A schema-driven green building assessment method and system

Abstract

The application relates to a Schema-driven green building evaluation method and system, and a core is an autonomous evaluation intelligent agent which takes an offline constructed Schema as an action instruction. In an evaluation process, the intelligent agent calls a knowledge base retrieval tool which internally executes a two-stage mixed retrieval strategy of coarse screening and fine selection according to the requirements of the Schema to accurately locate evidence, and finally guides a language model to complete constraint information extraction and report generation. Prototype system verification shows that the method effectively improves the automation degree, accuracy and traceability of report generation, and provides a low-cost and highly controllable solution for constructing a man-machine cooperative intelligent evaluation workflow.

Description

Technical Field

[0001] This invention relates to the field of building evaluation technology, and in particular to a schema-driven green building evaluation method and system. Background Technology

[0002] Currently, the demand for green building assessment and certification is growing. However, the writing of traditional assessment reports relies heavily on expert experience and requires manually sifting through a massive amount of heterogeneous project documents (such as design drawings, construction reports, and testing reports). This process is time-consuming and labor-intensive, and the assessment conclusions are easily influenced by subjective factors, making it difficult to guarantee the quality and consistency of the report format. This hinders the effective accumulation and reuse of assessment knowledge.

[0003] In recent years, artificial intelligence technologies, represented by Large Language Models (LLM), have provided new approaches to solving the aforementioned challenges. Agentic Retrieval Augmented Generation (AAG), as a cutting-edge paradigm, demonstrates great potential for solving complex, multi-step tasks by endowing systems with autonomous planning, tool invocation, and reflective optimization capabilities. However, applying it to generate highly professional and structured green building assessment reports still faces three major challenges:

[0004] 1) The "cold start" problem: The high heterogeneity and one-time nature of project documents make traditional model training methods expensive and unsuitable;

[0005] 2) Output uncontrollability: Without strong constraints, the open generation mode of the intelligent agent is prone to "illusion" problems such as content "drift", inconsistent formats, and omission of key information, making it difficult to meet the strict standards of professional reports;

[0006] 3) The lack of traceability of evidence: the assessment conclusions must be strongly related to the original evidence, but standard retrieval augmented generation (RAG) makes it difficult to accurately trace the source of the generated content and evidence, which affects the credibility of the report.

[0007] In summary, there is currently a lack of a green building assessment method and system to solve or partially solve the aforementioned problems. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a schema-driven green building assessment method and system, so as to solve or partially solve the problems of high green building assessment costs due to strong heterogeneity of project documents, uncontrollable output due to lack of constraints, and untraceable evidence.

[0009] The objective of this invention can be achieved through the following technical solutions:

[0010] One aspect of the present invention provides a schema-driven green building assessment method, comprising the following steps:

[0011] The system acquires report template data and unstructured assessment requirement data, parses, segments, and injects metadata through a schema engine to obtain structured schema data, and achieves offline schema construction.

[0012] Acquire data on green building projects to be evaluated, and construct a project knowledge base through parsing and vectorization.

[0013] Using a pre-built online autonomous evaluation agent, the evaluation steps are planned based on the schema data as action instructions;

[0014] The online autonomous evaluation agent invokes the built-in retrieval tool and uses a two-stage retrieval strategy of coarse screening and fine screening to locate the evidence required for evaluation from the project knowledge base;

[0015] The online autonomous evaluation agent extracts structured information that conforms to the schema constraints based on the evidence obtained from localization and the schema data using a language model.

[0016] The online autonomous evaluation agent evaluates and scores the extracted structured information by combining the preset scoring criteria in the schema data.

[0017] The online autonomous assessment agent, based on the report template in the schema, renders the scoring results and structured information into a green building assessment report, thereby realizing green building assessment.

[0018] As a preferred technical solution, the schema data includes a fields array and a display template. The fields array defines the key technical parameters and evaluation indicators to be extracted, and supports data types including text, numbers, options, and tables. It also includes natural language instructions and return format requirements for the language model. The display template is used to define the rendering format of the evaluation report and includes preset rendering macros.

