, contains, limit, dependency, and similarity;

[0019] 3) Attribute: including number, provision text, release date, and applicable scope.

[0020] Compared with the prior art, the present application has the following advantages: 1) Breakthrough the limitations of traditional RAG in processing complex semantic relationships of urban planning standards, realize the multi-hop reasoning ability across provisions through the weighted fusion strategy of "graph path correlation + semantic similarity". The system can accurately process complex constraint relationship judgment problems such as "conflict between different provisions, exception conditions or applicable scope", etc., and solve the long-standing rule association processing problem in urban planning field; 2) Significantly improve the explainability of the answer, through the combination of structured knowledge graph and "multiple clause summary method", the generated answer not only contains the conclusion, but also with reference provisions, evidence text and complete reasoning chain. This traceable and verifiable output method makes the planning decision process more transparent, and meets the rigorous requirements of urban planning professional work; 3) Optimize the work flow and decision efficiency of urban planners, through the intuitive interaction interface and knowledge graph visualization, users can quickly understand the relevance, applicable scope and potential conflicts of planning standards. The system supports multi-dimensional filtering and navigation, which converts the complex work of a large number of manual checking and comparison into efficient and accurate intelligent question and answer experience, greatly improves the professionalism and efficiency of planning and design work. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 is the overall architecture schematic diagram of the system module in the present application,

[0022] Figure 2It is an operation implementation example reference schematic diagram of the present application. DETAILED DESCRIPTION

[0023] The purpose of the present application is to provide a large language model and GraphRAG-based urban planning design standard intelligent question-answering method to overcome the shortcomings of the prior art.

[0024] The overall technical process of the present application includes five steps: first, obtain urban planning standard texts through data acquisition and cleaning, and perform structured preprocessing; second, call a large language model to complete entity recognition and relationship extraction, and construct a knowledge graph in the field of urban planning; third, generate semantic embeddings for each node in the graph, and establish a graph-enhanced retrieval index; fourth, encode and analyze the user's question, and perform semantic expansion retrieval in the graph to obtain relevant knowledge fragments; finally, generate question-answering content based on the GraphRAG mechanism, and output the article reference path and the visualization result of the knowledge graph, realizing intelligent question-answering and visualization of urban planning knowledge.

[0025] The data acquisition and preprocessing module is used for comprehensive acquisition of urban planning design standard literature, including national standards (such as national standard "GB"), industry standards, local standards, and group standards, etc. This module first performs data acquisition and cleaning, acquires urban planning design standard literature, collects and summarizes national standards, industry standards, local standards, etc. related to urban planning, and makes a csv format table, which contains the rule field, standard number, standard name, large category, medium category, specific rule issuance time, and issuance unit. Then, semantic chunking is performed, the text is segmented and semantic chunking is performed, providing a structured and standardized text basis for subsequent knowledge extraction and retrieval, to improve the overall stability and accuracy of the system.

[0026] The standard knowledge extraction and structuring module uses a large language model to deeply analyze the preprocessed planning standard text. First, key entities such as building height, volume rate, sunshine distance, fire safety specifications, etc. are automatically identified. Finally, this module combines context engineering and prompt templates to guide the large language model to extract entities and relationships. In order to enable GraphRAG to more accurately identify laws and regulations in the field of urban design, the present application classifies entity types: the extracted entities are mainly divided into three categories: meta-information category: standard number, standard name, issuance unit, issuance time, standard type; classification structure category: rule field, large category, medium category; technical rule category: specific rule, numerical constraint, building category, building level.

