Method and system for process consultation and responsibility association question and answer based on enterprise knowledge base

Abstract

The application provides a process consultation and responsibility association question-answering method and system based on an enterprise knowledge base, comprising: obtaining dialogue information of inquiring process consultation and responsibility association; querying in an enterprise knowledge base according to the dialogue information to obtain corresponding query results, wherein the enterprise knowledge base comprises a vector database for vector representation information, a document database for storing document content, and a knowledge graph for storing responsibility association; and analyzing the query results by a large model to generate and display the results of the final relevant process information and responsibility association information, so as to solve the problems of low efficiency of traditional query methods and unclear responsibility in enterprise knowledge management.

Description

Technical Field

[0001] This invention relates to the field of data management technology, and more specifically, to a process consultation and responsibility association question-and-answer method and system based on an enterprise knowledge base. Background Technology

[0002] As businesses expand and their processes become increasingly complex, employees frequently need to consult various rules, regulations, procedures, and responsibility assignment information in their daily work. Traditional information retrieval methods primarily rely on internal document management systems, static FAQ pages, or human consultation services. These methods suffer from low response efficiency, inaccurate information retrieval, and ambiguous responsibility attribution. In recent years, the development of artificial intelligence technologies, especially natural language processing and large language models, has provided new possibilities for enterprise knowledge management. However, existing systems often lack a deep understanding of the enterprise's specific knowledge structure, particularly the network of responsibility relationships. They cannot accurately link the system to the corresponding responsible department and contact person while answering process-related inquiries, leading to information fragmentation, unclear responsibilities, and impacting problem-solving efficiency. Summary of the Invention

[0003] In view of this, the present invention proposes a process consultation and responsibility association question-and-answer method and system based on enterprise knowledge base to solve the problems existing in the prior art.

[0004] To achieve the above objectives, this invention proposes a process consultation and responsibility association question-and-answer method and system based on an enterprise knowledge base, including: Obtain dialogue information regarding the inquiry process, consultation, and responsibility association; The query is performed in the enterprise knowledge base based on the dialogue information to obtain the corresponding query results. The enterprise knowledge base includes a vector database for vector representation information, a document database for storing document content, and a knowledge graph for storing responsibility associations. The query results are analyzed using a large model to generate and display the final relevant process information and responsibility association information.

[0005] Optionally, dialogue information can be obtained through a user interface, wherein the user interface is provided with prompts about the dialogue information, status prompts for displaying the progress of the question and answer, and a display area for displaying the results of the final relevant process information and responsibility association information.

[0006] Optionally, the construction process of the knowledge graph in the enterprise knowledge base includes: Obtain raw data, including rules and regulations, operation manuals, process descriptions, and organizational structure data; construct a knowledge graph based on the raw data using named entity recognition and relationship extraction methods, wherein the knowledge graph includes process nodes, department nodes, personnel nodes, rule nodes, and the semantic relationships between them.

[0007] Optionally, the construction process of the vector database and document database in the enterprise knowledge base includes: Acquire document data, wherein the document data includes rules and regulations, operation manuals and process descriptions, preprocess the document data, segment and vectorize the preprocessed document data, and obtain the vectorized representation of the segmented document data to construct a vector database; The document data is mapped and indexed to generate structured data for retrieval, thereby building a document database.

[0008] Optionally, the process of querying the enterprise knowledge base based on the dialogue information includes: The dialogue information is transformed to generate query information. Based on the query information, the enterprise knowledge base is searched using vector retrieval, keyword retrieval, and graph retrieval methods to generate corresponding search candidate results and relevance scores. A comprehensive score is calculated based on the relevance scores. The search candidate results are sorted and filtered based on the comprehensive scores to generate the final search results, i.e., the query results.

[0009] Optionally, the process of analyzing query results using a large model includes: The query results, dialogue information, and corresponding session context information are combined to form a large model prompt. The large model generates data based on the large model prompt, and finally obtains the relevant process information and responsibility association information.

