City updating intelligent expert system architecture and method based on large language model

Abstract

The invention discloses a city updating intelligent expert system architecture and method based on a large language model, and relates to the technical field of city planning and construction. The knowledge base management module supports efficient retrieval and application of professional domain knowledge; the expert agent group comprises a user-defined core expert and a system dynamic supplement expert; the expert team collaborative decision-making module is used for executing multiple rounds of hierarchical collaborative decision-making processes; the meta-agent module is responsible for project feature analysis and expert team dynamic configuration; the auxiliary agent module comprises a host agent which is responsible for guiding the decision making process; the recorder agent generates a decision summary report; the verification agent is responsible for verifying and evaluating the professional ability of the expert agent; and the scoring agent evaluates the decision quality. According to the method, the problems of dispersed professional knowledge, low cross-domain expert cooperation efficiency, high decision-making cost and the like in the existing city updating decision-making process are solved, and the decision-making efficiency and quality of city updating early-stage planning are comprehensively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of urban planning and construction. More specifically, it relates to an intelligent expert system architecture and method for urban renewal based on large language models. Background Art

[0002] Urban renewal is an important strategy for contemporary urban development, aiming to improve the quality of life of residents, stimulate regional vitality, and promote the sustainable development of cities by transforming and optimizing old urban areas. In this process, the preliminary planning determines core elements such as project positioning and renewal strategies, and is a key link to ensure the realization of renewal goals.

[0003] At the current stage, the preliminary planning method mainly needs to go through multiple links such as market research organized by architectural planning companies, scheme demonstration, policy coordination, and economic evaluation. Each link mainly relies on the participation of experts in different fields. This method not only has a long decision-making cycle and high costs, but also is difficult to quickly control the planning direction. Therefore, there is an urgent need for an efficient and intelligent preliminary planning method to integrate knowledge in multiple fields, provide scientific and accurate renewal plans, and meet the complex needs of current and future urban renewal. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an intelligent expert system architecture and method for urban renewal based on large language models. Through mechanisms such as dynamic expert team construction, hierarchical knowledge base management, and multi-stage collaborative decision-making, it solves problems such as scattered professional knowledge, low efficiency of cross-field expert collaboration, and high decision-making costs in the existing urban renewal decision-making process, and comprehensively improves the decision-making efficiency and quality of urban renewal preliminary planning.

[0005] To solve the above technical problems, the technical solutions adopted by the present invention are as follows:

[0006] An intelligent expert system for urban renewal based on large language models, comprising:

[0007] A knowledge base management module: adopting retrieval-augmented generation technology (RAG) to support the efficient retrieval and application of professional domain knowledge; the knowledge base management module is provided with a knowledge index unit and a permission management unit; the knowledge index unit maintains a FAISS index and a multi-level directory structure to support vectorized retrieval; the permission management unit dynamically allocates and manages knowledge access permissions based on expert role characteristics;

[0008] An expert intelligent agent group: including user-defined core experts and system-dynamically supplemented experts;

[0009] An expert team collaborative decision-making module: executing a multi-round hierarchical collaborative decision-making process, covering expert independent analysis, cross-discussion, and scheme optimization;

[0010] Meta-Agent Module: Responsible for project feature analysis and dynamic configuration of the expert team, and dynamically create expert agents;

[0011] Assistant Agent Module: Includes a host agent: responsible for process control, topic guidance, and integration of differences, and guiding the decision-making process; a recorder agent: real-time recording of discussion content and generating a decision summary report; a verification agent: responsible for verifying and evaluating the professional capabilities of expert agents; a scoring agent: responsible for real-time scoring of the discussion quality, dynamically optimizing weights, and evaluating decision quality.

[0012] A decision-making method for urban renewal projects using an intelligent expert system based on large language models, including the following steps:

[0013] S1: Obtain project basic information

[0014] Obtain the basic information of the urban renewal project, including multimodal data such as project text descriptions and site status photos; the basic information of the urban renewal project is divided into two categories: text materials and picture materials. Among them, text materials include project overview, land use status, surrounding environment, development demands, etc.; picture materials include site status photos, surrounding environment photos, etc.;

[0015] Perform standardization processing on the classified materials, specifically including: unifying the format of text materials and extracting key information; uniformly adjusting the resolution of picture materials;

[0016] Call the large language model to perform multimodal analysis on the standardized project materials, parse classification information such as project type, scale level, technical difficulty, and professional requirements in the text materials, and at the same time identify site features, surrounding environment, and existing problems in the picture materials;

[0017] Establish an index and store the standardized materials to achieve unified management of text and pictures for subsequent input;

[0018] S2: Analyze project features and construct an expert team

[0019] The meta-agent performs feature analysis on the project materials and constructs an expert team based on the analysis results, including selecting experts with high matching degrees from the preset core expert library, and system prompts for dynamically generating new experts according to the project type to create expert agents in specific fields; input the project feature description and the list of key technical requirements to the meta-agent, and the meta-agent selects the core expert agent with the highest matching degree from the preset expert library;

[0020] The verification agent conducts professional verification and ability assessment on the generated experts, dynamically assigns knowledge base permissions according to their roles, and forms a complete expert decision-making team; for the characteristics of the project, the meta-agent, based on the semantic parsing ability of the large language model itself, independently judges and identifies the professional fields that need to be supplemented, and dynamically supplements new expert agents based on the specialties of the selected expert library; specifically including generating system prompt words and assigning knowledge base permissions; the verification agent is responsible for detecting and calibrating the professional capabilities of the new expert agents to ensure that the newly generated experts have sufficient professionalism;

[0021] S3: Based on the expert team configuration plan, execute multi-round hierarchical collaborative decision-making, and the process is as follows:

[0022] The first stage: Independent analysis

[0023] Each expert agent obtains domain professional knowledge from the knowledge base based on the retrieval-augmented generation technology, conducts professional field analysis and puts forward preliminary analysis;

[0024] The second stage: Cross-supplementation and discussion

[0025] The host agent guides each expert agent to review and supplement the analysis of other experts in the first stage, and at the same time conducts discussions around key issues. The evaluation agent evaluates the depth and professionalism of the experts' speeches in real time, and the recorder agent tracks the discussion points and controversial issues;

[0026] The third stage: Disagreement resolution and solution optimization

[0027] The host agent extracts the existing controversial issues based on the records in the second stage, and organizes relevant experts to conduct a second-round argumentation from a professional perspective; the remaining experts score according to the argumentation process in three dimensions: sufficiency of evidence (40%), technical feasibility (30%), and implementation effect (30%); according to the scoring results, the experts with low scores need to adjust or make concessions to the opinions of the experts with high scores to reach a consensus on the issues; finally, the host agent summarizes the discussion results, sorts out the disagreement solutions, and outputs the optimized decision-making results.

[0028] Preferably, S2 also includes the following content:

[0029] S21: The meta-agent compares the project feature list with the professional coverage scope of the selected core experts, identifies the professional fields not covered, and generates a supplementary professional requirement list;

[0030] Among them, the professional coverage calculation formula is:

[0031]

[0032] Among them, C represents the professional set of core experts, which is preset by the system, R represents the professional set of project requirements, which is obtained when using GPT4o preprocessing, |C∩R| represents the size of the intersection of the two sets, and |R| represents the size of the project requirement set, which is judged by the meta-agent. If the Coverage Score is less than 0.8, new experts are activated and supplemented;

[0033] S22: Based on the technical characteristics of the uncovered professional field, the meta-agent uses the predefined expert role template to dynamically generate the system prompt words for the corresponding experts. The prompt words contain the following structure:

[0034] Expert_Prompt={Role_Definition+Professional_Background+

[0035] Knowledge_Scope+Decision_Rules}

[0036] Among them, Role_Definition defines the expert identity, Professional_Background sets the professional background, now_Scope defines the knowledge scope, and Decision_Rules specifies the decision rules;

[0037] S23: Based on the generated system prompt words, a new expert agent is dynamically instantiated by calling the large language model and assigning decision weights (determined by the meta-agent); the new agent is implemented by encapsulation class, supports accepting task input and generating responses that meet professional characteristics; after instantiation, the verification agent verifies its role to ensure that its response meets the expected professional characteristics and decision-making behavior;

[0038] The role verification specifically includes: asking questions based on the new expert system prompts to evaluate the professionalism, logic and decision consistency of the answers; when the total score reaches the preset threshold, the verification is passed; if the threshold is not reached, the communication meta-agent optimizes the system prompts and re-verifies until it passes;

[0039] S24: Assign corresponding knowledge base access rights to the newly generated expert agents that have passed the verification, and integrate the newly generated expert agents with the core expert agents to form a complete expert decision-making team.