[0019] As a preferred technical solution, during the rendering of the green building assessment report, the structured data extracted from the fields array is filled into the report frame according to a preset format by calling the display template in the schema data, thereby generating a green building assessment report that associates the assessment conclusions with the located original evidence.

[0020] As a preferred technical solution, the online autonomous evaluation agent performs a retrieval-extraction-reflection loop to extract structured information that conforms to schema constraints:

[0021] If the extracted information is incomplete or missing, the online autonomous evaluation agent determines that supplementary evidence is needed, regenerates the query statement, and re-executes the coarse-fine two-stage retrieval until the extracted information is complete or the preset maximum number of retrieval attempts is reached.

[0022] As a preferred technical solution, the coarse screening-fine screening two-stage retrieval strategy is as follows:

[0023] Coarse screening stage: Approximate nearest neighbor search is performed on the vector data in the project knowledge base using the Embedding model;

[0024] Selection stage: The Reranker model is used to re-rank the results obtained from the coarse screening.

[0025] As a preferred technical solution, after evaluating and scoring the extracted structured information, the method further includes:

[0026] If insufficient evidence is detected for a scoring item, the item is marked as a low-confidence item, and the user is prompted to conduct a manual review.

[0027] As a preferred technical solution, the following are also included:

[0028] Synchronize the green building assessment report and related evidence to the interactive assessment workbench;

[0029] Experts use the interactive evaluation platform to review, revise, and supplement evidence in the draft report.

[0030] Record expert review and modification actions and evidence supplementation behavior, and iteratively optimize the prompt words of the schema and language model.

[0031] As a preferred technical solution, the interactive evaluation workbench adopts a three-column layout, including a document management area, an evaluation item area, and an evaluation details area.

[0032] As a preferred technical solution, a language model is used to transform report template data and unstructured evaluation requirement data into structured schema data and store it in a rule base. The schema engine and the online autonomous evaluation agent are decoupled through the rule base.

[0033] Another aspect of the present invention provides a schema-driven green building assessment method system for implementing the aforementioned green building assessment method, the system comprising:

[0034] The offline schema engine module is used to obtain report template data and unstructured assessment requirement data, and then process them through the schema engine for parsing, chunking, and metadata injection to obtain structured schema data.

[0035] The online autonomous assessment agent module is used to acquire data on green building projects to be assessed, construct a project knowledge base through parsing and vectorization, plan assessment steps using the schema data as action instructions, call the built-in retrieval tool, and locate the evidence required for assessment from the project knowledge base through a two-stage retrieval strategy of coarse screening and fine screening. Based on the located evidence and the schema data, the module uses a language model to extract structured information that conforms to schema constraints, and evaluates and scores the extracted structured information in combination with the preset scoring criteria in the schema data. Based on the report template in the schema, the scoring results and structured information are rendered into a green building assessment report, thus realizing the green building assessment.

[0036] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0037] (1) Achieving efficient, strongly constrained and traceable green building assessment: In response to the current assessment problems of high assessment costs, lack of constraints leading to uncontrollable output and untraceable evidence, the offline schema engine of this invention uses a language model to preprocess unstructured assessment requirement data and expert report templates into schema data. The online autonomous assessment agent uses schema as action instructions to call the integrated coarse screening-fine selection two-stage retrieval tool to locate evidence, guide the language model to complete the fill-in-the-blank constraint information extraction, rather than open generation. Through strong schema constraints, the illusion of language model is avoided and a strong correlation between assessment conclusion and original evidence is achieved.

[0038] (2) Accurate evidence location and complete information extraction: In view of the problem that existing generation methods are prone to difficulty in balancing evidence recall and precision, as well as incomplete extraction of key parameters, this invention adopts a two-stage hybrid retrieval strategy of coarse screening and fine screening. First, the Embedding model is used to perform an approximate nearest neighbor search on the vector data in the project knowledge base. Then, the Reranker model is used to re-rank the results obtained from the coarse screening, and the evidence most relevant to the evaluation target is selected to balance recall and precision. Core evidence is quickly located from heterogeneous project documents, avoiding interference from irrelevant information. In addition, the iterative method of retrieval-extraction-reflection is combined to solve the problem of information omission caused by insufficient initial evidence, and ensure that the extracted key technical parameters are accurate and complete.