[0010] Optionally, the data generation process based on the prompts of the large model employs a constraint decoding method and a factual check; wherein, the constraint decoding method includes enhancing the probability weight of key information during the decoding process of data generation and detecting whether there are any omissions in the final generated result; the real-time check includes matching and data backtracking based on the query results according to the final generated result to correct the final generated result.

[0011] On the other hand, the present invention also provides a process consultation and responsibility association question-and-answer system based on an enterprise knowledge base for performing the above-described method.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention addresses the problems of low efficiency, poor accuracy, and unclear responsibilities associated with traditional query methods in enterprise knowledge management. It proposes an intelligent question-answering system based on the fusion of knowledge graphs and large-scale models. By constructing an enterprise-specific knowledge graph, the system achieves deep structuring and intelligent association of knowledge. Combined with a hybrid retrieval strategy, it accurately matches diverse query needs and leverages a large language model to achieve natural interaction and structured answer generation. The system possesses the ability to automatically extract and dynamically update responsibility information, supporting continuous knowledge accumulation and optimization, while meeting enterprise-level security and operational requirements. This solution not only significantly improves the efficiency and accuracy of information retrieval but also constructs a sustainably evolving intelligent enterprise knowledge ecosystem, providing core support for organizational digital transformation. Attached Figure Description

[0013] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is a schematic diagram of the interaction process in an embodiment of the present invention; Figure 2 This is a flowchart related to backend data processing in an embodiment of the present invention. Detailed Implementation

[0014] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0015] This embodiment proposes a process consultation and responsibility association question-and-answer method and system based on an enterprise knowledge base, such as... Figure 1-2 As shown, it includes: Inputting dialogue information related to process consultation and responsibility association results in output in the dialog box of the user interface. The dialog box is marked with the prompt "You can enter the questions you need to consult" and provides progress and status prompts. The user interface is an interface for employees to consult on the operation process. The operation process and regulations are displayed in the display area of ​​the user interface, and the contact person and contact information of the department in charge of the regulations related to responsibility association are output. Finally, the operation process is output.

[0016] The backend data processing of the user interface relies on the LightRAG API, which connects to the corresponding knowledge graph and large model API. The dialogue information mentioned above is used as user information input in the content, and can be supplemented with the corresponding documents. By combining the knowledge graph with the large model, the operation process and regulations, as well as the contact persons and contact information of the departments in charge of the regulations, are queried and generated, and then displayed in the user interface.

[0017] The specific interaction content is as follows: The company's new employee handbook and other procedural documents are pre-loaded inside the node machine, and the corresponding responsibility-related content is recorded. This content is stored in the form of a relevant detailed storage database and knowledge graph. In the user interface, users enter their operation and procedural questions in the dialog box below; clicking "Confirm" displays the query and analysis progress and status. The backend analyzes and processes the input dialogue information through the large model and the knowledge graph in the enterprise knowledge base, and then feeds the results back to the user interface. The display area shows the operation process and related regulations, and simultaneously outputs the contact person and contact information of the department responsible for the process and regulations.

[0018] The input options for the dialog box include mobile devices, computers, or external voice devices.

[0019] For the aforementioned large model and knowledge graph, the final result is input through a question-and-answer method, without involving file output or template input. A knowledge base for storing the knowledge graph can be generated in advance by inputting document data, and then accurate question-and-answer is provided.

[0020] Specifically, the user interface adopts a layered architecture design, divided into a visual presentation layer, an interaction logic layer, and a data communication layer. The visual presentation layer is developed based on the Vue 3 framework and the Element Plus component library, employing responsive design to ensure compatibility across different devices. The main interface is divided into three main areas: the top navigation area displays the system logo and user information; the left-hand function area provides history and frequently used query entry points; and the central main interaction area is the core operation area. The main interaction area is further divided into a question input area, a progress display area, and a result display area. The question input area provides text input boxes and voice input buttons. The text input box is equipped with intelligent suggestion functionality, dynamically recommending relevant questions based on the user's input. The progress display area uses a combination of progress bars and status text to display different stages of query processing in real time, including question analysis, knowledge retrieval, and answer generation. The result display area uses a tabbed design, displaying operation process steps, relevant rules and regulations, responsible department information, and contact details, supporting content expansion / collapse and interactive operations.