[0040] Preferably, the knowledge allocation collaboration mechanism in S24 specifically includes:

[0041] S241: Classify and label the content of the knowledge base by professional field, establish a knowledge domain label system, including main fields such as architectural design, structural transformation, historical protection, economic evaluation, etc., and their respective sub-fields; use the LLM method to extract domain labels from documents; when new content is added to the knowledge base, trigger the dynamic expansion mechanism and call the LLM (such as GPT4o) to automatically extract labels;

[0042] S242. First, calculate through OpenAI Embeddings to convert the role description of the agent and the knowledge base labels into vectors; then calculate the Role_Match value through the cosine similarity formula

[0043]

[0044] Among them, V1 represents the query vector, V1·V2 represents the dot product of the two vectors, and ||V1|| and ||V2|| represent the Euclidean norms of the two vectors respectively;

[0045] Next, calculate the relevance between the expert agent and each knowledge base. The relevance calculation formula is:

[0046] Domain Relevance = μ·Role Match+(1-μ)·Usage Score

[0047] Among them, Role_Match is the role matching degree, Usage_Score is the usage frequency with a default value of 0, and μ is the balance factor set to 0.7; generate an expert-knowledge base mapping matrix to guide subsequent permission allocation;

[0048] S243. Implement dynamic authorization for knowledge access according to the relevance. The access permission calculation formula:

[0049]

[0050] Among them, Domain Relevance: the matching degree with the knowledge base label; Performance Score: calculated according to the contribution degree of the agent; Usage Frequency: the access times of the knowledge base; the weights w1, w2, w3 can be adjusted through experiments, and the default values can be set to 0.5, 0.3, 0.2; Access_Level takes values of {3, 2, 1, 0} corresponding to full access, read-only access, restricted access, and no access rights respectively;

[0051] Conflict handling: When agent A has full access rights and agent B only has restricted access rights, first sort according to the level of access rights. For example, if A has full access rights and B has read-only access rights, then the priority of A is higher than that of B; secondly, when the priorities are the same, sort according to the decision-making weights;

[0052] S244. Permission Initialization and Optimization

[0053] For existing core expert agents, grant full access rights to the corresponding knowledge bases;

[0054] For newly generated expert agents, allocate initial access rights based on their role definitions. At the same time, record the usage frequency and decision-making contributions of the knowledge bases through system logs, and dynamically optimize the permissions. For example, knowledge bases with high-frequency access will be given higher permissions to the using experts.

[0055] Preferably, the organizational structure of the knowledge base specifically includes:

[0056] I. Hierarchical storage structure:

[0057] a: Vector data layer, using OpenAI Embeddings to embed vectors to represent knowledge content, and using FAISS index to store vector representations of various text documents

[0058] b: Raw data layer, using MongoDB to store professional knowledge in fields such as architectural design, structural transformation, historical protection, and economic evaluation according to professional fields;

[0059] c: Relational data layer, based on Neo4j to record knowledge associations, case associations, and decision-making experience paths between professional fields;

[0060] II. Index management mechanism:

[0061] a: Build a multi-level index structure, including professional field index, knowledge type index, and keyword index;

[0062] b: Maintain vector retrieval index, including regularly updating the FAISS index;

[0063] III. Access control mechanism: including access control based on professional fields, level setting of knowledge importance, and cache management of usage frequency.

[0064] Preferably, step S3 specifically includes:

[0065] S31: Each expert agent performs the first-stage independent analysis, respectively obtains professional knowledge in its own field from the knowledge base based on retrieval-enhanced generation technology, conducts diagnostic analysis in combination with project characteristics, identifies key problems and proposes preliminary suggestions, and forms a professional report including current situation assessment, problem analysis, and improvement ideas;

[0066] S32: Each expert agent conducts cross-discussion in the second stage. The host agent guides the expert agents to review and supplement the analyses of other experts. The evaluation agent assesses the speech quality of each expert and the overall quality of the current round. The recorder agent records the key points of the discussion, major disagreements, and consensus.

[0067] S33: The host agent organizes the solution optimization in the third stage. Based on the key points, problems, and disagreements recorded in the second round of discussion, the core problems are summarized and refined. The host agent organizes the experts to demonstrate the feasibility of various solution plans from a professional perspective, and resolves the disagreements through secondary demonstration and expert voting to reach a final consensus.

[0068] S34: The host agent generates a decision summary report, including the advantages and disadvantages analysis of each solution plan, expert opinion disagreements and consensus, and final suggestions. The recorder agent stores the key data, expert evaluation opinions, and solution optimization paths in the decision-making process into the knowledge base.

[0069] Preferably, step S31 specifically includes:

[0070] S311: Each expert agent performs retrieval enhancement generation based on its professional knowledge base, specifically including: converting the project feature list into a retrieval query vector, and retrieving relevant cases and professional knowledge in the knowledge base through vector similarity calculation. The similarity calculation formula is:

[0071]

[0072] where Q represents the query vector, D represents the document vector in the knowledge base, Q·D represents the dot product of the two vectors, and ||Q|| and ||D|| represent the Euclidean norms of the two vectors respectively.

[0073] S312: Sort the retrieval results by relevance, select the top k documents with the highest similarity scores, and embed them as context information into the expert input prompt. The prompt template for each expert is:

[0074] Expert_C={Role_Description+Retrieved_Knowledge+

[0075] Project_Features+Analysis_Task[+Constraints]}

[0076] where Role_Description is the expert role description, Retrieved_Knowledge is the retrieved knowledge content, Project_Features is the project feature, Analysis_Task is the analysis task requirement; Constraints is the constraint condition.

[0077] S313: Each expert agent generates an item analysis report from its respective professional perspective based on the prompt words embedded with retrieved knowledge, including current situation assessment, problem diagnosis, and improvement suggestions.

[0078] S314: The host agent summarizes the analysis reports of each expert to form a preliminary plan for the first round. The content of the plan includes specific transformation strategies and implementation suggestions for each professional field.

[0079] Preferably, step S32 specifically includes:

[0080] S321: At the beginning of the second stage, the host agent organizes cross-discussions. Specifically, first, a professional dependency matrix M is constructed based on the results of the first round, where M ij represents the degree of dependence of profession i on profession j, which is scored by the evaluation agent on the results of the first stage, and the value range is [0, 1].

[0081] Calculate the professional correlation degree according to the dependency matrix:

[0082]

[0083] Design the discussion order from high to low according to Relevance_Score; the host agent designs guiding questions based on the first-round plans of high-correlation expert agents to drive experts to conduct in-depth criticism and professional supplementation of the analysis reports of other experts.

[0084] S322: The evaluation agent gives an instant score to each speech of each expert in each round:

[0085] Expert_core = w1·Technical_Depth + w2·Cross_Domain _ Analysis + w3·Discussion_Completeness + w4·Discussion_Effect

[0086] where Technical_Depth is the score of discussion depth, Cross_Domain_Analysis is the score of cross-domain analysis, Discussion Completeness is the score of discussion integrity, Interaction_Effect is the score of interaction effect, w1 to w4 are weight coefficients, which are set during system initialization and satisfy Σwi = 1; the scores of each dimension are normalized to the [0, 1] interval, and a threshold is set.

[0087] The evaluation agent calculates the overall quality of the round based on the speech scores of all experts in the current round:

[0088]

[0089] Where N is the number of experts participating in the discussion. When any one of the following three conditions is met, the evaluation agent notifies the host agent to converge:

[0090] I. When Discussion_Quality > 0.8

[0091] II. When the change of Discussion_Quality for three consecutive rounds < the set value of 0.1

[0092] III. Exceeding the maximum number of discussion rounds (10 rounds)

[0093] If the convergence condition is not met, the evaluation agent will pass the scoring situation of the experts' speeches below the threshold in each round to the host agent before the start of the next round; the host agent adjusts the discussion based on the following strategies:

[0094] Low-score dimension guidance: If Technical_Depth is low, guide the corresponding expert to increase the depth of thinking. If Cross_Domain_Analysis is low, design cross-professional association questions. If Interaction_Efect is low, guide the experts to enhance interactivity;

[0095] Dimension priority adjustment: Dynamically adjust the key weight of the next round of discussion according to the scores, and give priority to focusing on weak links;

[0096] Feedback notification: Before starting the next round of discussion, the host agent notifies the experts of the current weak dimension and the proposed optimization direction;

[0097] S323: Recording and sharing mechanism of the recorder agent: The recorder agent tracks and records the key points of each round of discussion, stores them in a structured data format, and the recorded content includes: timing information (discussion time and discussion stage); issue number (Issue ID and professional field); record of experts' opinions (including expert ID, object of evaluation, professional divergence points and their argumentative supports); associated questions (including question description, involved profession); consensus progress (including key points of consensus reached, problems to be solved, professional opinions to be supplemented); when the host agent guides each expert to speak, the communication recorder agent synchronously displays the focus of the current discussion, the experts' opinions collected in this round, and the current divergence points;

[0098] When an expert is speaking, he can directly quote the opinions of other experts provided by the recorder to make supplements or optimizations; when the following situations occur, the recorder agent updates the displayed content: after a new expert speaks, a stage consensus is reached, and a new divergence appears;

[0099] For controversial issues, a multi-version recording mechanism is adopted to track the evolution of viewpoints. After reaching a consensus on the differences, the recorder agent performs version merging, retaining the final version of the content on which a consensus is reached and treating the rest as historical versions;

[0100] S324: Tasks after convergence: The recorder agent generates a summary report on the second-phase discussion. The report structure includes:

[0101] I. List of completed issues: The status of each issue and its final conclusion;

[0102] II. List of unfinished issues: Classification of problem types, responsible experts, and suggestions for follow-up handling;

[0103] III. Record of discussion progress: Record the changing trends of Discussion_Quality and each indicator;

[0104] IV. Update the knowledge base: Store the discussion content as structured data for subsequent use.

[0105] Preferably, step S33 specifically includes:

[0106] S331: The host agent extracts key points, differences, and the list of unfinished issues based on the second-phase summary report;

[0107] S332: The host agent first guides each expert to supplement and optimize the consensus-reached solution. Secondly, for controversial issues, it organizes the corresponding experts to conduct secondary demonstrations from a professional perspective, guides the other expert agents to score the professionalism and effectiveness of their demonstration processes. According to the scoring results, the low-scoring expert agents make concessions to the high-scoring expert agents to form a compromise solution and finally reach a consensus.

[0108] Preferably, step S34 specifically includes:

[0109] S341: The recorder agent systematically sorts out the decision-making process and generates a structured decision summary report. The report structure includes: solution summary, technical analysis, benefit evaluation, focus of disputes, and final suggestions;

[0110] S342: Record the solution optimization path diagram, record the evolution process from the initial solution to the final solution, including key modification points, adjustment bases, and improvement effects in each round of decision-making, and at the same time mark the decision-making contributions of each expert in different stages;

[0111] S343: Extract the experience knowledge in the decision-making process, including: solutions to key divergence points, quantitative indicators for expert evaluation criteria, effective strategies for solution optimization, etc., and store them structurally according to a preset knowledge template;

[0112] S344: The recorder updates the decision report and empirical knowledge to the system knowledge base, ignoring the content with a similarity higher than the threshold γ and storing the content lower than the threshold as new knowledge entries.