[0027] The application is carried out in the field of urban design, and the following prompt word files are mainly written:

[0028] 1) Relationship extraction: for the prompt words used for relationship extraction, first clearly inform the large language model that the task is to extract entities based on the given entity types, then guide the large language model to generate JSON output in a fixed format, and finally give output examples to strengthen the guiding effect, part of the example content is:

[0029] Example 1:

[0030] Entity_types:standard number, standard name, issue time, issuing unit, standard type, rule field, major category, subcategory, specific rule, numerical constraint, building category, building level

[0031] Text:

[0032] Rule field: building engineering, standard number: GB 50352-2019, standard name: “Unified Standard for Civil Building Design”, major category: building, subcategory: building type, specific rule: residential buildings with a building height not greater than 27.0m are low-rise or multi-story civil buildings, issue time: 2019.10.01, issuing unit: Ministry of Housing and Urban-Rural Development, standard type: national standard

[0033] Rule field: building engineering, standard number: GB 50352-2019, standard name: “Unified Standard for Civil Building Design”, major category: building, subcategory: building type, specific rule: public buildings with a building height not greater than 24.0m are low-rise or multi-story civil buildings, issue time: 2019.10.01, issuing unit: Ministry of Housing and Urban-Rural Development, standard type: national standard

[0034] Rule field: building engineering, standard number: GB 50352-2019, standard name: “Unified Standard for Civil Building Design”, major category: building, subcategory: building type, specific rule: single-story public buildings with a building height greater than 24.0m are low-rise or multi-story civil buildings, issue time: 2019.10.01, issuing unit: Ministry of Housing and Urban-Rural Development, standard type: national standard

[0035] Rule field: building engineering, standard number: GB 50352-2019, standard name: “Unified Standard for Civil Building Design”, major category: building, subcategory: building type, specific rule: residential buildings with a building height greater than 27.0m are high-rise civil buildings, issue time: 2019.10.01, issuing unit: Ministry of Housing and Urban-Rural Development, standard type: national standard

[0036] Regulations Area: Building Engineering; Standard Number: GB 50352-2019; Standard Name: Unified Standard for Design of Civil Buildings; Category: Architecture; Subcategory: Building Type; Specific Regulations: Non-single-story public buildings with a height greater than 24.0m and a height not exceeding 100.0m are classified as high-rise civil buildings; Issuance Date: October 1, 2019; Issuing Authority: Ministry of Housing and Urban-Rural Development; Standard Type: National Standard

[0037] Regulations Area: Building Engineering; Standard Number: GB 50352-2019; Standard Name: Unified Standard for Design of Civil Buildings; Category: Architecture; Subcategory: Building Type; Specific Rule: Buildings with a height greater than 100.0m are classified as super high-rise buildings; Issuance Date: October 1, 2019; Issuing Authority: Ministry of Housing and Urban-Rural Development; Standard Type: National Standard

[0038] Output:

[0039] ("entity"{tuple_delimiter}Architectural Engineering{tuple_delimiter}Rules Domain{tuple_delimiter}This standard belongs to the architectural engineering field and involves the classification and design of civil buildings)

[0040] {record_delimiter}

[0041] (GB 50352-2019 standard number {tuple_delimiter} This is the national standard number for the "Unified Standard for Design of Civil Buildings")

[0042] {record_delimiter}

[0043] ("entity"{tuple_delimiter}《Unified Standard for Design of Civil Buildings》{tuple_delimiter}Standard Name{tuple_delimiter}This is a national standard issued by the Ministry of Housing and Urban-Rural Development, which regulates the design and classification of civil buildings.)

[0044] {record_delimiter}

[0045] ("entity"{tuple_delimiter}Architecture{tuple_delimiter}Major Category{tuple_delimiter}This standard's major category is architecture, covering design specifications for residential and public buildings)

[0046] {record_delimiter}

[0047] 2) Community Building: To facilitate subsequent retrieval after community building, this invention has specifically tailored the prompts used. First, it clearly states that the background of the large language model task is urban design. Then, it guides the large language model to generate JSON output in a fixed format. Finally, it provides an output example to enhance the guidance effect. A portion of the example content is as follows:

[0048] {{

[0049] Title: GB 50352-2019 Classification Rules for the Height of Civil Buildings (Community)

[0050] Summary: This community is built around the national standard "Unified Standard for Design of Civil Buildings" (GB 50352-2019), covering the classification logic of residential and public buildings at different height ranges. The standard was issued by the Ministry of Housing and Urban-Rural Development and belongs to the national standard system in the field of building engineering. The rules form a clear logical chain through the relationship between building height and building levels, and are widely used in urban planning, building approval, and safety design.

