Enterprise-level knowledge management system based on large language model

Abstract

The invention discloses an enterprise-level knowledge management system based on a large language model. The enterprise-level knowledge management system comprises a knowledge storage module, a knowledge analysis module and a knowledge application module, the knowledge storage module is responsible for managing and storing knowledge from different sources, supporting uploading, updating, deleting and sharing of electronic documents and multimedia documents, and is provided with authority control; the knowledge analysis module is used for carrying out deep understanding and structured processing on text and multimedia contents through a special analyzer by utilizing an advanced large language model technology; the knowledge application module provides intelligent retrieval and intelligent question and answer functions and supports advanced applications such as behavior portrait analysis and standard process supervision. According to the method and the system, the acquisition, understanding and application efficiency of enterprise knowledge is remarkably improved by fusing a large language model technology and a traditional knowledge management method, an intelligent, automatic and safe knowledge management solution is provided for an enterprise, and value mining and innovative application of enterprise knowledge assets are effectively promoted.

Description

Technical Field

[0001] The present invention relates to an enterprise-level knowledge management system, and in particular to a knowledge management system based on a large language model, which is used to improve the storage, parsing and application efficiency of internal knowledge of an enterprise. Specifically, the present invention provides an integrated system framework that can process various types of knowledge documents, including but not limited to electronic documents and multimedia documents, and enhance the utilization value of knowledge through intelligent analysis and processing technology. Background Art

[0002] With the development of information technology, enterprises are facing the challenges of massive data management and knowledge accumulation. Traditional knowledge management systems are often limited to simple file storage and retrieval functions, lacking the ability to understand complex knowledge structures and efficient application methods. In recent years, the advancement of natural language processing (NLP) technology and machine learning algorithms has provided new solutions for knowledge management. In particular, the emergence of large language models enables the system to better understand and generate human language, thereby greatly improving the intelligence level of knowledge management systems.

[0003] The present invention is an innovative solution proposed in this context, which aims to combine the latest large language model technology with traditional knowledge management methods to provide enterprises with a more intelligent, flexible and secure knowledge management platform. The system not only supports efficient storage and access control of different types of knowledge documents, but also has powerful knowledge parsing capabilities and can automatically extract useful information from unstructured data. In addition, by introducing advanced retrieval, question-answering and behavior analysis and other application functions, the application scenarios of knowledge are further enhanced, helping enterprises optimize decision-making processes and improve overall operational efficiency. Summary of the invention

[0004] Purpose of the invention: The purpose of the present invention is to provide an enterprise-level knowledge management system based on a large language model, which improves the efficiency and quality of internal knowledge management in an enterprise by optimizing the knowledge storage, parsing and application processes. Specifically, the present invention aims to solve the following problems:

[0005] Knowledge storage and management: Currently, enterprises are faced with the challenge of storing and managing a large number of electronic documents and multimedia documents. The present invention aims to provide an efficient and secure storage module and a flexible authority management mechanism to ensure the security and usability of knowledge assets.

[0006] Knowledge parsing: Parsing documents of different formats and extracting valid information is a complex and time-consuming process. The present invention aims to automatically extract and integrate knowledge through advanced electronic document and multimedia document parsers for further application.

[0007] Knowledge application: Enterprises have an increasing demand for knowledge application, including intelligent retrieval and intelligent question and answer. The present invention aims to improve the accuracy and efficiency of knowledge retrieval through intelligent retrieval technology, including keyword retrieval, vector retrieval and hybrid retrieval, and intelligent question and answer system, including knowledge base question and answer and intelligent agent question and answer.

[0008] Multimedia document application: Multimedia documents contain rich unstructured information. The present invention aims to deeply mine and utilize the knowledge in multimedia documents through behavioral profiling analysis and standard process supervision to support more comprehensive enterprise decision-making and process optimization.

[0009] The present invention specifically provides an enterprise-level knowledge management system based on a large language model (an artificial intelligence model with a large number of parameters, the number of which is usually more than billions), including a knowledge storage module, a knowledge parsing module and a knowledge application module;

[0010] The knowledge storage module includes a knowledge authority management unit and a knowledge source management unit. The knowledge authority management unit is used to manage and store knowledge from different sources, and the knowledge source management unit is used to manage electronic documents and multimedia documents.

[0011] The knowledge parsing module includes an electronic document parser and a multimedia document parser, wherein the electronic document parser uses a large language model technology to process documents, extract key information and generate an index for search;

[0012] The multimedia document parser is used to parse multimedia documents in audio and video formats, convert speech into text, and extract information from the video using image recognition technology;

[0013] The knowledge application module includes an electronic document application unit and a multimedia document application unit. The electronic document application unit is used to provide intelligent retrieval and intelligent question-and-answer functions; the multimedia document application unit is used to perform behavioral profiling analysis and standard process supervision of multimedia documents.

[0014] The knowledge rights management unit supports the uploading, updating, deleting and sharing of electronic documents (such as PDF, Word, Markdown, TXT, HTML) and multimedia documents (audio and video files), and is provided with rights control to ensure information security: the knowledge rights management unit adopts a fine-grained rights control mechanism combining role-based access control (RBAC) and attribute-based access control (ABAC) to achieve document-level access control;

[0015] The system supports hierarchical management of permissions based on user roles, departments, job levels, etc., and can dynamically adjust permission policies;

[0016] The knowledge rights management unit uses TLS / SSL encryption protocol during data transmission (SSL / TLS protocol is a protocol used to provide secure communications on the Internet. SSL stands for Secure Sockets Layer, and TLS stands for Transport Layer Security) to ensure data transmission security; for sensitive data stored in the system, such as company strategy documents, proprietary technical details, customer confidential information, financial statements, etc., the system uses a private file cabinet mechanism for encrypted storage, uses encryption algorithms based on industry standards to encrypt sensitive data, and implements key segmentation management to further improve data security. In addition, the system has also established a data backup and recovery mechanism to respond to emergencies and data losses, and ensure the integrity and availability of key knowledge of the enterprise;

[0017] The knowledge rights management unit has a comprehensive operation record function, which permanently records the operations of the entire knowledge management process. The recorded content includes the operation time, operation type, operator identity, operation content and operation results. The recorded content is encrypted and stored to ensure that it cannot be tampered with, and is archived and backed up regularly;

[0018] The knowledge authority management unit integrates a behavior monitoring system to monitor user behavior in real time and identify abnormal operations; the system regularly detects overlapping authorities, omissions and lack of authorities, and optimizes authority allocation through authority conflict detection mechanisms and regular authority reviews; at the same time, the system establishes a security incident response mechanism to provide real-time alarms, emergency response and post-event tracing functions.

[0019] The formats of the electronic documents include PDF, Word, Markdown, TXT and HTML, and the formats of the multimedia documents include mp3, m4a, aac, wma, mp4 and wav.

[0020] The intelligent search function includes keyword search, vector search and hybrid search, and is sorted according to the relevance score to the question raised by the user; the hybrid search is a combination of BM25 search and vector search.

[0021] The intelligent question-answering function includes knowledge base question-answering and intelligent agent question-answering. It parses, blocks, and vectorizes documents and stores them in a vector database. It also performs intent recognition, intelligent retrieval, and rearrangement on user questions, and selects the top k (generally 5 to 10) document blocks to be sent to the large language model for answering.

[0022] The behavior portrait analysis includes: processing audio and video files through voice-to-text conversion, identifying speaker information and timestamps, and using a Voice Activity Detection (VAD) model to process long audio and video, and after writing prompt words, sending the chat record vector into a large language model to extract customer portraits;

[0023] The standard process supervision includes: dividing the chat records into blocks, vectorizing them, writing prompt words, and then sending the chat record vectors into a large language model for supervision.

