Risk adaptation type intelligent compilation method for construction full-content type

Abstract

The invention belongs to the technical field of building construction informatization, and particularly relates to a construction full content type-oriented risk adaptation type intelligent compilation method, which comprises the following steps of S1, carrying out two-dimensional classification on construction scheme contents into a text class, a table class and a picture class according to'type-universality ', and executing special preprocessing; according to the risk adaptation type intelligent compilation method for the construction full-content type, by setting coverage of'text + table + picture 'and automatic calculation of characteristic table parameters, the efficiency is improved by 80%, drawing guidance is provided for characteristic pictures, the manual supplementary workload is reduced by 90%, the limitation that only text is processed in the prior art is solved, full-content coverage is achieved, and the risk adaptation type intelligent compilation method for the construction full-content type is achieved. And the manual supplement effect is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of construction information technology, and in particular to a risk-adaptive intelligent preparation method for construction full-content types. BACKGROUND

[0002] The construction technical scheme is the core file for guiding the safe and efficient implementation of the project, and its content covers text description, parameter table and professional drawings, and needs to be strictly adapted to the project risk level. In the prior art, patent application number 202510839037.1 proposes an intelligent preparation scheme of "RAG+ large model fine-tuning of text paragraph", but still has the following technical bottlenecks that cannot be solved:

[0003] Limited content processing range: only focusing on the intelligent generation of "text paragraphs", not covering "parameter tables" (such as material demand plans, equipment configuration tables) and "professional pictures" (such as construction site layout diagrams, process flowcharts) which account for 30%-50% of the construction scheme, and a large amount of manual supplement is needed, with limited efficiency improvement;

[0004] Lack of risk adaptation in RAG retrieval: using fixed cosine similarity retrieval logic, without adjusting the retrieval priority in combination with the construction project risk level (such as special risk deep foundation pit, primary risk high formwork), high-risk projects are easy to retrieve non-key knowledge, the standard reference accuracy is only about 80%, and there is a safety hazard;

[0005] Strong generalization of large model fine-tuning: the fine-tuning data cleaning only contains general steps of "denoising, segmentation", without optimization for the characteristics of construction documents "term-intensive, parameter-repeated-annotation-rich"; the reward model only generalizes the "winning / losing answer", without strengthening the "standard clause accurate reference, risk parameter integrity" and other construction core needs;

[0006] Missing full-process support: no real-time intelligent question-answering function, and the preparation personnel need to manually consult the specifications when encountering technical difficulties (such as special working condition parameter determination); no large model selection and adaptation scheme is provided, and construction enterprises (especially small and medium-sized enterprises) have high deployment costs for single models, making it difficult to implement;

[0007] Low knowledge update efficiency: no incremental update mechanism is designed, and after new specifications are released or new excellent historical schemes are added, the entire index library needs to be reconstructed, the adaptation cycle is long, and it cannot meet the high-frequency update needs of construction specifications.

[0008] The prior art cannot meet the preparation needs of "full content, high risk, low cost, and fast adaptation" of the construction scheme, and there is an urgent need for a risk-adaptive intelligent preparation method for construction full-content types. SUMMARY

[0009] In order to solve the technical problem that the existing construction scheme compilation cannot meet the compilation requirements of full content, high risk, low cost and fast adaptation of the construction scheme, the application provides a risk adaptation type intelligent compilation method for the full content type of construction.

[0010] The risk adaptation type intelligent compilation method for the full content type of construction comprises the following steps:

[0011] S1: classifying the construction scheme content into text, table and picture according to the two-dimensional classification of type and generality, and performing special preprocessing;

[0012] S2: constructing a dynamic weight RAG index library based on project risk level division, performing similarity retrieval with risk factors, and outputting the first k retrieval results;

[0013] S3: performing construction field deep customization fine-tuning on the basic large model, including construction exclusive unsupervised fine-tuning and construction core demand supervised fine-tuning;

[0014] S4: splicing the preprocessing content vector, RAG retrieval result vector and risk level vector to form a context, generating text, table and picture content by type, and performing compliance verification;

[0015] S5: providing full-process multi-round intelligent question and answer support, and realizing knowledge base incremental update.

