Large-scale intelligent processing methods and algorithm models for office automation scenarios

By combining the LoRA low-rank adaptive algorithm and the bge-m3 vectorization model with the RAG retrieval enhancement algorithm, the problems of insufficient model accuracy and high resource cost in office automation scenarios are solved, realizing lightweight, automated and efficient document processing and question answering capabilities.

CN123117325APending Publication Date: 2026-11-17CNOOC INFORMATION TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611204993.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-10
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies lack dedicated large-scale model fine-tuning algorithms in office automation scenarios, resulting in insufficient model accuracy, inadequate semantic understanding, reliance on manual generation of meeting minutes, low efficiency in processing multi-format documents, and high model training and inference costs, making it difficult to implement in a lightweight manner.

Method used

The LoRA low-rank adaptive algorithm is used to fine-tune a general large language model for the office domain. Combined with the bge-m3 vectorization representation model and the RAG retrieval enhancement algorithm, it can realize document vectorization, speech-to-text structuring, multi-format document parsing and summary generation, and support lightweight model training and accurate question answering.

Benefits of technology

It achieves lightweight model fine-tuning, improves the efficiency of intelligent office processing, enhances accurate retrieval and question-answering capabilities, reduces computing resource requirements, and supports automatic processing of multi-format documents and full-process automation without human intervention.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117325A_ABST
    Figure CN123117325A_ABST
Patent Text Reader

Abstract

This invention discloses a large-scale intelligent processing method and algorithm for office automation scenarios, including the following steps: decomposing office policy documents into knowledge units and representing them as text vectors; fine-tuning a general-purpose large language model for office domains using the LoRA low-rank adaptive algorithm; achieving accurate question answering of office policy knowledge based on the RAG retrieval enhancement algorithm; automatically generating meeting minutes through speech-to-text and text structure extraction algorithms; and performing content parsing, summary generation, and content question answering on multi-format office documents. This invention achieves functions such as vectorized construction of an office policy knowledge base, lightweight adaptation of a large domain model, accurate intelligent question answering, automatic generation of meeting recording minutes, and office document summarization and content question answering. This method does not rely on specific hardware or system environments, features a lightweight model, low training overhead, and zero inference latency, and can significantly improve the efficiency of office text and speech processing.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the fields of large model algorithms, natural language processing, speech processing, and office intelligent processing technology, and in particular relates to large model intelligent processing methods and algorithm models for office automation scenarios. Background Technology

[0002] With the continuous growth in the number of enterprise office documents and policies, traditional text processing methods can no longer meet the intelligent needs of modern office work. Existing technologies mainly suffer from the following algorithmic-level deficiencies: I. There is a lack of dedicated large-scale model fine-tuning algorithms for the office domain, resulting in insufficient accuracy of general-purpose large language models in tasks such as policy-related question-and-answer and minutes generation in office scenarios. While full fine-tuning can improve model performance in specific domains, it requires calculating and storing an update matrix of the same size for each weight matrix in the model, leading to high computational and storage costs. Existing efficient parameter fine-tuning methods such as Adapter Tuning and Prompt Tuning, although reducing training costs to some extent, still suffer from accuracy loss and insufficient generalization ability in adapting to multi-task scenarios in the office domain.

[0003] Second, office policy documents are characterized by numerous technical terms and complex semantic structures. Traditional keyword matching retrieval methods struggle to capture the deep semantic information of the text, resulting in slow query speeds and low recall rates. Existing vectorized models often suffer from input length limitations and insufficient semantic representation when processing long text office documents, making it difficult to meet the high-precision retrieval needs in office scenarios.

[0004] Third, there is a lack of intelligent extraction and structured generation algorithms after the meeting audio is transcribed into text. Current technology can only provide the original verbatim transcript after the meeting recordings are transcribed from speech recognition, lacking the ability to intelligently analyze and structure the transcribed text. The compilation of meeting minutes still relies on manual reading of the transcribed text, manual extraction of key points and tasks, which is time-consuming, labor-intensive, and prone to missing crucial information.

[0005] Fourth, there is a lack of unified parsing, summarizing, and content question-answering algorithms for multi-format office documents. Office documents come in a variety of formats, and existing document processing tools often only support text extraction from a single format, lacking a unified document parsing framework. The absence or insufficient accuracy of document summarization and content question-answering functions leads to low efficiency and poor standardization in office document processing.

