Assembly process procedure auxiliary compilation method based on large model

Abstract

The invention belongs to the technical field of intelligent manufacturing of aeronautical manufacturing enterprises, and relates to an assembly process procedure auxiliary compilation method based on a large model, which comprises the following steps: data preparation and processing: collecting and preprocessing an assembly process procedure text file; carrying out algorithm model selection and fine tuning, using an assembly process procedure text as a training data set and a test data set, and selecting different types of large models to carry out fine tuning; service publishing: deploying a plurality of large models with the best evaluation indexes as services, publishing the services, and externally providing available interface services; constructing an RAG knowledge base, vectorizing the assembly process procedure text, and constructing a vector knowledge base; intelligent question and answer: performing knowledge question and answer based on the RAG knowledge base; and assembly process procedure file generation: constructing the assembly process procedure file through text generation. According to the method, the functions of semantic retrieval intelligent question answering of the assembly technological procedure knowledge base, automatic assembly technological procedure file generation, information recording and the like are achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent manufacturing in aviation manufacturing enterprises and relates to a method for auxiliary compilation of assembly process specifications based on a large model. Background Art

[0002] Assembly technology is the general term for the benchmarks, methods, and technologies used in each assembly stage, developed by the process department based on product structure, technical requirements, and production scale. Assembly process specification documents clearly define the assembly process and operating methods for parts and components, and are an important basis for organizing production and guiding on-site operations. Assembly process specification documents include the file name, process step number, process step name, and operation details. With the continuous advancement of informatization, aviation manufacturing companies have accumulated a large amount of assembly process data. However, due to the lack of in-depth data analysis, process data is difficult to find, data utilization is low, and the value of the data is not effectively utilized.

[0003] A large model is a pre-trained model trained using a large amount of data, with millions to billions of parameters. The large model can be applied to the field of natural language processing to automatically generate text (such as patent CN117909990A). Large model fine-tuning is a secondary training based on the pre-trained model using professional field datasets, which can better adapt to specific downstream tasks. Commonly used large model fine-tuning evaluation indicators include bilingual evaluation bleu and recall-based summary evaluation rouge. bleu is used to evaluate the similarity of n consecutive word sequences between the text generated by the large model and the reference text, and rouge is used to evaluate the degree of summary matching between the generated text and the reference text. Prompt word engineering is an interaction with the large model that enables the large model to accurately answer user instructions. Enhanced retrieval rag is an information retrieval technology based on the large model. After retrieving information from an external knowledge base, it is input as a prompt to the large model to enhance the large model's ability to process professional knowledge.

[0004] The current compilation of assembly process specifications is mainly based on manual experience, supplemented by knowledge engineering and machine learning reasoning. The compilation of process specifications is labor-intensive, involves a lot of repetitive work, is labor-intensive, and has high labor costs. The quality of the compilation of process specifications depends on the level of the process personnel themselves, and it is difficult to maintain consistency in the compilation quality, resulting in a waste of human resources. Chinese patent CN113901647B and Chinese patent CN108491376B disclose a method for compiling process specifications based on machine learning, but due to the poor generalization ability of machine learning reasoning, effective knowledge precipitation and sharing cannot be formed. The above reasons make the compilation of assembly process specifications difficult, error-prone, and inefficient, and cannot meet the daily work needs of process personnel. Therefore, there is a need to provide a new auxiliary compilation method for assembly process specifications. Summary of the Invention

[0005] In order to solve the problems of difficulty in compiling assembly process specifications, ineffective utilization of historical data value, difficulty in knowledge management, and poor generalization ability, the present invention proposes an auxiliary compilation method for assembly process specifications based on a large model, which realizes the functions of semantic retrieval and intelligent question answering of the assembly process specification knowledge base, automatic generation of assembly process specification files, and information recording.

[0006] The technical solutions of the present invention are as follows:

[0007] A method for assisting in compiling assembly process specifications based on a large model, the steps are as follows:

[0008] Step S1: Data preparation and processing, collecting assembly process specification text files and preprocessing;

[0009] Step S2: Algorithm model selection and fine-tuning, using the assembly process specification text as the training dataset and test dataset, and selecting different types of large models for fine-tuning;

[0010] Step S3: Service publishing: deploy multiple large models with the best evaluation indicators as services and publish them, providing externally available interface services;

[0011] Step S4: Build a RAG knowledge base, vectorize the assembly process specification text, and build a vector knowledge base;

[0012] Step S5: Intelligent question answering, performing knowledge question answering based on the RAG knowledge base;

[0013] Step S6: Generating an assembly process specification file, and constructing an assembly process specification file through text generation.

