Large model question answering system implementation method based on enterprise knowledge and business system

Abstract

The invention discloses a large model question-answering system implementation method based on enterprise knowledge and a business system, which comprises the following steps of: performing business intention analysis on a natural language question input by a user to generate business context information; based on the service context information, performing multi-source retrieval in a pre-constructed enterprise knowledge base and a butted service system to obtain a plurality of knowledge fragments; based on the natural language problem, the knowledge fragment and the service context information, according to the entity relationship in the enterprise knowledge base, fusing to generate a structured evidence chain; and providing the evidence chain and a preset business rule to the large language model as constraint conditions, constraining the reasoning process, and generating a final answer consistent with the evidence. Relates to the technical field of enterprise intelligent question answering and knowledge management. The method avoids the defects that output content contains factual errors, outdated information or imaginary conclusions and the like. Efficient retrieval and credible generation of enterprise knowledge are realized, and the accuracy, interpretability and business closed-loop capability of the question-answering system are improved.

Description

Technical Field

[0001] This invention relates to the field of enterprise intelligent question answering and knowledge management technology, and more specifically to a method for implementing a large-scale question answering system based on enterprise knowledge and business systems. Background Technology

[0002] Currently, in enterprise-specific scenarios, general-purpose search tools and large-scale models have significant limitations when dealing with professional knowledge question answering. This limitation primarily manifests as a pronounced "illusion" defect in the output content, meaning the output may contain factual errors, outdated information, or fabricated conclusions. This problem is particularly problematic in industries with extremely high accuracy requirements, such as finance, healthcare, and education, where such issues can lead to serious consequences. For example, in finance, models may incorrectly provide risk assessments or investment advice for financial products, leading to flawed corporate decisions; in healthcare, they may generate incorrect disease diagnoses or treatment plans, causing serious consequences for patients; and in education, they may provide incorrect knowledge, resulting in teaching accidents or misunderstandings. This large-scale model illusion problem severely impacts the accuracy and reliability of professional knowledge retrieval for enterprises, failing to meet their high-quality knowledge requirements in specialized fields.

[0003] On the other hand, internal enterprise knowledge typically originates from multiple heterogeneous systems, such as CRM, ERP, and SCM, characterized by diverse data formats, scattered storage, and semantic inconsistencies. Traditional knowledge base construction methods heavily rely on manual sorting and ETL processes, resulting in long knowledge entry cycles, delayed updates, and difficulty in adapting to rapid changes in enterprise business. Furthermore, most existing question-and-answer systems lack an understanding of the enterprise's organizational structure, business processes, and business rules, failing to achieve cross-system information fusion and scenario-based responses. This forces employees to frequently switch between different systems for queries, leading to low work efficiency.

[0004] Therefore, how to build a professional question-and-answer system that can be deeply integrated with enterprise multi-source knowledge and business systems to achieve high accuracy, interpretability, and support for closed-loop business scenarios is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of the above problems, the present invention proposes a method for implementing a large-scale question-answering system based on enterprise knowledge and business systems to overcome or at least partially solve the above problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] This invention provides a method for implementing a large-scale question-answering system based on enterprise knowledge and business systems, comprising the following steps: The system parses the natural language questions input by the user to understand their business intent and generates business context information. Based on the aforementioned business context information, multi-source retrieval is performed in the pre-built enterprise knowledge base and the connected business systems to obtain multiple knowledge fragments; Based on the natural language question, the knowledge fragment, and the business context information, and according to the entity relationships in the enterprise knowledge base, a structured chain of evidence is generated. The evidence chain and preset business rules are provided as constraints to the large language model to constrain the reasoning process and generate a final answer consistent with the evidence.

[0008] Furthermore, it also includes: The system presents the final answer to the user and recommends subsequent actions based on the evidence chain and preset business process template; at the same time, it updates the enterprise knowledge base based on the user's feedback on the answer or recommended actions.

[0009] Furthermore, the step of parsing the natural language questions input by the user to generate business context information specifically includes: The intent type of the natural language question is identified; the intent type includes knowledge query, business operation guidance, data report query, question feedback, and casual conversation. By combining user identity information with the current dialogue context, determine the business domain and business scenario to which the current intent type belongs; Based on preset structured slots and preset business rules, the business entity parameters required for the business scenario are extracted from the natural language problem; Based on the intent type, business domain, business scenario, and business entity parameters, business context information is generated.