[0006] Fifth, the high cost of model training and inference latency make lightweight deployment difficult. Training and inference of large language models require substantial computing resources, and full fine-tuning is costly. In actual deployments in office scenarios, excessively high inference latency severely impacts user experience, hindering lightweight deployment.

[0007] Therefore, there is an urgent need for a large-scale model fine-tuning algorithm, a retrieval enhancement algorithm, and a document and speech intelligent processing method and model for office automation. Summary of the Invention

[0008] The problem this invention aims to solve is to provide a large-scale intelligent processing method and algorithm model for office automation scenarios. This method integrates lightweight model fine-tuning, accurate question and answer retrieval, automatic meeting minutes, and intelligent document parsing, thereby improving the efficiency of intelligent office processing.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a large-model intelligent processing method for office automation scenarios, comprising the following steps: S1: Decompose the office system documents into knowledge units, and use a vectorized representation model to transform each knowledge unit into a fixed-dimensional text vector; S2: The LoRA low-rank adaptive algorithm is used to fine-tune the general large language model for the office domain; S3: Accurate Q&A on Office System Knowledge Based on RAG Search Enhancement Algorithm; S4: Automatically generate meeting minutes using speech-to-text and structured text extraction algorithms; S5: Performs content parsing, summary generation, and content Q&A on multi-format office documents.

[0010] Furthermore, in S1, the vectorization representation model is the bge-m3 vectorization model, which converts unstructured office text into fixed-dimensional text vectors, establishes a vector index based on all the text vectors, and constructs a vector knowledge base.

[0011] Furthermore, in S2, the pre-training weights of the general large language model are fixed, and low-rank matrices A and B are injected into the general large language model. The low-rank matrices A and B are fine-tuned using office domain corpus. After training, the low-rank matrices A and B are merged with the pre-training weights to obtain an office domain-adapted large model.

[0012] Furthermore, the number of parameters in the low-rank matrices A and B is reduced from d×d to 2×r×d, where d is the dimension of the pre-trained weights of the general large language model, and r is the rank of the low-rank matrix and r < d / r. <d。

[0013] Furthermore, in S3, the RAG retrieval enhancement algorithm includes obtaining a natural language question input by the user, converting the natural language question into a question vector using the vectorized representation model, performing a similarity search based on the question vector, obtaining a retrieval fragment associated with the natural language question, concatenating the retrieval fragment with the natural language question in context, inputting it into the office domain adaptation big model, generating an accurate answer, and outputting the original text location information corresponding to the retrieval fragment.

[0014] Furthermore, the similarity retrieval adopts a vector similarity calculation method to calculate the similarity between the question vector and the text vector of each knowledge unit, and the knowledge unit with a similarity exceeding a preset threshold is determined as the retrieval segment.

[0015] Furthermore, in step S4, the meeting recording file is acquired, audio noise reduction preprocessing is performed on the meeting recording file, the noise-reduced audio is transcribed into initial text using a speech-to-text model, and automatic error correction and punctuation are added to obtain the meeting text. Based on a preset prompt word template, key information is extracted from the meeting text, including meeting points, conclusions, and to-do items, and meeting minutes are generated according to a standard meeting minutes template.

[0016] Furthermore, the speech-to-text model is the SenseVoice model; the key information extraction is performed based on a preset prompt word engineering algorithm, which has a built-in output template and supports constraints on output format, output length and professionalism.

[0017] Furthermore, in step S5, multiple format office documents are acquired, the full text content of the office documents is extracted, a summary is automatically generated from the full text content, a natural language question from the user regarding the office documents is acquired, and an accurate answer is given to the natural language question based on the full text content, and the document segment corresponding to the answer is located; the format of the office documents includes, but is not limited to, PDF, DOCX, and DOC formats.