[0113] The beneficial effects of adopting the above technical solutions are as follows:

[0114] 1. Achieved the construction of an intelligent expert team: The present invention dynamically constructs and verifies an expert team based on project characteristics through a meta-agent, and ensures the professionalism and decision-making ability of the expert team through calculating the professional coverage and the role verification mechanism. Compared with the existing expert construction method that relies on manual selection, the present invention effectively overcomes the problem of subjective randomness, realizes the flexibility of team construction, and significantly improves the cross-domain team matching efficiency.

[0115] 2. Innovatively designed a hierarchical knowledge base structure: Adopting a three-layer architecture of a vector data layer, a raw data layer, and a relational data layer, combined with the FAISS indexing technology, realizes the efficient retrieval and precise application of professional knowledge. The raw data layer is classified and stored by professional fields to ensure data integrity; the relational data layer supports the interconnection and interoperability of cross-domain knowledge by recording the association relationships between knowledge. Through a dynamic permission allocation mechanism, it is ensured that the expert agent can accurately access knowledge resources based on its professional field, significantly improving the knowledge acquisition efficiency and the scientificity of decision-making.

[0116] 3. Established an efficient collaborative decision-making mechanism: The present invention proposes a three-stage hierarchical collaborative decision-making process of "independent analysis - cross-discussion - solution optimization": In the first stage, through the Retrieval-Augmented Generation (RAG) technology, it supports the independent analysis of the expert agent; in the second stage, the discussion order is optimized based on the professional dependency matrix; in the third stage, a scoring-driven disagreement resolution mechanism (the scoring dimensions include 40% for sufficiency of arguments, 30% for technical feasibility, and 30% for implementation effect) is adopted to achieve efficient cross-domain collaboration. Compared with the traditional manual decision-making method, the present invention significantly improves the expert collaboration efficiency and the scientificity of decision-making.

[0117] 4. Provided complete decision traceability ability: The recorder agent of the present invention uses a structured data format to track and record the entire decision-making process, including expert opinions, the evolution of disagreements, and the formation of consensus, making each decision-making plan fully traceable. Through the multi-version recording mechanism, the complete evolution path from the initial disagreement to the final consensus is saved, which not only supports the backtracking of historical plans but also provides decision-making references for similar scenarios, significantly improving the decision reuse efficiency.

[0118] 5. A closed-loop system for continuous optimization is formed: Through the knowledge update mechanism, the empirical knowledge generated in the decision-making process is structurally stored and updated to the knowledge base according to a preset template, realizing the continuous accumulation and optimization of the system's capabilities, and improving the efficiency and quality of subsequent decision-making. Compared with the traditional static knowledge base, the present invention not only improves the efficiency and quality of subsequent decision-making, but also enables the system to have the ability of self-iteration in long-term applications.

[0119] 6. Through the collaborative operation of multiple agents, the present invention effectively solves the problems existing in the prior art, such as scattered professional knowledge, low efficiency of cross-domain expert collaboration, and difficulty in tracing decision-making schemes, providing intelligent technical support for the preliminary planning of urban renewal projects. The system not only significantly improves the decision-making efficiency and scientificity, but also provides a promotable technical framework for collaborative decision-making in future complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0120] Figure 1 is the system architecture diagram of the present invention;

[0121] Figure 2 is the flowchart of the decision-making method of the present invention;

[0122] Figure 3 is the agent collaboration diagram;

[0123] Figure 4 is the example diagram of the large model API setting in Embodiment 1

[0124] Figure 5 is the example diagram of the input project materials in Embodiment 1.

[0125] Figure 6 is the example diagram of expert selection and speech in Embodiment 1.

[0126] Figure 7 is the example diagram of expert interaction in Embodiment 1. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0127] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0128] The system framework of the present invention is as Figure 1 shown, including the following five modules:

[0129] 1. Knowledge Base Management Module: Adopting Retrieval-Augmented Generation (RAG) technology, it supports the efficient retrieval and application of professional domain knowledge. Knowledge Indexing Unit: Maintains FAISS index and multi-level directory structure, supporting vectorized retrieval; Permission Management Unit: Dynamically allocates and manages knowledge access permissions based on expert role characteristics.

[0130] 2. Expert Agent Group: Includes user-defined core experts and system-dynamically supplemented experts.

[0131] 3. Expert Team Collaborative Decision-Making Module: Executes a multi-round hierarchical collaborative decision-making process, covering expert independent analysis, cross-discussion, and solution optimization.

[0132] 4. Meta-Agent Module: Responsible for project feature analysis and dynamic configuration of the expert team.

[0133] 5. Auxiliary Agent Module: Includes a Host Agent (responsible for guiding the decision-making process), a Recorder Agent (responsible for recording decision-making content), a Verification Agent (responsible for verifying the expert capabilities of agents), and a Scoring Agent (responsible for evaluating decision-making quality).

[0134] Using the above system for urban project update decision-making includes the following steps (as Figure 2 shown):

[0135] S1: Obtain project basic information (the process is as Figure 3 shown)

[0136] Obtain the basic information of the urban renewal project, including multi-modal data such as project text descriptions, current site photos, etc. Specifically, it includes the following steps:

[0137] S11: The basic information of the urban renewal project is divided into two categories: text materials and picture materials; among them, the text materials include project overview, land use status, surrounding environment, development demands, etc.; the picture materials include current site photos, surrounding environment photos, etc.;

[0138] S12: Standardize the classified materials, specifically including: unifying the format of text materials and extracting key information; uniformly adjusting the resolution of picture materials;

[0139] S13: Invoke the large language model to perform multi-modal analysis on the preprocessed project materials, parse classification information such as project type, scale level, technical difficulty, and professional requirements in the text materials, and at the same time identify site features, surrounding environment, and existing problems in the picture materials;

[0140] S14: Index and store the preprocessed materials to achieve unified management of text and pictures for subsequent input; the materials are stored in the database storage.

[0141] Text data: Stored in MongoDB, with each record using project_id as the primary key.

[0142] Image data: The image files are stored in the file system, and the path information is recorded in the image_data field of the database.

[0143] The format is as follows:

[0144]

[0145]

[0146] Provide a REST API for the meta-agent to call, for example:

[0147] 1. Get project data:

[0148] Request example: GET / projects / {project_id}

[0149] {

[0150] "project_id": "001",

[0151] "text_data": {...},

[0152] "image_data": [...],

[0153] "analysis_results": {...}

[0154] }

[0155] 2. Get multi-modal analysis results:

[0156] Request example: GET / projects / {project_id} / analysis

[0157] {

[0158] "analysis_results": {

[0159] "project_features": [...],

[0160] "technical_needs": [...],

[0161] "site_analysis": {...}

[0162] }

[0163] }

[0164] S2: Analyze project characteristics and build an expert team

[0165] The meta-agent analyzes the characteristics of project materials and builds an expert team based on the analysis results, including selecting experts with high matching degrees from a preset core expert library and generating system prompts for dynamically creating new experts according to the project type to create expert agents in specific fields.

[0166] The verification agent conducts professional verification and ability assessment on the generated experts, and dynamically assigns knowledge base permissions according to their roles to form a complete expert decision-making team;

[0167] Specifically, it includes the following steps

[0168] S21: Input the project feature description and the list of key technical requirements into the meta-agent, and the meta-agent selects the core expert agent with the highest matching degree from the preset expert library.

[0169] S22: For the project characteristics, the meta-agent, based on the semantic parsing ability of the large language model itself, independently judges and identifies the professional fields that need to be supplemented, and dynamically supplements new expert agents based on the specialties of the selected expert library. Specifically, it includes generating system prompts and assigning knowledge base permissions; the verification agent is responsible for detecting and calibrating the professional capabilities of the new expert agents to ensure that the newly generated experts have sufficient professionalism.

[0170] Specifically, it includes the following steps:

[0171] S221: The meta-agent compares the project feature list with the professional coverage scope of the selected core experts, identifies the uncovered professional fields, and generates a supplementary professional requirement list;

[0172] The professional coverage calculation formula is as follows:

[0173]

[0174] Among them, C represents the professional set of core experts, which is preset by the system, R represents the professional set of project requirements, which is obtained when preprocessing with GPT4o, |C∩R| represents the size of the intersection of the two sets, |R| represents the size of the project requirement set, which is judged by the meta-agent. If the Coverage Score < 0.8, then activate the supplement of new experts.

[0175] S222: According to the technical characteristics of the uncovered professional fields, the meta-agent dynamically generates system prompts for the corresponding experts using predefined expert role templates. The prompts include the following structure:

[0176] xpert_Prompt = {Role_Definition + Professional_Background +

[0177] Knowledge_Scope+Decision_Rules}

[0178] For example: {

[0179] "Role_Definition":"Energy Optimization Expert",

[0180] "Professional_Background":"Expert in renewable energy systems and passive heating technologies",

[0181] "Knowledge_Scope":"Thermal insulation materials, energy simulation, and building design",

[0182] "Decision_Rules":"Prioritize energy efficiency and occupant comfort while minimizing costs"

[0183] }

[0184] Among them, Role_Definition defines the expert identity, Professional_Background sets the professional background, Knowledge_Scope defines the knowledge scope, and Decision_Rules stipulates the decision rules;

[0185] S223: Based on the generated system prompt words, dynamically instantiate a new expert agent by calling a large language model and assign decision weights (judged by the meta-agent). The new agent is implemented in the way of encapsulating classes, supporting accepting task inputs and generating responses that conform to professional characteristics. After instantiation, the verification agent conducts role verification on it to ensure that its responses conform to the expected professional characteristics and decision-making behaviors.