[0024] The knowledge base question answering comprises the following steps:

[0025] Step 1-1, loading documents: using an electronic document parser to load documents of different formats (such as PDF, Word, HTML, etc.), the electronic document parser can recognize and process the content of the document and convert the document into a unified text format, specifically including the following steps:

[0026] Step 1-1-1: Intelligently identify document formats: The system uses a deep learning model to analyze file header information, content features and metadata to identify document formats and achieve automatic diversion processing;

[0027] Step 1-1-2: Multi-channel parallel parsing: According to the recognition results of the document format, different types of documents are assigned to the corresponding electronic document parsers to achieve parallel processing and greatly improve processing efficiency;

[0028] Step 1-1-3: Text content extraction and structuring: For text documents, use natural language processing technology to extract text content while retaining the document's hierarchical structure, format information, and semantic relationships; for imaged PDF or scanned documents, integrate OCR technology (Optical Character Recognition) to achieve high-precision text recognition and extraction; for multimedia documents, use speech recognition and video content analysis technology to extract text information from audio and video;

[0029] Step 1-1-4: Semantic understanding and knowledge graph construction: Use deep learning models to understand the extracted text content, identify key concepts, entity relationships, and topic structures, and build a document-level knowledge graph;

[0030] Step 1-1-5: Intelligent document segmentation: Based on the semantic understanding results, the document is intelligently segmented into semantically complete and logically independent paragraphs or sentences, laying the foundation for subsequent precise retrieval and question-answering;

[0031] Step 1-1-6: Multimodal vectorization: Use the pre-trained multimodal embedding model to fuse text content, hierarchical structure, format information, semantic relations, and multimedia features to generate a unified vector representation that captures the comprehensive semantic information of the document;

[0032] Step 1-1-7: Incremental knowledge update: vectorize the newly added document knowledge and add it to the vector library corresponding to the current document, process the newly added documents in real time, dynamically update the knowledge base, and maintain the timeliness and integrity of the system knowledge;

[0033] Step 1-2, document chunking: Chunk the loaded document into smaller, logically independent paragraphs or sentences. This step helps improve processing efficiency and ensures that each chunk can be processed and retrieved separately. It includes the following steps:

[0034] Step 1-2-1: Document structure analysis: Use natural language processing technology to perform structural analysis on the document and identify structural elements such as titles, paragraphs, and lists;

[0035] Step 1-2-2: Semantic unit identification: Use a deep learning model to perform semantic analysis on the document content and identify the smallest semantically complete unit;

[0036] Step 1-2-3: Perform multi-scale blocking, including large-scale blocking, medium-scale blocking and small-scale blocking;

[0037] The large-scale segmentation includes: based on the document structure, according to the chapter and paragraph marks (such as the title directory of the document), using the chapter and paragraph marks as separators, and after reading the separators, preliminarily segmenting the document into a large chapter or paragraph;

[0038] The small-scale block segmentation is to subdivide the result of the large-scale block segmentation: firstly, the chapters or paragraphs obtained by the large-scale block segmentation are divided into several sentences according to separators such as ".;!?", and the sentences are also added with a metadata to mark which large chapter the sentences belong to; secondly, the cosine similarity algorithm is used to determine whether there is a connection between the sentences in each chapter. If there is a connection, the two sentences are spliced ​​to obtain a longer small paragraph, and finally a semantically complete sentence or phrase is obtained;

[0039] Step 1-2-4: Dynamically adjust the chunk size: Dynamically adjust the chunk size based on the document type, content complexity, and system performance requirements; specifically: For highly structured documents (such as technical manuals), prefer smaller chunks (50-100 words). For narrative documents (such as FAQs), increase the chunk size appropriately (100-200 words). For complex academic documents, use larger chunks (200-300 words).

[0040] Step 1-2-5: Semantic relevance evaluation: Use the semantic similarity algorithm to calculate the semantic relevance between two adjacent blocks. If the semantic relevance is higher than a preset threshold (such as 0.8), merge the two adjacent blocks.

[0041] Step 1-2-6: Build a block index: Generate a unique identifier for each block and establish a mapping relationship with the original document to facilitate subsequent retrieval and backtracking;

[0042] Step 1-2-7: Block quality control: Design a quality assessment model to evaluate the block results; the assessment indicators include semantic integrity, logical independence, and information density; if a block does not meet the quality requirements, the system will automatically adjust the block strategy and re-execute the block process;

[0043] Step 1-3, document vectorization: Use Youdao's open-source pre-trained embedding model BCE-Embedding, which performs well in Chinese and English and cross-language, ensuring the semantic representation ability of documents in general fields; use the pre-trained embedding model to vectorize document blocks, each document block is converted into a fixed-length vector representation, and a document vector is obtained, which captures the semantic information of the document block;

[0044] Step 1-4, storing document vectors: storing the document vectors obtained in step 1-3 in a vector database. The vector database is an efficient vector search engine that can quickly retrieve similar vectors, thereby achieving efficient document retrieval;

[0045] Steps 1-5, processing user questions: Based on the customer data accumulation, a part of the data is extracted to train the intent recognition model. The model uses the Bert model as the base, and its training data format is generally (user question, yes / no). After obtaining the user's question, the intent recognition model will first determine whether the user's question is meaningful; after determining that the user's question is meaningful, the system will combine the current knowledge base document and use the large language model to determine whether the user's question can be answered in the knowledge base; then, the pre-trained embedding model is used to convert the user's question into a vector form to obtain the user's question vector, so that more similar related fragments can be recalled in the vectorized document vector;

[0046] Step 1-6, intelligent retrieval: perform intelligent retrieval in the vector database, calculate the cosine similarity between the document vector and the user question vector, thereby obtaining the similarity score of each document vector, sort all document vectors from high to low according to the similarity score with the user question vector, and recall all the document vectors. The calculation formula is:

[0047]

[0048] Among them, A i represents the i-th document vector, B i represents the i-th user question vector, n is the total number of document vectors, P i is the similarity score;

[0049] Steps 1-7, re-ranking related documents: Use the open source re-ranking model BGE-Reranker. The Reranker model can analyze the correlation between questions and documents very well through specially trained neural networks. Through steps 1-6, the sorted document vectors are obtained. Considering that the cosine similarity score is effective in most cases, but it does not work well for sparse vectors, it is necessary to load the re-ranking model to sort the document vectors after the documents are sorted, ensuring that the most relevant document blocks are ranked first;

[0050] Step 1-8, select the best document: select the top k document vectors with the highest similarity from the re-ranked document vectors as the best documents. The top K document vectors are the information sources that best match the user's question.

[0051] Step 1-9, generate answer: input the user's original question, best document and prompt information into the large language model, and the large language model generates the final answer and provides it to the user.

[0052] The intelligent agent question answering comprises the following steps:

[0053] Step 2-1, user problem decomposition and vectorization, includes the following steps:

[0054] Step 2-1-1, decomposing the problem: After identifying the user's question intention and determining whether it is meaningful, the ReAct framework will be used to decompose the user's question. First, each document in the knowledge base is set as a tool. After receiving the user's question, the large language model will consider what steps to take in response to the previous input, and then select one of several tools to operate and generate parameters for the tool, thereby splitting the user's question into several small sub-problems; this step aims to break down complex problems into multiple simpler and more specific sub-problems to ensure that each sub-problem can be handled independently and accurately understood.

[0055] Step 2-1-2, question vectorization: Use the pre-trained embedding model to vectorize each sub-question after decomposition. Each sub-question is converted into a fixed-length vector representation to obtain a sub-question vector. These vectors capture the semantic information of the sub-question for subsequent retrieval and matching.

[0056] Step 2-2, obtaining the document fragment corresponding to each sub-question, includes the following steps:

[0057] Step 2-2-1, loading documents: loading relevant documents stored in the system, the relevant documents including FAQ, manuals, technical documents, etc.; using an electronic document parser to convert documents in different formats (such as PDF, Word, HTML, etc.) into a unified text format;

[0058] Step 2-2-2, document chunking: split the loaded document into logically independent paragraphs or sentences; this step helps improve processing efficiency and ensures that each document chunk can be processed and retrieved separately;

[0059] Step 2-2-3, document vectorization, uses the pre-trained embedding model to vectorize the document blocks. Each document block is converted into a fixed-length vector representation to obtain document vectors, which capture the semantic information of the document blocks.

[0060] Step 2-2-4, storing document vectors: storing the document vectors obtained in step 2-2-3 in a vector database;

[0061] Step 2-2-5, intelligent retrieval: perform intelligent retrieval in the vector database, calculate the similarity between the document vector and the sub-question vector, take the document vector with the highest similarity to the sub-question vector as the most relevant document vector, and recall the most relevant document vector.