[0018] Furthermore, this invention also provides a large-scale intelligent processing algorithm model for office automation scenarios, which runs the aforementioned large-scale intelligent processing method for office automation scenarios, including: The document vectorization representation model is used to obtain a collection of office system documents. According to the preset splitting rules, each office system document is automatically split into several knowledge units by chapter. After the text content of each knowledge unit is reviewed and corrected, the vectorization representation model is used to convert each knowledge unit into a fixed-dimensional text vector. Based on all text vectors, a vector index is built to construct a vector knowledge base. LoRA fine-tuning model is used to obtain a general large language model. The pre-training weights of the general large language model are fixed, and low-rank matrix A and low-rank matrix B are injected into the general large language model. The low-rank matrix A and low-rank matrix B are fine-tuned and trained using office domain corpus. After training, the low-rank matrix A and low-rank matrix B are merged with the pre-training weights to obtain an office domain adapted large model. The RAG retrieval enhancement question answering model is used to obtain natural language questions input by users, convert the natural language questions into question vectors using the vectorized representation model, perform similarity retrieval in the vector knowledge base based on the question vectors, obtain retrieval fragments associated with the natural language questions, and input the retrieval fragments and the natural language questions into the office domain adaptation model after context concatenation, generate accurate answers and output the original text location information corresponding to the retrieval fragments. The speech-to-text and minutes generation model is used to acquire meeting recording files, perform audio noise reduction preprocessing on the meeting recording files, use a speech-to-text model to transcribe the noise-reduced audio into initial text and perform automatic error correction and punctuation addition to obtain meeting text, extract key information from the meeting text based on a preset prompt word template, extract meeting key points, conclusions and to-do items, and generate meeting minutes according to a standard meeting minutes template; The document parsing and summarizing model is used to acquire office documents, extract the full text content of the office documents, automatically generate a summary of the full text content, acquire the natural language questions of users regarding the office documents, accurately answer the natural language questions based on the full text content, and locate the document fragments corresponding to the answers.

[0019] The advantages and positive effects of this invention are: 1. The model of this invention is lightweight. It adopts the LoRA low-rank adaptive fine-tuning algorithm to compress the number of parameters from d×d to 2×r×d (r << d), which greatly reduces the number of training parameters and computational overhead. After training, the low-rank matrix is ​​merged with the original weights, and no additional delay is added during inference.

[0020] 2. This invention offers more accurate question-and-answer results. It employs the RAG retrieval enhancement algorithm combined with the original text of the policy for contextual enhancement, resulting in traceable and highly credible answers. The bge-m3 vectorization model supports both dense and sparse dual-path retrieval, taking into account semantic understanding and keyword matching, further improving retrieval recall and question-and-answer accuracy.

[0021] 3. This invention automates the entire process, from speech-to-text conversion and meeting minutes generation to document parsing and summary generation, all without human intervention, significantly improving office processing efficiency.

[0022] 4. This invention is highly versatile, does not depend on specific hardware and system environments, and can be quickly ported and reused in the office scenarios of different enterprises.

[0023] 5. This invention has good scalability, supports incremental updates of the vector knowledge base and multi-scenario algorithm extensions, and can flexibly adapt to ever-changing office needs. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall process of an embodiment of the present invention.

[0025] Figure 2 This is a detailed schematic diagram illustrating the implementation of the LoRA low-rank adaptive fine-tuning algorithm in this invention.

[0026] Figure 3 This is a schematic diagram of the execution flow of the RAG retrieval enhanced question answering algorithm according to an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram of the complete process of the meeting minutes generation algorithm according to an embodiment of the present invention.

[0028] Figure 5 This is a schematic diagram of the execution flow of the multi-format document parsing and summarizing algorithm according to an embodiment of the present invention. Detailed Implementation

[0029] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] The embodiments of the present invention will be further described below with reference to the accompanying drawings: like Figure 1 As shown, the intelligent processing method for large models in office automation scenarios includes the following steps.

[0031] S1: Decompose office policy documents into knowledge units, and use a vectorized representation model to transform each knowledge unit into a fixed-dimensional text vector. Establish a vector index and construct a vector knowledge base. Preferably, the vector knowledge base contains no fewer than 300 office policy documents, supporting dynamic updates and incremental index construction.