[0186] Among them, the role verification specifically includes: the verification agent evaluates the decision-making behavior of the expert by setting targeted questions, including the following process: setting targeted questions based on the new expert system prompt words and evaluating the professionalism, logic, and decision-making consistency of their answers. When the total score reaches a preset threshold (such as 80 points), the verification passes and it enters the multi-agent framework. If the threshold is not reached, the communication meta-agent optimizes the system prompt words and verifies again until it is qualified.

[0187] S224: Assign the access rights of the corresponding knowledge base to the newly generated expert agent that has passed the verification, and integrate the newly generated expert agent with the core expert agent to form a complete expert decision-making team.

[0188] The knowledge distribution and collaboration mechanism specifically includes the following:

[0189] 1. Classify and label the content of the knowledge base according to professional fields, and establish a knowledge domain label system, including main fields such as architectural design, structural transformation, historical protection, economic evaluation, etc., as well as their respective sub-fields; use the LLM method to extract domain labels from documents.

[0190] Supplement: Dynamic expansion mechanism: When new content is added to the knowledge base, trigger the dynamic expansion mechanism and call LLM (such as GPT4o) to automatically extract labels.

[0191] 2. First, calculate through OpenAI Embeddings to convert the role descriptions of the agents and the knowledge base labels into vectors. Then calculate the Role_Match value through the cosine similarity formula

[0192]

[0193] Among them, V1 represents the query vector, V1·V2 represents the dot product of the two vectors, and ||V1|| and ||V2|| represent the Euclidean norms of the two vectors respectively;

[0194] Next, calculate the correlation degree between the expert agent and each knowledge base. The correlation degree calculation formula is:

[0195] Domain Relevance = μ·Role Match+(1 - μ)·Usage Score

[0196] Among them, Role_Match is the role matching degree, Usage_Score is the usage frequency, which is defaulted to 0, and μ is the balance factor set to 0.7; generate an expert-knowledge base mapping matrix to guide subsequent permission allocation;

[0197] 3. Implement dynamic authorization for knowledge access according to the correlation degree. The access permission calculation formula:

[0198]

[0199] Among them, Domain Relevance: The matching degree with the knowledge base label. Performance Score: Calculated according to the contribution degree of the agent. Usage Frequency: The access times of the knowledge base. The weights w1, w2, w3 can be adjusted through experiments, and the default values can be set to 0.5, 0.3, 0.2

[0200] The Access_Level values are {3, 2, 1, 0}, corresponding to full access, read-only access, restricted access, and no access rights respectively.

[0201] Conflict handling: When Agent A has full access rights and Agent B only has restricted access rights, first sort according to the level of access rights. For example, if A has full access rights and B has read-only access rights, then the priority of A is higher than that of B. Secondly, when the priorities are the same, sort according to the decision weights.

[0202] 4. Permission Initialization and Optimization

[0203] For existing core expert agents, grant full access rights to the corresponding knowledge bases;

[0204] For newly generated expert agents, allocate initial access rights based on their role definitions. At the same time, record the usage frequency and decision contributions of the knowledge bases through system logs, and dynamically optimize the permissions. For example, knowledge bases with high-frequency access will be given higher permissions to the using experts. Example of the record format:

[0205] {

[0206] "agent_id":"Expert_123",

[0207] "knowledge_base":"Thermal Insulation",

[0208] "access_count":15,

[0209] "decision_contribution":8

[0210] }

[0211] Among them, the organizational structure of the knowledge base specifically includes:

[0212] 1: Hierarchical storage structure, including:

[0213] a: Vector data layer, using OpenAI Embeddings to embed vector representations of knowledge content, and using FAISS index to store vector representations of various text documents

[0214] b: Raw data layer, using MongoDB to store professional knowledge in fields such as architectural design, structural transformation, historical protection, and economic evaluation according to professional fields;

[0215] c: Relational data layer, based on Neo4j to record knowledge associations, case associations, and decision experience paths between professional fields;

[0216] 2: Index management mechanism, including:

[0217] a: Construct a multi-level index structure, including a professional field index, a knowledge type index, and a keyword index;

[0218] b: Maintain a vector retrieval index, including regularly updating the FAISS index

[0219] 3: Access control mechanism, including access control based on professional fields, level setting of knowledge importance, and cache management of usage frequency.

[0220] S3: Based on the expert team configuration plan, execute a multi-round hierarchical collaborative decision-making process (the process is as Figure 5 shown):

[0221] S31: The first stage: Independent analysis

[0222] Each expert agent obtains domain-specific professional knowledge from the knowledge base based on retrieval-augmented generation technology, conducts professional field analysis, and puts forward preliminary analysis.

[0223] Specifically, each expert agent executes the independent analysis in the first stage, respectively obtains the professional knowledge of its own field from the knowledge base based on retrieval-augmented generation technology, conducts diagnostic analysis in combination with project characteristics, identifies key problems, and puts forward preliminary suggestions, forming a professional report including current situation assessment, problem analysis, and improvement ideas.

[0224] The implementation steps are as follows:

[0225] S311: Each expert agent performs retrieval-augmented generation based on its professional knowledge base, specifically including: converting the project feature list into a retrieval query vector, and retrieving relevant cases and professional knowledge in the knowledge base through vector similarity calculation. The similarity calculation formula is:

[0226]

[0227] Among them, Q represents the query vector, D represents the document vector in the knowledge base, Q·D represents the dot product of the two vectors, and ||Q|| and ||D|| respectively represent the Euclidean norms of the two vectors;

[0228] S312: Sort the retrieval results by relevance, select the top k documents with the highest similarity scores, and embed them as context information into the expert input prompt words; The prompt word template for each expert is:

[0229] Expert_C={Role_Description+

[0230] Retrieved_Knowledge+

[0231] Project_Features+

[0232] Analysis_Task[+

[0233] Constraints]

[0234] }

[0235] Among them, Role_Description is the expert role description, Retrieved_Knowledge is the retrieved knowledge content, Project_Features is the project features, Analysis_Task is the analysis task requirements; Constraints is the constraint conditions;

[0236] S313: Each expert agent generates a project analysis report from its own professional perspective based on the prompt words embedded with the search knowledge, including status assessment, problem diagnosis and improvement suggestions;

[0237] S314: The host agent summarizes the analysis reports of the experts and forms a preliminary plan for the first round. The plan includes specific transformation strategies and implementation suggestions for each professional field.

[0238] S32: Phase 2: Cross-supplementation and discussion

[0239] The host agent guides each expert agent to review and supplement the analysis of other experts in the first phase, and discuss key issues. The evaluation agent evaluates the depth and professionalism of the expert's speech in real time, and the recorder agent tracks the key points of the discussion and the topics of disagreement.

[0240] Specifically, the host intelligent agent organized the third phase of solution optimization. Based on the key points, problems and differences recorded in the second round of discussions, it summarized and refined the core issues, organized experts to demonstrate the feasibility of various solutions from a professional perspective, resolved differences through secondary demonstration and expert voting, and reached a final consensus.

[0241] The specific implementation steps are:

[0242] S321: The second stage begins, and the host agent organizes cross-discussion. Specifically, first, a professional dependency matrix M is constructed based on the results of the first round, where M ij Indicates the degree of dependence of major i on major j. The evaluation agent scores the results of the first stage, and the value range is [0,1];

[0243] Calculate the professional relevance based on the dependency matrix:

[0244]

[0245] Design the discussion order from high to low according to Relevance_Score; the host agent designs guiding questions based on the first-round solutions of highly relevant expert agents, driving experts to conduct in-depth criticism and professional supplementation of the analysis reports of other experts.

[0246] S322: The evaluation agent gives an instant score to each speech of each expert in each round:

[0247] Expert_core = w1·Technical_Depth + w2·Cross_Domain_Analysis + w3·Discussion_Completeness + w4·Interaction_Effect

[0248] Among them, Technical_Depth is the score of discussion depth, Cross_Domain_Analysis is the score of cross-domain analysis, Discussion Completeness is the score of discussion integrity, Interaction_Efect is the score of interaction effect, and w1 to w4 are weight coefficients, which are set during system initialization. And it satisfies Σwi = 1; the scores of each dimension are normalized to the [0,1] interval, and a threshold is set.

[0249] The evaluation agent calculates the overall quality of the round based on the speech scores of all experts in the current round:

[0250]

[0251] where N is the number of experts participating in the discussion

[0252] When any one of the following three conditions is met, the evaluation agent notifies the host agent to converge.

[0253] 1. When Discussion_Quality > 0.8

[0254] 2. The change of Discussion_Quality for 3 consecutive rounds < the set 0.1

[0255] 3. Exceed the maximum number of discussion rounds (10 rounds)

[0256] If the convergence condition is not met, the evaluation agent passes the speech scores of experts below the threshold in each round to the host agent before the start of the next round. The host agent adjusts the discussion based on the following strategy:

[0257] 1. Low - dimensional guidance: If Technical_Depth is low, guide the corresponding experts to increase the depth of thinking. If Cross_Domain_Analysis is low, design cross - professional related questions. If Interaction_Effect is low, guide the experts to enhance interactivity.

[0258] 2. Dimension priority adjustment: Dynamically adjust the key weight of the next round of discussion according to the scores, and give priority to weak links.

[0259] 3. Feedback notification: Before starting the next round of discussion, the host agent notifies the experts of the current weak dimension and the proposed optimization direction.