[0014] Step S1 is as follows: Collect assembly process specification files. The assembly process specification files need to include file name, process step number, process step name and operation content, etc. The number is between 6,000 and 10,000. For the collected files, perform text data preprocessing, such as deleting duplicate values, special characters, spaces, line breaks, html tags, url links and incomplete data, etc., and convert numbers into string types. Two formats of files need to be prepared, namely txt files and json format files; the txt file is the training data required for enhanced retrieval, and needs to be saved in utf-8 encoding format. The file name is assembly process specification file name file name , file content doc = step num +step name +op content , step num is the step number, step name is the name of the process step, op contentThe json format file is the data required for fine-tuning the large model, which contains two types of question-answer pairs. One type is the answer pair based on the step name, with the format of {instruction:"step name", input:file name ,output:step num +step name}, the other is a response pair based on the operation content, the format is {instruction:”operation content”, input:file name +step num +step name ,output∶op content}.

[0015] Step S2 is as follows: Divide the JSON format fine-tuning dataset prepared in step S1 into a training dataset and a validation dataset in proportion. Select multiple open source or closed source large models for fine-tuning training. After initializing the commonly used hyperparameters, use the selected multiple models to perform multiple rounds of lora fine-tuning training based on the Ascend server graphics processing unit. Continuously adjust the size of the hyperparameters of each model, such as mixed precision mixed_precision, cutoff word length cutoff_len, batch size batchsize, training round epoch, learning rate learningrate, low rank matrix lora_rank, scaling factor lora_alpha and other model parameters. Model evaluation value is the measurement standard, where bleu is the bidirectional evaluation metric, rouge is the summary evaluation metric, and α is a number between 0 and 1. When the model's evaluation value v exceeds 97, training of this model is stopped and training continues with the next model. Finally, the m models with the best evaluation performance are selected from the multiple optimized models as the subsequent service release models.

[0016] Step S3 is as follows: The service is published using the m selected models with the best evaluation performance, providing an interface for external applications. It receives JSON-formatted data, such as the role, content, sampling flag do_sample, sampling parameter temperature, sampling parameter top-k, sampling probability top-p, maximum word length max_tokens, and streaming output stream. It returns JSON-formatted data, such as the model name model_name and reply information.

[0017] Step S4 is as follows: Based on the length of the assembly process specification document, the text segmentation length is determined. The text parsing method is to extract key content such as text, tables, and images from the file format, and slice it according to the text segmentation length or paragraph markers to form text paragraphs for subsequent retrieval and question-answering. After uploading the dataset file required for enhanced retrieval in step S1 using an open source or closed source text embedding model, the text is parsed and word vectors are constructed. The word vector u = [u1, u2…, un], where u is an n-dimensional matrix and n is the representable dimension. These text paragraphs and the generated word vector u are then stored in a vector database for query.

[0018] Step S5 is as follows: Write a web interface, the interface content includes a text box text, a chat box chatbot, a history record history, a submit button submit, a radio button checkbox, etc. Use the checkbox to select intelligent question and answer or assembly process procedure file generation. If you choose intelligent question and answer, you can enter text in the text, call the RAG vector knowledge base, and obtain the knowledge base information. The combined information t1 = text + prompt, where text is the text box input information and prompt is the knowledge base information. Send a post request for the combined information t1 to the model service published in step S3. The request contains role, content, do_sample, temperature, top-k, top-p, max_tokens, stream, etc., and obtain the model names and response information returned by multiple large models, etc., and combine these return values ​​into where model_name i is the name of the i-th model, reply i The response information of the i-th model is shown in the chatbot, and m is the number of models. t2 is displayed in the chatbot for the user to select. Partial or complete copying is possible. If the assembly process specification file is generated, proceed to step 6.