[0016] Preferably, in the step S1, the special preprocessing comprises: the text preprocessing adopts a construction field special tokenizer, and the tokenization formula is

[0017] Tokenizer const is a BERT pre-training based construction special tokenizer, I is an input text, t n is a Token containing a construction term, i n is a corpus index;

[0018] In the table preprocessing, the characteristic table parameters are calculated according to the formula , V struc is a structure volume, p mat is a material density, and d loss is a construction loss coefficient;

[0019] In the picture preprocessing, the characteristic picture output contains drawing instructions of size, proportion and key annotations.

[0020] Preferably, in the step S2, the project risk level is divided into four levels, and the four levels one-to-one correspond to the knowledge graph weight W K and the four unstructured document library weights W D .

[0021] Similarity search introduces risk factor γ risk , take 1.2, take 1.0, take 0.9, take 0.8, and modify the similarity formula ;

[0022] The search results are sorted in descending order of Sim dyn , take the top 10, take the top 8, and take the top 6.

[0023] Preferably, in step S3, the construction-specific unsupervised fine-tuning loss function is , λ=0.3 is the construction term penalty coefficient, Dict const is the construction term dictionary;

[0024] The reward model loss function for construction core demand supervised fine-tuning is , K=5, y w is the winning answer containing specification references, parameters, and risk measures, and y l is the losing answer.

[0025] Preferably, in step S4, the context construction formula is , V pre is the preprocessed content vector, V RAG is the RAG search result vector, and V risk is the risk level vector.

[0026] After content generation, check "specification reference integrity 100%, table parameter logic, and picture format compliance", and only output if the pass rate is ≥98%.

[0027] Preferably, in step S5, the context weight of multi-round intelligent question answering is calculated as , α=0.7, T=8 is the maximum number of dialogue rounds, and t is the current round.

[0028] The knowledge base incremental update formula is , only new data is vectorized and indexed, and the consistency verification conflict rate is ≤2%.

[0029] Preferably, in the construction loss coefficient δ loss , take 3% for steel reinforcement, 2% for concrete, and 5% for formwork; the resolution requirement of the characteristic picture drawing guide is ≥300dpi.

[0030] Preferably, the project risk level division is based on the "Building Construction Safety Risk Classification and Control Regulations":

[0031] The special level is that the deep foundation pit is greater than or equal to 20m or the main tower height is greater than or equal to 200m, the first level is that the deep foundation pit is 10-20m or the main tower height is 100-200m, the second level is that the deep foundation pit is 5-10m, and the third level is that the deep foundation pit is less than 5m.

[0032] Preferably, the construction exclusive unsupervised fine-tuning base model is Qwen2.5-7B, and the training parameters are: batch size 16, learning rate 2e-5, and iteration number 50 rounds;

[0033] The labeled data of the supervised fine-tuning is completed by 2 senior engineers and 3 employees with more than 8 years of experience, the training iteration is 30 rounds, and the learning rate is 1e-5.

[0034] Preferably, the base large model is selected from at least one of Qwen2.5, Byte Bao and Baidu Wenxin Yiyang;

[0035] The knowledge base incremental update trigger condition is: within 7 days after the new specification is released, within 5 days after the new excellent historical scheme is accepted, and when the cumulative number of expert experience summaries is 30.

[0036] The beneficial effects in the application are:

[0037] 1. By setting the coverage of "text + table + picture", the characteristic table parameter is automatically calculated, the efficiency is improved by 80%, the characteristic picture provides drawing guidance, and the manual supplementary workload is reduced by 90%, solving the limitation of the prior art that only text is processed, achieving full content coverage, and reducing manual supplementation.

[0038] 2. By setting the risk adaptation retrieval, the specification accuracy is improved, in use, the specification reference accuracy of the special risk project is 98%, compared with the prior art of 80%, avoiding the safety hazards caused by inaccurate retrieval due to the accuracy of 80% of the prior art, and meeting the construction risk control requirements.