[0032] Specifically, the vectorization representation model in this embodiment is the bge-m3 vectorization model. bge-m3 is a multilingual embedding model with multifunctional, multilingual, and multi-granular characteristics, supporting dense retrieval, word matching, and multi-vector interaction. This model converts unstructured text into dense vectors of fixed dimensions, supports input lengths of up to 8K tokens, and can meet the vectorization needs of long text office documents. bge-m3 also supports the generation of dense and sparse vectors. Dense vectors are suitable for cosine similarity retrieval, while sparse vectors are suitable for word or keyword-aware retrieval.

[0033] This invention establishes a vector index based on the bge-m3 model, which supports high recall and high precision similarity retrieval, providing an efficient vector retrieval foundation for RAG question answering systems.

[0034] S2: The LoRA low-rank adaptive algorithm is used to fine-tune the general-purpose large language model for the office domain. This embodiment uses the Qwen large language model as the core inference model.

[0035] LoRA (Low-Rank Adaptation) is a parameter-efficient fine-tuning method. Its core idea is to treat the weight matrix of the original pre-trained model as fixed values ​​during fine-tuning, adapting to the new task only by training low-rank matrices A and B. Specifically, the pre-trained weights of a general-purpose language model are fixed, and low-rank matrices A and B are injected into the model. Office-specific corpora are used to fine-tune low-rank matrices A and B. After training, low-rank matrices A and B are merged with the pre-trained weights to obtain an office-specific adapted model.

[0036] Preferably, the number of parameters in low-rank matrices A and B is reduced from d×d to 2×r×d, where d is the dimension of the pre-trained weights of the general large language model, and r is the rank of the low-rank matrix and r < d / r. <d。

[0037] This invention applies the LoRA algorithm to fine-tuning of large-scale office models, supporting rapid adaptation to multiple tasks such as office policy Q&A, meeting minutes generation, and document processing. Detailed implementation of the LoRA low-rank adaptive fine-tuning algorithm is as follows... Figure 2 As shown.

[0038] S3: Accurate Q&A of Office System Knowledge Based on RAG Retrieval Enhancement Algorithm. The RAG retrieval enhancement algorithm includes: acquiring the natural language question input by the user; converting the natural language question into a question vector using a vectorized representation model; performing similarity retrieval based on the question vector to obtain retrieval fragments associated with the natural language question; concatenating the retrieval fragments with the natural language question contextually and inputting them into a large-scale office domain adaptation model; generating accurate answers and outputting the original text location information corresponding to the retrieval fragments. Preferably, the similarity retrieval uses a vector similarity calculation method to calculate the similarity between the question vector and the text vectors of each knowledge unit; knowledge units with similarity exceeding a preset threshold are identified as the retrieval fragments.

[0039] Specifically, RAG (Retrieval-Augmented Generation) is an architectural pattern that enhances prompts by retrieving relevant information, thereby generating responses. Its core idea is to retrieve relevant information from an external knowledge base before sending prompts to the large language model, thus enabling the model to generate more accurate and relevant answers using information from the external knowledge base.

[0040] The execution flow of the RAG algorithm in this invention is as follows: user inputs a natural language question → question vectorization → relevant fragments are retrieved from the vector knowledge base → the retrieved fragments are concatenated with the question context → the answer is generated by the large-scale model adapted to the office domain → the original text is associated with the output.

[0041] The execution flow of the RAG retrieval enhanced question answering algorithm is as follows: Figure 3 As shown, this algorithm balances versatility with domain-specific expertise, generating factually accurate outputs for office-specific policies outside the model's training data, significantly improving question-answering accuracy.

[0042] S4: Automatically generate meeting minutes using speech-to-text and structured text extraction algorithms. Obtain the meeting recording file, perform audio noise reduction preprocessing, and use a speech-to-text model to transcribe the noise-reduced audio into initial text, performing automatic error correction and punctuation addition to obtain the meeting text. Based on a preset prompt word template, extract key information from the meeting text, including key points, conclusions, and to-do items, and generate meeting minutes according to a standard meeting minutes template.

[0043] Preferably, the speech-to-text model is the SenseVoice model. Key information extraction is based on a preset prompt word engineering algorithm. The prompt word engineering algorithm has a built-in output template and supports constraints on output format, output length and professionalism.

[0044] Specifically, SenseVoice is a speech understanding engine designed to understand the real-world context of meetings. It can not only accurately recognize speech content, but also support multilingual recognition and speaker differentiation.