[0019] Step S6 is as follows: Using the embedding model of step S4 and the interface service of step S3, enter the name of the assembly process specification file you want to generate in the text box of step S5. Use the embedding model to convert the file name content text into word vectors, and then search in the RAG vector database to compare cosine similarity and cosine similarity. Among them, feature1 is the text vector value, feature2 is the vector value stored in the vector database, the symbol · represents the dot product of the vector, ||feature1|| and ||feature2|| represent the modulus of the vector, and when the similarity value sim>s (s is the set similarity threshold), the prompt text information prompt of the corresponding vector is extracted from the database, and the prompt word project is used to combine the text content content and the prompt information prompt into a string text1 = {content+prompt+"Based on the above content, generate the work step number and work step name. The format is: work step number. Work step name."}, which is sent to multiple large model services and returns multiple response texts containing the work step number and work step name. From the multiple large model response texts, the similarity of each two response messages is calculated, and the cosine similarity of each two response messages is calculated. Among them, feature i is the response information vector value of the large model i, feature j is the response information vector value of the large model j, the symbol represents the dot product of the vector, ||feature i || and ||feature j || represents the modulus of the vector. Select max(sim ij ) is used as the response message reply1max. Use the prompt word project to combine into a new string text2 = {content + reply1max + "Based on the above file name, work step number, and work step name, generate the operation content. The generated format is: work step number, sent to multiple model services, and returned multiple response texts containing work step number, work step name, and operation content. From the multiple large model response texts, calculate the similarity of each two response messages and calculate max(sim ij ), select any one of the two corresponding models as the response information reply2max. Use the prompt word project to combine into a new string text3 = {content + reply2max + "Generate file content according to the assembly process specification file format"}, send it to any model service, and return the response text. Save the response text as a docx file with the file name in content. The file content is multiple step numbers, step names, and operation contents. If the file is generated successfully, use the database to record information such as content, text1, reply1, reply1max, text2, reply2, reply2max, text3, reply3, and time, provide the user with a download function, and end the program. Otherwise, restart step S6.

[0020] Beneficial effects of the present invention:

[0021] To address the difficulties, errors, and low efficiency of compiling assembly process specifications, this paper proposes a large-scale model-based, assisted compilation method for assembly process specifications. This method leverages large-scale model fine-tuning, enhanced retrieval, and prompt word engineering to effectively utilize the value of historical data, simplify compilation, and achieve high accuracy and efficiency.

[0022] This method uses large-model fine-tuning technology to embed assembly knowledge into the model, strengthening existing knowledge and improving the utilization of existing data. Prompt word engineering reduces output deviations from large models and improves interaction quality. By integrating multiple models and cross-validating output results, model generalization is enhanced. Enhanced retrieval (RAG) technology addresses the difficulties and high error rates associated with updating assembly process knowledge. Recording text information in a database facilitates subsequent knowledge base query and maintenance, as well as the construction of fine-tuned datasets.

[0023] In summary, the auxiliary compilation method of assembly process specifications provided by the present invention can solve the problems existing in the existing assembly process specification compilation process, reduce the workload of manual compilation of assembly process specifications, make the compilation of assembly process specifications more reliable, efficient and convenient, and improve the compilation quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Flow chart of the method of the present invention.

[0025] Figure 2 A functional processing flow chart is generated for the assembly process specification file of the present invention.

[0026] Figure 3 Assist in compiling a functional structure diagram for the assembly process specification of the present invention. DETAILED DESCRIPTION

[0027] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0028] The following is an example of the method for assisting in compiling assembly process specifications, which is mainly implemented using python 3.10, llama-factory and ragflow. Figure 1 and 3 As shown, the implementation process is as follows:

[0029] Step S1, data preparation and processing: From the assembly process system database, export a CSV file containing the file name, step number, step name, and operation content, totaling 130,000 entries. Use pandas to preprocess the text data in the CSV file, such as removing duplicate values, special characters, spaces, line breaks, HTML tags, URL links, and incomplete data, and converting numbers to string types, obtaining 120,173 data entries. Save them in two file formats: txt and json. The txt file is the training data required for enhanced retrieval and needs to be saved in utf-8 encoding format. The file name is the assembly process specification file name file_name, and the file content doc = step_num + step_name + op_content, where step_num is the step number, step_name is the step name, and op_content is the operation content. A total of 9,932 txt files were exported. The json format file is the data required for fine-tuning the large model, saved as gongyi.json, and contains two types of question-answer pairs. One type of step name answer pair is in the format of {instruction:"step name",input:file_name,"output":step_num+step_name}; the other type of operation content answer pair is in the format of {instruction:"operation content",input:file name +step num +step name ,output∶op content}.