[0051] "rating": 8.7,

[0052] "rating_explanation": "This rule system is mandatory nationwide and directly affects building classification, approval processes, and safety design standards."

[0053] "findings": [

[0054] {{

[0055] "summary": "Residential buildings with a height not exceeding 27.0m are classified as low-rise or multi-story civil buildings".

[0056] "explanation": "According to GB 50352-2019, when the height of a residential building does not exceed 27.0m, it should be classified as a low-rise or multi-story civil building. This classification affects key parameters such as the building's structural design, fire compartmentation, and evacuation routes, and is the basic technical boundary for residential building development. This rule is also used in urban planning to control building density and population capacity. [Data: Entities (1, 3, 7, 8); Relationships (101, 103, 104)]"

[0057] }},

[0058] {{

[0059] "summary": "Public buildings with a height of up to 24.0m are classified as low-rise or multi-story civil buildings",

[0060] "explanation": "Public buildings with a height of up to 24.0m are classified as low-rise or multi-story civil buildings. This rule applies to building types such as office buildings, schools, and hospitals, affecting their structural load, personnel density, and evacuation strategies. Due to the higher number of people typically accommodated in public buildings, this height restriction plays a crucial role in safety design.[Data: Entities (1, 4, 7, 8); Relationships (102, 106, 107)]"

[0061] }} ]

[0063] }}

[0064] The atlas construction and storage module constructs a city planning standard knowledge graph based on the extraction results, organizes entities, relationships, and context semantic units into nodes and edges. The constructed graph structure has the following contents: nodes: including standards, provisions, terms, and indicators; edges: with reference, inclusion, restriction, dependence, similarity, and other relationship types; attributes: including number, provision text, release date, applicable scope, etc. The constructed graph is stored using a high-performance graph database (Neo4j), supporting fast querying, updating, and expansion. Finally, the node and edge embedding calculation is performed to generate semantic vectors to support retrieval. Through the dual embedding of structural hierarchy and semantic similarity, the association reasoning between standard provisions is realized.

[0065] The semantic retrieval and GraphRAG enhancement module combines knowledge graph and text vector retrieval to construct an enhanced question and answer mechanism based on GraphRAG. The core is a retrieval fusion strategy based on "graph path relevance and semantic similarity weighting". The knowledge retrieval is mainly realized through the following four steps:

[0066] 1) User question analysis: Encode the user question into a semantic vector, combine entity recognition and term standardization, extract explicit entities, aliases, and domain keywords, and map them to the standard nodes in the knowledge graph as graph search anchors. This step aims to retain both semantic vectors and symbolic anchors, allowing the search to rely on semantic similarity and utilize the precise positioning of graph structure, reducing semantic drift and the problem of multiple names for the same thing; 2) Graph semantic neighborhood expansion: In the knowledge graph, expand the semantic neighborhood around the anchor point to obtain related nodes, relationship edges, and evidence text units within a limited number of jumps. Combine community division results to prioritize internal expansion within relevant communities while focusing on cross-community bridge edges; 3) Subgraph screening: Weighted fusion scoring of candidate nodes and text based on "semantic similarity + graph path relevance". Introduce node importance and evidence coverage as additional weighting factors. By dynamically adjusting the weight parameters, the most suitable fusion strategy is selected for different types of questions (fact-based, reasoning-based, and comprehensive). High-quality candidate subgraphs are selected as structured context for the question, and this step supports complex multi-hop reasoning, such as cross-document constraint relationship judgment; 4) Structured prompt generation: Convert candidate subgraphs into structured prompt information, such as presenting in the form of "entity - relationship - evidence" chains, and attaching original text units as evidence. Input the structured prompt (node summary, relationship chain, and evidence text) into a large language model to generate answers, supporting multi-hop reasoning and traceable references.