[0021] The interaction logic layer handles front-end business logic processing, including input validation, state management, and user behavior tracking. The input validation module checks the format and filters user input to ensure its validity and security. The state management module maintains the overall interface's interactive state, controlling the display, hiding, and state switching of various components. The user behavior tracking module records the user's action sequence on the interface, including clicks, scrolling, and input; this data is used to analyze user habits and optimize interface design. The data communication layer encapsulates API communication logic with the backend, uniformly handling request sending, response receiving, and error handling, using Promise chaining to ensure the sequential execution of asynchronous operations. The interface also implements an offline caching mechanism, caching critical data locally when the network is unstable to ensure the availability of basic functions.

[0022] The backend system adopts a microservice architecture, divided into five core service modules: API gateway service, business logic service, data retrieval service, intelligent processing service, and data persistence service. The API gateway service serves as the system's unified entry point, responsible for request routing, load balancing, security authentication, and rate limiting. The business logic service framework includes core functional modules such as user management, session management, access control, and business rule processing. The data retrieval service builds a full-text search engine to achieve efficient retrieval of structured documents; it also integrates the Neo4j client to provide knowledge graph query capabilities. The intelligent processing service specifically handles AI-related tasks such as natural language understanding, knowledge reasoning, and large-scale model interaction. The data persistence service manages multiple data storage methods, including MySQL and Redis, ensuring data reliability and consistency through relevant databases.

[0023] The backend system employs a containerized deployment approach, with each service packaged as an independent container image and orchestrated by a container manager. Inter-service communication utilizes both RESTful APIs and message queues; synchronous calls use the HTTP protocol, while asynchronous processing uses RabbitMQ message queues. Configuration management employs a centralized configuration center, supporting configuration differences across different environments and dynamic configuration updates. The monitoring system monitors service health status, performance metrics, and business metrics in real time. The logging system uses the ELK technology stack to centrally collect and analyze the runtime logs of each service, supporting distributed tracing and problem diagnosis. Security mechanisms include JWT-based authentication, role-based access control, encrypted data transmission, and operation audit logs to ensure system security and compliance.

[0024] The front-end and back-end communication utilizes a dual-channel communication mechanism built upon HTTP / HTTPS and WebSocket protocols. For routine data requests, the system employs a RESTful API design, defining a unified resource access interface. The API design follows a resource-oriented principle, abstracting business objects as resources and operating on them through standard HTTP methods. All API requests are processed by a unified interceptor for authentication, parameter validation, and logging. The response format uses a unified JSON structure, including status codes, message descriptions, and business data. For scenarios with high real-time requirements, such as chat progress updates and real-time message pushes, the system establishes a WebSocket long-lived connection channel to achieve bidirectional real-time communication.

[0025] Multi-layered security measures are implemented during communication. The transport layer encrypts all communication content to prevent eavesdropping or tampering during transmission. The application layer implements request signature verification; each request includes a timestamp and digital signature to prevent replay attacks and parameter tampering. Session management adopts a stateless design, maintaining user identity through encrypted JWT tokens to avoid server-side session storage burden. Traffic control mechanisms include interface-level and user-level rate limiting to prevent malicious requests and system overload. An error handling mechanism defines a complete error code system, with the frontend adopting appropriate recovery strategies based on different error types. For performance optimization, an HTTP caching strategy is implemented to cache static resources and some query results; connection pooling is used to manage database and HTTP connections, improving connection reuse; request merging and batch processing are implemented to reduce network round trips.

[0026] In the enterprise knowledge base for data storage, the knowledge graph is constructed based on existing enterprise documents and database information, employing a semi-automated construction process. First, raw data is extracted from the enterprise document management system, process management system, and human resources system, including rules and regulations documents, process descriptions, organizational structure data, and job descriptions. Then, natural language processing technology is used for information extraction. Named entity recognition models are used to identify entities such as personnel, departments, processes, and regulations in the documents, while relation extraction models are used to identify the attribution of responsibility, process sequence, and referencing relationships between entities. The extracted knowledge triples are manually verified and corrected according to rules before being imported into the Neo4j graph database for storage. The graph schema design defines node types and relation types. Node types include process nodes, department nodes, personnel nodes, and regulation nodes, each containing specific attribute fields. Relationship types define semantic connections between entities, such as "belongs to," "responsible for," "references," and "next step," and each relation type can include attributes such as weight and timeliness.