[0039] (3) High efficiency of human-machine collaboration: In view of the problem that pure machine automatic evaluation is prone to bias and lacks continuous learning ability, this invention constructs a human-machine collaborative workflow. The machine automatically completes document parsing, knowledge base construction and report draft generation, and then synchronizes the report draft and related evidence to the interactive evaluation workbench. The system records the expert's review and modification operations and interaction behavior in the workbench, and uses this data for schema iteration optimization and language model prompt word improvement. Attached Figure Description

[0040] Figure 1 This is a flowchart of the schema-driven green building assessment method in the embodiment;

[0041] Figure 2 This is a schematic diagram illustrating the architecture and workflow of the green building intelligent assessment system in the embodiment;

[0042] Figure 3 This is a schematic diagram of the interactive evaluation workbench interface in the embodiment;

[0043] Figure 4 This is a schematic diagram of the schema-driven green building assessment system in the embodiment. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0045] Example 1

[0046] To address the issues of low efficiency, strong subjectivity, and poor consistency in traditional green building assessment reports, this embodiment provides a schema-driven green building assessment method. This method combines the autonomous actions of intelligent agents with rigorous schema planning, transforming uncontrollable open generation into controllable completion. While ensuring the accuracy and format consistency of the report content, it achieves a strong correlation between assessment conclusions and evidence. This research provides a new technological paradigm for solving the problem of automatic report generation in professional fields and has significant theoretical and practical value for improving the intelligence level of the green building assessment industry.

[0047] The core of this method is an autonomous evaluation agent that uses an offline-built schema as its action instructions. During the evaluation process, based on the schema's requirements, the agent invokes an internal knowledge base retrieval tool that employs a two-stage hybrid retrieval strategy of coarse screening and fine screening to accurately locate evidence, ultimately guiding the language model to complete the extraction of constraint information and report generation.

[0048] To achieve effective separation between knowledge modeling and online evaluation, this method employs a dual-engine architecture combining an offline schematization engine (corresponding to step S1) with an online autonomous evaluation agent (corresponding to steps S2-S7). The architecture and core workflow are as follows: Figure 2 As shown, the process begins with a user uploading a document and initiating an evaluation via the network. In the "Document Processing and Knowledge Base Construction" phase, the system uses MinerU to parse the document, vectorizes it using an Embedding model, and stores the processed data in a persistent layer (including Minio, PostgreSQL, and a vector database). When the evaluation task starts, the autonomous evaluation agent in the "Core Evaluation Module" receives instructions via "Schema Injection" and invokes a "Search Tool" (which integrates hybrid retrieval and Reranker ranking) to accurately locate evidence. Finally, the agent completes fill-in-the-blank extraction and intelligent scoring, and then submits the structured data to the "Automatic Report Rendering" module, which uses a Jinja2 template to generate the final report.

[0049] See Figure 1 The method includes the following steps:

[0050] Step S1: Obtain report template data and unstructured assessment requirement data, and process them to obtain structured schema data.

[0051] Preferably, report template data and unstructured assessment requirement data can be obtained through an application programming interface (API) on the network.

[0052] Specifically, this step is used for knowledge modeling. Before the assessment task begins, it uses a Large Language Model (LLM) to preprocess the unstructured clauses of the "Green Building Evaluation Standard" and expert report templates into structured, machine-readable schema data, and stores it in a rule base. Schema data is the core of this method, and it mainly consists of two parts:

[0053] The `fields` array: This part is the "instruction set" for the LLM to perform its tasks, defining all the key technical parameters (technical measures) and evaluation metrics (self-evaluation) that need to be extracted. It supports multiple data types such as text (string), number, choices, and table, providing precise and structured targets for subsequent information extraction tasks of the intelligent agent.

[0054] `display_template` (i.e., the display template): This section defines how the extracted and evaluated data is rendered into the final report. This template is not provided to the LLM. It uses Jinja2 syntax and preset rendering macros (such as `render_boolean()`, `render_choices()`, `render_table()`, etc.) to format the data obtained from the fields into standardized report paragraphs.