[0010] Furthermore, the process of constructing the enterprise knowledge base includes: Based on the enterprise's organizational structure and pre-set core business processes, heterogeneous data from multiple business systems are topologically analyzed. Based on the position of the heterogeneous data in the business decision chain and the frequency of cross-departmental calls after topological sorting, the priority of the heterogeneous data for entering the database is determined. After the heterogeneous data is entered into the database, business semantic analysis is performed to extract business scenario metadata containing entities, relationships, and associations, and a knowledge graph is constructed as an enterprise knowledge base.

[0011] Furthermore, the topology sorting includes extracting, cleaning, and mapping the heterogeneous data based on a preset enterprise-level business terminology standard.

[0012] Furthermore, the knowledge graph has a two-layer structure, specifically including a bottom general semantic layer and an upper business constraint layer. The business constraint layer corrects and constrains the underlying semantic relationships by injecting preset business rules.

[0013] Furthermore, the dynamic maintenance process of the enterprise knowledge base includes: Record the source path of knowledge fragments referenced during the question-and-answer process; when the cumulative negative user feedback for the same knowledge fragment reaches a preset threshold, automatically trigger the verification and update process for that knowledge fragment; and generate content completion suggestions for the enterprise knowledge base based on the current dialogue context.

[0014] Furthermore, based on the natural language question, the knowledge fragment, and the business context information, and according to the entity relationships in the enterprise knowledge base, a structured chain of evidence is generated; including: Based on the semantic relevance of each knowledge fragment to the natural language problem and the business scenario fit of each knowledge fragment to the business context information, a comprehensive score is calculated for each knowledge fragment; and the knowledge fragments are re-ranked based on the comprehensive score. Based on the entity relationships in the pre-built enterprise knowledge base, the reordered knowledge fragments are organized into a chain of evidence with logically supporting relationships; duplicate content from different sources is merged, and all source annotations are retained.

[0015] Furthermore, the comprehensive score is expressed as: Score(ri) = α·Rel(ri, q) + β·Trust(si) + γ·Comp(ri) + δ·Biz(Context q Tag ri ) Where Score(ri) represents the fit; Rel(ri, q) represents the semantic similarity between the i-th knowledge fragment ri and the natural language question q; Trust(si) represents the data source corresponding to the knowledge fragment ri; Comp(ri) represents the completeness of the knowledge fragment ri; Context q This represents the business context information corresponding to the current natural language question q; Tag ri The Biz (Context) label represents the business context attached to the knowledge fragment ri; q Tag ri ) represents the business context information. q Business Context Tags ri The degree of matching between them; α, β, γ and δ are all weight coefficients, and α+β+γ+δ=1.

[0016] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method for implementing a large-scale question-answering system based on enterprise knowledge and business systems, which has the following beneficial effects: This invention effectively avoids defects such as factual errors, outdated information, or fabricated conclusions in the output content by deeply integrating a large language model with multi-source heterogeneous knowledge and business processes of an enterprise. It pioneers a business-constrained reasoning framework, controlling the reasoning process of the large model within enterprise rules and evidence chains, thereby generating highly accurate, interpretable answers that can be directly applied to business scenarios. This not only significantly improves the efficiency and accuracy of employees acquiring knowledge but also forms a closed loop of "question-answer-feedback-knowledge optimization," enabling continuous accumulation and intelligent evolution of enterprise knowledge. Simultaneously, it ensures data security through access control, building a secure, reliable, and business-aware intelligent assistant for the enterprise. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the implementation method of a large-scale question-answering system based on enterprise knowledge and business systems provided in this embodiment of the invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention discloses a method for implementing a large-scale question-answering system based on enterprise knowledge and business systems, such as... Figure 1 As shown, it includes the following steps: S1. Parse the natural language questions input by the user to understand their business intent and generate business context information; S2. Based on business context information, perform multi-source retrieval in the pre-built enterprise knowledge base and the connected business systems to obtain multiple knowledge fragments; S3. Based on natural language questions, knowledge fragments, and business context information, and according to entity relationships in the enterprise knowledge base, a structured chain of evidence is generated. S4. Provide the evidence chain and preset business rules as constraints to the largest language model to constrain the reasoning process and generate a final answer consistent with the evidence; S5. Present the final answer to the user and recommend subsequent operations based on the chain of evidence and preset business process templates; at the same time, update the enterprise knowledge base based on user feedback on the answer or recommended operations.

[0021] Next, each of the above steps will be explained in detail.