[0260] S323: Recording and sharing mechanism of the recorder agent: The recorder agent tracks and records the key points of each round of discussion, stores them in a structured data format, and the recorded content includes:

[0261] 1. Timing information (discussion time and discussion stage);

[0262] 2. Issue number (Issue ID and professional field);

[0263] 3. Record of experts' opinions (including expert ID, object of evaluation, professional divergence points and their argument supports);

[0264] 4. Related questions (including question description and involved professions);

[0265] 5. Consensus progress (including key points of consensus reached, problems to be solved, and professional opinions to be supplemented),

[0266] When the host agent guides each expert to speak, the communication recorder agent synchronously displays: 1. The focus of the current discussion topic; 2. The experts' opinions collected in this round; 3. The current divergence points.

[0267] When experts are speaking, they can directly quote the opinions of other experts provided by the recorder for supplementation or optimization. When the following situations occur, the recorder agent updates the displayed content: 1. After a new expert speaks; 2. Reaching a stage consensus; 3. Appearing new divergences

[0268] For topics with divergences, a multi - version recording mechanism is used to track the evolution process of viewpoints. When the divergence reaches a consensus, the recorder agent performs version merging, retains the final version of the content that reaches a consensus, and the rest as historical versions.

[0269] S324: Tasks after convergence: The recorder agent generates a summary report of the second - stage discussion, and the report structure includes:

[0270] 1. List of completed issues: The status of each issue and its final conclusion.

[0271] 2. List of Outstanding Issues: Classification of problem types, responsible experts, and suggestions for follow-up handling.

[0272] 3. Record of Discussion Progress: Record the changing trends of Discussion_Quality and each indicator.

[0273] 4. Update the Knowledge Base: Store the discussion content as structured data for subsequent use.

[0274] S33: The Third Stage: Disagreement Resolution and Solution Optimization

[0275] S331: The host agent extracts key points, disagreements, and the list of outstanding issues based on the summary report of the second stage.

[0276] S332: The host agent first guides each expert to supplement and optimize the consensus solution. Secondly, for the issues with disagreements, the corresponding experts are organized to conduct a secondary argumentation from a professional perspective. The remaining experts score the sufficiency of evidence (40%), technical feasibility (30%), and implementation effect (30%) in three dimensions according to the argumentation process. According to the scoring results, the experts with low scores need to adjust or make concessions to the opinions of the experts with high scores to form a compromise solution to reach a consensus on the issues. Finally, the host agent summarizes the discussion results, organizes the disagreement resolution solutions, and outputs the optimized decision results. Specifically, the host agent generates a decision summary report, including the advantages and disadvantages analysis of each solution, expert opinion disagreements and consensus, and final suggestions. The recorder agent stores the key data, expert evaluation opinions, and solution optimization paths in the decision-making process in the knowledge base.

[0277] S34: The host agent generates a decision summary report, including the advantages and disadvantages analysis of each solution, expert opinion disagreements and consensus, and final suggestions. The recorder agent stores the key data, expert evaluation opinions, and solution optimization paths in the decision-making process in the knowledge base. Specifically, the following steps are included:

[0278] S341: The recorder agent systematically sorts out the decision-making process and generates a structured decision summary report. The report structure includes: solution summary, technical analysis, benefit evaluation, focus of disputes, and final suggestions;

[0279] S342: Record the solution optimization path diagram, record the evolution process from the initial solution to the final solution, including the key modification points, adjustment basis, and improvement effects in each round of decision-making, and at the same time mark the decision-making contributions of each expert at different stages;

[0280] S343: Extract the empirical knowledge in the decision-making process, including: solutions to key disagreement points, quantitative indicators of expert evaluation criteria, effective strategies for solution optimization, etc., and store them in a structured manner according to the preset knowledge template;

[0281] S344: The recorder updates the decision report and empirical knowledge to the system knowledge base, ignoring the content with a similarity higher than the threshold γ and storing the content lower than the threshold as new knowledge entries. The knowledge update specifically includes: triggered when the recorder agent submits the final decision report; first, suspending all access operations in the relevant knowledge domain; then, performing data preprocessing and text vectorization; updating the three-layer storage structure (vector layer, original layer, relationship layer); finally, synchronously updating the index system, resuming access to the knowledge domain, recording the update log, and triggering the permission reset of the relevant knowledge domain.

[0282] The standardized agent interaction mechanism of the present invention specifically includes:

[0283] 1. Communication message format specification

[0284]

[0285]

[0286] 2. Message priority management mechanism

[0287] The present invention establishes a five-level message priority system, which specifically includes:

[0288] The first-level priority: used to process system control commands, error exception reports, and the process control instructions of the host;

[0289] The second-level priority: used to process the immediate evaluation results of the evaluation agent, the consensus information of the recorder, and emergency professional opinions;

[0290] The third-level priority: used to process regular expert discussion messages, the guiding questions of the host, and the real-time record update of the recorder;

[0291] The fourth-level priority: used to process knowledge base query requests, historical record retrievals, and non-urgent clarification requests;

[0292] The fifth-level priority: used to process background analysis data, statistical information updates, and log records.

[0293] 3. Message flow control mechanism

[0294] The present invention adopts a message processing model based on a state machine, which specifically includes:

[0295] (1) Message state definition: The message processing process is divided into four states: waiting (WAITING), processing (PROCESSING), completed / failed (COMPLETED / FAILED), and confirmed (CONFIRMED);

[0296] (2) Message flow rules:

[0297] 1. Ensure the priority processing right of the host agent's message

[0298] 2. Require the expert agent's speech to be confirmed by the host

[0299] 3. Ensure the real-time feedback of the evaluation results

[0300] 4. Maintain the instant synchronization of record updates

[0301] (3) Basic processing principles:

[0302] 1. Implement a message confirmation mechanism

[0303] 2. Ensure the sequentiality of messages on the same topic

[0304] 3. Support the parallel processing ability of messages

[0305] 4. Process control mechanism

[0306] The process control of the present invention adopts a multi-level design:

[0307] (1) Stage control strategy:

[0308] - First-stage independent analysis control: Allow each expert agent to conduct independent analysis simultaneously without interference. Each expert agent can independently access the knowledge base and form a preliminary analysis report

[0309] - Second-stage discussion control: The host agent strictly controls the speaking order to ensure that each expert's analysis opinion is orderly examined and supplemented by other experts

[0310] - Third-stage consensus control: The host agent selects the experts with differences to speak in turn. For the topics with differences, version management is adopted to record the complete evolution process from the initial differences to the final consensus

[0311] (2) Processing thread division:

[0312] - Decision main thread: Responsible for promoting the three-stage decision-making process and executing the process control instructions of the host agent

[0313] - Evaluation thread: Responsible for the evaluation agent to conduct real-time scoring on the speaking quality of experts in each round of discussion and calculate Discussion_Quality

[0314] - Recording thread: Responsible for the recorder agent to record the discussion content in real time, track the consensus formation process, and execute version management

[0315] - Knowledge retrieval thread: Support the expert agent to conduct real-time knowledge retrieval during the discussion to provide professional support for the argumentation

[0316] 5. Exception handling mechanism

[0317] 1. If the message times out without confirmation, trigger re - sending.

[0318] 2. When there is a priority conflict, forcefully interrupt the low - priority task.

[0319] 3. When there is a resource conflict, use a task lock to avoid parallel conflicts.

[0320] In some embodiments, the present invention can be implemented in the following different ways:

[0321] Regarding the implementation method of the knowledge base management module:

[0322] The first implementation method adopts the current three - layer storage structure, including a vector data layer, an original data layer, and a relational data layer;

[0323] The second implementation method adopts a graph database storage structure, where:

[0324] Use nodes to store knowledge entities, including professional knowledge, cases, etc.;

[0325] Use edges to store the relationships between entities, including association degree, dependence degree, etc.;

[0326] Support complex knowledge association queries and path analysis based on the graph structure;

[0327] The third implementation method adopts a hybrid storage architecture, where:

[0328] Structured data is stored in a relational database;

[0329] Unstructured data is stored in a document - type database;

[0330] The knowledge graph is used to store the association relationships in the professional field.

[0331] Regarding the construction method of the expert agent:

[0332] The first implementation method calls closed - source large models / open - source large models such as GPT4 and claude to build expert roles; the second implementation method uses large models based on vertical domains to build experts, where:

[0333] a) Use models fine - tuned in professional fields such as architecture and planning;

[0334] b) Provide more professional domain - knowledge support;

[0335] The third implementation method adopts a hybrid expert system, where:

[0336] a) Combine a rule engine and a large language model;

[0337] b) The rules engine processes explicit professional specifications and standards;

[0338] c) The large language model processes open-ended questions.

[0339] Regarding the implementation method of the expert verification mechanism:

[0340] The first implementation method adopts the current score evaluation mechanism;

[0341] The second implementation method adopts an adversarial verification mechanism, where:

[0342] a) Set up a dedicated interrogation agent;

[0343] b) Verify the expert's ability through professional question challenges and debates;

[0344] c) Examine the accuracy and professional depth of the expert's answers;

[0345] The third implementation method adopts case practice verification, where:

[0346] a) Use historical cases for ability testing;

[0347] b) Evaluate the accuracy of the expert's analysis and decision-making;

[0348] c) Compare the expert's solution with the actual solution.

[0349] Regarding the implementation method of the collaborative decision-making process:

[0350] The first implementation method adopts the current three-round decision-making mechanism;

[0351] The second implementation method adopts a dynamic iterative decision-making mechanism, where:

[0352] a) Cancel the fixed round setting;

[0353] b) Dynamically adjust the number of discussion rounds according to the complexity of the problem;

[0354] c) Adopt a real-time convergence judgment mechanism;

[0355] The third implementation method adopts a grouped parallel decision-making mechanism, where:

[0356] a) Group the experts for parallel discussions;

[0357] b) Each group independently forms a preliminary solution;

[0358] c) Integrate the final solution through cross-group discussions.