[0030] Step S2: Algorithm Model Selection and Fine-tuning: Place the fine-tuning dataset (gongyi.json) saved in Step 1 in the data folder of llama-factory. Create a new file named gongyi in dataset_info.json. Set val_size to 0.2 in train_bash.py and split the fine-tuning dataset into training and validation datasets with an 8:2 ratio. Select a fine-tuning model, such as Alibaba Tongyi Qianwen Qwen1.5-7b and Qwen2-7b, Tsinghua ChatGlm2-6b and ChatGlm3-6b, or Meta's Llama3-8b. Initialize common hyperparameters for model training to form a tuple p(mx, cut, ba, ep, lr, rank, alpa), where mx is mixed precision, cut is the cutoff length, ba is the batch size, ep is the number of training epochs, lr is the learning rate, rank is the low-rank matrix lora_rank, and alpa is the scaling factor lora_alpha. Use the gongyi.json file in step 1 to fine-tune the training and validation datasets. Select the tuple p for the hyperparameter. Run ep LoRa fine-tuning training on the aforementioned models using the train_bash.py file in the llama-factory package on an Ascend 910B server with eight 32GB NPUs.

[0031] Continuously adjust the size of each parameter in the tuple p to the evaluation value of the model is the metric, where bleu is the bidirectional evaluation metric, rouge is the summary evaluation metric, and α is a decimal between 0 and 1. When the model's evaluation value v > 97, training of this model is stopped and training of the next model continues. During actual training, for the fine-tuning dataset in step 1, when mx = fp16, cut = 2048, ba = 4, ep = 100, lr = 4e-5, rank = 8, alpa = 16, and α = 0.5, the evaluation values ​​v of the models qwen2-7b, chatglm3-6b, and llama3-8b were all greater than 97%. These three models were then merged with the original large model to form a new model, which was used as the model for subsequent service release.

[0032] Step S3: Service Publishing: Using the llama-factory api_demo.py program, based on the Ascend server NPU environment, publish the three models selected in Step 2 as services, providing an interface call service for external applications. This interface receives post request data from the web page containing chat message information in JSON format, such as the role, content, sampling flag do_sample, sampling parameter temperature, sampling parameter top-k, sampling probability top-k, maximum word length max_tokens, and streaming output stream. It returns JSON-formatted data, such as the model name model_name and a reply message.

[0033] Step S4: Build the RAG knowledge base: Using ragflow to build the knowledge base requires determining the text segmentation length chunks = 2048. The text parsing method is to extract key content such as text, tables, and images from the file format, and slice it according to the text segmentation length or paragraph markers to form text paragraphs for subsequent retrieval and question-answering. Using Alibaba qwen2-7b-instruct-q5_k_m as the embedding model, after uploading the dataset file required for the enhanced rag search in step S1, perform text parsing and construct word vectors. These text paragraphs and the generated word vectors are then stored in the vector database for query.

[0034] Step S5, Intelligent Q&A: Use the Python package gradio to create a web interface. The interface content includes a text box, a chatbot, a history record, a submit button, and a checkbox. The text field is used to enter text, the chatbot displays input and response information, the history field is used to record previous historical conversations, the submit field is used to submit user input text, and the checkbox field is used to select intelligent Q&A or assembly process specification file generation. If intelligent Q&A is selected, text can be entered in the text field and the rag vector knowledge base is called to obtain knowledge base information. This information is then combined with the original text to form the combined information t1 = text + prompt, where text is the text box input information and prompt is the knowledge base information. A post request for the combined information t1 is sent to the model service published in step S3. The request contains the following parameters: role = user, content = t1, do_sample = true, temperature = 0.75, top-k = 1, top-p = 1, max_tokens = 2048, stream = true, etc. Get the model name and response information returned by large models such as Qwen2-7b, ChatGlm3-6b and Llama3-8b, and combine these return values ​​into where model_name i is the name of the i-th model, reply i For the response information of the i-th model, m is 3. t2 is displayed in the chatbot for the user to select. Partial or full copying is possible. If the assembly process specification file is generated, proceed to step 6.