[0022] In step S1 above, the natural language query input by the user is parsed for business intent to generate business context information; wherein the user input supports both text and voice input. Specifically, this embodiment of the invention abandons the simple "business consultation / small talk" classification and proposes a multi-level intent parsing model that is deeply integrated with the enterprise's organizational structure and business processes; specifically including: Level 1 Analysis: Intent Type Recognition. Lightweight models (such as FastText) or rules are used to quickly identify the intent type of natural language questions; these intent types include knowledge queries, business operation guidance, data report queries, problem feedback, and casual conversation, etc. Level 2 Analysis: Business Domain and Scenario Localization. For core types such as knowledge queries and business operation guidance, combining user identity information (such as department and role) with the current dialogue context, a dedicated classification model (such as a fine-tuned BERT) fine-tuned from business texts such as enterprise work orders and customer service logs is used to determine the business domain and business scenario to which the current intent type belongs; for example: sales-quotation scenario, after-sales-fault handling scenario, human resources-onboarding scenario, and finance-reimbursement process scenario.

[0023] Level 3 Analysis: Specific Tasks and Parameter Extraction. After determining the business scenario, the system uses pre-defined structured slots and business rules to extract the necessary business entity parameters from the natural language questions. For example, in an after-sales fault handling scenario, the system will proactively extract key business entities such as "product model," "fault code," and "contract number." When key business entity parameters are missing, the system will not directly reject the request but will instead initiate a clarifying multi-turn dialogue, prompting the user to supplement the necessary information based on the business process logic.

[0024] After the above multi-layered parsing, business context information is generated based on intent type, business domain, business scenario, and business entity parameters.

[0025] Compared to existing technologies that simply perform downstream processing based on intent tags, this invention transfers the parsed and generated business context information to the downstream processing module, thereby improving the processing accuracy of downstream tasks.

[0026] In step S2 above, based on business context information, multi-source retrieval is performed in the pre-built enterprise knowledge base and the connected business systems to obtain multiple knowledge fragments; specifically: (1) The construction process of the above-mentioned enterprise knowledge base includes: 1) Obtain heterogeneous data from multiple business systems (such as CRM, ERP, SCM), including internal data such as product manuals, customer service records, and training materials, as well as external data such as industry reports. This heterogeneous data includes, but is not limited to, text files, PDFs, Word documents, and Excel spreadsheets. The collected data is first transmitted to the ETL processing unit. This unit, based on the enterprise's organizational structure and pre-defined core business processes (such as leads-opportunities-contracts-delivery-service), performs a topological sorting of the heterogeneous data from multiple business systems. Specifically, based on pre-defined enterprise-level business terminology standards, it extracts, cleans, and maps the heterogeneous data; among which: Extracting heterogeneous data: Accurately extracting fields and content related to business knowledge through preset rules and algorithms; Heterogeneous data cleaning: Remove duplicate, erroneous, or irrelevant information from the data to ensure its accuracy and consistency. Specifically, the cleaning process targets the dirty data characteristics of enterprise data, such as inconsistent field naming between systems (e.g., "Customer ID" vs. "Customer_No"), abbreviations of business terms (e.g., "KYC" stands for "Know Your Customer"), and historical obsolete fields generated due to business process changes. Automatic mapping and normalization are achieved by establishing an enterprise-level business terminology standard library. Mapping heterogeneous data: Converting data of different formats into a standard format that the system can process, laying the foundation for subsequent knowledge processing.

[0027] The embodiments of the present invention also include: using natural language processing technology to segment the long text knowledge in the heterogeneous data after the above topological sorting into fine-grained knowledge units, and semantically annotating them to generate a tagged index system for subsequent retrieval and matching operations.

[0028] 2) Based on the position of the heterogeneous data in the business decision-making chain and the frequency of cross-departmental calls after topological sorting, the priority of heterogeneous data for data entry is determined; for example, frequently used product configuration rules and recently updated service SOP (Standard Operating Procedure) documents are given the highest priority. Compared with the existing technology that simply determines the entry priority based on data volume or source, the mechanism for determining the entry priority in this embodiment of the invention can dynamically reflect the distribution of knowledge value in the actual business operation of the enterprise, ensuring that, under limited storage and computing resources, the core knowledge content that has the greatest impact on business decisions and is used most frequently is loaded first. 3) This invention utilizes natural language processing (NLP) technology to perform business semantic analysis on heterogeneous data after it has been entered into the database. Unlike general knowledge graphs that only focus on entity-relationship pairs, this invention constructs a "business context-enhanced knowledge graph." Specifically, after extracting entities (such as "Product A" and "Customer B") and relationships (such as "purchase" and "consultation"), business scenario metadata is forcibly associated to generate the knowledge. This business scenario metadata includes, but is not limited to, the department to which the knowledge belongs, the applicable process stage, the relevant policy document number, the effective start and end time, and the data source system and version. For example, a knowledge point stating "Product A has a 10% discount" will be associated with contextual tags such as "Sales Policy V2.1," "Applicable to Q3-Q4 of 2024," and "Approver: Zhang San."