[0055] Through this design, the Schema clearly separates "what to extract" from "how to present," serving not only as instructions for the agent's actions but also as a "blueprint" to ensure accurate output, consistent format, and correct evaluation logic. Table 1 uses the specific evaluation item "recycled aggregate replacement rate" as an example to analyze in detail the role of each component element within the Schema.

[0056] Table 1. Examples and Explanations of Core Schema Fields

[0057]

[0058] The next steps, S2-S7, are executed by the online autonomous evaluation agent, which is responsible for online inference. It receives project documents uploaded by the user, uses the offline-generated schema as action instructions, autonomously plans evaluation steps, invokes internal tools, and generates evaluation report paragraphs in real time. The online autonomous evaluation agent and the offline schema engine are decoupled through a shared PostgreSQL rule base, ensuring that knowledge updates do not affect the stability of the online agent and improving the system's modularity and scalability.

[0059] To ensure the stability and controllability of the system, the online evaluation follows a fixed process driven by code, and calls the core evaluation module at key information acquisition and decision-making stages. This module is a collaborative effort between a first agent responsible for extracting structured data and a second agent responsible for scoring.

[0060] Step S2, Document Preprocessing: Obtain data on the green building projects to be evaluated, and construct a project knowledge base through parsing and vectorization.

[0061] After users upload documents for green building projects to be evaluated, the system's Document Service automatically executes a fixed preprocessing workflow, including: calling MinerU for deep parsing, intelligent chunking, and metadata injection. The processed document chunks and their embedding vectors are stored in the Milvus vector database, building a project-specific knowledge base.

[0062] Step S3, Evaluation Cycle Initiation: Utilize a pre-built online autonomous evaluation agent to plan evaluation steps using schema data as action instructions.

[0063] When a user selects an evaluation item, the system loads the corresponding schema and begins to iterate through each field defined in the schema that needs to be populated.

[0064] Step S4, Retrieval: Using a two-stage retrieval strategy of coarse screening and fine screening, locate the evidence needed for evaluation from the project knowledge base.

[0065] The agent generates a query based on the field description and invokes its core tool—the knowledge base retrieval tool. Internally, this tool executes a two-stage retrieval strategy of "coarse screening + fine screening": first, it performs vector retrieval using the Qwen3-Embedding-0.6B model to ensure recall (coarse screening); then, it uses the Qwen3-Reranker-0.6B model to re-rank the preliminary results to improve precision (fine screening); finally, it retrieves the most relevant evidence from the knowledge base.

[0066] Step S5, Extraction: Based on the evidence obtained from the location and the Schema data, extract structured information that conforms to the Schema constraints using a language model.

[0067] The agent combines the acquired evidence and schema definition into a Prompt, which guides the LLM to extract constraint information.

[0068] Preferably, the process after extraction also includes reflection and decision-making: the agent analyzes the extraction results. If the results are incomplete (e.g., a required field is empty), it determines that supplementary evidence is needed, regenerates the query, and returns to step S4 to continue retrieval. This loop continues until all information has been completely extracted or the maximum number of attempts has been reached, forming a "retrieval-extraction-reflection" loop.

[0069] Specifically, in steps S4-S5, the intelligent agent is invoked to execute the task. For each field in the schema, the main process "outsources" the task to the autonomous evaluation intelligent agent. After receiving the task, the intelligent agent will start its internal iterative loop.

[0070] Step S6, Automatic Scoring: The online autonomous evaluation agent evaluates and scores the extracted structured information by combining the preset scoring criteria in the schema data.

[0071] Once all fields have been extracted, the main process hands over the complete structured data and the scoring criteria defined in the schema to the scoring service. This service calculates the scores according to preset rules and formulas and provides preliminary scores. If the system detects insufficient evidence for certain scoring items, it will mark them and prompt the user for manual review.

[0072] Step S7, Report Rendering: The online autonomous assessment agent renders the scoring results and structured information into a green building assessment report based on the report template in the schema, thereby realizing the green building assessment.

[0073] The main process collects all extracted data and the final score, and then hands it over to the Report Service. This service uses the Jinja2 template engine to automatically render and generate the final report paragraphs.

