Report automatic generation retrieval enhancement optimization method and system based on local knowledge base
Through the automatic report generation method based on the local knowledge base, the problem of incomplete extraction of non-structural documents and insufficient search accuracy is solved, and more accurate and efficient search results are achieved, enhancing user experience and large-scale model output controllability.
Patent Information
- Application Number
- CN202410954562.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-17
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2044-07-17
AI Technical Summary
The existing search enhancement generation method is incomplete for the extraction of non-structured document tables, pictures and other information. There are differences between user questions and actual searches, the search accuracy and correlation are insufficient, the search results are too redundant, and there is a certain format of document output. flaw.
Using the automatic report generation method based on the local knowledge base, the generation structure is enhanced by reconstructing and searching, the structure of unstructured files is standardized, the charts and picture information is extracted, and user input is disassembled and reconstructed, and the improved TextRank algorithm and Neo4j diagrams are used. The database performs keyword extraction and intent extraction, filters and compresses the search results based on the text information comprehensive index, and generates text with a specific format through a specific template.
It improves the accuracy and relevance of the search results, reduces the redundant context of model input, improves the efficiency and user experience of large-scale model output, and ensures the controllability and accuracy of the output.
Smart Images

Figure CN118981476A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of natural language processing, and in particular to a method and system for automatically generating retrieval enhancement and optimization of reports based on a local knowledge base. Background Art
[0002] With the rapid development of natural language processing technology, especially breakthroughs in deep learning and large language models, it has become possible to understand and generate human language. Today, large language models have excellent text classification, question-answering, and generation and summary capabilities. In actual use, due to the black-box characteristics of large language models, they are prone to hallucinations and are not accurate enough in answering private domain knowledge and real-time messages. They are difficult to apply to companies with data privacy. Retrieval Augmented Generation (RAG) is a method aimed at improving the quality and efficiency of information retrieval. The retrieval results are usually used as supplementary knowledge for large language models to improve model capabilities.
[0003] However, the existing retrieval enhancement generation has the following problems during use: (1) The retrieval enhancement generation is not perfect in extracting information such as tables and pictures in unstructured documents; (2) There are differences between user questions and actual search questions; (3) The retrieval accuracy and relevance are insufficient, and the retrieval results are too redundant; (4) There are defects in the output of documents in specific formats. To this end, this application proposes a retrieval enhancement optimization method and system for automatic report generation based on a local knowledge base, and proposes a reconstruction of the retrieval enhancement generation architecture. Summary of the invention
[0004] The purpose of the present invention is to provide a retrieval enhancement optimization method and system based on the automatic generation of local knowledge base reports, especially to standardize the structure of unstructured files, extract chart and image information, disassemble, reconstruct and expand user input, and generate documents in a specific format based on local knowledge. It can effectively solve the problems of imperfect extraction of information such as tables and pictures in unstructured documents by retrieval enhancement generation; differences between user questions and actual search questions; insufficient retrieval accuracy and relevance, and overly redundant retrieval results; and defects in the output of documents in a specific format.
[0005] To achieve the above object, the present invention provides the following technical solution: a retrieval enhancement method for automatic report generation based on a local knowledge base, comprising:
[0006] A new architecture for reconstructing retrieval enhancement generation is proposed; knowledge base documents are preprocessed to extract table and image information and format documents in a unified manner; a question reconstruction method is used to disassemble, reconstruct and expand the original user input question into four parts: retrieval keywords, user intentions, question rewriting and question expansion; among them, the retrieval keywords are applied with keyword retrieval algorithm, and question rewriting and question expansion are applied with vector semantic retrieval; the two search results are filtered, reordered, compressed and merged to form a logical context; the context is used as external knowledge, and a text with a specific format is generated based on this knowledge and following a known template.
[0007] Furthermore, the document preprocessing uses OpenCV to detect and extract the positioning of the wireframe incomplete table, including:
[0008] First, convert the pages to be extracted from the PDF into images;
[0009] Process the image and binarize the grayscale image;
[0010] Use the binarized image to obtain the horizontal and vertical lines in the table, using the morphological functions Erosion and Dilation in OpenCV;
[0011] The obtained horizontal and vertical lines are crossed to obtain the intersection point, and the obtained intersection point contour is judged whether it is a table, and finally the contour is approximated and converted into a rectangle.