[0062] Step 2-2-6, re-rank relevant documents: Use the re-ranking model to re-rank the recalled document vectors; the re-ranking model is fine-tuned according to the relevance of the document block to the sub-question to ensure that the most relevant document block is ranked first;

[0063] Step 2-2-7, select the best document: select the top k document vectors with the highest relevance from the re-ranked document vectors as the information source that best matches the sub-question;

[0064] Step 2-3, integrating the answers to two or more sub-questions to generate the final answer, includes the following steps:

[0065] Step 2-3-1, generate answers to sub-questions: Input each sub-question and the corresponding document fragment into the large language model to generate specific answers for each sub-question. The large language model generates detailed and accurate answers.

[0066] Step 2-3-2, generate the final answer: Integrate each sub-question and its corresponding specific answer, as well as the unsplit user question, into the pre-written prompt words, load them into the large language model together, and generate the user answer; the final answer not only contains the answers to all sub-questions, but also uses the prompt word engineering to ensure the overall fluency and consistency of the answer.

[0067] The behavior profiling analysis includes the following steps:

[0068] Step 3-1, speech-to-text: The system uses two speech-to-text models. One is the Mandarin model, and the other is the dialect model. The Mandarin model uses Alibaba's series of open source models FunASR, which include speech-to-text, speaker recognition, and timestamp prediction; the dialect model is the Wuxi dialect model trained by the Qiaozhiyin team of Jiangsu Yongshan Qiao Archives Management Service Co., Ltd. This model was trained locally by collecting 21 hours of Wuxi dialect. Both of the above speech-to-text models can process audio and video files of less than 2 hours. Among them, the cross entropy loss function l used by the dialect model pred for:

[0069]

[0070] in represents the concatenation result of the kth document block, P k is the generation probability, t k is the one-hot vector of the number of tokens in the k-th document block, W 1 , W 2 is the weight, b 1 、b 2 is bias;

[0071] Step 3-2, chat record segmentation: The speech-to-text result obtained in step 3-1 is segmented into blocks; due to the context length limit of the model, when processing chat records that exceed the context, they will be truncated into several dialogue blocks within a fixed length range (generally 1204 to 2048 characters) according to the speaker identity information to ensure that the correlation between different blocks is not affected;

[0072] Step 3-3: Write prompt words for extracting customer portraits based on business needs and analysis objectives, and specify the key information and features that need to be extracted from chat records. For example, in the banking field, the customer portraits required and analyzed include "assets" and "investment preferences", and the prompt words written tend to extract these tags based on chat records.

[0073] In step 3-4, the conversation block is converted into a vector representation using the Embedding model, and the prompt words in step 3-3 are sent together with the vector representation into the large language model. The powerful natural language processing capabilities of the large language model are used to extract the customer's portrait information from the chat history, including the customer's behavior patterns, preferences, and needs.

[0074] The standard process supervision includes the following steps:

[0075] Step 4-1, chat record segmentation: The chat records in the multimedia document are segmented. Due to the context length limit of the model, when processing chat records that exceed the context, they will be truncated within a fixed length range according to the speaker identity information to ensure that the correlation between different segments is not affected for subsequent analysis;

[0076] Step 4-2, chat record vectorization: Use the Embedding model to convert the conversation blocks into vector representations for easy processing and analysis by machine learning models;

[0077] Step 4-3, according to the company's standard process and supervision needs, write prompt words for supervising employees, clarify the key information and features that need to be extracted from the chat records; at the same time, define the format of the extraction results to ensure the structuring and consistency of the supervision results. For example, bank employees are not allowed to use words of impatience during communication with customers, and when writing prompt words, they will tend to be in this direction, that is, to determine whether the employee has said impatient words. At the same time, the system will also write relevant prompts for returning the original text and reasons in the prompt words, and return a triple such as (whether impatience appears, the original text that appears, and the reason for judgment);

[0078] Step 4-4: Send the vector representation obtained in step 4-2 and the prompt words written in step 4-3 to the large language model. Utilize the powerful natural language processing capabilities of the large language model to monitor the behavior of employees from the chat records, identify whether it complies with standard procedures, and generate a supervision report.

[0079] The present invention has the following beneficial effects:

[0080] (1) Improving knowledge management efficiency: By integrating the knowledge storage module and the knowledge parsing module, the present invention can automatically process and store documents in various formats, greatly reducing the time and effort of manual document processing and improving the overall efficiency of knowledge management.

[0081] (2) Enhanced knowledge management security: Data is encrypted during transmission to prevent theft by others. Fine-grained permission control ensures that knowledge assets are not operated without authorization. For sensitive and important knowledge, enterprises can choose to encrypt and store it in private filing cabinets, thereby better protecting the information security of corporate knowledge.

[0082] (3) Operation traceability: Operation records are permanently retained for each process and link of knowledge management, effectively ensuring the compliance of operations and the integrity of data.

[0083] (4) Optimizing the knowledge retrieval experience: Intelligent retrieval functions, including keyword retrieval, vector retrieval, and hybrid retrieval, enable users to find the knowledge they need more quickly. In addition, the intelligent question-answering system provides a more natural and efficient way of interaction, allowing users to get answers directly by asking questions.

[0084] (5) Improving the value of knowledge application: Through the knowledge application module, the present invention can support the in-depth application of electronic documents and multimedia documents, such as behavioral profiling analysis and standard process supervision, helping enterprises to better utilize knowledge for decision support and process optimization.

[0085] (6) Promote knowledge sharing and innovation: The sharing mechanism of the present invention encourages the circulation of knowledge within the organization and promotes cross-departmental and cross-team cooperation. This knowledge sharing and exchange helps stimulate innovative thinking and promote the progress and development of corporate culture.

[0086] (7) Reduce maintenance costs: By using lightweight local large language models and advanced parsing technology, the present invention reduces dependence on external professional tools and services, and reduces the cost of long-term maintenance and updates.

[0087] (8) Improving system scalability: The architecture design of the present invention takes into account future expansion needs, allowing the system to easily integrate new technologies and functions to adapt to the ever-changing needs of the enterprise. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and other advantages of the present invention will become more clear.

[0089] Figure 1 This is the structural composition diagram of this system.

[0090] Figure 2 The figure is a flowchart of the system of the present invention. DETAILED DESCRIPTION

[0091] The present invention specifically provides an enterprise-level knowledge management system based on a large language model (an artificial intelligence model with a large number of parameters, the number of which is usually more than billions), including a knowledge storage module, a knowledge parsing module and a knowledge application module;

[0092] The knowledge storage module includes a knowledge authority management unit and a knowledge source management unit. The knowledge authority management unit is used to manage and store knowledge from different sources, and the knowledge source management unit is used to manage electronic documents and multimedia documents.

[0093] The knowledge parsing module includes an electronic document parser and a multimedia document parser, wherein the electronic document parser uses a large language model technology to process documents, extract key information and generate an index for search;

[0094] The multimedia document parser is used to parse multimedia documents in audio and video formats, convert speech into text, and extract information from the video using image recognition technology;

[0095] The knowledge application module includes an electronic document application unit and a multimedia document application unit. The electronic document application unit is used to provide intelligent retrieval and intelligent question-and-answer functions; the multimedia document application unit is used to perform behavioral profiling analysis and standard process supervision of multimedia documents.

[0096] The knowledge rights management unit supports the uploading, updating, deleting and sharing of electronic documents (such as PDF, Word, Markdown, TXT, HTML) and multimedia documents (audio and video files), and is provided with rights control to ensure information security: the knowledge rights management unit adopts a fine-grained rights control mechanism combining role-based access control (RBAC) and attribute-based access control (ABAC) to achieve document-level access control;

[0097] The system supports hierarchical management of permissions based on user roles, departments, job levels, etc., and can dynamically adjust permission policies;

[0098] The knowledge rights management unit uses TLS / SSL encryption protocol during data transmission (SSL / TLS protocol is a protocol used to provide secure communications on the Internet. SSL stands for Secure Sockets Layer, and TLS stands for Transport Layer Security) to ensure data transmission security; for sensitive data stored in the system, such as company strategy documents, proprietary technical details, customer confidential information, financial statements, etc., the system uses a private file cabinet mechanism for encrypted storage, encrypts sensitive data based on industry-standard encryption algorithms, and further improves data security through key segmentation management. In addition, the system has also established a data backup and recovery mechanism to respond to emergencies and data losses, and ensure the integrity and availability of key knowledge of the enterprise;

[0099] The knowledge rights management unit has a comprehensive operation record function, which permanently records the operations of the entire knowledge management process. The recorded content includes the operation time, operation type, operator identity, operation content and operation results. The recorded content is encrypted and stored to ensure that it cannot be tampered with, and is archived and backed up regularly;

[0100] The knowledge authority management unit integrates a behavior monitoring system to monitor user behavior in real time and identify abnormal operations; the system regularly detects overlapping authorities, omissions and lack of authorities, and optimizes authority allocation through authority conflict detection mechanisms and regular authority reviews; at the same time, the system establishes a security incident response mechanism to provide real-time alarms, emergency response and post-event tracing functions.