[0074] Through the above design, the role of the agent is precisely limited to the stage that requires "intelligent decision-making" (i.e., judging whether the evidence is sufficient and conducting iterative retrieval), while the macroscopic stability and consistency of the entire evaluation process are guaranteed by the fixed process defined by the code.

[0075] Preferably, it also includes a human-machine collaborative workflow (not shown in the figure):

[0076] Step S8, Automated Review First: After evaluation experts upload project documents, steps S1-S7 automatically complete document parsing, knowledge base construction, and the generation of a preliminary report draft. This shortens the report writing cycle from weeks to hours, laying an efficient foundation for subsequent manual review.

[0077] Step S9, Manual Review and Multi-person Collaboration: The generated report draft and related evidence are presented on an interactive evaluation workbench that supports multi-person collaboration (e.g., Figure 3 (As shown). This workbench is the core hub of human-machine collaboration. Its design achieves transparency and controllability in the evaluation process through scientific functional zoning.

[0078] The workbench interface uses a three-column layout, which reflects the logical flow of the evaluation process:

[0079] The "Document Management" panel on the left serves as the foundation of evidence, centrally managing all project source files. The system clearly informs experts of the knowledge base's construction progress through status indicators (such as "Parsing" or "Completed"), and activates the evaluation process upon completion of document processing, ensuring that evaluations are always based on complete and up-to-date evidence.

[0080] The central "Evaluation Board" area is the main workspace for experts. The system breaks down complex evaluation criteria into a clearly hierarchical list of evaluation items (Specialty → Indicator → Evaluation Item), and uses multi-dimensional filtering functions to help experts quickly locate and process tasks. Each evaluation item is accompanied by a clear status label (such as "Pending," "Low Confidence," "Completed"), making the overall project progress and key milestones requiring manual intervention readily apparent.

[0081] The "Evidence Panel" on the right is a central showcase of the "machine-based review" results and a focal point for human-computer interaction. It not only presents the machine's initial assessment results for the selected clauses (including scores and confidence levels), but more importantly, it displays key extracted summaries and original text evidence supporting the conclusions in a structured manner. This design greatly enhances the traceability and transparency of the assessment process.

[0082] When experts need to conduct an in-depth review of a "low-confidence" or complex evaluation item, they can access a dedicated Evaluation Item Detail View page. This page offers more granular human intervention capabilities, including evidence management operations such as adding, deleting, and modifying machine-extracted evidence, as well as seeking advice from the AI ​​assistant through conversational interaction.

[0083] Step S10 Closed-Loop Learning: All modifications, supplementary evidence, and interactions made by experts during the review process will be recorded by the system and used for future schema iterations and prompt word optimization, forming a closed loop of continuous learning and improvement.

[0084] This interactive design, combining an overview (workbench) with detailed examination (details page), creates an efficient, transparent, and controllable human-computer collaborative workflow. It successfully transforms the role of assessment experts from "information filterers" to "final decision-makers," significantly improving assessment efficiency while ensuring the professional quality of the final report.

[0085] To verify the core role of the schema in constraining agent behavior and ensuring generation quality, a conceptual ablation analysis was conducted. The key value is demonstrated by comparing the system performance with and without schema constraints. Detailed comparison results are shown in Table 2.

[0086] Table 2. Examples and Explanations of Core Schema Fields

[0087]

[0088] As shown in Table 2, when the strong constraints of the schema are removed, and the agent is given only a high-level objective (e.g., "evaluate a clause") and is allowed to freely call tools and generate reports, a significant "illusion" problem occurs. Specifically, this manifests as follows: 1) The generated content deviates from the evaluation requirements, containing speculative information unrelated to the evidence; 2) The output format is inconsistent, making it difficult to meet the normative requirements of professional reports; 3) Key information points (such as specific values ​​and judgment conclusions) are easily omitted.

[0089] In contrast, under the strong constraints of the schema, the agent's task changes from an open and uncontrollable generative exploration to a deterministic "structured fill-in-the-blank" process. Since the output framework (JSON structure) and content type are predefined, the agent's "degrees of freedom" are greatly restricted, and it can only find information that conforms to the schema requirements from the provided evidence to fill in the blanks.