[0035] Step S6: Generate assembly process specification file (such as Figure 2 As shown in the figure): Select the embedded model Qwen2-7b-instruct-q5_k_m in step S4 and the three model services after fine-tuning in step 3. Enter the name of the assembly process specification file you want to generate in the text box text in step S5. Click the Submit button and use Qwen2-7b-instruct-q5_k_m to convert the file name content text into a word vector. Then search in the rag vector database and compare the cosine similarity and cosine similarity. Among them, feature1 is the text vector value, feature2 is the vector value stored in the vector database, the symbol · represents the dot product of the vector, ||feature1|| and ||feature2|| represent the modulus of the vector, and when the similarity value sim>0.8, the prompt text information prompt of the corresponding vector is extracted from the database. Using the prompt word project, the text content content and the prompt information prompt are combined into a string text1 = {content+prompt+"Based on the above content, generate the work step number and work step name. The format is: work step number. Work step name."}, which is sent to the three model services and returns the response text reply1 containing the work step number and work step name. Based on the response text reply1, the returned response information is calculated for similarity, and the cosine similarity of the response information is Among them, feature i is the response information vector value of the large model i, feature j is the response information vector value of the large model j, the symbol represents the dot product of the vector, ||feature i || and ||feature j || represents the modulus of a vector. Calculate max(sim ij ), Llama3-8b, ChatGlm3-6b sim ij Maximum, select the return value of model Llama3-8b as the reply information reply1max.

[0036] Combine into a new string text2 = {content + reply1max + "Based on the above file name, work step number, and work step name, generate the operation content. The generation format is: work step number. , send it to the three model services, and return the reply text reply2 containing multiple work step numbers, work step names, and operation content. Based on the reply text reply2, calculate max(sim ij ), Llama3-8b and Qwen2-7b sim ijMaximum, select the return value of model Qwen2-7b as the reply information reply2max. Combine into a new string text3 = {content + reply2max + "Generate file content according to the assembly process specification file format."}, send it to the ChatGlm3-6b model service, and return the reply text reply3. Use python-docx to save reply3 as a docx file. The file name is the file name in content. The file content is multiple step numbers, step names and operation contents. If the file is generated successfully, use the oracle 11g database to record information such as content, text1, reply1, reply1max, text2, reply2, reply2max, text3, reply3 and time, use the file component of the python package gradio to provide the user with a download function, and end the program. Otherwise, restart step S6.

Claims

1. A method for assisting in compiling assembly process specifications based on a large model, characterized in that: Here are the steps: Step S1: Data preparation and processing, collecting assembly process specification text files and preprocessing; Step S2: Algorithm model selection and fine-tuning, using the assembly process specification text as the training dataset and test dataset, and selecting different types of large models for fine-tuning; Step S3: Service publishing: deploy multiple large models with the best evaluation indicators as services and publish them, providing externally available interface services; Step S4: Build a RAG knowledge base, vectorize the assembly process specification text, and build a vector knowledge base; Step S5: Intelligent question answering, performing knowledge question answering based on the RAG knowledge base; Step S6: Generating an assembly process specification file, and constructing an assembly process specification file through text generation.

2. The method for assisting in compiling assembly process specifications based on a large model according to claim 1, characterized in that: Step S1 is as follows: Collect assembly process specification files, which need to include file name, process step number, process step name, and operation content; perform text data preprocessing on the collected files, including removing duplicate values, special characters, spaces, line breaks, HTML tags, URL links, and incomplete data, and converting numbers into string types; prepare two file formats: txt file and json file; txt file is the training data required for enhanced retrieval and needs to be saved in utf-8 encoding format. The file name is assembly process specification file name file name , file content doc = step num +step name +op content , step num is the step number, step name is the name of the process step, op content The json format file is the data required for fine-tuning the large model, which contains two types of question-answer pairs. One type is the answer pair based on the step name, with the format of {instruction:"step name", input:file name ,output:step num +step name }, the other is a response pair based on the operation content, the format is {instruction:”operation content”, input:file name +step num +step name ,output∶op content }.

3. The method for assisting in compiling assembly process specifications based on a large model according to claim 1, characterized in that: Step S2 is as follows: divide the JSON format fine-tuning dataset prepared in step S1 into a training dataset and a validation dataset in proportion; select multiple open source or closed source large models for fine-tuning training, initialize common hyperparameters, use the selected multiple models, and perform multiple rounds of Lora fine-tuning training based on the Ascend server graphics processing unit; continuously adjust the size of the hyperparameters of each model; the evaluation value of the model is the measurement standard, where bleu is the two-way evaluation indicator, rouge is the summary evaluation indicator, and α is a number between 0 and 1. When the evaluation value v of the model is greater than 97, stop training the model and continue training the next model. Finally, select the m models with the best evaluation performance from multiple optimized models as the subsequent service release models.