[0101] The formats of the electronic documents include PDF, Word, Markdown, TXT and HTML, and the formats of the multimedia documents include mp3, m4a, aac, wma, mp4 and wav.

[0102] The intelligent search function includes keyword search, vector search and hybrid search, and is sorted according to the relevance score to the question raised by the user; the hybrid search is a combination of BM25 search and vector search.

[0103] The intelligent question-answering function includes knowledge base question-answering and intelligent agent question-answering. It parses, blocks, and vectorizes documents and stores them in a vector database. It also performs intent recognition, intelligent retrieval, and rearrangement on user questions, and selects the top k (generally 5 to 10) document blocks to be sent to the large language model for answering.

[0104] The behavior portrait analysis includes: processing audio and video files through voice-to-text conversion, identifying speaker information and timestamps, and using a Voice Activity Detection (VAD) model to process long audio and video, and after writing prompt words, sending the chat record vector into a large language model to extract customer portraits;

[0105] The standard process supervision includes: dividing the chat records into blocks, vectorizing them, writing prompt words, and then sending the chat record vectors into a large language model for supervision.

[0106] The knowledge base question answering comprises the following steps:

[0107] Step 1-1, loading documents: using an electronic document parser to load documents of different formats (such as PDF, Word, HTML, etc.), the electronic document parser can recognize and process the content of the document and convert the document into a unified text format, specifically including the following steps:

[0108] Step 1-1-1: Intelligently identify document formats: The system uses a deep learning model to analyze file header information, content features and metadata to identify document formats and achieve automatic diversion processing;

[0109] Step 1-1-2: Multi-channel parallel parsing: According to the recognition results of the document format, different types of documents are assigned to the corresponding electronic document parsers to achieve parallel processing and greatly improve processing efficiency;

[0110] Step 1-1-3: Text content extraction and structuring: For text documents, use natural language processing technology to extract text content while retaining the document's hierarchical structure, format information, and semantic relationships; for imaged PDF or scanned documents, integrate OCR technology (Optical Character Recognition) to achieve high-precision text recognition and extraction; for multimedia documents, use speech recognition and video content analysis technology to extract text information from audio and video;

[0111] Step 1-1-4: Semantic understanding and knowledge graph construction: Use deep learning models to understand the extracted text content, identify key concepts, entity relationships, and topic structures, and build a document-level knowledge graph;

[0112] Step 1-1-5: Intelligent document segmentation: Based on the semantic understanding results, the document is intelligently segmented into semantically complete and logically independent paragraphs or sentences, laying the foundation for subsequent precise retrieval and question-answering;

[0113] Step 1-1-6: Multimodal vectorization: Use the pre-trained multimodal embedding model to fuse text content, hierarchical structure, format information, semantic relations, and multimedia features to generate a unified vector representation that captures the comprehensive semantic information of the document;

[0114] Step 1-1-7: Incremental knowledge update: vectorize the newly added document knowledge and add it to the vector library corresponding to the current document, process the newly added documents in real time, dynamically update the knowledge base, and maintain the timeliness and integrity of the system knowledge;

[0115] Step 1-2, document chunking: Chunk the loaded document into smaller, logically independent paragraphs or sentences. This step helps improve processing efficiency and ensures that each chunk can be processed and retrieved separately. It includes the following steps:

[0116] Step 1-2-1: Document structure analysis: Use natural language processing technology to perform structural analysis on the document and identify structural elements such as titles, paragraphs, and lists;

[0117] Step 1-2-2: Semantic unit identification: Use a deep learning model to perform semantic analysis on the document content and identify the smallest semantically complete unit;

[0118] Step 1-2-3: Perform multi-scale blocking, including large-scale blocking, medium-scale blocking and small-scale blocking;

[0119] The large-scale segmentation includes: based on the document structure, according to the chapter and paragraph marks (such as the title directory of the document), using the chapter and paragraph marks as separators, and after reading the separators, preliminarily segmenting the document into a large chapter or paragraph;

[0120] The small-scale block segmentation is to subdivide the result of the large-scale block segmentation: firstly, the chapters or paragraphs obtained by the large-scale block segmentation are divided into several sentences according to separators such as ".;!?", and the sentences are also added with a metadata to mark which large chapter the sentences belong to; secondly, the cosine similarity algorithm is used to determine whether there is a connection between the sentences in each chapter. If there is a connection, the two sentences are spliced ​​to obtain a longer small paragraph, and finally a semantically complete sentence or phrase is obtained;

[0121] Step 1-2-4: Dynamically adjust the chunk size: Dynamically adjust the chunk size based on the document type, content complexity, and system performance requirements; specifically: For highly structured documents (such as technical manuals), prefer smaller chunks (50-100 words). For narrative documents (such as FAQs), increase the chunk size appropriately (100-200 words). For complex academic documents, use larger chunks (200-300 words).

[0122] Step 1-2-5: Semantic relevance evaluation: Use the semantic similarity algorithm to calculate the semantic relevance between two adjacent blocks. If the semantic relevance is higher than a preset threshold (such as 0.8), merge the two adjacent blocks.

[0123] Step 1-2-6: Build a block index: Generate a unique identifier for each block and establish a mapping relationship with the original document to facilitate subsequent retrieval and backtracking;

[0124] Step 1-2-7: Block quality control: Design a quality assessment model to evaluate the block results; the assessment indicators include semantic integrity, logical independence, and information density; if a block does not meet the quality requirements, the system will automatically adjust the block strategy and re-execute the block process;

[0125] Step 1-3, document vectorization: Use Youdao's open-source pre-trained embedding model BCE-Embedding, which performs well in Chinese and English and cross-language, ensuring the semantic representation ability of documents in general fields; use the pre-trained embedding model to vectorize document blocks, each document block is converted into a fixed-length vector representation, and a document vector is obtained, which captures the semantic information of the document block;

[0126] Step 1-4, storing document vectors: storing the document vectors obtained in step 1-3 in a vector database. The vector database is an efficient vector search engine that can quickly retrieve similar vectors, thereby achieving efficient document retrieval;

[0127] Steps 1-5, processing user questions: Based on the customer data accumulation, a part of the data is extracted to train the intent recognition model. The model uses the Bert model as the base, and its training data format is generally (user question, yes / no). After obtaining the user's question, the intent recognition model will first determine whether the user's question is meaningful; after determining that the user's question is meaningful, the system will combine the current knowledge base document and use the large language model to determine whether the user's question can be answered in the knowledge base; then, the pre-trained embedding model is used to convert the user's question into a vector form to obtain the user's question vector, so that more similar related fragments can be recalled in the vectorized document vector;

[0128] Step 1-6, intelligent retrieval: perform intelligent retrieval in the vector database, calculate the cosine similarity between the document vector and the user question vector, thereby obtaining the similarity score of each document vector, sort all document vectors from high to low according to the similarity score with the user question vector, and recall all the document vectors. The calculation formula is:

[0129]

[0130] Among them, A i represents the i-th document vector, B i represents the i-th user question vector, n is the total number of document vectors, P i is the similarity score;

[0131] Steps 1-7, re-ranking related documents: Use the open source re-ranking model BGE-Reranker. The Reranker model can analyze the correlation between questions and documents very well through specially trained neural networks. Through steps 1-6, the sorted document vectors are obtained. Considering that the cosine similarity score is effective in most cases, but it does not work well for sparse vectors, it is necessary to load the re-ranking model to sort the document vectors after the documents are sorted, ensuring that the most relevant document blocks are ranked first;

[0132] Step 1-8, select the best document: select the top k document vectors with the highest similarity from the re-ranked document vectors as the best documents. The top K document vectors are the information sources that best match the user's question.