[0029] 4) Based on the extracted entity, relationship, and business scenario metadata, a knowledge graph is constructed as an enterprise knowledge base. The knowledge graph is constructed using a two-layer structure: a bottom general semantic layer and an upper business constraint layer. The general semantic layer is built based on a general pre-trained model. The business constraint layer corrects and constrains the underlying semantic relationships by injecting preset business rules (such as "after-sales issues must be associated with specific contract numbers") to ensure that the graph logic conforms to the actual operation of the enterprise.

[0030] Furthermore, this enterprise knowledge base employs an efficient data storage structure, supporting the storage and rapid retrieval of large-scale knowledge data. The knowledge base utilizes a hybrid indexing mechanism, including inverted indexes and vector indexes. The inverted index is built based on keywords in knowledge fragments and their associated metadata fields (such as business domains, document types, and permission tags), supporting rapid filtering of candidate knowledge units through keywords or structured attributes. The vector index encodes knowledge fragments into semantic vectors and establishes an efficient near-nearest neighbor index to calculate their semantic similarity to user questions, achieving semantic-level relevance matching. By combining the precise filtering capability of the inverted index with the semantic generalization capability of the vector index, the system can efficiently recall highly relevant results from large-scale knowledge data, significantly improving question-answering response speed, retrieval accuracy, and user experience.

[0031] (2) The dynamic maintenance process of the above-mentioned enterprise knowledge base includes: 1) Record detailed information for each knowledge entry, including the content, time, and responsible person. The ledger data is stored in a dedicated database, allowing managers to easily query and trace the knowledge change process at any time, effectively avoiding content duplication or omissions when maintained by multiple people.

[0032] 2) Record the source path of knowledge fragments referenced during the Q&A process; when negative user feedback on the same knowledge fragment accumulates to a preset threshold, the verification and update process for that knowledge fragment is automatically triggered; this update strategy can be flexibly adjusted according to the speed of business development and industry dynamics to ensure that the knowledge in the knowledge base always remains timely and accurate. Update operations include adding new knowledge, correcting and deleting old knowledge, etc., to ensure the high quality of the knowledge base content; Specifically, when a knowledge fragment is used multiple times in a short period of time to answer similar questions, but the generated answers receive a large number of user "dislikes" or "inaccurate" feedback, the system automatically marks the fragment as "high risk pending verification" and triggers an alarm notification to the corresponding knowledge responsible person.

[0033] 3) Generate content completion suggestions for the enterprise knowledge base based on the current dialogue context; for example, by analyzing the context of the user's follow-up questions, the system can identify missing "implicit knowledge" or "knowledge gaps" in the knowledge base and automatically generate knowledge completion suggestions.

[0034] 4) Optimize the index structure of the enterprise knowledge base based on user feedback. For example, adjust the word segmentation strategy of the inverted index to improve the accuracy of retrieval, and optimize the calculation method of the vector index to reduce retrieval latency.

[0035] 5) Introduce a caching mechanism to cache hot knowledge in memory, further improving retrieval speed and ensuring that the knowledge base can still operate efficiently under high concurrency access.

[0036] By constructing and dynamically maintaining the aforementioned enterprise knowledge base, a high-quality knowledge base can be efficiently built and maintained, providing a solid knowledge foundation for the intelligent question-and-answer system. This ensures that the system can provide users with accurate and fast question-and-answer services, and can be continuously optimized and updated as business and technology develop.