[0045] The algorithm execution flow is as follows: audio noise reduction preprocessing → SenseVoice model transcribing into initial text → automatic error correction and punctuation addition → obtaining meeting text → extracting key information (meeting points, conclusions, to-do items) based on prompt word templates → generating structured minutes according to standard meeting minutes templates → supporting editing and export. The complete process of the meeting minutes generation algorithm is as follows: Figure 4 As shown.

[0046] S5: Performs content parsing, summary generation, and content-based question answering for multi-format office documents. It acquires office documents in various formats, extracts the full-text content, automatically generates summaries from the full-text content, obtains user-generated natural language questions about the documents, and provides accurate answers to these questions based on the full-text content, locating the corresponding document segments. Office document formats include, but are not limited to, PDF, DOCX, and DOC formats. Preferably, it uses prompt word engineering algorithms to constrain the output format, content length, and level of professionalism, improving the output quality and standardization of question answering, minutes, summaries, and reports.

[0047] Specifically, the document parsing algorithm execution flow is as follows: Upload office documents in PDF / DOCX / DOC formats → System extracts full-text content → Content structuring processing → Automatic generation of summaries and reviews → User asks natural language questions about the document content → Accurately answers questions based on the document content and locates the corresponding document segments. The execution flow of multi-format document parsing and summarizing algorithms is as follows: Figure 5 As shown.

[0048] This invention also provides a large-scale intelligent processing algorithm model for office automation scenarios, which runs the aforementioned large-scale intelligent processing method for office automation scenarios, including: The document vectorization representation model is used to obtain a collection of office system documents. According to the preset splitting rules, each office system document is automatically split into several knowledge units by chapter. After the text content of each knowledge unit is reviewed and corrected, the vectorization representation model is used to convert each knowledge unit into a fixed-dimensional text vector. Based on all text vectors, a vector index is built to construct a vector knowledge base. LoRA fine-tuning model is used to obtain a general large language model. The pre-training weights of the general large language model are fixed, and low-rank matrix A and low-rank matrix B are injected into the general large language model. The low-rank matrix A and low-rank matrix B are fine-tuned and trained using office domain corpus. After training, the low-rank matrix A and low-rank matrix B are merged with the pre-training weights to obtain an office domain adapted large model. The RAG retrieval enhancement question answering model is used to obtain natural language questions input by users, convert the natural language questions into question vectors using the vectorized representation model, perform similarity retrieval in the vector knowledge base based on the question vectors, obtain retrieval fragments associated with the natural language questions, and input the retrieval fragments and the natural language questions into the office domain adaptation model after context concatenation, generate accurate answers and output the original text location information corresponding to the retrieval fragments. The speech-to-text and minutes generation model is used to acquire meeting recording files, perform audio noise reduction preprocessing on the meeting recording files, use a speech-to-text model to transcribe the noise-reduced audio into initial text and perform automatic error correction and punctuation addition to obtain meeting text, extract key information from the meeting text based on a preset prompt word template, extract meeting key points, conclusions and to-do items, and generate meeting minutes according to a standard meeting minutes template; The document parsing and summarizing model is used to acquire office documents, extract the full text content of the office documents, automatically generate a summary of the full text content, acquire the natural language questions of users regarding the office documents, accurately answer the natural language questions based on the full text content, and locate the document fragments corresponding to the answers.

[0049] The present invention will be further described below with reference to specific embodiments: S1: Implementation of the vectorization algorithm for office policy documents.

[0050] 317 office policy documents were collected from the enterprise office system. They were automatically identified and split into several knowledge units according to chapter titles. Each knowledge unit was manually reviewed and the text content was corrected. The bge-m3 model was used to perform vectorization transformation on each knowledge unit after review and correction. A vector index was built based on the text vectors of all knowledge units to construct a vector knowledge base. The vector index supports retrieval.

[0051] S2: Implementation of the LoRA large model fine-tuning algorithm.

[0052] A general-purpose large language model is loaded as the base model. The pre-trained weights of the general-purpose large language model are fixed, and low-rank matrices A and B are injected. The low-rank matrices A and B are fine-tuned and trained using office domain corpus. After training, the low-rank matrices A and B are merged with the original pre-trained weights. The merged model does not add any extra latency during inference.