[0039] 3. By setting real-time question answering and low-cost landing, in use, the response time is less than 1 second through multiple rounds of question answering, supporting small and medium-sized enterprises to select low-cost models, solving the problems of support missing and high cost in the prior art, and through fast knowledge updating, adapting to specification iteration, the knowledge base incremental updating period is shortened to 1 day, meeting the high-frequency updating demand of construction specifications, and improving the timeliness of the scheme. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 A schematic diagram of a risk adaptation type intelligent preparation method for construction full content type is provided for the application;

[0041] Figure 2 A risk adaptation dynamic weight RAG retrieval principle diagram of a risk adaptation type intelligent preparation method for construction full content type is provided for the application;

[0042] Figure 3 A construction field customization fine-tuning flowchart of a risk adaptive intelligent preparation method for the construction full content type proposed in the present application;

[0043] Figure 4 A full content type differentiation processing flowchart of a risk adaptive intelligent preparation method for the construction full content type proposed in the present application;

[0044] Figure 5 A risk adaptive intelligent preparation method for the construction full content type proposed in the present application Figure 1 An enlarged view of the structure at A in the middle;

[0045] Figure 6 A risk adaptive intelligent preparation method for the construction full content type proposed in the present application Figure 1 An enlarged view of the structure at B in the middle. DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments.

[0047] Referring to Figures 1-6 , a risk adaptive intelligent preparation method for the construction full content type, comprising the following steps:

[0048] S1: classifying the construction scheme content into text type (general / specific), table type (general / specific), and picture type (general / specific) according to the two-dimensional classification of “type-generalness”, and performing special preprocessing;

[0049] In the step S1, the special preprocessing includes: the text preprocessing adopts a construction field special tokenizer, and the tokenization formula is

[0050] Tokenizer const is a construction special tokenizer based on BERT pre-training, I is an input text, t n is a Token containing a construction term, i n is a corpus index;

[0051] In the table preprocessing, the specific table parameter is calculated according to the formula , V struc is a structure volume, p mat is a material density, and d loss is a construction loss coefficient;

[0052] In the picture preprocessing, the specific picture output contains drawing instructions with size, proportion, and key annotation requirements.

[0053] Further, the construction loss coefficient δ loss Among them, the steel takes 3%, the concrete takes 2%, and the formwork takes 5%; the resolution requirement of the characteristic picture drawing guide is ≥300dpi, and the format is AutoCAD. dwg or PNG / JPG.

[0054] S2: Based on the project risk level division, a dynamic weight RAG index library is constructed, a similarity retrieval with risk factors is performed, and the top k retrieval results are output;

[0055] In step S2, the project risk level is divided into four levels, and the four levels are one-to-one corresponding to the knowledge graph weight W K , and the four unstructured document library weights W D .

[0056] The four levels of project risk level are special, first, second, and third, corresponding to the knowledge graph weight W K 0.8, 0.6, 0.4, 0.3, respectively, and the unstructured document library weight W D 0.2, 0.4, 0.6, 0.7, respectively;

[0057] The similarity retrieval introduces a risk factor γ risk , the special level takes 1.2, the first level takes 1.0, the second level takes 0.9, and the third level takes 0.8, and the modified similarity formula is .

[0058] The retrieval results are sorted in descending order of Sim dyn , the top 10 are taken for the special level, the top 8 are taken for the first level, and the top 6 are taken for the second level and below.

[0059] Further, the project risk level division is based on the “Building Construction Safety Risk Classification and Control Regulations”: special for deep foundation pit ≥20m or main tower height ≥200m, first for deep foundation pit 10-20m or main tower height 100-200m, second for deep foundation pit 5-10m, and third for deep foundation pit <5m.

[0060] S3: Perform deep customization fine-tuning of the basic large model in the construction field, including construction exclusive unsupervised fine-tuning and construction core demand supervised fine-tuning;

[0061] In step S3, the loss function of the construction exclusive unsupervised fine-tuning is , λ=0.3 is the construction term penalty coefficient, and Dict const is the construction term dictionary;

[0062] The reward model loss function of the construction core demand supervised fine-tuning is , K=5, and y wFor the winning answer containing specification references, parameters, and risk measures, y l For the losing answer.