[0359] Regarding the implementation method of the evaluation mechanism:

[0360] The first implementation uses the current multi-dimensional scoring mechanism;

[0361] The second implementation uses a case-based evaluation mechanism, where:

[0362] a) Establish a standard case library;

[0363] b) Evaluate the quality of the solution by comparing it with similar cases;

[0364] c) Consider the feedback on the practical effect;

[0365] The third implementation uses a hierarchical evaluation mechanism, where:

[0366] a) Professional evaluation: Evaluate the application of professional knowledge;

[0367] b) Innovation evaluation: Evaluate the innovation degree of the solution;

[0368] c) Feasibility evaluation: Evaluate the possibility of the solution being implemented.

[0369] Regarding the implementation method of the knowledge update mechanism:

[0370] The first implementation uses the current similarity-based update mechanism; The second implementation uses an incremental learning mechanism, where:

[0371] a) Continuously collect new knowledge in the decision-making process;

[0372] b) Regularly update the knowledge base;

[0373] c) Dynamically adjust the importance weights of knowledge;

[0374] The third implementation uses dynamic updates of the knowledge graph, where:

[0375] a) Real-time update of knowledge entities and relationships;

[0376] b) Automatically discover new knowledge associations;

[0377] c) Optimize the organization of the knowledge structure.

[0378] Regarding the overall implementation method of the system:

[0379] The first implementation uses the current large language model-based implementation; The second implementation uses a hybrid architecture implementation, where:

[0380] a) The rule engine processes the standard process;

[0381] b) The machine learning model processes data analysis;

[0382] c) The large language model processes natural language interaction;

[0383] d) The knowledge graph supports knowledge management;

[0384] The third implementation method is realized by adopting a microservices architecture, where:

[0385] a) Each functional module is independently deployed;

[0386] b) Services interact through APIs;

[0387] c) It supports flexible expansion and update.

[0388] Taking a historical block renovation project as an example

[0389] Step S1 specifically includes:

[0390] S11: Divide the original materials into two categories: text and pictures. Among them, the text includes project overview documents, planning documents, research reports, etc.; the pictures include on-site current situation photos and aerial photos.

[0391] S12: Unify the format of the text and extract key information (such as extracting project background, renovation objectives, surrounding facility data, etc.), and uniformly adjust the resolution of the pictures;

[0392] S13: Index and store the preprocessed materials to achieve unified management of text and pictures, and establish a hierarchical storage structure of "project number / material type / date";

[0393] Step S2 specifically includes:

[0394] S21: Invoke the large language model to comprehensively analyze the text and picture materials, obtain classification information such as project scale, number of buildings, technical difficulty, professional requirements, etc., identify site characteristics, surrounding environment and main problems, and finally generate a project feature description and a list of key requirements.

[0395] S22: Input the project feature description and the list of key technical requirements into the meta-agent. The meta-agent selects the 3 core expert agents with the highest matching degree from the preset expert library, including: historical building protection expert (matching degree 0.92), urban renewal planning expert (matching degree 0.88), and architectural design expert (matching degree 0.85);

[0396] S23: If the project requirements exceed the coverage of the existing experts, the meta-agent automatically generates new expert agents (such as economic evaluation, traffic planning, etc.) based on the uncovered professional fields, and through system prompt words, role verification and permission configuration methods, ensure that these new experts have sufficient professionalism and decision consistency.

[0397] Step S23 specifically includes:

[0398] S231: The meta-agent compares the project feature list with the professional coverage of the selected core experts, identifies the uncovered professional fields, generates a supplementary professional requirement list, and calculates the current professional coverage rate as 67%;

[0399] S232: According to the characteristics of the uncovered professional fields, use the preset expert role template to generate system prompt words

[0400] S233: Based on the generated system prompt words, create a new expert agent, verify the agent and conduct role verification on it, specifically including 3 typical case analysis questions and 2 professional judgment questions, evaluate problem recognition (score 0.88), solution generation (score 0.85), and answer depth (score 0.82). After 3 rounds of optimization, the total score reaches 85 points and successfully passes the verification;

[0401] S234: Assign corresponding knowledge base access permissions and initial weights to the newly generated expert agent that has passed the verification, and integrate 4 newly generated expert agents (economic evaluation, transportation planning, landscape design, social research) with 3 core experts to form a complete expert decision-making team;

[0402] The knowledge distribution and collaboration mechanism of S234 specifically includes:

[0403] Classify and label the content of the knowledge base by professional fields, establish a knowledge domain label system, including main fields such as architectural design, structural transformation, historical protection, economic evaluation, etc., and their respective sub-fields; calculate the correlation degree between the expert agent and each knowledge domain. The correlation degree calculation formula is:

[0404] Domain Relevance = μ·RoleMatch+(1-μ)·Usage Score

[0405] Among them, Role_Match is the role matching degree determined by evaluating the agent (the economic evaluation expert has a 0.95 for the commercial real estate knowledge domain), Usage_Score is the historical usage effect score (initially set to 0.85), and μ is the balance factor set to 0.7; generate an expert-knowledge domain mapping matrix to guide subsequent permission allocation;

[0406] Domain Relevance = 0.7·0.95+(1-0.7)·0.85 = 0.92

[0407] Then implement dynamic authorization for knowledge access:

[0408]

[0409] Taking an economic evaluation expert as an example, after calculation, their access level to the commercial real estate knowledge domain is 2 (read-only access); for the newly generated expert agent, based on the professional background defined in its system prompt, the initial access permissions are automatically configured; perform knowledge access optimization: record the use of the expert knowledge base, including access frequency, retrieval effectiveness, and decision-making contributions;

[0410] Step S31 specifically includes:

[0411] S311: Each expert agent performs retrieval enhancement generation based on its professional knowledge base. Taking the historical building protection expert as an example, convert the list of 52 historical building features into a retrieval query vector, and retrieve relevant cases and professional knowledge in the knowledge base through vector similarity calculation. The similarity calculation formula is shown above. The system filters out 15 cases with a similarity greater than 0.7, among which the similarity with the Fuzhou Three Lanes and Seven Alleys renovation project is the highest, reaching 0.86;

[0412] S312: Sort the retrieval results by relevance, select the top 3 cases with the highest similarity scores, and embed them as context information into the expert input prompt.

[0413] S313: Each expert agent generates a project analysis report from its own professional perspective based on the prompt embedded with retrieval knowledge. The report of the historical protection expert includes content such as a building value assessment matrix, a hierarchical protection strategy (12 buildings with strict protection, 28 buildings with moderate renovation, and 12 buildings with functional transformation), and adaptive renovation suggestions;

[0414] S314: The host agent summarizes the analysis reports of the 7 experts to form a preliminary plan for the first round. The plan content includes 52 specific renovation strategies and 86 implementation suggestions in each professional field;

[0415] Step S32 specifically includes:

[0416] S321: At the beginning of the second stage, the host agent organizes cross-discussions. For example, the dependence degree between historical protection and architectural design is 0.85, and the dependence degree between commercial evaluation and architectural design is 0.78; calculate Relevance_Score according to the dependence relationship matrix: and design the discussion order from high to low; the host agent designs guiding questions based on the first-round plan of the highly relevant expert agent, such as "Please ask the historical protection expert to evaluate whether the protection measures in the architectural design plan are sufficient", to drive the experts to conduct in-depth criticism and professional supplementation;

[0417] S322: The evaluation agent instantly scores the speeches of each expert in each round: Taking the third round of speeches by historical preservation experts as an example, Technical_Depth = 0.92, Cross_Domain_Analysis = 0.88, Discussion_Completeness = 0.85, Interaction_Effect = 0.82; the scores of each dimension are standardized to the interval [0,1];

[0418] Expert_core=0.4·0.92+0.3·0.88+0.2·0.85+0.82·0.1=0.884

[0419] 0.884>0.7

[0420] The expert's answer is satisfactory and no adjustment is required.

[0421] S323: Calculate the overall quality of the round based on the speech scores of all experts in the round. At the same time, the evaluation agent passes the score of each expert to the host agent. The host agent guides the experts to answer based on the score. After 8 rounds of discussion, Discussion_Quality reaches 0.83, reaching the second round convergence condition.

[0422] S324: The recorder agent tracks and records the key points of each round of discussion, and stores them in a structured data format. The record content includes: time series information; topic number; expert opinion record; related issues (type: balance between protection and renewal, involving two majors of historical protection and architectural design, and specific description of the impact of renovation depth); consensus progress (consensus has been reached: the protection method of 12 strictly protected buildings, and to be resolved: the renovation depth of 28 moderately updated buildings). For topics with differences, a multi-version recording mechanism is used to track the evolution of opinions. For example, the renovation depth is gradually adjusted from the initial 30% to 20%. When it is finally determined to be 12%-20%, the recorder agent performs version merging, retains the final version, and the rest are kept as historical versions;

[0423] S325: Post-convergence task: The recorder agent generates a discussion summary report. The report structure includes: 1. Completed topic list: 48 out of 52 buildings have been determined for protection and renewal plans; 2. Unfinished topic list: There are differences in the renovation plans of 4 buildings. It is recommended that architectural design experts focus on the discussion; 3. Discussion progress record: Discussion_Quality is improved from the initial 0.72 to 0.83, and the scores of each expert are stable above 0.8; 4. Update the knowledge base: store the decision-making experience of 48 successful cases in the knowledge base;

[0424] Step S33 specifically includes:

[0425] S331: The host agent extracts 3 core disagreement points (depth of historical building renovation, commercial format positioning, traffic organization plan), 12 consensus points (such as principles for controlling building style), and 4 unfinished issues based on the discussion records of the second stage.