[0012] Furthermore, the improved TextRank algorithm is applied to extract search keywords in the question reconstruction method, including:
[0013] Using keyword part-of-speech consistency and sliding window methods to construct the initial undirected graph can accurately capture keyword relationships and improve the semantic expression ability of the graph, reducing the computational complexity of subsequent algorithms;
[0014] Introducing a dynamic damping factor to satisfy the non-periodicity and irreducibility of the state transfer matrix can balance the scores between nodes of different categories and improve the accuracy and robustness of the results;
[0015] Introducing cosine similarity as the weight of the edge between two nodes to speed up Markov convergence, can more richly represent the relationship between nodes and handle more complex text relationships, and can significantly improve the accuracy of keyword extraction;
[0016] Adding the L1 regularization term will cause some parameter weights to become zero, which helps to build a sparse graph structure, allowing the algorithm to pay more attention to important nodes and edges and extract more representative keywords, while also improving computational efficiency.
[0017] Furthermore, the extraction of question intent in the question reconstruction module includes:
[0018] Use the Neo4j graph database to pre-build a database for question intent classification;
[0019] By using the Cypher query language, various graph operations can be performed efficiently to retrieve intents, sub-intents, and related intents;
[0020] Using semantic similarity as the judgment standard, the retrieved intent and the original question are calculated using cosine similarity, and the node intent with the highest similarity is taken as the final question intent;
[0021] By operating the question intention node, its sub-intentions and related intentions can be retrieved;
[0022] Furthermore, the search results are filtered, various text characteristics are fully considered, and a comprehensive index of text information is proposed, including:
[0023] Text repetition rate: The text vector is composed of the frequency of each word, and the cosine similarity is used to calculate the repetition rate of two texts;
[0024] The information entropy per unit word is calculated to indicate the amount of information contained in the text;
[0025] Calculate text similarity using the embedded text vectors;
[0026] The text information comprehensive index can be expressed as a linear combination of unit word information entropy and text similarity. First, the text repetition rate of the retrieval results is calculated, and the texts with higher repetition rates are preliminarily filtered. Then, the information comprehensive index of each text is calculated. Texts with too low information comprehensive index can be discarded or compressed.
[0027] A retrieval enhancement system for automatically generating local knowledge base reports, comprising:
[0028] The question reconstruction module determines whether the local knowledge base content needs to be used. If the local knowledge base search is required, the gap between user input and search is narrowed, and the question is reconstructed. The four parts of the question reconstruction are stored in the form of key-value pairs, including:
[0029] Retrieve keywords, extract the search keywords given by the user input and the most representative keywords of the entire input;
[0030] User intent, extracting the purpose and requirements of user input and turning redundant user input into simple semantic instructions;
[0031] Question rewriting: enriching and specifying user input, completing incomplete user input, and removing redundant parts;
[0032] Question expansion: expand the original question from one question to multiple questions, including: word reorganization, grammatical structure transformation, and intention expansion;
[0033] The retrieval processing module is used to retrieve documents related to the question, including:
[0034] Segment the documents and use the embedding model to vectorize the segmented documents and questions;
[0035] Find the k most similar sentence blocks by calculating cosine similarity;
[0036] The post-retrieval processing module is used to perform preliminary filtering and compression on the retrieval results, including:
[0037] Calculate the repetition rate between sentence blocks, preliminarily filter out sentence blocks with higher repetition rates, and then calculate the comprehensive index of text information of the remaining sentence blocks, and discard or compress sentence blocks with lower indexes.
[0038] The text generation module obtains the knowledge base context from the retrieval processing module and generates the user intent from the question reconstruction module. Through organic combination and in combination with specific format templates, the text generation capability of the large language model is used to generate the final text.