[0063] Further, the construction-specific unsupervised fine-tuning base model is Qwen2.5-7B, with training parameters: batch size 16, learning rate 2e-5, and iteration number 50 rounds; the labeled data for supervised fine-tuning is completed by 2 senior engineers and 3 staff with more than 8 years of experience, with training iterations of 30 rounds and a learning rate of 1e-5.

[0064] S4: splice the preprocessed content vector, RAG retrieval result vector, and risk level vector to form a context, generate text, table, and picture content by type, and perform compliance verification;

[0065] In the step S4, the context construction formula is , V pre is the preprocessed content vector, V RAG is the RAG retrieval result vector, V risk is the risk level vector.

[0066] After content generation, verify "specification reference integrity 100%, table parameter logic, and picture format compliance", and only if the verification pass rate is ≥98% can the output be output.

[0067] S5: Provide full-process multi-round intelligent question and answer support, while realizing incremental updating of the knowledge base.

[0068] In the step S5, the context weight of multi-round intelligent question and answer is calculated as , α = 0.7, T = 8 is the maximum number of dialogue rounds, and t is the current round;

[0069] The incremental updating formula of the knowledge base is Only new data is subjected to vectorization and index addition, and the consistency verification conflict rate is ≤2%.

[0070] The base large model is selected from at least one of Qwen2.5, Byte Bean, and Baidu Ernie; the incremental updating trigger condition of the knowledge base is: within 7 days after the new specification is released, within 5 days after the new excellent historical solution is accepted, and when the cumulative number of expert experience summaries reaches 30.

[0071] By setting coverage "text + table + picture", the characteristic table parameter is automatically calculated, achieving an efficiency improvement of 80%, the characteristic picture provides drawing guidance, and the manual supplement workload is reduced by 90%, solving the limitation of the prior art that only handles text, achieving full content coverage, and reducing manual supplement.

[0072] By setting up risk adaptation retrieval, the accuracy of the specification is improved. In use, the specification reference accuracy of special risk items is 98%, which is higher than the existing technology of 80%. The safety hazards caused by inaccurate retrieval due to the existing technology of 80% accuracy are avoided, and the construction risk control requirements are met.

[0073] By setting up real-time question answering and low-cost landing, in use, the response time of multiple rounds of question answering is less than 1 second, supporting small and medium-sized enterprises to choose low-cost models (such as Byte Bean Bag, with a 60% reduction in deployment cost), solving the problems of existing technology support missing and high cost, and through fast knowledge update, adapting to specification iteration, the knowledge base incremental update cycle is shortened to 1 day (the existing technology needs 7 days), meeting the high-frequency update demand of construction specifications, and improving the timeliness of the scheme.

[0074] Embodiment 1

[0075] A risk adaptation type intelligent compilation method for construction full content type, comprising the following steps:

[0076] S1: Construction full content classification and special preprocessing, receiving user input compilation instructions, classifying construction scheme content according to "type-universality" two-dimensional classification and performing special preprocessing, covering text, table, picture full content type:

[0077] S11: Two-dimensional classification definition:

[0078] Text type: divided into general text (such as quality assurance system, acceptance standard clauses) and characteristic text (such as construction difficulties, process adaptation instructions);

[0079] Table type: divided into general table (such as work efficiency analysis table, risk factor list, fixed format) and characteristic table (such as material demand plan, equipment configuration table, parameters need to be calculated according to the project);

[0080] Picture type: divided into common pictures (such as pile protection schematic diagram, rainproof shed standard diagram, reusable) and characteristic pictures (such as engineering structure schematic diagram, construction site layout diagram, need to be professionally drawn);

[0081] S12: Special preprocessing operation:

[0082] Text preprocessing: use construction field special word segmentation tool (based on BERT pre-training, training data is 100,000 construction texts), word segmentation formula is:

[0083]

[0084] Tokenizer const Tokenizer is a construction special word segmentation tool based on BERT pre-training, I is the input text, t nFor Token containing construction terms (such as "hydraulic climbing formwork" and "maintenance temperature difference"), i n Indexing the corpus.

[0085] Table preprocessing: fixed format for general table call template library; characteristic table extracts parameters to calculate factors (such as material requirement = structure volume × material density × loss coefficient), formula is: , where V struc is the structure volume (user input), p mat is the material density (knowledge base call), and d loss is the construction loss coefficient (such as 3% for steel reinforcement and 2% for concrete).