[0426] S332: For the issues with disagreements, the host agent organizes a second-round argumentation by experts. For example, regarding the disagreement on the depth of historical building renovation (the conservation experts advocate <15%, and the commercial experts suggest >30%), the two experts are required to provide their respective arguments (conservation case data vs economic feasibility analysis), and the other 5 experts rate from three dimensions: sufficiency of arguments (0.85), technical feasibility (0.78), and implementation effect (0.82). According to the comprehensive scores, the suggestion of the conservation experts is adopted, but a flexible clause is added, setting a renovation depth range of 12% - 20%.

[0427] Step S34 specifically includes:

[0428] S341: The recorder agent systematically sorts out the decision-making process and generates a structured decision summary report. The report structure includes: plan summary (the "one building, one strategy" renovation plan for 52 historical buildings), technical analysis (conservation technical route, renovation process selection, material application suggestions), benefit evaluation (social, economic, and environmental benefit analysis), focus of disputes (the disagreement process on renovation depth, format selection, and traffic organization), and final suggestions (zoned and graded renovation strategies).

[0429] S342: Record the plan optimization path diagram, recording the evolution process from the initial plan to the final plan, including: 7 professional suggestions formed by each expert's independent analysis in the first round, the adjustment basis and improvement effects of 48 key modification points in 8 cross-discussions in the second round, and at the same time mark the decision-making contribution degrees of 7 experts in 3 stages.

[0430] S343: Extract the key experience points in the decision-making process, including: solutions to key disagreement points, quantitative indicators for expert evaluation criteria, and effective strategies for plan optimization. Structurally store them according to the preset knowledge template.

[0431] S344: The recorder updates the decision report and experience knowledge to the system knowledge base. Among them, for the 3 experience points with a similarity higher than the threshold γ (set to 0.8) with the existing content, they are merged and optimized, and the remaining 5 points are stored as new knowledge entries.

[0432] In some embodiments, the S344 knowledge update trigger mechanism specifically includes: being triggered when the recorder agent submits the final decision report; first suspending all access operations in the relevant knowledge domain; then performing data preprocessing and text vectorization on the key points of experience; updating the three-layer storage structure; finally synchronously updating the FAISS index system, restoring access to the knowledge domain, recording the update log, and triggering the permission reset of the relevant knowledge domain.

[0433] The above content is a further detailed description of the present invention in combination with specific / preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several alternatives or modifications can be made to these described embodiments, and these alternative or modified forms should all be regarded as belonging to the protection scope of the present invention. In the description of this specification, the description with reference to terms such as "an embodiment", "some embodiments", "preferred embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples. Although the embodiments of the present invention and their advantages have been described in detail, it should be understood that various changes, substitutions, and alterations can be made herein without departing from the scope defined by the appended claims.

Claims

1. An intelligent expert system for urban renewal based on a large language model, characterized in that: include: Knowledge base management module: uses retrieval enhancement generation technology (RAG) to support efficient retrieval and application of professional field knowledge; The knowledge base management module is equipped with a knowledge indexing unit and a rights management unit; the knowledge indexing unit maintains the FAISS index and multi-level directory structure and supports vectorized retrieval; the rights management unit dynamically allocates and manages knowledge access rights based on expert role characteristics; Expert agent group: including user-defined core experts and system dynamic supplementary experts; Expert team collaborative decision-making module: executes multiple rounds of hierarchical collaborative decision-making processes, covering independent analysis by experts, cross-discussion and solution optimization; Meta-agent module: responsible for project feature analysis and dynamic configuration of expert teams, and dynamic creation of expert agents; Auxiliary agent module: including moderator agent: responsible for process control, topic guidance and disagreement integration, and guiding the decision-making process; recorder agent: real-time recording of discussion content and generation of decision summary report; Verification agent: responsible for verifying and evaluating the professional capabilities of the expert agent; Scoring agent: responsible for real-time scoring of discussion quality, dynamic optimization of weights, and evaluation of decision quality.

2. A decision-making method for urban renewal projects using an urban renewal intelligent expert system based on a large language model, characterized in that: The following steps are involved: S1: Obtain basic project information Obtain basic information on urban renewal projects, including multimodal data such as project text descriptions and site status photos; urban renewal project basic information is divided into two categories: text information and picture information. Text information includes project overview, land use status, surrounding environment, development demands, etc.; picture information includes site status photos, surrounding environment photos, etc.; Standardize the classified data, including: unify the format of text data, extract key information; adjust the resolution of image data; Use the large language model to conduct multimodal analysis on the standardized project data, analyze the project type, scale, technical difficulty, professional requirements and other classified information in the text data, and identify the site characteristics, surrounding environment and existing problems in the picture data; Index and store the standardized data to achieve unified management of text and images for easy subsequent input; S2: Analyze project characteristics and build an expert team The meta-agent performs feature analysis on the project data and builds an expert team based on the analysis results, including selecting highly matched experts from the preset core expert database and dynamically generating system prompts for new experts based on the project type to create expert agents in specific fields; the project feature description and key technical requirements list are input to the meta-agent, which selects the core expert agent with the highest match from the preset expert database; The verification agent verifies the professionalism and evaluates the capabilities of the generated experts, and dynamically allocates knowledge base permissions according to their roles to form a complete expert decision-making team; based on the characteristics of the project, the meta-agent, based on the semantic parsing capabilities of the large language model itself, independently determines and identifies the professional fields that need to be supplemented, and dynamically supplements new expert agents based on the expertise of the selected expert database; specifically, it generates system prompt words and allocates knowledge base permissions; the verification agent is responsible for testing and calibrating the professional capabilities of the new expert agents to ensure that the newly generated experts are sufficiently professional. S3: Based on the expert team configuration plan, multiple rounds of hierarchical collaborative decision-making are performed. The process is as follows: Phase 1: Independent Analysis Each expert agent acquires domain expertise from the knowledge base based on retrieval-enhanced generation technology, performs professional domain analysis, and proposes preliminary analysis; Phase 2: Cross-supplementation and discussion The host agent guides each expert agent to review and supplement the analysis of other experts in the first phase, and discuss key issues. The evaluation agent evaluates the depth and professionalism of the expert's speech in real time, and the recorder agent tracks the key points of the discussion and the topics of disagreement; Phase 3: Disagreement resolution and solution optimization Based on the records of the second stage, the host agent extracts the existing controversial issues and organizes relevant experts to conduct a secondary argumentation from a professional perspective; the remaining experts score the argumentation process based on three dimensions: sufficiency of arguments (40%), technical feasibility (30%) and implementation effect (30%); based on the scoring results, low-scoring experts need to adjust or make concessions to the opinions of high-scoring experts in order to reach a consensus on the topic; finally, the host agent summarizes the discussion results, organizes solutions to controversial issues, and outputs optimized decision results.

3. According to claim 2, a decision-making method for urban renewal projects using an urban renewal intelligent expert system based on a large language model is characterized in that: S2 also includes the following: S21: The meta-agent compares the project feature list with the professional coverage of the selected core experts, identifies uncovered professional areas, and generates a list of supplementary professional requirements; The calculation formula for professional coverage is: Among them, C represents the professional set of core experts, which is preset by the system, R represents the professional set of project requirements, which is obtained when using GPT4o preprocessing, |C∩R| represents the size of the intersection of the two sets, and |R| represents the size of the project requirement set, which is judged by the meta-agent. If CoverageScore < 0.8, new experts are activated and supplemented; S22: Based on the technical characteristics of the uncovered professional field, the meta-agent uses the predefined expert role template to dynamically generate the system prompt words for the corresponding experts. The prompt words contain the following structure: Expert_Prompt={Role_Definition+Professional_Background+Knowledge_Scope+Decision_Rules} Among them, Role_Definition defines the expert identity, Professional_Background sets the professional background, now_Scope defines the knowledge scope, and Decision_Rules specifies the decision rules; S23: Based on the generated system prompt words, a new expert agent is dynamically instantiated by calling the large language model and assigning decision weights (determined by the meta-agent); the new agent is implemented by encapsulation class, supports accepting task input and generating responses that meet professional characteristics; after instantiation, the verification agent verifies its role to ensure that its response meets the expected professional characteristics and decision-making behavior; The role verification specifically includes: asking questions based on the new expert system prompts to evaluate the professionalism, logic and decision consistency of the answers; when the total score reaches the preset threshold, the verification is passed; if the threshold is not reached, the communication meta-agent optimizes the system prompts and re-verifies until it passes; S24: Assign corresponding knowledge base access rights to the newly generated expert agents that have passed the verification, and integrate the newly generated expert agents with the core expert agents to form a complete expert decision-making team.

4. A decision-making method for urban renewal projects using an urban renewal intelligent expert system based on a large language model according to claim 3, characterized in that: The knowledge distribution collaboration mechanism in S24 specifically includes: S241: Classify and annotate the knowledge base content according to professional fields, and establish a knowledge domain label system, including main fields such as architectural design, structural renovation, historical protection, economic evaluation, and their respective sub-fields; use the LLM method to extract field labels from documents; when new content is added to the knowledge base, the dynamic expansion mechanism is triggered, and LLM (such as GPT4o) is called to automatically extract labels; S242. First, convert the agent's role description and knowledge base label into a vector using OpenAI Embeddings; then calculate the Role_Match value using the cosine similarity formula Where V1 represents the query vector, V1·V2 represents the dot product of two vectors, ||V1|| and ||V2|| represent the Euclidean norms of the two vectors respectively; Next, the correlation between the expert agent and each knowledge base is calculated. The correlation calculation formula is: Domain Relevance=μ·Role Match+(1-μ)·Usage Score Among them, Role_Match is the role matching degree, Usage_Score is the usage frequency, which defaults to 0, and μ is the balance factor, which is set to 0.7; an expert-knowledge base mapping matrix is ​​generated to guide subsequent permission allocation; S243. Implement dynamic authorization of knowledge access based on relevance, and the formula for calculating access rights is: Where Domain Relevance: the degree of match with the knowledge base label; Performance Score: calculated based on the contribution of the agent; Usage Frequency: the number of visits to the knowledge base; weights w1, w2, w3 can be adjusted through experiments, and the default values ​​can be set to 0.5, 0.3, 0.2; Access_Level values ​​are {3, 2, 1, 0}, corresponding to full access, read-only access, restricted access, and no access rights, respectively; Conflict handling: When agent A has full access rights and agent B has only limited access rights, first sort them according to the level of access rights. For example, if A has full access rights and B has read-only access rights, then A has a higher priority than B. Secondly, when the priorities are the same, sort them according to the decision weight. S244. Permission initialization and optimization For existing core expert agents, full access rights to the corresponding knowledge base are granted; Initial access permissions are assigned to newly generated expert agents based on their role definitions. The frequency of use of the knowledge base and decision-making contributions are recorded through system logs to dynamically optimize permissions. For example, a knowledge base that is accessed frequently will be granted higher permissions to use as an expert.