[0053] S3: Implementation of the RAG intelligent question answering algorithm.

[0054] The user inputs a natural language question, which is then converted into a question vector using the bge-m3 model. The cosine similarity between the question vector and the text vectors of each knowledge unit in the vector knowledge base is calculated. Knowledge units with similarity exceeding a preset threshold are identified as retrieval fragments, enabling similarity retrieval. The retrieved relevant policy fragments are then concatenated with the original user question to construct enhanced prompts. These prompts are then input into a large-scale office domain adaptation model, which generates an accurate answer and associates it with the original policy text. The results are then output and can be viewed at a specific location.

[0055] S4: Implementation of the meeting minutes generation algorithm.

[0056] Users upload meeting recordings in mp3 / wav / m4a format. The uploaded meeting recordings undergo audio noise reduction preprocessing. The SenseVoice model is used to transcribe the noise-reduced audio into the original text, perform automatic error correction, add punctuation marks, extract key points, conclusions, and to-do items based on prompts, and generate standard meeting minutes that can be edited and exported.

[0057] S5: Implementation of document parsing and summarizing algorithms.

[0058] Users upload office documents in PDF / DOCX format. The system calls the corresponding parsing engine to extract the full text content based on the document format. The extracted full text is then processed for content structuring. Based on preset prompt word templates, the system automatically generates summaries and reviews of the full text content. Users can ask questions about the document content in natural language, and the system accurately answers and locates relevant segments based on the document content.

[0059] It also includes a step for verifying the effectiveness of the algorithm.

[0060] Specifically, this includes verifying the functions of intelligent question answering, minutes generation, and document parsing; verifying the LoRA fine-tuning efficiency and inference latency; verifying the RAG question answering accuracy and policy recall rate; verifying the compatibility of parsing multiple document formats; and ensuring that the algorithm is stable, efficient, and reusable.

[0061] The advantages and positive effects of this invention are: 1. The model of this invention is lightweight. It adopts the LoRA low-rank adaptive fine-tuning algorithm to compress the number of parameters from d×d to 2×r×d (r << d), which greatly reduces the number of training parameters and computational overhead. After training, the low-rank matrix is ​​merged with the original weights, and no additional delay is added during inference.

[0062] 2. This invention offers more accurate question-and-answer results. It employs the RAG retrieval enhancement algorithm combined with the original text of the policy for contextual enhancement, resulting in traceable and highly credible answers. The bge-m3 vectorization model supports both dense and sparse dual-path retrieval, taking into account semantic understanding and keyword matching, further improving retrieval recall and question-and-answer accuracy.

[0063] 3. This invention automates the entire process, from speech-to-text conversion and meeting minutes generation to document parsing and summary generation, all without human intervention, significantly improving office processing efficiency.

[0064] 4. This invention is highly versatile, does not depend on specific hardware and system environments, and can be quickly ported and reused in the office scenarios of different enterprises.

[0065] 5. This invention has good scalability, supports incremental updates of the vector knowledge base and multi-scenario algorithm extensions, and can flexibly adapt to ever-changing office needs.

[0066] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for intelligent processing of large models for office automation scenarios, characterized in that: Includes the following steps, S1: Decompose the office system documents into knowledge units, and use a vectorized representation model to transform each knowledge unit into a fixed-dimensional text vector; S2: The LoRA low-rank adaptive algorithm is used to fine-tune the general large language model for the office domain; S3: Accurate Q&A on Office System Knowledge Based on RAG Search Enhancement Algorithm; S4: Automatically generate meeting minutes using speech-to-text and structured text extraction algorithms; S5: Performs content parsing, summary generation, and content Q&A on multi-format office documents.

2. The method of claim 1, wherein the method is applied to an office automation scenario. In S1, the vectorization representation model is the bge-m3 vectorization model, which converts unstructured office text into fixed-dimensional text vectors, establishes a vector index based on all the text vectors, and constructs a vector knowledge base.

3. The method of claim 1 or 2, wherein the method is applied to an office automation scenario. In step S2, the pre-training weights of the general large language model are fixed, and low-rank matrices A and B are injected into the general large language model. The low-rank matrices A and B are fine-tuned using office domain corpus. After training, the low-rank matrices A and B are merged with the pre-training weights to obtain an office domain-adapted large model.