[0027] Specifically, the construction of a knowledge graph first requires connecting to the enterprise's existing multi-source heterogeneous data systems. This includes PDF and Word documents of regulations stored in the document management system, BPMN process definitions in the process management system, organizational structure data in the human resources system, and question-and-answer records in the historical work order system. This data forms the raw corpus foundation of the knowledge graph. In the data preprocessing stage, the system performs format parsing and noise removal on the documents, converting files of different formats into standardized text. Simultaneously, it performs chapter structure recognition and semantic segmentation to ensure that the basic units for subsequent processing have complete semantic context.

[0028] The information extraction phase employs a combined approach of deep learning and natural language processing. A named entity recognition model scans the text, automatically identifying entities such as departments (HR, Finance, IT Support), personnel (Manager Zhang, Supervisor Li), processes (onboarding approval, expense reimbursement), and regulations (Employee Handbook Chapter 5, Information Security Management System). These entities form the candidate set of nodes for the knowledge graph. Subsequently, the relation extraction model analyzes the semantic relationships between entities, identifying attribution relationships (e.g., an employee belongs to a department), responsibility relationships (e.g., a department is responsible for a process), referencing relationships (e.g., a process is based on a regulation), and temporal relationships (e.g., the order of process steps), forming a set of triples representing entity relationships or entity attribute values. These triples are then filtered by manual verification rules and sampled by business experts before being imported into the Neo4j graph database for storage.

[0029] In terms of graph pattern design, the system defines a rich system of node types and relationship types. Node types cover categories such as process nodes, department nodes, personnel nodes, rule nodes, and document nodes. Each type of node has standardized attribute fields. For example, personnel nodes include name, position, employee ID, contact information, and office location, while process nodes include process name, applicable scenarios, processing time limits, and required materials. Relationship types meticulously depict various business connections in enterprise operations, including belonging relationships connecting personnel and departments, responsible relationships connecting departments and processes, formulating relationships connecting departments and rules, referencing relationships connecting processes and rules, next-step relationships connecting the sequence of process steps, and supervising relationships connecting superior and subordinate departments. Some relationships also have weight attributes indicating the strength of the association or time-limited attributes indicating the validity period.

[0030] The knowledge graph of the enterprise knowledge base is deployed in the data persistence layer and provides data support for the backend data retrieval service. Built on the Neo4j graph database engine, it is deployed in a containerized environment in a cluster mode, forming a complementary storage matrix with the full-text search engine and vector database. Neo4j instances run as independent service containers on the Kubernetes orchestration platform, storing graph data through persistent volumes, while read / write separation and replica set mechanisms ensure high availability. The graph data is kept synchronized with the original document library and vector embedding library. When the enterprise releases new regulations or adjusts its organizational structure, the data persistence service triggers an incremental update process, and newly extracted triples are merged into the Neo4j graph in real-time or near real-time, ensuring that the knowledge base remains consistent with the enterprise's actual business status. During the query phase, the Neo4j client in the data retrieval service interacts with the graph engine through the Cypher query language, supporting complex multi-hop relationship traversal and path search.

[0031] The enterprise knowledge base also includes a vector database and a document database. The construction of the vector database begins with document collection and preprocessing. Documents in various formats, such as regulations, operation manuals, and process descriptions, are retrieved from the enterprise document management system. These documents are then converted into plain text using text extraction technology and subsequently cleaned and standardized. In the document segmentation stage, long texts are divided into appropriately sized text blocks according to semantic boundaries, ensuring that each text block contains complete semantic information. The vectorization process uses a pre-trained embedding model to convert each text block into a high-dimensional vector representation. These vectors maintain semantic similarity relationships in mathematical space. Finally, the vectors, along with their corresponding metadata, are stored in a dedicated vector database, and an efficient index structure is built to support rapid similarity retrieval.