5. The decision-making method for urban renewal projects using an urban renewal intelligent expert system based on a large language model according to claim 2 is characterized in that: The organizational structure of the knowledge base specifically includes:

1. Hierarchical storage structure: a: Vector data layer, using OpenAI Embeddings to embed vectors to represent knowledge content, and using FAISS index to store vector representations of various text documents b: The original data layer uses MongoDB to store professional knowledge in areas such as architectural design, structural renovation, historical preservation, and economic assessment by professional field; c: Relational data layer, based on Neo4j, records the knowledge associations, case associations and decision-making experience paths between professional fields; 2. Index management mechanism: a: Build a multi-level index structure, including professional field index, knowledge type index and keyword index; b: Maintain the vector retrieval index, including regularly updating the FAISS index; 3. Access control mechanism: including access control based on professional fields, level setting of knowledge importance and cache management of usage frequency.

6. According to the method for making decisions on urban renewal projects using an urban renewal intelligent expert system based on a large language model as claimed in claim 2, it is characterized in that: Step S3 specifically includes: S31: Each expert agent performs the first phase of independent analysis, obtains professional knowledge in its own field from the knowledge base based on retrieval enhancement generation technology, conducts diagnostic analysis based on project characteristics, identifies key issues and makes preliminary suggestions, and forms a professional report that includes status assessment, problem analysis and improvement ideas; S32: Each expert agent performs the second stage of cross-discussion. The host agent guides the expert agent to review and supplement the analysis of other experts. The evaluation agent evaluates the quality of each expert's speech and the overall quality of the round. The recorder agent records the key points of the discussion, key differences and consensus. S33: The host agent organizes the third phase of solution optimization. Based on the key points, problems and differences recorded in the second round of discussion, it summarizes and refines the core issues, organizes experts to demonstrate the feasibility of various solutions from a professional perspective, resolves differences through secondary demonstration and expert voting, and reaches a final consensus; S34: The host agent generates a decision summary report, which includes an analysis of the advantages and disadvantages of each plan, expert opinions and consensus, and final recommendations; the recorder agent stores the key data of the decision-making process, expert evaluation opinions, and plan optimization paths into the knowledge base.

7. According to the method for making decisions on urban renewal projects using an urban renewal intelligent expert system based on a large language model as claimed in claim 6, it is characterized in that: Step S31 specifically includes: S311: Each expert agent performs retrieval enhancement generation based on its professional knowledge base, specifically including: converting the project feature list into a retrieval query vector, and retrieving relevant cases and professional knowledge in the knowledge base through vector similarity calculation. The similarity calculation formula is: Where Q represents the query vector, D represents the document vector in the knowledge base, Q·D represents the dot product of the two vectors, and ||Q|| and ||D|| represent the Euclidean norms of the two vectors respectively; S312: Sort the search results by relevance, select the first k documents with the highest similarity scores, and embed them into the expert input prompt words as context information; the prompt word template of each expert is: Expert_C={Role_Description+Retrieved_Knowledge+ Project_Features+Analysis_Task[+Constraints]}, where Role_Description is the expert role description, Retrieved_Knowledge is the retrieved knowledge content, Project_Features is the project features, Analysis_Task is the analysis task requirements; Constraints is the constraint conditions; S313: Each expert agent generates a project analysis report from its own professional perspective based on the prompt words embedded with the search knowledge, including status assessment, problem diagnosis and improvement suggestions; S314: The host agent summarizes the analysis reports of the experts and forms a preliminary plan for the first round. The plan includes specific transformation strategies and implementation suggestions for each professional field.

8. According to the method for making decisions on urban renewal projects using an urban renewal intelligent expert system based on a large language model as claimed in claim 6, it is characterized in that: Step S32 specifically includes: S321: The second stage begins, and the host agent organizes cross-discussion. Specifically, first, a professional dependency matrix M is constructed based on the results of the first round, where M ij Indicates the degree of dependence of major i on major j. The evaluation agent scores the results of the first stage, and the value range is [0,1]; Calculate the professional relevance based on the dependency matrix: Design the discussion order from high to low Relevance_Score; the host agent designs guiding questions based on the first round of proposals of the highly relevant expert agents, leading the experts to conduct in-depth criticism and professional supplements on the analysis reports of other experts; S322: The evaluation agent instantly scores each speech of each expert in each round: Expert_core=w1·Technical_Depth+w2·Cross_Domain_Analysis+w3·Discussion_Completeness+w4·Interaction_Effect Among them, Technical_Depth is the discussion depth score, Cross_Domain_Analysis is the cross-domain analysis score, Discussion Completeness is the discussion completeness score, Interaction_Efect is the interaction effect score, w1 to w4 are weight coefficients, which are set when the system is initialized and satisfy Σwi=1; the scores of each dimension are standardized to the [0,1] interval, and the threshold is set; The evaluation agent calculates the overall quality of the round based on the speech scores of all experts in the round: Where N is the number of experts participating in the discussion. When any of the following three conditions is met, the evaluation agent notifies the host agent to converge:

1. Discussion_Quality>0.8 2. Discussion_Quality changes for 3 consecutive rounds < setting 0.1 3. Exceeding the maximum number of discussion rounds (10 rounds) If the convergence condition is not met, the evaluation agent will pass the expert speech scores below the threshold in each round to the host agent before the next round begins; the host agent adjusts the discussion based on the following strategy: Guidance for low-scoring dimensions: If Technical_Depth is low, guide the corresponding experts to increase the depth of their thinking; if Cross_Domain_Analysis is low, design cross-disciplinary related questions; if Interaction_Efect is low, guide the experts to enhance interactivity; Dimension priority adjustment: dynamically adjust the focus weight of the next round of discussion based on the scores, giving priority to weak links; Feedback notification: Before starting the next round of discussion, the host agent informs the experts of the current weak dimensions and suggested optimization directions; S323: Recording and sharing mechanism of recorder agent: the recorder agent tracks and records the key points of each round of discussion, and stores them in a structured data format. The recorded content includes: time sequence information (discussion time and discussion stage); issue number (Issue ID and professional field); expert opinion record (including expert ID, review object, professional disagreement and its supporting arguments); related issues (including problem description and involved professions); consensus progress (including key points that have been reached, problems to be solved, and professional opinions that need to be supplemented); when the host agent guides each expert to speak, the communication recorder agent simultaneously displays the focus of the current discussion, the expert opinions collected in this round, and the current disagreement points; When an expert is speaking, he or she can directly quote other expert opinions provided by the recorder to supplement or optimize them. The recorder agent updates the displayed content when the following situations occur: after a new expert speaks, when a phased consensus is reached, or when new differences arise. For topics with differences, a multi-version recording mechanism is used to track the evolution of opinions. When a consensus is reached, the recorder agent performs version merging, retaining the final version of the content that has reached consensus, and the rest are kept as historical versions; S324: Post-convergence task: The recorder agent generates a summary report of the second phase of discussion. The report structure includes:

1. List of completed topics: the status of each topic and its final conclusion; 2. List of unfinished topics: classification of problem types, responsible experts and follow-up suggestions; 3. Discussion progress record: record the change trend of Discussion_Quality and each indicator; 4. Update the knowledge base: store the discussion content as structured data for subsequent use.

9. According to claim 6, a decision-making method for urban renewal projects using an urban renewal intelligent expert system based on a large language model is characterized in that: Step S33 specifically includes: S331: The moderator agent extracts the key points, disagreements, and a list of unfinished topics based on the summary report of the second stage; S332: The host agent first guides the experts to supplement and optimize the consensus plan. Secondly, for the issues with differences, it organizes the corresponding experts to conduct a secondary argument from a professional perspective, and guides the remaining expert agents to score the professionalism and effectiveness of their argumentation process. According to the scoring results, the low-scoring expert agents make concessions to the high-scoring expert agents to form a compromise plan and finally reach a consensus.

10. The decision-making method for urban renewal projects using an urban renewal intelligent expert system based on a large language model according to claim 6 is characterized in that: Step S34 specifically includes: S341: The recorder agent systematically sorts out the decision-making process and generates a structured decision summary report. The report structure includes: plan summary, technical analysis, benefit evaluation, controversial points and final recommendations; S342: Record the solution optimization path map, record the evolution process from the initial solution to the final solution, including the key modification points, adjustment basis and improvement effect in each round of decision-making, and mark the decision-making contribution of each expert at different stages; S343: Extract the experiential knowledge in the decision-making process, including solutions to key divergence points, quantitative indicators of expert evaluation criteria, effective strategies for solution optimization, etc., and store them in a structured manner according to the preset knowledge template; S344: The recorder updates the decision report and experience knowledge to the system knowledge base, wherein the content with similarity higher than the threshold γ is ignored, and the content lower than the threshold is stored as a new knowledge entry.

Images