[0133] Step 1-9, generate answer: input the user's original question, best document and prompt information into the large language model, and the large language model generates the final answer and provides it to the user.

[0134] The intelligent agent question answering comprises the following steps:

[0135] Step 2-1, user problem decomposition and vectorization, includes the following steps:

[0136] Step 2-1-1, decomposing the problem: After identifying the user's question intention and determining whether it is meaningful, the ReAct framework will be used to decompose the user's question. First, each document in the knowledge base is set as a tool. After receiving the user's question, the large language model will consider what steps to take in response to the previous input, and then select one of several tools to operate and generate parameters for the tool, thereby splitting the user's question into several small sub-problems; this step aims to break down complex problems into multiple simpler and more specific sub-problems to ensure that each sub-problem can be handled independently and accurately understood.

[0137] Step 2-1-2, question vectorization: Use the pre-trained embedding model to vectorize each sub-question after decomposition. Each sub-question is converted into a fixed-length vector representation to obtain a sub-question vector. These vectors capture the semantic information of the sub-question for subsequent retrieval and matching.

[0138] Step 2-2, obtaining the document fragment corresponding to each sub-question, includes the following steps:

[0139] Step 2-2-1, loading documents: loading relevant documents stored in the system, the relevant documents including FAQ, manuals, technical documents, etc.; using an electronic document parser to convert documents in different formats (such as PDF, Word, HTML, etc.) into a unified text format;

[0140] Step 2-2-2, document chunking: split the loaded document into logically independent paragraphs or sentences; this step helps improve processing efficiency and ensures that each document chunk can be processed and retrieved separately;

[0141] Step 2-2-3, document vectorization, uses the pre-trained embedding model to vectorize the document blocks. Each document block is converted into a fixed-length vector representation to obtain document vectors, which capture the semantic information of the document blocks.

[0142] Step 2-2-4, storing document vectors: storing the document vectors obtained in step 2-2-3 in a vector database;

[0143] Step 2-2-5, intelligent retrieval: perform intelligent retrieval in the vector database, calculate the similarity between the document vector and the sub-question vector, take the document vector with the highest similarity to the sub-question vector as the most relevant document vector, and recall the most relevant document vector.

[0144] Step 2-2-6, re-rank relevant documents: Use the re-ranking model to re-rank the recalled document vectors; the re-ranking model is fine-tuned according to the relevance of the document block to the sub-question to ensure that the most relevant document block is ranked first;

[0145] Step 2-2-7, select the best document: select the top k document vectors with the highest relevance from the re-ranked document vectors as the information source that best matches the sub-question;

[0146] Step 2-3, integrating the answers to two or more sub-questions to generate the final answer, includes the following steps:

[0147] Step 2-3-1, generate answers to sub-questions: Input each sub-question and the corresponding document fragment into the large language model to generate specific answers for each sub-question. The large language model generates detailed and accurate answers.

[0148] Step 2-3-2, generate the final answer: Integrate each sub-question and its corresponding specific answer, as well as the unsplit user question, into the pre-written prompt words, load them into the large language model together, and generate the user answer; the final answer not only contains the answers to all sub-questions, but also uses the prompt word engineering to ensure the overall fluency and consistency of the answer.

[0149] The behavior profiling analysis includes the following steps:

[0150] Step 3-1, speech-to-text: The system uses two speech-to-text models. One is the Mandarin model, and the other is the dialect model. The Mandarin model uses Alibaba's series of open source models FunASR, which include speech-to-text, speaker recognition, and timestamp prediction; the dialect model is the Wuxi dialect model trained by the Qiaozhiyin team of Jiangsu Yongshan Qiao Archives Management Service Co., Ltd. This model was trained locally by collecting 21 hours of Wuxi dialect. Both of the above speech-to-text models can process audio and video files of less than 2 hours. Among them, the cross entropy loss function l used by the dialect model pred for:

[0151]

[0152] in represents the concatenation result of the kth document block, P k is the generation probability, t k is the one-hot vector of the number of tokens in the k-th document block, W 1 , W 2 is the weight, b 1 、b 2 is bias;

[0153] Step 3-2, chat record segmentation: The speech-to-text result obtained in step 3-1 is segmented into blocks; due to the context length limit of the model, when processing chat records that exceed the context, they will be truncated into several dialogue blocks within a fixed length range (generally 1204 to 2048 characters) according to the speaker identity information to ensure that the correlation between different blocks is not affected;

[0154] Step 3-3: Write prompt words for extracting customer portraits based on business needs and analysis objectives, and specify the key information and features that need to be extracted from chat records. For example, in the banking field, the customer portraits required and analyzed include "assets" and "investment preferences", and the prompt words written tend to extract these tags based on chat records.

[0155] In step 3-4, the conversation block is converted into a vector representation using the Embedding model, and the prompt words in step 3-3 are sent together with the vector representation into the large language model. The powerful natural language processing capabilities of the large language model are used to extract the customer's portrait information from the chat history, including the customer's behavior patterns, preferences, and needs.

[0156] The standard process supervision includes the following steps:

[0157] Step 4-1, chat record segmentation: The chat records in the multimedia document are segmented. Due to the context length limit of the model, when processing chat records that exceed the context, they will be truncated within a fixed length range according to the speaker identity information to ensure that the correlation between different segments is not affected for subsequent analysis;

[0158] Step 4-2, chat record vectorization: Use the Embedding model to convert the conversation blocks into vector representations for easy processing and analysis by machine learning models;

[0159] Step 4-3, according to the company's standard process and supervision needs, write prompt words for supervising employees, clarify the key information and features that need to be extracted from the chat records; at the same time, define the format of the extraction results to ensure the structuring and consistency of the supervision results. For example, bank employees are not allowed to use words of impatience during communication with customers, and when writing prompt words, they will tend to be in this direction, that is, to determine whether the employee has said impatient words. At the same time, the system will also write relevant prompts for returning the original text and reasons in the prompt words, and return a triple such as (whether impatience appears, the original text that appears, and the reason for judgment);

[0160] Step 4-4: Send the vector representation obtained in step 4-2 and the prompt words written in step 4-3 to the large language model. Utilize the powerful natural language processing capabilities of the large language model to monitor the behavior of employees from the chat records, identify whether it complies with standard procedures, and generate a supervision report.

[0161] In an embodiment of the present invention, an enterprise-level knowledge management system based on a large language model is provided. Figure 1 The following is a structural diagram of the system, including: knowledge storage module, knowledge analysis module, and knowledge application module. The knowledge storage module is used to manage and store knowledge from different sources, supports uploading, updating, deleting and sharing of electronic documents (such as PDF, Word, Markdown, TXT, HTML) and multimedia documents (such as MP3, M4A, AAC, WMA, MP4, WAV and other audio and video documents), and has permission control to ensure information security.

[0162] The knowledge parsing module is used to utilize advanced large language model technology to process documents in PDF, Word, Markdown, TXT, HTML and other formats through a specialized electronic document parser, extract key information and generate a searchable index; process audio and video files through a multimedia document parser, convert speech into text, and use image recognition technology to extract important information from the video.

[0163] The knowledge application module is used to provide intelligent retrieval and intelligent question-and-answer functions. Intelligent retrieval integrates keyword retrieval, vector retrieval and hybrid retrieval (BM25 retrieval + vector retrieval), and is sorted according to relevance scores; intelligent question-and-answer is divided into knowledge base question-and-answer and intelligent agent question-and-answer, which parses, blocks, vectorizes documents and stores them in a vector database, combines user question intention recognition, intelligent retrieval, and rearrangement, and selects the first k document blocks to be sent to the large language model for answering; behavioral portrait analysis processes audio and video files through voice-to-text conversion, identifies speaker information and timestamps, and uses the VAD model to process long audio and video, writes prompt words, and then sends the chat record vector to the large language model to extract customer portraits; standard process supervision blocks and vectorizes chat records, writes prompt words, and then sends the chat record vector to the large language model for supervision.

[0164] like Figure 2 The flowchart of the system of the present invention is shown. Specifically, the following steps are included:

[0165] Step 1-1: Start the system, enter the knowledge storage module, and define knowledge permissions, including upload, update, delete, share, etc. This process ensures that different users' access rights to knowledge can be differentiated and managed according to permission levels.