[0032] The construction of the document database begins with the design of the index structure, defining the field mapping relationships for documents, including core fields such as title, content, department, publication date, and document type. A dedicated word segmenter and analyzer are configured for Chinese text. The inverted index construction is the core step, breaking down document content into independent lexical units through word segmentation, establishing a mapping relationship between words and document locations, and forming an efficient retrieval data structure. During the data import phase, preprocessed documents are loaded into the retrieval system in batches, and the system automatically completes index construction and optimization. Finally, performance tuning is performed, including sharding strategy configuration, replica settings, and cache optimization, to ensure the system can support high-concurrency query requirements.

[0033] The construction of the three databases requires coordination to form a complete knowledge service system. A data synchronization mechanism ensures information consistency by synchronizing data from the original data source to the three databases through a unified ETL pipeline, using transactional operations to guarantee the atomicity of data updates. An index linkage design allows indexes from different databases to work together; for example, storing vector ID references in the full-text search engine and recording the vectorized state of documents in the graph database. Consistency maintenance strategies include periodically comparing data content, monitoring synchronization latency, and establishing an exception handling mechanism to ensure that the data in the three databases remains synchronized at all times. A performance monitoring system monitors each stage of the construction process in real time, promptly identifying and resolving problems to ensure the stable operation of the entire knowledge base system.

[0034] The large-scale model system uses locally deployed models such as Qwen, ChatGLM, or GPT as its foundation, improving their performance in enterprise scenarios through domain-adaptive training. The training process consists of two phases: the first phase involves continued pre-training on a large number of enterprise documents, allowing the model to learn enterprise-specific terminology and expressions; the second phase involves supervised fine-tuning on labeled question-answer pairs to optimize the model's question-answering capabilities. Model deployment utilizes the vLLM inference engine, supporting efficient inference under high concurrency. The system designs specialized prompt word templates for the model, organizing user questions, retrieved contextual information, and responsibility-related information in a specific format to guide the model in generating structured answers. The system also implements a model monitoring mechanism to track the model's response time, generation quality, and resource usage in real time.

[0035] In the data retrieval service, after a user submits a query, the system initiates a multi-stage search and analysis process. First, the query understanding module analyzes the user's question, identifying the query intent and key entities. Intent identification uses a BERT-based classification model to categorize queries into types such as process consultation, regulation inquiry, and responsibility confirmation. Entity identification uses a sequence labeling model to extract specific process names, department names, regulation numbers, etc., mentioned in the question. Based on the analysis results, the system generates multiple query variations, including synonym replacement, question decomposition, and contextual supplementation.

[0036] Next, the system executes three retrieval operations in parallel: vector retrieval searches for semantically similar document fragments in the vector database and calculates relevance using cosine similarity; keyword retrieval performs a full-text search on the search server and calculates the matching degree using the BM25 algorithm; and graph retrieval traverses the relevant knowledge network in Neo4j to find relevant entities and relationships. Each retrieval method returns a set of candidate results and their relevance scores.

[0037] Next, the results from multiple search paths are integrated. First, the scores are standardized to eliminate differences between different scoring systems. Then, a weighted fusion algorithm is used to calculate the overall score for each candidate result, with the weight allocation dynamically adjusted according to the query type: process queries prioritize graph retrieval results, concept queries prioritize vector retrieval results, and fact queries prioritize keyword retrieval results. The fused results are then sorted by their overall scores, and the most relevant items are selected as the final search results.

[0038] The search results are then fed into an intelligent processing service, which organizes the search results, user questions, and dialogue context into prompts and submits them to a large model to generate the final answer. The generation process employs constraint decoding technology to ensure that the answer includes necessary responsibility information and contact details. The system also performs factual checks on the generated content and compares it with the search results to ensure the accuracy and consistency of the answer.

[0039] Within the intelligent processing service, a multi-layered prompt word context framework is constructed. This framework includes not only the query semantics of the user's current input question but also the continuity information from historical dialogues. More importantly, it organically integrates entity relationships retrieved from the knowledge graph, semantically similar text fragments returned from the vector database, and precise regulations obtained from keyword retrieval. Based on a pre-designed domain template, different types of retrieval evidence are categorized and embedded into different blocks of prompt words, forming an input sequence with a clearly structured guidance. This enables the large model to clearly identify which content belongs to background knowledge, which belongs to constraints, and which are the core questions that require reasoning and answers.