[0039] Compared with the prior art, the present invention has the following beneficial effects:
[0040] 1. The present invention provides a retrieval enhancement optimization method and system based on the automatic generation of local knowledge base reports, which improves the previous problems of inaccurate retrieval of local documents, low relevance, and large model hallucination. Among them, an innovative question reconstruction module is proposed to reconstruct user input, which improves the clarity and comprehensibility of the questions, and also improves the accuracy and relevance of the answers, enhancing the user experience. Among them, the keyword extraction uses an improved TextRank algorithm. The improved algorithm has a faster convergence speed, more accurate keyword extraction, and higher computational efficiency. At the same time, the search results are filtered and compressed, which greatly shortens the context of the model input and improves the efficiency of large model output. Specific prompt templates are applied to ensure controllable output while exploring the potential of large model text output.
[0041] 2. The method designed by the present invention first extracts and vectorizes the local document structure. In the previous RAG structure, user input is usually directly used for retrieval. The present technical solution adds a question reconstruction module to apply the original question to other situations, improve the retrieval accuracy and breadth, and identify whether retrieval is needed and what to retrieve without fine-tuning the model, thereby adding intelligent performance to the RAG structure. In the output stage, controllable output is achieved through customized templates and local knowledge. The present method has superior performance in both retrieval results and the final generated text. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 It is a schematic diagram of the overall process of the present invention;
[0043] Figure 2 for Figure 1 Flowchart of the problem reconstruction module;
[0044] Figure 3 for Figure 1 Flow chart of the retrieval processing module;
[0045] Figure 4 for Figure 1 Flowchart of post-retrieval processing module;
[0046] Figure 5 Generate a module flow chart for the text. DETAILED DESCRIPTION
[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0048] See also Figure 1-Figure 5 The present invention provides a technical solution: a search enhancement optimization method and system based on automatic generation of local knowledge base reports, referring to Figure 1 , is a schematic diagram of the retrieval enhancement optimization process automatically generated based on the local knowledge base report, including a question reconstruction module, a retrieval processing module, a retrieval post-processing module, and a text generation module, wherein the user input is directly connected to the question reconstruction module, the question after user input enters the question reconstruction module, the reconstructed partial result enters the retrieval processing module, the retrieval processing module generates relevant context fragments and combines them with the other reconstructed part to enter the text generation module, in the text generation module, different prompt templates are selected and applied according to different contents to generate text content based on the local knowledge base and with different characteristics.
[0049] Document preprocessing includes extracting the structure and content of unstructured documents such as word and pdf documents, and finally storing them in json data format, including the following steps:
[0050] For word documents, the python-docx library is used to preprocess the word documents, such as extracting the number of document pages, headers, footers, images, tables, styles, etc. as the basic information of the document, and extracting the text content of each chapter as the json data body;
[0051] The pdfplumber library is used to extract text, images, and regular tables from the PDF document. Then the PDF is traversed twice, and the OCR technology is used to extract text and locate and extract irregular tables from the PDF file. The obtained information is stored in a JSON file in the form of a link path.
[0052] The image data is stored in an external folder and stored in the json data in index form;
[0053] Json data is divided into two parts. The first part is the basic information of the document, including the number of document pages, headers, footers, and styles. The second part is the main content, which is stored in paragraphs. Each paragraph will be marked with the chapter to which it belongs.
[0054] Reference Figure 2 , is the question reconstruction module, which is the first module that the user inputs into. This module reconstructs the original question into four parts, namely, search keywords, user intention, question rewriting, and question expansion.
[0055] In question reconstruction, the first step is to extract the keywords of the original question, including the following steps:
[0056] First, use the jieba library to perform word segmentation on the original question. After word segmentation, load the stop word list of Harbin Institute of Technology, which can be filtered and expanded according to needs. Then, remove the stop words from the result after word segmentation.
[0057] Traverse the remaining words, determine the part of speech of each word and mark them, and use a dictionary to store each word and part of speech;
[0058] At this point, taking "What will the weather be like tomorrow" as an example, the specific structure of the question vocabulary is:
[0059] {"tomorrow":"t","weather":"n","how is it":"r"}.
[0060] Among them, t is a time word, n represents a noun, and r represents a pronoun.
[0061] Based on the consistency of part of speech and combined with the sliding window method, the initial undirected graph is constructed, that is, two vocabulary nodes with the same part of speech are connected by a line, or two words in the same sliding window are connected by a line, and the weight on each line is calculated by the formula:
[0062]
[0063] Calculated, among which Represents the vector of the word after the embedding model.