[0166] Step 1-2: Import data from knowledge sources. The data can be electronic documents or multimedia documents (such as audio and video). Different types of documents enter different parsing channels to ensure that subsequent processing is optimized for the data type.

[0167] Step 2-1: Enter the knowledge analysis module, the system loads data from the knowledge base and determines the document format. When a user makes a query, the system first extracts relevant information based on the existing electronic document knowledge base.

[0168] Step 2-2: The system determines the data format. If the document is text, it will be parsed by the electronic document parser. The parser will identify the text, paragraph structure, semantics and other information in the document and generate a data stream that can be processed by the system.

[0169] Step 2-3: If the document is an audio or video document, it is processed by a multimedia document parser to extract the audio information, video content and metadata.

[0170] Step 3-1: Enter the knowledge application module, and the system starts processing different types of document applications.

[0171] Step 3-1-1: Electronic document application: Users can perform intelligent retrieval through the system, including keyword retrieval, vector retrieval and hybrid retrieval. Different retrieval methods provide more accurate and efficient information positioning. At the same time, knowledge base question and answer and intelligent question and answer can also be applied to electronic documents.

[0172] Step 3-1-1-1: Knowledge base question and answer When a user asks a query, the system first extracts relevant information based on the existing electronic document knowledge base. Through natural language processing and semantic understanding, the system generates answers based on the knowledge base content. This question-and-answer system can cover the specific details and logic of the document and provide accurate information retrieval.

[0173] Step 3-1-1-2: In the intelligent agent question answering, the system first receives the question raised by the user and breaks it down into multiple more specific sub-questions through a large language model to ensure that each sub-question can be processed independently. Next, the system vectorizes these sub-questions and generates a fixed-length vector representation to capture their semantic information. Subsequently, the system loads the relevant documents, converts files of different formats into a processable text format, and divides them into independent paragraphs or fragments. These document fragments are also vectorized, and the generated vectors are stored in the vector database. By calculating the similarity between the sub-question vector and the document fragment vector, the system performs intelligent retrieval to find the most relevant document fragment, and further optimizes it through the rearrangement model to ensure that the most matching information is displayed first. Finally, the system matches the relevant document fragments with the sub-questions and generates specific answers, and then integrates the answers to all sub-questions to generate a logically coherent, comprehensive and fluent final answer.

[0174] Step 3-1-2: Multimedia document application: In the multimedia document application, users can analyze voice, image and other information in multimedia documents, including user behavior profile analysis and standard process supervision.

[0175] Step 3-1-2-1: In the user behavior portrait analysis, the system uses the speech-to-text model to process the audio and video files, identify the speaker information and timestamp, and improve the efficiency and accuracy of long-term audio and video processing through the VAD model. Next, the speech-to-text results are divided into blocks according to the speaker and timestamp to generate independent dialogue blocks. Then, according to business needs, the prompt words for extracting customer portraits are written to clarify the key information that needs to be extracted. Finally, the chat record vectors and prompt words after the blocks are input into the large language model together, and its natural language processing capabilities are used to extract customer portrait information, such as behavior patterns, preferences, and needs.

[0176] Step 3-1-2-2: In standard process supervision, the system divides the chat records in the multimedia document into blocks according to the timestamp and conversation content, and generates multiple independent conversation blocks for subsequent analysis. Next, natural language processing technology is used to vectorize the chat records after segmentation for processing by the machine learning model. Then, according to the company's standard processes and supervision requirements, prompts for supervising employees are written, and the format of extracted information is clarified to ensure the structured and consistent results. Finally, the vectorized chat records and the written prompts are input into the large language model, and the model is used to analyze employee behavior, identify whether it complies with the company's standard processes, and generate a supervision report.

[0177] In order to more clearly illustrate the technical solution of this embodiment, the specific implementation method of the enterprise-level knowledge management system based on the large language model is described in detail below in combination with an actual application case of a large manufacturing enterprise.

[0178] 1. Deployment and data import of knowledge storage module: After the system is deployed, the company's existing document data must first be imported into the knowledge storage module. The types of documents owned by the company include PDF, Word, audio (MP3 format) and video (MP4 format), with a total data volume of 2TB. The administrator classifies these documents by type and usage scenario, mainly into the following categories: technical manuals, customer feedback, operation guides, and internal training materials.

[0179] 1.1 Data types: PDF documents (5,000 copies, approximately 800 GB), Word documents (2,000 copies, approximately 200 GB), audio files (300 hours, approximately 500 GB), and video files (100 hours, approximately 500 GB).

[0180] 1.2 Permission control: The system assigns role-based access rights to different departments. For example, the technical support department can access all technical manuals and operation guides, but cannot view sensitive customer feedback records; while management can access all documents, including customer feedback.

[0181] 1.3 Result: All documents were successfully imported into the system, and the system generated metadata for the documents, including document type, upload time, file size, etc. The documents were stored in encrypted form to ensure data security.

[0182] 2. Application of knowledge parsing module and data processing: After the system is connected to the knowledge parsing module, it starts to parse and process documents of different formats.

[0183] 2.1 Electronic document analysis: Taking the "Product Model A Operation Manual" (PDF format, 100 pages) as an example, the system uses an electronic document parser to parse the PDF file and extract the technical steps, warning information and experimental data tables in the document. The structure of this document is relatively complex, containing mixed pictures, tables and text. The system uses OCR technology to recognize the text in the picture, with an accuracy of 98.5% and a recognition speed of about 1 second per page. It takes a total of 100 seconds to complete the analysis of the entire document.

[0184] Parsing parameters: The character recognition accuracy of the OCR model is 98.5%, and the processing speed is 1 page / second; the table extraction accuracy is 95%, and the extraction accuracy of complex tables (such as tables with merged cells) is 90%.

[0185] Storage results: The system generates an index entry for the document, and extracts key information including: motor maintenance steps, voltage adjustment methods, troubleshooting suggestions, etc. The extracted table data includes technical specification parameters, such as "the maximum power of the motor is 5kW and the rated voltage is 220V."

[0186] 2.2 Multimedia document parsing: For customer feedback audio (MP3 format, 30 minutes long), the system uses Alibaba FunASR speech-to-text model to convert customer conversations in the audio file into text with a recognition accuracy of 95%, and extracts the speaker’s identity and timestamp.

[0187] Speech recognition parameters: The speech recognition accuracy of the speech-to-text model is 95%, and the processing time is 1 / 10 of the multimedia duration, that is, one minute of audio is processed within 6 seconds.

[0188] Result: The system generated customer feedback content in text form and annotated the speaker's identity and timestamp information.

[0189] 3. Application of knowledge application module: In actual application, technical support personnel use the system's intelligent retrieval and intelligent question-and-answer functions to quickly solve technical problems raised by customers.

[0190] 3.1 Intelligent retrieval: When a customer asked, "How to perform daily maintenance on the motor of product model A?", the technical support staff entered the keyword "motor maintenance" through the system. The system uses a combination of the BM25 algorithm and vector retrieval to perform retrieval. First, the BM25 retrieval algorithm calculates the relevance score between each document and the keyword, and preliminarily screens out the top 20 documents. Then, the system uses the BCE-Embedding embedding model to vectorize the document, calculates the cosine similarity between the document vector and the question vector, and finally returns the top 5 most relevant documents and sorts the results.

[0191] Search parameters: The BM25 weight parameter is set to 1.2, and the cosine similarity threshold of vector search is 0.85. The overall response time of smart search is 2 seconds.

[0192] Result: The system returned 5 documents related to "motor maintenance". Page 25 of the "Product Model A Operation Manual" contains detailed motor maintenance steps: "Check the motor voltage every week to see if it is stable. The rated voltage is 220V and the allowable fluctuation range is ±5V." Based on the search results, the technical support staff provided specific operational suggestions to the customer.

[0193] 3.2 Intelligent Question and Answer: When a customer asks a complex question, "What is the difference between the motor debugging methods of product models A and B?", the system breaks the question down into two sub-questions through the ReAct framework: "Motor debugging methods for product model A" and "Motor debugging methods for product model B". Each sub-question is vectorized separately, and relevant documents are retrieved separately. Then the system integrates the answers to the two sub-questions and provides them to the customer.