[0040] Constraint decoding technology is employed to ensure the structure and integrity of the output content, with control signals applied in real time during model decoding. A dynamic keyword set is maintained, containing key entities such as the responsible department name, contact person name, phone number, and email address extracted from the knowledge graph. At each step of model generation, the decoder calculates the association probability between the current candidate word and these key entities and forcibly increases the probability weight of key information by adjusting the logits distribution. Simultaneously, the system parses a predefined output template structure, monitoring the current content paragraph during generation. When the model is about to enter the responsibility information description paragraph, a structural constraint mechanism is automatically triggered to ensure that required fields such as contact information are not omitted.

[0041] To ensure high consistency between the generated content and the underlying knowledge base, entity extraction is performed on the generated text, extracting key information such as all names, departments, contact information, and process steps. This information is then precisely matched and fuzzy similarity calculated against the original search results to check for any illusions or factual biases. For structured data in the knowledge graph, the system performs rigorous consistency verification, such as verifying whether mentioned contacts actually belong to the corresponding departments and whether phone numbers completely match database records. Secondly, the system introduces a semantically entailed verification model to determine whether the generated explanatory text is semantically supported by the retrieved regulations and clauses, ensuring the rationality of the reasoning process. If inconsistent or low-confidence generated content is found, the system initiates a correction mechanism, iteratively optimizing by adding hints or retrieving more background knowledge, ultimately forming a reliable answer that both matches the user's query intent and strictly adheres to the company's knowledge standards.

[0042] After the search and analysis are completed, the system returns the processing results to the front-end interface in a standardized data format. The output data adopts a hierarchical structure, including the answer body layer, supporting evidence layer, responsibility information layer, and operation suggestion layer. The answer body layer contains the generated text answer, using different structured formats according to the question type: process consultation questions use step-by-step explanations, regulation query questions use clause-based analysis, and responsibility confirmation questions use list-based display. The supporting evidence layer provides the original document fragments on which the answer is based. Each fragment is labeled with the source document, relevant chapter, and confidence score, and users can expand to view detailed information.

[0043] The responsibility information layer specifically handles responsibility-related content, including the name of the responsible department, the name, position, contact information (phone, email, workstation), and current availability. The system not only provides direct responsibility information but also offers backup contacts and supervisor information based on organizational structure and business processes, forming a complete chain of responsibility. For each contact, the system indicates recommended contact times and suggested contact methods, making intelligent recommendations based on calendar data and historical response patterns.

[0044] The operation suggestion layer provides further action guidance based on the query content, including application portals for relevant processes, form templates that need to be filled out, and suggestions for subsequent steps. The system will also recommend relevant rules, regulations, and training materials based on the user's historical queries and departmental characteristics to help users gain a deeper understanding of related issues.

[0045] After receiving data, the front-end interface performs visual rendering based on data type and user preferences. Text content features clear layout and highlights key information, using different colors and fonts. Process information is presented intuitively in flowchart format, supporting interactive expansion and step navigation. Responsibility information uses a card-style design, displaying contact person's avatar, basic information, and status indicator. All output content supports interactive operations; users can click to view detailed information, copy contact information, save useful content, or provide feedback on the quality of the responses. The system also records user actions on output content for continuous optimization of output strategies and personalized recommendations.

[0046] This invention addresses the pain points of traditional query methods in enterprise knowledge management, such as low response efficiency, inaccurate information retrieval, and ambiguous responsibility attribution. It provides a complete intelligent solution, achieving significant technological improvements and application value. The system constructs a dedicated enterprise knowledge graph, transforming scattered rules, processes, and organizational structures into a structured semantic network. This enables machines to deeply understand and intelligently connect various business processes. When answering process inquiries, it automatically associates responsible departments, specific contacts, and complete operational instructions, forming an executable solution. Employing a hybrid architecture integrating vector retrieval, keyword retrieval, and graph retrieval, the system accurately matches query needs in different scenarios, significantly improving the accuracy and recall rate of complex queries. Leveraging the natural interaction and intelligent generation capabilities of a large language model, the system supports multi-turn natural language dialogue, generating personalized and structured professional answers based on user identity and context, greatly enhancing user experience and problem-solving efficiency.