[0086] Picture preprocessing: common pictures are retrieved from the knowledge base (format is PNG / JPG, resolution ≥ 300 dpi); characteristic pictures generate "drawing guidelines" (including size, proportion, and key labeling requirements, such as "the main tower spray system layout diagram needs to mark the wind buckle position").

[0087] S2: Indexing and retrieval of dynamic weight RAG adapted to risk, based on project risk level, constructing a dynamic weight RAG retrieval system to improve the retrieval accuracy of high-risk projects:

[0088] S21: Project risk level division and weight mapping:

[0089] According to the "Provisions on the Classification and Control of Construction Safety Risks", project risks are divided into special level (such as deep foundation pit ≥ 20m, main tower height ≥ 200m), first level (deep foundation pit 10-20m, main tower height 100-200m), second level (deep foundation pit 5-10m), and third level (deep foundation pit < 5m), with weight mapping rules as follows:

[0090] Risk level Knowledge graph (specification / parameters) weight W K ]] Unstructured document library (historic solution) weights W D ]] Search priority Special class 0.8 0.2 Norm clause > historical cases First class 0.6 0.4 Norm clause = historical cases Second class 0.4 0.6 Historical cases > norm clause Third class 0.3 0.7 Historical cases dominant

[0091] S22: Dynamic weight RAG retrieval process:

[0092] Vector embedding: use pre-trained Sentence-BERT model in construction field (vector dimension 768), embedding formula is:

[0093] , where E const is the construction-specific embedding model, and T is the Token sequence after tokenization.

[0094] Indexing and storage: historical schemes are divided into blocks according to "chapter-risk level" (such as "deep foundation pit support-special risk"), and are stored in Chroma vector database after vectorization, with indexing formula as:

[0095] , where Vc,risk Risk is the index vector with risk label. tag Risk level label (e.g., "special level").

[0096] Dynamic weight retrieval: joint "content vector + risk level" retrieval, similarity calculation introduces risk factor γ risk (special level takes 1.2, first level takes 1.0, second level takes 0.9, and third level takes 0.8), and the correction formula is:

[0097] , sorted in descending order of Sim dyn , and the top k (k=10 for special level, k=8 for first level, and k=6 for second level and below) retrieval results are taken, and the formula is:

[0098] .

[0099] S3: Fine-tuning of deep customized large models in the construction field, according to the characteristics of strong dependence on construction specifications and strict parameter requirements, a "two-stage + double constraint" fine-tuning strategy is designed:

[0100] S31: First stage: unsupervised fine-tuning for construction (adapt language habits)

[0101] Data selection: select excellent construction plans in the past 5 years (5000), current national / industry standards (300, such as GB50204, JGJ276), and senior engineer experience summary (200, with more than 10 years of experience);

[0102] Special data cleaning:

[0103] a1: remove construction document-specific noise (such as repeated "approval opinions" and "header and footer");

[0104] a2: extract core process parameters (such as "curing time ≥ 14 days" and "wind resistance level ≥ 6"), and form a parameter dictionary;

[0105] a3: standardize construction terminology (such as "climbing formwork" to "hydraulic climbing formwork" and "concrete" to "concrete");

[0106] Fine-tuning training: use Qwen2.5-7B as the base model, and the training task is "construction text continuation". The loss function adds a construction terminology penalty term (to avoid generating non-standard terminology), and the formula is ;

[0107] where λ=0.3 is the penalty coefficient, Dictconst is the construction terminology dictionary, and I(·) is the indicator function (1 for non-terminology and 0 otherwise); training parameters: batch size 16, learning rate 2e-5, and iteration number 50 rounds.