4. The method for assisting in compiling assembly process specifications based on a large model according to claim 3, characterized in that: In step S2, the hyperparameters include mixed precision mixed_precision, cutoff word length cutoff_len, batch size batchsize, training round epoch, learning rate learningrate, low rank matrix lora_rank and scaling factor lora_alpha.

5. The method for assisting in compiling assembly process specifications based on a large model according to claim 1, characterized in that: Step S3 is as follows: Use the m models with the best evaluation performance to publish services and provide interface connections to external applications; receive JSON format data, including role role, content content, sampling tag do_sample, sampling parameter temperature, sampling parameter top-k, sampling probability top-p, maximum word length max_tokens, and streaming output stream; return JSON format data, including model name model_name and reply information reply.

6. The method for assisting in compiling assembly process specifications based on a large model according to claim 1, characterized in that: Step S4 is specifically as follows: according to the text length of the assembly process specification file, the text segmentation word length is determined. The text parsing method is to extract key content including text, tables and images according to the file format, and slice it according to the text segmentation word length or paragraph mark to form text paragraphs that can be used for subsequent retrieval and question-answering; using an open source or closed source text embedding model, after uploading the dataset file required for the enhanced retrieval in step S1, text parsing and word vector construction are performed, the word vector u = [u1, u2…, un], u is an n-dimensional matrix, and n is the representable dimension. These text paragraphs and the generated word vector u are then stored in a vector database for query.

7. The method for assisting in compiling assembly process specifications based on a large model according to claim 1, characterized in that: Step S5 is as follows: write a web interface, the interface content includes a text box text, a chat box chatbot, a history record history, a submit button submit, and a radio button checkbox; select smart question and answer or assembly process specification file generation through the checkbox; if smart question and answer is selected, enter text in the text box, call the RAG vector knowledge base, obtain the knowledge base information, and combine the information t1 = text + prompt, where text is the text box input information and prompt is the knowledge base information; send a post request of the combination information t1 to the model service published in step S3, the request contains role, content, do_sample, temperature, top-k, top-p, max_tokens, stream, obtain the model names and response information returned by multiple large models, and combine these return values ​​into where model_name i is the name of the i-th model, reply i is the response information of the i-th model, m is the number of models, and t2 is displayed in the chatbot for the user to select. Partial or full copying is possible; If you choose to generate the assembly process specification file, proceed to step 6.

8. The method for assisting in compiling assembly process specifications based on a large model according to claim 1, characterized in that: Step S6 is as follows: using the embedding model of step S4 and the interface service of step S3, enter the name of the assembly process specification file to be generated in the text box of step S5; use the embedding model to convert the file name content text into word vectors, and then search in the RAG vector database to compare the cosine similarity and cosine similarity. Where feature1 is a text vector value, feature2 is a vector value stored in the vector database, the symbol · represents the dot product of the vectors, ||feature1|| and ||feature2|| represent the modulus of the vectors, and when the similarity value sim>s, where s is the set similarity threshold, the prompt text message prompt corresponding to the vector is extracted from the database. Using the prompt word project, the text content and the prompt message prompt are combined into a string text1= {content+prompt+ "Based on the above content, generate the work step number and work step name; the format is: work step number; work step name;"}, send it to multiple large model services, and return multiple response texts containing the work step number and work step name; From multiple large model response texts, the similarity calculation is performed for each two response messages, and the cosine similarity of each two response messages is calculated. Among them, feature i is the response information vector value of the large model i, feature j is the response information vector value of the large model j, the symbol represents the dot product of the vector, ||feature i || and ||feature j || represents the modulus of the vector; select max(sim ij ) as the response message reply1max; using the prompt word project, combine it into a new string text2 = {content + reply1max + "" based on the above file name, work step number, and work step name, generate the operation content; the generated format is: work step number, sent to multiple model services, and return multiple response texts containing the work step number, work step name, and operation content; From multiple large model response texts, calculate the similarity between each two response messages and calculate max(sim ij ), select any one of the two corresponding models as the response information reply2max; use the prompt word project to combine into a new string text3 = {content + reply2max + "Generate file content according to the assembly process specification file format"}, send it to any model service, and return the response text; save the response text as a docx file with the file name in content, and the file content is multiple step numbers, step names and operation contents; if the file is successfully generated, use the database to record content, text1, reply1, reply1max, text2, reply2, reply2max, text3, reply3 and time information, provide the user with a download function, and end the program; otherwise, restart step S6.

Images