[0047] The system features a dedicated responsibility association engine that automatically extracts and dynamically updates responsibility chain information, verifies accuracy in real time, and integrates enterprise communication status to ensure the timeliness and availability of responsibility information. Simultaneously, as a continuous learning knowledge platform, the system records and analyzes user interaction data to continuously optimize the quality of the knowledge base, identify knowledge gaps, and accumulate organizational experience, achieving systematic knowledge accumulation and self-evolution. In terms of security and management, the system provides role-based access control, comprehensive operation auditing, data isolation mechanisms, and flexible deployment options, fully meeting enterprise-level security compliance and operational management needs.

[0048] Overall, this system not only solves the efficiency and accuracy problems in traditional information retrieval, but more importantly, it builds a complete intelligent ecosystem for enterprise knowledge, transforming scattered knowledge assets into operable and iterative intelligent service capabilities. This significantly improves organizational operational efficiency, decision-making quality, and knowledge transfer effectiveness, providing a solid and reliable technical foundation for enterprises' digital transformation and intelligent upgrading.

[0049] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A process consultation and responsibility association question-and-answer method based on enterprise knowledge base, characterized in that, include: Obtain dialogue information regarding the inquiry process, consultation, and responsibility association; The query is performed in the enterprise knowledge base based on the dialogue information to obtain the corresponding query results. The enterprise knowledge base includes a vector database for vector representation information, a document database for storing document content, and a knowledge graph for storing responsibility associations. The query results are analyzed using a large model to generate and display the final relevant process information and responsibility association information.

2. The method according to claim 1, characterized in that, Dialogue information is obtained through a user interface, which includes prompts about the dialogue information, status prompts for displaying the progress of the question and answer, and a display area for showing the results of the final related process information and responsibility association information.

3. The method according to claim 1, characterized in that, The construction process of the knowledge graph in the enterprise knowledge base includes: Obtain raw data, including rules and regulations, operation manuals, process descriptions, and organizational structure data; construct a knowledge graph based on the raw data using named entity recognition and relationship extraction methods, wherein the knowledge graph includes process nodes, department nodes, personnel nodes, rule nodes, and the semantic relationships between them.

4. The method according to claim 1, characterized in that, The construction process of the vector database and document database in the enterprise knowledge base includes: Acquire document data, wherein the document data includes rules and regulations, operation manuals and process descriptions, preprocess the document data, segment and vectorize the preprocessed document data, and obtain the vectorized representation of the segmented document data to construct a vector database; The document data is mapped and indexed to generate structured data for retrieval, thereby building a document database.

5. The method according to claim 1, characterized in that, The process of searching the enterprise knowledge base based on dialogue information includes: The dialogue information is transformed to generate query information. Based on the query information, the enterprise knowledge base is searched using vector retrieval, keyword retrieval, and graph retrieval methods to generate corresponding search candidate results and relevance scores. A comprehensive score is calculated based on the relevance scores. The search candidate results are sorted and filtered based on the comprehensive scores to generate the final search results, i.e., the query results.

6. The method according to claim 1, characterized in that, The process of analyzing query results using a large model includes: The query results, dialogue information, and corresponding session context information are combined to form a large model prompt. The large model generates data based on the large model prompt, and finally obtains the relevant process information and responsibility association information.

7. The method according to claim 6, characterized in that, The data generation process based on the large model prompts employs a constraint decoding method and factual checks. The constraint decoding method includes increasing the probability weight of key information during the data generation decoding process and detecting whether there are any omissions in the final generated result. The real-time checks include matching the final generated result with the query result and backtracking the data to correct the final generated result.

8. A process consultation and responsibility association question-and-answer system based on an enterprise knowledge base, characterized in that: Used to perform the method described in any one of claims 1-7.

Images