[0108] S32: Second stage: Construction core needs supervised fine-tuning (strengthen professional);

[0109] Label data construction: 2 construction experts (senior engineers) + 3 experienced editors (employed for ≥8 years) label "construction problem-answer" samples (10,000), label standards:

[0110] Answers must contain "specification clause number (such as GB50204-2015 Article 7.4.1) + technical parameters + risk control measures";

[0111] Priority selection of high-risk scene samples (such as "special risk deep foundation pit dewatering parameters");

[0112] Reward model (RM) training: with "specification reference integrity (weight 0.4), parameter accuracy (0.3), risk measure integrity (0.3)" as evaluation indicators, the loss function formula is: , where C(K,2) is the number of combinations of selecting 2 from K=5 answers, y w is the winning answer (satisfies 3 indicators), y l is the inferior winning answer (missing ≥1 indicator), r θ (·) is the reward model score, and σ is the sigmoid function;

[0113] Fine-tuning optimization: based on the InstructGPT framework, fine-tune the unsupervised model with labeled samples, train for 30 iterations, and the learning rate is 1e-5.

[0114] S4: Full content type scheme generation and differential processing, integrating RAG retrieval results and fine-tuned models to generate "text-table-picture" full content schemes and perform differential adaptation:

[0115] S41: Context construction: concatenate "preprocessed content vector + RAG retrieval result vector + risk label", formula: , where V pre is the preprocessed content vector, V RAG is the RAG retrieval result vector, and V risk is the risk level vector (such as [1,0,0,0] for special level).

[0116] S42: Generation of sub-type content:

[0117] Text generation: call construction-specific prompt word templates (example: "role: bridge senior engineer; goal: generate special risk main tower concrete curing text; constraints: include curing time, temperature difference control, windproof measures, quote 1 specification per point"), input fine-tuned model, generate parameters: temperature coefficient 0.2 (ensure stability), maximum length 1000 characters, repetition penalty 1.3;

[0118] Table generation: general table filling template; characteristic table based on S12 P cal Automatic calculation of parameters, generation of MarkDown format (supporting Excel export);

[0119] Picture generation: common picture directly inserted; characteristic picture output "drawing guide + blank template" (such as AutoCAD format. dwg), prompting user to draw and import after drawing;

[0120] S43: Content verification: automatically verify "specification reference integrity (100% required), table parameter logic (such as material demand > 0), picture format compliance (resolution ≥ 300 dpi)", and pass rate ≥ 98% before output.

[0121] S5: Multi-round intelligent question and answer and knowledge base incremental update in the whole process, providing real-time technical support and efficient knowledge update, solving the problems of existing technical support missing and slow update:

[0122] S51: Multi-round intelligent question and answer (adapt to construction technology problems):

[0123] Problem analysis: use construction intent recognition model (training data 20,000 construction problems) to analyze "core entities (such as climbing system), problem types (such as safety threshold), risk levels (such as special level)";

[0124] Context management: use Redis to cache nearly 8 rounds of dialogue, and the context weight is calculated by exponential decay, the formula is: , where α = 0.7 (recent dialogue weight is higher), T = 8 (maximum round), t is the current round.

[0125] Answer generation: combine the dynamic weight retrieval result of S2 to generate an answer containing "conclusion + specification reference + case reference" (example: "Climbing system stops working when wind speed is 6 or above (reference JGJ276-2012 5.3.4), reference 2023 special risk main tower scheme (number QT-2023-05), instantaneous wind speed ≥ 10.8 m / s, stop working");

[0126] S52: Knowledge base incremental update:

[0127] Update trigger: within 7 days after the release of new specifications, within 5 days after the acceptance of new excellent historical schemes, and when the cumulative number of expert experience summaries reaches 30;

[0128] Incremental index: only new data is executed for vectorization and index addition, without the need to reconstruct the whole library, the formula is: , where Index old is the original index, V new is the new data vector, and Risknew New data risk label;

[0129] Consistency check: check the conflict between new data and original knowledge (such as specification parameter change), adopt the principle of "new specification priority", and the conflict rate is less than or equal to 2% to complete the update.

[0130] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can make equivalent replacement or change according to the technical scheme and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.