[0064] Calculate the score of each vocabulary node using the formula
[0065] Perform iterative calculations.
[0066] The introduction of dynamic damping factor, two-node text similarity and L1 regularization term greatly improves the accuracy of keyword extraction.
[0067] After the iteration converges, the k words with the highest scores are selected as retrieval keywords.
[0068] For the extraction of user intent in question reconstruction, the intent classification graph is used. Specifically, a graph database of intent classification is first constructed. The graph contains common question intents. Each node represents an intent, and each edge represents the relationship between intents. Each edge or node has a category attribute. The relationship between intents includes but is not limited to sub-intent, similar intent, and extended intent. Question intent extraction uses Cypher query language to retrieve multiple alternative intents through fuzzy query, and the alternative intents are calculated with the original question using a similarity algorithm. The intent node with the highest similarity is retained and selected as the final question intent.
[0069] Question expansion in question reconstruction is carried out from three aspects: 1. Lexical reorganization; 2. Grammatical transformation; 3. Intent expansion. Lexical reorganization includes synonym replacement, order exchange, and word removal. Grammatical transformation mainly transforms grammatical structure. Intent expansion can query similar or extended intents through the retrieved intent nodes.
[0070] The original question is reconstructed into four parts. The reconstructed text is in the form of a json data set, as follows:
[0071] {"rag_keywords":"content","user_purpose":"content","query_rewrite":"content","expand_input":"content"}
[0072] After reconstruction, rag_keywords, user_purpose, query_rewrite, and expand_input represent search keywords, user intent, question rewriting, and question expansion, respectively.
[0073] Reference Figure 3 , is a flowchart of the retrieval processing module. The module receives the reconstructed rag_keywords, query_rewrite, and expand_input parts, and outputs the relevant fragments retrieved from the local document. The module includes the following steps:
[0074] Step S310, load local knowledge base documents, including pdf, word, txt, csv, html and other types, and read local documents into memory.
[0075] Step S320, divide the loaded document into blocks using a recursive segmentation method, that is, using a set of separators to layer and iterate, and layer the input text into smaller blocks. If the segmented text does not initially produce blocks of the required size or structure, the method will recursively call the result block using different separators or standards until the required block size or structure is generated.
[0076] Step S330: After the block division is completed, the fine-tuned BGE-M3-Embedding model is used to vectorize the block content.
[0077] Step S340, index the Faiss database, using IndexFlatl2 index here, and then store the vectorized data into the database.
[0078] Step S350, search for fragments, perform a hybrid search based on BM25 and Embedding, that is, rag_keywords applies keyword search, query_rewrite and expand_input apply vector search, and finally merge the two search results.
[0079] Reference Figure 4 , is a post-retrieval processing module for processing the retrieval results. The steps are as follows:
[0080] Step S410, firstly perform preliminary filtering on the search results, and remove sentence blocks with high repetition rate by calculating the repetition rate between sentence blocks. The repetition rate is calculated using the cosine similarity formula, in which the vector is composed of the word frequency of the sentence block.
[0081] Step S420, re-ranking the search results of the preliminary filtering, using the invented text information comprehensive index indicator, the formula is as follows:
[0082]
[0083] Where H(x) represents the information entropy of a unit word. Indicates the relevance of this statement block to the original question.
[0084] Texts with higher comprehensive information indexes are placed at the front and back ends, with sentence blocks with lower indexes placed in the middle.
[0085] Step S430, performing text compression on the sorted sentence blocks, and removing stop words from each sentence block.
[0086] Step S440, summarize the processed sentence blocks, and use the large language model to summarize multiple sentence blocks into a logical sentence as external supplementary knowledge of the large model.
[0087] Reference Figure 5 , is the text generation module, which receives the output content of the post-retrieval processing module and uses it as the external knowledge of the large model. It also proofreads the output content of the large model to achieve the purpose of controllable generation.
[0088] The processed search text is combined with the question intent in the reconstruction module to form the model input. For different types of text generation, such as paper evaluation, company weekly reports, marketing copywriting, etc., which have specific formats and require domain knowledge, they need to be combined with templates of this type to form the final model input.