[0194] Question and answer parameters: The vectorization dimension of each sub-question is 768, and the cosine similarity threshold between the sub-question and the document is 0.80. The overall response time of the intelligent agent question and answer is 3.5 seconds.

[0195] Result: The system generated the answer: "The motor commissioning steps for model A include: 1) connect the power supply; 2) adjust the voltage to 220V; 3) test the operating conditions; 4) record the current and voltage data. The motor commissioning steps for model B are similar to those for model A, but its rated voltage is 110V, so the voltage setting should be 110V during commissioning."

[0196] 3.3 Behavioral Profile Analysis in Multimedia Document Applications:

[0197] The system generates a behavioral profile of the customer by converting the audio recordings of customer feedback into text and combining it with the system's behavioral profile analysis function.

[0198] Speech-to-text parameters: The system processes a 30-minute audio clip, and the recognition accuracy of the speech-to-text model is 95%, completing the transcription within 3 minutes.

[0199] Behavior profile analysis: The system analyzed the customer's call content and identified that the main issue the customer was concerned about was "motor failure." The customer profile generated by the system showed: "This customer is sensitive to product quality and has high expectations for after-sales service. It is recommended to provide more technical support and quick response in subsequent services."

[0200] Results: The behavioral profile analysis results helped the company optimize its service strategy for this customer, and the after-sales department developed a quick response plan based on the customer profile.

[0201] 3.4 Application of standard process supervision:

[0202] The system is used to monitor whether the work processes of enterprise employees meet standard specifications. In a call between a customer service representative and a customer, the call recording is first converted into text, and the preset prompt words are used to determine whether the customer service representative follows the standard process.

[0203] Supervision parameters: The VAD model has a block time interval of 0.5-1 minute. The call recording file will be divided into several small blocks according to the block time interval. After completing the text conversion of each small segment of speech, the text of each segment will be spliced. The prompt words are set to "customer emotion", "problem focus", and "whether to use polite language".

[0204] Result: The system generated a supervision report, pointing out that the customer service staff did not use the word "please" during a call and their tone was rather blunt, suggesting that they improve their communication skills and modify the standard operating procedures. Through this supervision module, the company improved its customer service training program.

[0205] 4. Final effect and technical problem solving

[0206] Through the enterprise-level knowledge management system based on the large language model of the present invention, the enterprise achieves the following improvements:

[0207] Improved knowledge retrieval efficiency: The intelligent retrieval function combined with BM25 and vector retrieval enables technical support personnel to find the required documents within seconds, and the average response time is shortened from 10 minutes to 2 seconds.

[0208] Intelligent question-and-answer function improves customer service quality: Through the intelligent question-and-answer function, the system can quickly generate accurate answers in complex problem scenarios.

[0209] Behavioral profiling to optimize customer service strategies: The system helps companies customize service plans and improve overall customer satisfaction by analyzing behavioral patterns in customer feedback.

[0210] Standard process supervision regulates employee behavior: The standard process supervision function helps enterprises monitor employees' service behaviors in real time, ensure compliance with corporate service standards, and reduce violations in the service process.

[0211] In a specific implementation, the present application provides an enterprise-level knowledge management system based on a large language model, which can be stored in a computer storage medium, such as a disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), etc., and the computer storage medium can store a computer program. When the computer program is executed by a data processing unit, it can implement the functions of each module proposed in the present invention and some or all of the steps in the embodiments. Through the computer program, the knowledge storage, analysis, and application modules in the present invention can be automatically operated to realize the analysis and management of electronic documents and multimedia documents, intelligent question and answer, behavior profiling analysis, standard process supervision and other functions.

[0212] Those skilled in the art can clearly understand that the technical solution of the present invention can be implemented by combining a computer program with a general hardware platform, specifically including a server, a PC, a single-chip microcomputer or other computing devices. The technical solution, in essence or the improved part of the prior art, can be embodied in the form of a computer program, which can be stored in a storage medium and contains multiple instructions, which drive a device including a data processing unit to execute various embodiments of the present invention or part of its functions.

[0213] The present invention proposes an enterprise-level knowledge management system based on a large language model, and provides multiple implementation methods. The above description is only a preferred embodiment. For ordinary technicians in this field, several improvements or modifications can be made without departing from the basic principles of the present invention. These improvements and modifications also fall within the scope of protection of the present invention. The unclear parts can be implemented by existing technologies.

Claims

1. An enterprise-level knowledge management system based on a large language model, characterized in that: It includes knowledge storage module, knowledge analysis module and knowledge application module; The knowledge storage module includes a knowledge authority management unit and a knowledge source management unit. The knowledge authority management unit is used to manage and store knowledge from different sources, and the knowledge source management unit is used to manage electronic documents and multimedia documents. The knowledge parsing module includes an electronic document parser and a multimedia document parser, wherein the electronic document parser uses a large language model technology to process documents, extract key information and generate an index for search; The multimedia document parser is used to parse multimedia documents in audio and video formats, convert speech into text, and extract information from the video using image recognition technology; The knowledge application module includes an electronic document application unit and a multimedia document application unit. The electronic document application unit is used to provide intelligent retrieval and intelligent question-and-answer functions; the multimedia document application unit is used to perform behavioral profiling analysis and standard process supervision of multimedia documents.

2. The system according to claim 1, characterized in that The knowledge rights management unit supports uploading, updating, deleting and sharing of electronic documents and multimedia documents, and is provided with rights control to ensure information security: the knowledge rights management unit adopts a fine-grained rights control mechanism combining role-based access control and attribute-based access control to achieve document-level access control; The system supports hierarchical management of permissions based on user roles, departments, job levels, etc., and can dynamically adjust permission policies; The knowledge rights management unit adopts TLS / SSL encryption protocol during data transmission to ensure data transmission security; For sensitive data stored in the system, the system uses a private file cabinet mechanism for encrypted storage, uses encryption algorithms to encrypt sensitive data, and implements key segmentation management; the system also establishes a data backup and recovery mechanism; The knowledge rights management unit has an operation record function, which permanently records the operations of the entire knowledge management process. The recorded content includes the operation time, operation type, operator identity, operation content and operation results. The recorded content is encrypted and stored to ensure that it cannot be tampered with, and is archived and backed up regularly; The knowledge rights management unit integrates a behavior monitoring system for real-time monitoring of user behavior and identifying abnormal operations; The system regularly detects overlapping permissions, missed permissions, and lack of permissions, and optimizes permission allocation through permission conflict detection mechanisms and regular permission reviews; at the same time, the system establishes a security incident response mechanism to provide real-time alarms, emergency response, and post-event tracing capabilities.

3. The system according to claim 2, characterized in that The formats of the electronic documents include PDF, Word, Markdown, TXT and HTML, and the formats of the multimedia documents include mp3, m4a, aac, wma, mp4 and wav.

4. The system according to claim 3, characterized in that The intelligent search function includes keyword search, vector search and hybrid search, and is sorted according to the relevance score to the question raised by the user; the hybrid search is a combination of BM25 search and vector search.

5. The system according to claim 4, characterized in that The intelligent question-answering function includes knowledge base question-answering and intelligent agent question-answering. The documents are parsed, segmented, and vectorized and stored in a vector database. The user questions are then identified for intent, intelligently retrieved, and rearranged. The top k document blocks are selected and sent to a large language model for answering.

6. The system according to claim 5, characterized in that The behavior portrait analysis includes: processing audio and video files through voice-to-text conversion, identifying speaker information and timestamps, and using an audio truncation model to process long audio and video, and after writing prompt words, sending the chat record vector into a large language model to extract customer portraits; The standard process supervision includes: dividing the chat records into blocks, vectorizing them, writing prompt words, and then sending the chat record vectors into a large language model for supervision.