[0090] This mechanism fundamentally guarantees the reliability of generated content in multiple dimensions, including content fidelity, format consistency, information integrity, and evidence traceability, proving that schema-driven approaches are an effective way to solve the problem of uncontrollability in professional report generation.

[0091] The following section will use Article 7.2.17 of the "Green Building Evaluation Standard" (draft for approval), "Reasonable selection of building structural materials and components," as an example to intuitively illustrate the workflow and effects of this method.

[0092] 1. Offline Schema Generation: The offline engine first transforms the assessment requirements and report template of this clause into a structured schema. This schema defines several fields that need to be extracted, such as green_labeled_materials (green building material types) and green_material_ratio (green building material application ratio), and includes the final report rendering template display_template.

[0093] 2. The agent performs "retrieval-extraction-reflection": During online evaluation, the main process assigns the task of the "green building materials application ratio" field to the agent.

[0094] The agent's initial search only found the description "This project uses green building materials", and the extraction results were empty.

[0095] After reflecting on the information, the agent realized that there was insufficient information and generated a new query for "green building material application ratio - specific value" to search again. Finally, it located the key evidence in the "Material List" that "the application ratio of green building materials has reached 55%" and successfully extracted "green_material_ratio":55.

[0096] 3. Automatic scoring by the system: After all information has been extracted, the system receives the data "green_material_ratio":55 and the scoring criteria in the schema (e.g., "Application ratio not less than 50% earns 8 points"). The scoring module determines 55% > 50% based on the rules and gives a conclusion of 8 points.

[0097] 4. Automated report rendering: Finally, the main process collects all data and scores, calls the report service, and uses the display_template in the schema to render and generate the final report paragraphs. The style is completely consistent with the report style expected by the user, and evidence tracing is automatically completed.

[0098] In summary, this method effectively solves the efficiency, consistency, and traceability challenges in traditional evaluation processes by constructing an autonomous evaluation agent with schema as its action command, combined with a two-stage hybrid retrieval strategy of "coarse screening + fine screening" and a constrained generation mechanism. The design philosophy of this research does not pursue complete autonomy, but rather precisely utilizes the agent's reasoning ability to solve the most critical cognitive aspects, such as judging the sufficiency of evidence, within a stable, code-driven workflow, thus achieving a balance between system robustness and intelligence. By combining the agent's autonomous actions with rigorous schema planning, this method provides a highly controllable and low-cost solution for the automatic generation of professional domain reports, demonstrating significant application potential in human-machine collaborative workflows.

[0099] Example 2

[0100] Based on Example 1, this example provides a schema-driven green building assessment method system to implement the aforementioned green building assessment method. The system includes:

[0101] (1) Offline Schema engine module, used to obtain report template data and unstructured evaluation requirement data, and process them by parsing, segmenting and injecting metadata through the Schema engine to obtain structured Schema data;

[0102] (2) The online autonomous assessment agent module is used to acquire data of green building projects to be assessed, construct a project knowledge base through parsing and vectorization, plan assessment steps with Schema data as action instructions, call the built-in retrieval tool, and locate the evidence required for assessment from the project knowledge base through a two-stage retrieval strategy of coarse screening and fine screening. Based on the located evidence and the Schema data, the language model is used to extract structured information that conforms to the Schema constraints. Combined with the preset scoring criteria in the Schema data, the extracted structured information is evaluated and scored. Based on the report template in the Schema, the scoring results and structured information are rendered into a green building assessment report to realize the green building assessment.

[0103] Prototype system verification shows that this method effectively improves the automation, accuracy and traceability of report generation, providing a low-cost and highly controllable solution for building a human-machine collaborative intelligent assessment workflow.