[0089] The text generation module will determine whether a template should be used and what type of template to use based on the question intent. If no template is used, the large model will only answer questions based on the retrieved sentence blocks and questions. If a template is needed, the text will be output in the template format. Taking an excellent paper evaluation as an example, the following shows the paper evaluation template format:
[0090] {
[0091] {<#system>: You are a senior paper review expert. You can output relevant content according to the known information and strictly follow the prompt format.},
[0092] {<#Known information>: content generated by the post-processing module},
[0093] {<#prompt>: "<Name> completed the graduation project <thesis title> under the diligent / careful guidance of <instructor>. The thesis adopted <technical method> and used <technology and tools>, including <technical implementation details>, to complete / implement / build <system / task>. During the graduation project, the student was proactive and able to think deeply, and correctly completed the content and requirements of the graduation project assignment; he was able to skillfully use relevant tools, and his guiding ideology and design plan were correct during the design process. He had strong independent work ability, could integrate theory with practice, and applied the knowledge he learned to solve practical problems well; the content of the thesis was correct, the structure was clear, the concepts were clear, the logic was clear, and the text was fluent. This thesis was rated as excellent"},
[0094] {<#Question Intent>: "Give me a paper evaluation based on the paper content."}
[0095] }
[0096] The final model input is composed of four parts. The first part <#system> is the role definition, which fixes the big model role and explains the task requirements. The second part <#known information> is the corpus obtained from the retrieval module as the external knowledge source of the big model. The third part <#prompt> is the format template of the output text. The above gives a template for the evaluation of excellent papers. The big model will output in the corresponding format according to the given template. The fourth part <#problem intention> is the user's purpose and task, which is generated by the problem reconstruction module. Different problem intentions can lead to changes in <#system>, <#prompt> and <#known information>.
[0097] Here, the locally deployed large language model is applied, and the final template is used as the model input. The model will generate a paper evaluation as required. The paper evaluation after the large model output will enter the proofreading model. The proofreading model will check whether the output contains errors or sensitive content, and detect whether the output is consistent with the template to achieve the purpose of controllable generation. Compared with the stiff and empty evaluation sentences generated by other AI, R-RAG+LLMs demonstrates its powerful text generation and summarization capabilities. Not only does it not have stiff clichés, it can also point out the technical tools and implementation details used by the author in the paper, analyze the article, and explain the defects and deficiencies, and provide some improvement directions.
[0098] The above description is only an embodiment of the present specification and is not intended to limit the present specification. For those skilled in the art, the present specification may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present specification shall be included in the scope of the claims of the present specification.
Claims
1. A report automatic generation retrieval enhancement optimization method based on a local knowledge base, characterized in that: include: S11 knowledge base document processing, extracting document structure and content, parsing the information presented in tables and images, and formatting the documents uniformly; S12 question reconstruction module breaks down the original question into four parts: search keywords, user intent, question rewriting, and question expansion; S13 applies the keyword search method to the search keyword, applies vector semantic search to question rewriting and question expansion, calculates the semantic richness and question relevance of the two search results respectively, filters the search results, and finally merges the two results; S14 reorders and compresses the final search results, and forms a logical context content by supplementing the grammatical structure; S15 organically combines the user intention with the context content, and applies a specific template to generate the required text content based on the local knowledge base.
2. The method according to claim 1, characterized in that: The step S11 locates the table in the PDF document, including: converting the PDF into a picture; processing the picture and identifying the horizontal and vertical lines of the table; extracting the intersection of the horizontal and vertical lines; approximating a rectangle through the outline of the intersection; and extracting data from the rectangle.
3. The method according to claim 1, characterized in that The step S12 proposes a reconstructed retrieval enhancement controllable generation R-RACG architecture based on the retrieval enhancement generation RAG architecture; The controllable generation in the R-RACG architecture uses a trained pluggable proofreading model to detect and mark potential false or erroneous information in the content generated by the large language model, filter the content before output, and gradually approach the expected goal through multiple rounds of feedback.