7. The system according to claim 6, characterized in that The knowledge base question answering comprises the following steps: Step 1-1, loading documents: using an electronic document parser to load documents of different formats, the electronic document parser can identify and process the content of the document and convert the document into a unified text format, specifically including the following steps: Step 1-1-1: Intelligently identify document formats: The system uses a deep learning model to analyze file header information, content features and metadata to identify document formats and achieve automatic diversion processing; Step 1-1-2: Multi-channel parallel parsing: According to the recognition results of the document format, different types of documents are assigned to the corresponding electronic document parsers to achieve parallel processing; Step 1-1-3: Text content extraction and structuring: For text documents, use natural language processing technology to extract text content while retaining the document's hierarchical structure, format information, and semantic relationships; for imaged PDF or scanned documents, integrate OCR technology to achieve high-precision text recognition and extraction; for multimedia documents, use speech recognition and video content analysis technology to extract text information from audio and video; Step 1-1-4: Semantic understanding and knowledge graph construction: Use deep learning models to understand the extracted text content, identify key concepts, entity relationships, and topic structures, and build a document-level knowledge graph; Step 1-1-5: Intelligent document segmentation: Based on the semantic understanding results, the document is intelligently segmented into semantically complete and logically independent paragraphs or sentences; Step 1-1-6: Multimodal vectorization: Use the pre-trained multimodal embedding model to fuse text content, hierarchical structure, format information, semantic relations, and multimedia features to generate a unified vector representation that captures the comprehensive semantic information of the document; Step 1-1-7: Incremental knowledge update: vectorize the newly added document knowledge and add it to the vector library corresponding to the current document, process the newly added documents in real time, and dynamically update the knowledge base; Step 1-2, document segmentation: segment the loaded document into smaller, logically independent paragraphs or sentences; specifically, the steps are as follows: Step 1-2-1: Document structure analysis: Use natural language processing technology to perform structural analysis on the document and identify titles, paragraphs, and lists; Step 1-2-2: Semantic unit identification: Use a deep learning model to perform semantic analysis on the document content and identify the smallest semantically complete unit; Step 1-2-3: Perform multi-scale blocking, including large-scale blocking, medium-scale blocking and small-scale blocking; The large-scale segmentation includes: based on the document structure, according to the chapter and paragraph marks, using the chapter and paragraph marks as separators, and after reading the separators, preliminarily segmenting the document into a large chapter or paragraph; The small-scale block segmentation is to subdivide the result of the large-scale block segmentation: first, the chapters or paragraphs obtained by the large-scale block segmentation are divided into several sentences according to the separators, and the sentences are simultaneously added with a metadata for marking which large chapter the sentences belong to; secondly, the cosine similarity algorithm is used to determine whether there is a connection between the sentences in each chapter. If there is a connection, the two sentences are spliced ​​to obtain a longer small paragraph, and finally a semantically complete sentence or phrase is obtained; Step 1-2-4: Dynamically adjust the block size: Dynamically adjust the block size according to the document type, content complexity and system performance requirements; Step 1-2-5: Semantic relevance evaluation: Use the semantic similarity algorithm to calculate the semantic relevance between two adjacent blocks. If the semantic relevance is higher than the preset threshold, merge the two adjacent blocks. Step 1-2-6: Build a block index: Generate a unique identifier for each block and establish a mapping relationship with the original document; Step 1-2-7: Block quality control: Design a quality assessment model to evaluate the block results; if the block does not meet the quality requirements, the system will automatically adjust the block strategy and re-execute the block process; Step 1-3, document vectorization: Use the pre-trained embedding model to vectorize the document block. Each document block is converted into a fixed-length vector representation to obtain a document vector, which captures the semantic information of the document block. Step 1-4, storing document vectors: storing the document vectors obtained in steps 1-3 in a vector database; Step 1-5, processing user questions: After obtaining the user's question, the intent recognition model will first be used to determine whether the user's question is meaningful. After determining that the user's question is meaningful, the system will combine the current knowledge base documents and use the large language model to determine whether the user's question can be answered in the knowledge base. Then, the pre-trained embedding model is used to convert the user's question into a vector form to obtain the user's question vector. Step 1-6, intelligent retrieval: perform intelligent retrieval in the vector database, calculate the cosine similarity between the document vector and the user question vector, thereby obtaining the similarity score of each document vector, sort all document vectors from high to low according to the similarity score with the user question vector, and recall all the document vectors. The calculation formula is: Among them, A i represents the i-th document vector, B i represents the i-th user question vector, n is the total number of document vectors, P i is the similarity score; Steps 1-7, re-ranking related documents: Load the re-ranking model to re-rank the document vectors after document sorting to ensure that the most relevant document blocks are ranked first; Step 1-8, select the best document: select the top k document vectors with the highest similarity from the re-ranked document vectors as the best documents. The top K document vectors are the information sources that best match the user's question. Step 1-9, generate answer: input the user's original question, best document and prompt information into the large language model, and the large language model generates the final answer and provides it to the user.

8. The system according to claim 7, characterized in that The intelligent agent question answering comprises the following steps: Step 2-1, user problem decomposition and vectorization, includes the following steps: Step 2-1-1, deconstruct the problem: After identifying the user's problem intent and determining whether it is meaningful, use the ReAct framework to deconstruct the user's problem and split it into several small sub-problems; Step 2-1-2, problem vectorization: Use the pre-trained embedding model to vectorize each sub-problem after decomposition. Each sub-problem is converted into a vector representation of a fixed length to obtain a sub-problem vector; Step 2-2, obtaining the document fragment corresponding to each sub-question, includes the following steps: Step 2-2-1, loading documents: loading relevant documents stored in the system, the relevant documents including FAQ, manuals, technical documents; using an electronic document parser to convert documents of different formats into a unified text format; Step 2-2-2, document segmentation: split the loaded document into logically independent paragraphs or sentences; Step 2-2-3, document vectorization, use the pre-trained embedding model to vectorize the document block, each document block is converted into a fixed-length vector representation to obtain the document vector; Step 2-2-4, storing document vectors: storing the document vectors obtained in step 2-2-3 in a vector database; Step 2-2-5, intelligent retrieval: perform intelligent retrieval in the vector database, calculate the similarity between the document vector and the sub-question vector, take the document vector with the highest similarity to the sub-question vector as the most relevant document vector, and recall the most relevant document vector; Step 2-2-6, re-rank relevant documents: Use the re-ranking model to re-rank the recalled document vectors to ensure that the most relevant document blocks are ranked first; Step 2-2-7, select the best document: select the top k document vectors with the highest relevance from the re-ranked document vectors as the information source that best matches the sub-question; Step 2-3, integrating the answers to two or more sub-questions to generate the final answer, includes the following steps: Step 2-3-1, generate sub-question answers: input each sub-question and the corresponding document fragment into the large language model to generate specific answers for each sub-question; Step 2-3-2, generate the final answer: Integrate each sub-question and its corresponding specific answer, as well as the unsplit user question, into the pre-written prompt words, load them into the large language model together, and generate the user answer.

9. The system according to claim 8, characterized in that The behavior profiling analysis includes the following steps: Step 3-1, speech-to-text conversion: Two speech-to-text conversion models are used, one is a Mandarin model and the other is a dialect model. The cross entropy loss function used by the dialect model is l pred for: in represents the concatenation result of the kth document block, P k is the generation probability, t k is the one-hot vector of the number of tokens in the k-th document block, W 1 , W 2 is the weight, b 1 、b 2 is bias; Step 3-2, chat record segmentation: The speech-to-text result obtained in step 3-1 is segmented into blocks; when processing chat records that exceed the context, they are segmented into several dialogue blocks within a fixed length range according to the speaker identity information; Step 3-3: Write prompt words for extracting customer portraits based on business needs and analysis objectives, and clarify the key information and features that need to be extracted from chat records; In step 3-4, the conversation block is converted into a vector representation using the Embedding model, and the prompt words in step 3-3 are sent together with the vector representation into the large language model. The large language model is used to extract the customer's portrait information from the chat record, including the customer's behavior patterns, preferences, and needs.

10. The system according to claim 9, characterized in that The standard process supervision includes the following steps: Step 4-1, chat record segmentation: The chat records in the multimedia document are segmented and processed. When processing chat records that exceed the context, they are truncated within a fixed length range according to the speaker identity information; Step 4-2, chat record vectorization: Use the Embedding model to convert the conversation block into a vector representation; Step 4-3: Write prompts for supervising employees according to the company's standard processes and supervision requirements, and specify the key information and features that need to be extracted from the chat records; at the same time, define the format of the extraction results to ensure the structuring and consistency of the supervision results; In step 4-4, the vector representation obtained in step 4-2 and the prompt words written in step 4-3 are sent to the large language model. The large language model is used to monitor the behavior of employees from the chat records, identify whether it complies with the standard process, and generate a supervision report.

Images