[0067] The question and answer generation and result explanation module inputs the knowledge fragments returned by GraphRAG and the user question into a large language model to generate accurate and structured answers. After receiving the structured context, the large language model performs multi-hop reasoning along the provided reasoning path, enabling complex cross-document constraint judgment. The invention also redesigns the prompt words used during answer generation, requiring the large language model to provide no information beyond the search materials, significantly reducing the illusion of RAG and the large language model, and ensuring that the accuracy of the answers meets the requirements of urban design; Additionally, the language model is provided with examples, with sources labeled after each answer, allowing users to view the source thread and increasing the model's explainability. Examples include:

[0068] For example:

[0069] "Building height not exceeding 27.0m residential building should be classified as low-rise or multi-storey civil building [Data: Reports (2, 7, 64, 46, 34, +more)]. Building height not greater than 24.0m public building is low-rise or multi-storey civil building [Data: Reports (1, 3)]"

[0070] where 1, 2, 3, 7, 34, 46, and 64 represent the id (not the index) ofthe relevant data record.

[0071] Do not include information where the supporting evidence for it isnot provided.

[0072] Limit your response length to {max_length} words.

[0073] Language models output answers that not only contain conclusions but also include references to articles, evidence texts, reasoning chains, and confidence levels, making the results traceable and interpretable.

[0074] Example: For the question "Should a residential building with a height of 25 meters be classified as a low-rise or multi-story building or a high-rise building? What are the corresponding standard names?", the system can generate the answer: "According to the Unified Standard for Civil Building Design, residential buildings with a height of not more than 27.0 meters are classified as low-rise or multi-story buildings, and a 25-meter residential building should be classified as a low-rise or multi-story building."

[0075] System interaction and visualization module: This module provides an intuitive human-computer interaction interface, supporting natural language questioning, result browsing, and knowledge graph navigation. It has the following functions: 1) Interaction function: supports natural language questioning (such as the input box in Figure 2 , result browsing; 2) Visual traceability: users can view the source articles, reasoning paths, and related entity relationships of the answers through the interface; 3) Graph navigation: the structure of the knowledge graph can be explored in a graphical manner, supporting multi-dimensional filtering and visual analysis, such as navigation by standard type, article theme, or constraint relationship.

[0076] It should be noted that the above embodiments are not intended to limit the scope of protection of the present application, and any equivalent transformations or substitutions made on the basis of the above technical solutions fall within the scope of protection of the claims of the present application.

Claims

1. An intelligent question-answering system for urban planning and design standards based on a large language model and GraphRAG, characterized in that, The system includes Data Acquisition and Preprocessing Module: This module is used to comprehensively collect urban planning and design standard documents, including national standards, industry standards, local standards, and group standards from multiple sources. Standard Knowledge Extraction and Structuring Module: This module uses a large language model to perform deep analysis of the planning standard text, automatically identifying key entities and the constraints, references and logical dependencies between clauses. The extraction results are transformed into structured knowledge units, which include entities, relationships, attributes and their source clause index information. Graph Construction and Storage Module: This module constructs a knowledge graph of urban planning standards based on the extraction results. It organizes entities, relationships, and contextual semantic units into nodes and edges. The knowledge graph can reflect the hierarchical structure, constraint logic, and reference chain between articles, and supports community division and topic aggregation. The completed graph will be stored in a high-performance graph database, supporting fast query, update, and expansion, ensuring that the system can still maintain high efficiency when facing large-scale standard documents. Semantic Retrieval and GraphRAG Enhancement Module: This module combines knowledge graph and text vector retrieval to construct an enhanced question-answering mechanism based on GraphRAG. Specifically, the system first converts user questions into semantic vectors, locates relevant nodes in the knowledge graph, expands their neighborhood, and then generates candidate knowledge subgraphs through a weighted fusion strategy of "semantic similarity and graph path relevance." The question-answering generation and result interpretation module uses candidate subgraph content as contextual prompts input into the large language model to generate accurate and structured answers. These answers not only include conclusions but also cited references, evidence text, and reasoning chains, ensuring the traceability and interpretability of the results. For complex questions, the system can output multi-dimensional analysis results. System Interaction and Visualization Module: This module provides an intuitive human-computer interaction interface, supporting natural language questioning, result browsing, and knowledge graph navigation. Users can view the source text of the answer, reasoning path, and related entity relationships through the interface, and explore the structure of the knowledge graph in a graphical way. The interactive interface also supports multi-dimensional filtering and visualization analysis.