4. The method according to claim 1, characterized in that: The step S12 is a question reconstruction module, wherein the search keyword extraction uses an improved TextRank algorithm, including: Where: S(V i ) represents the score of the i-th keyword node; D(V i ) represents the dynamic damping factor of the node; C(E ij ) represents the cosine similarity between node i and node j; In(V i ) indicates the node set consisting of the i-node; Out(V j ) represents the node set pointed to by node j; |Out(V j )| represents the number of nodes in the collection; ∥S∥1 represents the L1 regularization term, and λ represents the regularization strength parameter; The improved TextRank algorithm first uses the Jieba library for word segmentation when constructing an undirected graph. It then constructs the initial undirected graph by analyzing whether the keyword parts of speech are consistent and using the sliding window method. Considering the consistency of the parts of speech can avoid some meaningless associations. The sliding window method can enhance the connection strength between related words and improve the semantic accuracy of the undirected graph. The damping factor is improved to the dynamic damping factor D(V i ), dynamically adjust according to the importance of the node to prevent the algorithm from falling into an infinite loop or the influence of isolated nodes being too large, balance the scores between nodes of different categories, the dynamic damping factor is related to the similarity between nodes and the score of the node, and the logistic function is used to map the damping factor to the interval [0,1]; The text similarity between two nodes is introduced when calculating the score. Here, the embedding vector of the pre-trained embedding model is used to calculate the similarity between nodes, thereby introducing context information and accelerating the Markov convergence speed. The L1 regularization term is introduced to prevent the scores from being overly concentrated on a few nodes. By penalizing the absolute value and the value of the scores and encouraging sparsity, important features can be automatically screened out, making the scoring more effective and reasonable.
5. The method according to claim 1, characterized in that The step S12 is a question reconstruction module, in which the Neo4j graph database is used to extract the question intent, construct an intent classification graph, each node represents an intent, construct a hierarchical relationship of intent, clarify the relationship between the main intent, sub-intent, and subdivided intent, and model the cross-relationship of intent, identify the cross-relationship and mutual influence between different intents, and use the Cypher query language to query the intent node, calculate the semantic similarity between the original question and the intent node, and the one that reaches 0.9 or above is taken as the final question intent.
6. The method according to claim 1, characterized in that The step S13 retrieves the result and filters it. First, the repetition rate of each sentence block with the rest of the sentence blocks is calculated. If the repetition rate of two sentence blocks is high, the sentence block with the highest text information comprehensive index is retained. The text information comprehensive index is defined, including: Among them, α1+α2=1 Where: C(x) represents the comprehensive index of text information of sentence block x; H(x) represents the information entropy of a unit word; Indicates the relevance of this statement block to the original question, where is the vector after word embedding; α1 and α2 represent weights; Represents the repetition rate of text x1 and x2, where is a vector consisting of the frequency of each word in the sentence block; The repetition rate between sentence blocks is calculated using the cosine similarity calculation formula, where the two sentence block vectors are composed of word frequencies.
7. The method according to claim 1, characterized in that In the reordering step S14, the sentence blocks are sorted according to the text information comprehensive index defined in claim 6, using a bidirectional comprehensive index sorting, that is, the sentence blocks with high comprehensive indexes are located at the front and back ends, and the sentence blocks with low comprehensive indexes are stored in the middle.
8. The method according to claim 1, characterized in that The large model prompt template in step S15 adopts multiple candidate templates, and intelligently selects different templates or does not use templates according to different question intentions to generate text content in different formats.
9. A system for implementing the method according to any one of claims 1 to 8, comprising: Document preprocessing module, question reconstruction module, retrieval processing module, retrieval post-processing module and text generation module.
Citation Information
Patent Citations
Retrieval method and retrieval system for knowledge graph in design field of water conservancy industry
CN115982322A
Medical auxiliary question and answer method and system based on knowledge calibration and retrieval enhancement
CN117573843A
Government affair intelligent response device and method based on intention recognition and large language model
CN118035419A
Internet mass data accurate search method and system based on AI technology
CN118093982A
Unsupervised focus-driven graph-based content extraction
US20230114425A1
Cited By
Information extraction system and method based on layout analysis and multiple modes
CN119377388A
Data security document generation method and system based on RAG technology
CN119475436A
Retrieval generation optimization method based on long-term enhancement, medium and system
CN119494410A
Optimization method and system for generating mathematical formula format based on large language model
CN119647410A
Private domain-oriented large language model personalized question and answer method and related products
CN119719287A