4. The method of claim 3, wherein the method is applied to an office automation scenario. The number of parameters in the low-rank matrices A and B is reduced from d×d to 2×r×d, where d is the dimension of the pre-trained weights of the general large language model, and r is the rank of the low-rank matrix and r < d / r. <d。 5. The method of claim 3, wherein the method is applied to an office automation scenario. In S3, the RAG retrieval enhancement algorithm includes obtaining a natural language question input by the user, converting the natural language question into a question vector using the vectorized representation model, performing a similarity search based on the question vector, obtaining a retrieval fragment associated with the natural language question, concatenating the retrieval fragment with the natural language question in context, inputting it into the office domain adaptation model, generating an accurate answer, and outputting the original text location information corresponding to the retrieval fragment.

6. The method of claim 5, wherein the method is applied to an office automation scenario. The similarity retrieval uses a vector similarity calculation method to calculate the similarity between the question vector and the text vector of each knowledge unit, and the knowledge units with similarity exceeding a preset threshold are identified as the retrieval segments.

7. The large model intelligent processing method for office automation scenarios according to claim 1 or 2, characterized in that: In step S4, the meeting recording file is acquired, audio noise reduction preprocessing is performed on the meeting recording file, the noise-reduced audio is transcribed into initial text using a speech-to-text model, and automatic error correction and punctuation are added to obtain the meeting text. Based on a preset prompt word template, key information is extracted from the meeting text, including meeting points, conclusions, and to-do items, and meeting minutes are generated according to a standard meeting minutes template.

8. The method of claim 7, wherein the method is applied to an office automation scenario. The speech-to-text model is the SenseVoice model; the key information extraction is based on a preset prompt word engineering algorithm, which has a built-in output template and supports constraints on output format, output length and professionalism.

9. The method of claim 1 or 2, wherein the method is applied to an office automation scenario. In step S5, multiple format office documents are acquired, the full text content of the office documents is extracted, a summary is automatically generated from the full text content, a natural language question from the user regarding the office documents is acquired, and an accurate answer is given to the natural language question based on the full text content, and the document segment corresponding to the answer is located; the format of the office documents includes, but is not limited to, PDF, DOCX, and DOC formats.

10. A large model intelligent processing algorithm model for office automation scenarios, characterized in that: The method for intelligent processing of large models for office automation scenarios as described in any one of claims 1 to 9 includes, The document vectorization representation model is used to obtain a collection of office system documents. According to the preset splitting rules, each office system document is automatically split into several knowledge units by chapter. After the text content of each knowledge unit is reviewed and corrected, the vectorization representation model is used to convert each knowledge unit into a fixed-dimensional text vector. Based on all text vectors, a vector index is built to construct a vector knowledge base. LoRA fine-tuning model is used to obtain a general large language model. The pre-training weights of the general large language model are fixed, and low-rank matrix A and low-rank matrix B are injected into the general large language model. The low-rank matrix A and low-rank matrix B are fine-tuned and trained using office domain corpus. After training, the low-rank matrix A and low-rank matrix B are merged with the pre-training weights to obtain an office domain adapted large model. The RAG retrieval enhancement question answering model is used to obtain natural language questions input by users, convert the natural language questions into question vectors using the vectorized representation model, perform similarity retrieval in the vector knowledge base based on the question vectors, obtain retrieval fragments associated with the natural language questions, and input the retrieval fragments and the natural language questions into the office domain adaptation model after context concatenation, generate accurate answers and output the original text location information corresponding to the retrieval fragments. The speech-to-text and minutes generation model is used to acquire meeting recording files, perform audio noise reduction preprocessing on the meeting recording files, use a speech-to-text model to transcribe the noise-reduced audio into initial text and perform automatic error correction and punctuation addition to obtain meeting text, extract key information from the meeting text based on a preset prompt word template, extract meeting key points, conclusions and to-do items, and generate meeting minutes according to a standard meeting minutes template; The document parsing and summarizing model is used to acquire office documents, extract the full text content of the office documents, automatically generate a summary of the full text content, acquire the natural language questions of users regarding the office documents, accurately answer the natural language questions based on the full text content, and locate the document fragments corresponding to the answers.