[0104] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A Schema-driven green building assessment method, characterized in that, The method comprises the following steps: Obtaining report template data and unstructured evaluation requirement data, and performing parsing, blocking and metadata injection through a Schema engine to obtain structured Schema data, and realizing offline Schema construction; Obtaining green building project data to be evaluated, and constructing a project knowledge base through parsing and vectorization processing; Using a pre-constructed online autonomous evaluation agent to plan an evaluation step based on the Schema data as an action instruction; The online autonomous evaluation agent calls a built-in retrieval tool to locate the required evidence for evaluation from the project knowledge base through a coarse screening-fine selection two-stage retrieval strategy; The online autonomous evaluation agent extracts structured information conforming to the Schema constraint based on the located evidence and the Schema data using a language model; The online autonomous evaluation agent combines the scoring standard preset in the Schema data to evaluate and score the extracted structured information; The online autonomous evaluation agent renders the scoring results and the structured information into a green building evaluation report based on the report template in the Schema, and realizes green building evaluation.

2. The Schema-driven green building assessment method of claim 1, wherein, The Schema data comprises a fields array and a display template, wherein the fields array defines key technical parameters and evaluation indexes to be extracted, and the supported data types include text, numerical value, options and table, and the natural language instruction for the language model and the return format requirement are included; the display template is used to define the rendering format of the evaluation report, and includes a preset rendering macro.

3. The Schema-driven green building assessment method of claim 2, wherein, In the process of rendering the green building evaluation report, the structured data extracted in the fields array is filled into the report framework according to the preset format by calling the display template in the Schema data, and a green building evaluation report associated with the located original evidence and evaluation conclusion is generated.

4. The Schema-driven green building assessment method of claim 1, wherein, The online autonomous evaluation agent performs a retrieval-extraction-reflection cycle to extract structured information conforming to the Schema constraint: If the extracted information is incomplete or has missing items, the online autonomous evaluation agent determines that the evidence needs to be supplemented, generates a new query statement and re-executes the coarse screening-fine selection two-stage retrieval until the extracted information is complete or the maximum number of preset retrieval attempts is reached.

5. The Schema-driven green building assessment method of claim 1, wherein, The coarse screening-fine selection two-stage retrieval strategy is as follows: Coarse screening stage: performing approximate nearest neighbor search on the vector data in the project knowledge base using an Embedding model; Fine selection stage: reordering the results obtained in the coarse screening stage using a Reranker model.

6. The Schema-driven green building assessment method of claim 1, wherein, After the extracted structured information is evaluated and scored, the following steps are further included: If it is detected that the evidence of a scored item is insufficient, the scored item is marked as a low confidence item, and the user is prompted to perform manual review.

7. The Schema-driven green building assessment method of claim 1, wherein, Further comprising: Synchronizing the green building evaluation report and the associated evidence to an interactive evaluation workbench; Experts review, modify and supplement evidence for the report draft through the interactive evaluation workbench; Recording the review and modification operations and evidence supplement behavior of the experts, and iteratively optimizing the prompt words of the Schema and the language model.

8. The Schema-driven green building assessment method of claim 7, wherein, The interactive evaluation workbench adopts a three-column layout, and includes a document management area, an evaluation item area, and an evaluation detail area.

9. The Schema-driven green building assessment method of claim 1, wherein, The language model is used to convert report template data and unstructured evaluation requirement data into structured Schema data and store the data in a rule library, and the Schema engine and the online autonomous evaluation agent are decoupled through the rule library.

10. A Schema driven green building assessment method system characterized in that, The system is used for implementing the green building evaluation method according to any one of claims 1-9, and the system comprises: An offline Schema engine module is configured to acquire report template data and unstructured evaluation requirement data, and to perform parsing, block division and metadata injection through a Schema engine to obtain structured Schema data. An online autonomous evaluation agent module is configured to acquire green building project data to be evaluated, to construct a project knowledge base through parsing and vectorization processing, to plan an evaluation step based on the Schema data as an action instruction, to call a built-in retrieval tool, to locate required evidence for evaluation from the project knowledge base through a coarse screening-fine selection two-stage retrieval strategy, to extract structured information meeting a Schema constraint based on the located evidence and the Schema data, to combine the structured information with a preset scoring standard in the Schema data to evaluate and score the extracted structured information, to render the scoring result and the structured information into a green building evaluation report based on a report template in the Schema, and to implement green building evaluation.

Images