2. The intelligent question-answering system for urban planning and design standards based on a large language model and GraphRAG as described in claim 1, characterized in that, The data acquisition and preprocessing module extracts the main text content from a standardized database or PDF document, and removes headers, footers, annotations, blank lines, and symbols; it identifies the article number and chapter level based on regular expressions and semantic rules; and it performs segmentation and semantic block division (chunking) to provide a foundation for subsequent structuring and retrieval.

3. The intelligent question-answering system for urban planning and design standards based on a large language model and GraphRAG as described in claim 1, characterized in that, The main entity types in the standard knowledge extraction and structuring module are divided into three categories: meta-information type, including standard number, standard name, issuing unit, issuance time, and standard type; classification structure type, including rule domain, major category, and intermediate category; and technical rule type, including specific rules, numerical constraints, building category, and building level.

4. The intelligent question-answering system for urban planning and design standards based on a large language model and GraphRAG as described in claim 1, characterized in that, The prompts for knowledge extraction and graph construction have been reconstructed. Based on the characteristics of the urban planning field, important information such as buildings and thresholds are extracted to improve the accuracy of numerical queries. The specialized prompt template is used to guide the Large Oracle Model (LLM) to generate entities and relations. The extraction results are checked for consistency and the output is in structured JSON format.

5. The intelligent question-answering system for urban planning and design standards based on a large language model and GraphRAG as described in claim 1, characterized in that, The graph construction and storage module uses a graph database for storage, performs embedding calculations on nodes and edges, generates semantic vectors to support retrieval, and realizes the association reasoning between standard texts through dual embedding of structural hierarchy and semantic similarity. The graph structure constructed in the graph construction and storage module has the following structure: Nodes: contain four attributes: standards, clauses, terms, and indicators; Edges can be defined as references, containment, restrictions, dependencies, or similarities. Attributes include number, clause text, publication date, and scope of application.

6. The intelligent question-answering system for urban planning and design standards based on a large language model and GraphRAG as described in claim 2, characterized in that, The semantic retrieval, combined with the GraphRAG module, utilizes knowledge graphs and text vector retrieval to construct a GraphRAG question-answering mechanism, specifically including the following steps: Step 1: Encode user questions into semantic vectors and extract explicit entities and keywords as anchor points for graph retrieval. Step 2: Expand the semantic neighborhood of the anchor point in the knowledge graph with a finite number of hops to obtain relevant nodes, relation edges, and their supporting text units. Control noise and scope by using node importance and relation confidence. Step 3: Perform a weighted fusion scoring of "semantic similarity and graph path relevance" on candidate nodes and text to select high-quality candidate subgraphs as the structured context of the problem. Step 4: Convert candidate subgraphs into structured hints, input them into the large model to generate answers, and support multi-hop reasoning and traceable references.

7. An intelligent question-answering method for urban planning and design standards based on a large language model and GraphRAG, characterized in that, The method of using the intelligent question-answering system for urban planning and design standards based on large language models and GraphRAG as described in any one of claims 1-5 includes the following steps: S1. Data Acquisition and Cleaning: Obtain standard urban planning text and perform structured preprocessing; S2. Knowledge Extraction and Graph Construction: Utilize a large language model to complete entity recognition and relation extraction, constructing a knowledge graph; context engineering was performed for urban planning design, redesigning the prompt words and entity types for knowledge extraction using the large language model; S3. Vectorized Representation and Graph Index Generation: Generate semantic embeddings for each node and establish a graph-enhanced retrieval index; S4. User Question Parsing and Retrieval Enhancement: Encode questions and perform semantic expansion retrieval in the graph to obtain relevant knowledge fragments; S5. Response Generation and Visualization: Generate responses based on the GraphRAG context, outputting the text reference path and graph visualization results.

Images