Claims

1. A risk adaptation type intelligent preparation method for construction full content type, characterized in that, The method comprises the following steps: S1: The construction scheme content is classified into text, table and picture according to the two-dimensional classification of "type-universality", and special preprocessing is performed; S2: Based on the project risk level division, a dynamic weight RAG index library is constructed, a similarity search with risk factors is performed, and the top k search results are output; S3: The basic large model is subjected to deep customization fine-tuning in the construction field, including construction-specific unsupervised fine-tuning and construction core demand supervised fine-tuning; S4: The preprocessed content vector, RAG search result vector and risk level vector are spliced to form a context, and text, table and picture content are generated by type, and compliance verification is performed; S5: Provide full-process multi-round intelligent question and answer support, and realize incremental updating of the knowledge base.

2. The risk-adapted intelligent preparation method for the whole content type of construction according to claim 1, characterized in that: In the step S1, the special preprocessing includes: the text preprocessing adopts a construction field special word segmentation device, and the word segmentation formula is , Tokenizer const is a BERT pre-training based construction-specific tokenizer, I is an input text, t n is a Token containing construction terminology, i n is a corpus index; Characteristic table parameters in table pre-processing according to formula Calculate, V struc For structural volume, p mat For material density, d loss For construction loss coefficient; In the picture preprocessing, the characteristic picture output includes size, proportion, key label requirements and drawing guide.

3. The risk-adapted intelligent preparation method for the construction-oriented full content type according to claim 1, characterized in that: The project risk level in the step S2 is divided into four levels, the four levels one by one correspond to the knowledge graph weight W K , and the four unstructured document library weights W D ; Similarity search introduces risk factor γ risk , the first level of 1.2, 1.0, 0.9, 0.8, three levels of correction similarity formula is ; Search results are sorted in descending order of Sim dyn order.

4. The risk-adapted intelligent preparation method for the construction-oriented full content type according to claim 1, characterized in that: In the step S3, the construction-specific unsupervised fine-tuning loss function is , λ = 0.3 is a construction term penalty coefficient, Dict const is a construction term dictionary; The construction core needs a supervised fine-tuned reward model loss function , K = 5, y w is the winning answer containing norm references, parameters, risk measures, and y l is the losing answer.

5. The risk-adapted intelligent preparation method for the construction-oriented full content type according to claim 1, characterized in that: The context construction formula in the step S4 is , pre is a preprocessed content vector, V RAG is a RAG retrieval result vector, V risk is a risk level vector; After the content is generated, the "specification reference integrity 100%", "table parameter logic" and "picture format compliance" are checked, and the pass rate is ≥98% before the output.

6. The risk-adapted intelligent preparation method for the whole content type of construction according to claim 1, characterized in that: In the step S5, the context weight of the multi-round intelligent question answering is calculated according to , wherein α=0.7, T=8 is the maximum dialogue round, and t is the current round. The knowledge base incremental updating formula is Only vectorization and index addition are performed on new data, and the consistency check conflict rate is less than or equal to 2%.

7. The risk-adapted intelligent preparation method for the construction-oriented full content type according to claim 2, characterized in that: The construction loss coefficient δ loss Among them, the steel takes 3%, the concrete takes 2%, and the formwork takes 5%; the resolution requirement of the characteristic picture drawing guide is ≥300dpi.

8. The risk-adapted intelligent preparation method for the construction-oriented full content type according to claim 3, characterized in that: The project risk level division is based on the "Building Construction Safety Risk Classification and Control Regulations": Special level for deep foundation pit ≥20m or main tower height ≥200m, first level for deep foundation pit 10-20m or main tower height 100-200m, second level for deep foundation pit 5-10m, third level for deep foundation pit <5m.

9. The risk-adapted intelligent preparation method for the construction-oriented full content type according to claim 4, characterized in that: The basic model of the construction-specific unsupervised fine-tuning is Qwen2.5-7B, and the training parameters are: batch size 16, learning rate 2e-5, and iteration number 50 rounds; The labeled data for supervised fine-tuning is completed by 2 senior engineers and 3 personnel with more than 8 years of experience, and the training iteration is 30 rounds, and the learning rate is 1e-5.

10. The risk-adapted intelligent preparation method for the construction-oriented full content type according to claim 1, characterized in that: The incremental updating of the knowledge base is triggered when the following conditions are met: within 7 days after the new specification is released, within 5 days after the new excellent historical scheme is accepted, and when the cumulative number of expert experience summaries is 30.

Images