[0037] In step S3 above, a structured chain of evidence is generated based on natural language questions, knowledge fragments, and business context information, and according to entity relationships in the enterprise knowledge base; specifically, this includes: (1) Calculate the comprehensive score of each knowledge segment based on the semantic relevance of each knowledge segment to the natural language problem and the business scenario fit of each knowledge segment to the business context information; and re-rank each knowledge segment based on the comprehensive score. The above comprehensive score is expressed as follows: Score(ri) = α·Rel(ri, q) + β·Trust(si) + γ·Comp(ri) + δ·Biz(Context q Tag ri ) Where Score(ri) represents the fit; Rel(ri, q) represents the semantic similarity between the i-th knowledge fragment ri and the natural language question q; Trust(si) represents the data source corresponding to the knowledge fragment ri; Comp(ri) represents the completeness of the knowledge fragment ri; Context q This represents the business context information corresponding to the current natural language question q; Tag ri The Biz (Context) label represents the business context attached to the knowledge fragment ri; q Tag ri ) represents the business context information. q Business Context Tags ri The degree of matching between them; α, β, γ and δ are all weight coefficients, and α+β+γ+δ=1.

[0038] Traditional search-enhanced generation (RAG) systems typically rely solely on semantic similarity to rank knowledge fragments, easily leading to the problem of "semantically related but misaligned with business needs." For example, when a salesperson inquires about product presentation scripts, the system might prioritize parameter specifications because technical documents contain more keywords, even if semantically similar, failing to meet actual business requirements. In this invention, however, by explicitly introducing Biz (Context...)... q Tag ri This allows the system to identify the business context of a user's query (such as "sales," "expenses," or "IT support") and prioritize retrieving knowledge sources highly aligned with that context (such as sales manuals, financial regulations, and maintenance guides). Even if a technical document has a slight semantic advantage, its low Biz (Context) score due to mismatched business tags can still be mitigated. q Tag ri The score will be effectively suppressed, thereby ensuring that the final generated answer is not only accurate in content, but also in line with the actual needs of user roles and business processes.

[0039] (2) Based on the entity relationships in the pre-built enterprise knowledge base, the reordered knowledge fragments are organized into a chain of evidence with logically supporting relationships; for example, when answering a complex product configuration question, the chain of evidence may include: basic product specifications, a list of configurable options, pricing policies corresponding to different configurations, and related compatibility instructions. Multiple sources of duplicate content are merged at the chain of evidence level, and all sources are marked.

[0040] In step S4 above, the evidence chain and preset business rules are provided as constraints to the large language model to constrain the reasoning process and generate a final answer consistent with the evidence.

[0041] Specifically, this invention proposes the "Business Constraint Chain Thinking (BC-CoT)" method, which injects the retrieved evidence chain and business rules into the reasoning process of a large model.

[0042] (1) Dedicated Prompt Engineering Framework: Design a thought chain prompt template with a fixed structure to force the large model to think in the following steps: 1) Scenario restatement and confirmation: First, restate the identified business scenario and core user needs to ensure alignment.

[0043] 2) Evidence citation and credibility assessment: List each knowledge fragment and, based on its source and timeliness, self-assess its credibility and applicability in this business scenario.

[0044] 3) Application of business rules: Call or reference business rules, systems or constraints related to this scenario in the knowledge base, such as, "According to the Marketing Department's 2024 policy, discount applications for this type of product require director-level approval."

[0045] 4) Conflict Detection and Resolution: Check for contradictions between pieces of evidence or between evidence and business rules. If contradictions exist, make judgments and selections based on predefined conflict resolution strategies such as "rules take precedence over personal experience," "highly credible sources take precedence," and "latest update time takes precedence."

[0046] 5) Answer Synthesis and Boundary Declaration: Based on the above analysis, synthesize the final answer. The conditions for the answer to take effect, its scope of application, and the core evidence and rules upon which it relies must be clearly declared. Any parts that exceed the scope of the knowledge base or involve uncertainty must be explicitly declared.

[0047] (2) Explainable output: The complete BC-CoT thinking process described above can be selectively presented to users in the form of a structured summary (not the original token stream) as the basis for reasoning the answer, which greatly enhances credibility and explainability and allows users to quickly locate potential problems.

[0048] Based on the output of the BC-CoT module, the final answer is generated. The prompt templates are aligned with the company's common document styles and communication standards; for example, answers from the technical department tend to be structured and clause-based, while answers from the sales department tend to summarize key points and highlight value. Furthermore, the final answer supports formats such as images and tables. For answers involving processes and approval paths, the system automatically calls the built-in flowchart generation component to convert text descriptions into standard business process diagrams (BPMN). In practical applications, adaptive layouts can be implemented to ensure a seamless experience across multiple platforms.

[0049] In step S5 above, the final answer is presented to the user, and subsequent operations are recommended based on the chain of evidence and a preset business process template. For example, after answering "How to apply for software permissions," the system can directly recommend a quick link to fill out the "Permission Application Form" or call the relevant approval process robot. In addition, the enterprise knowledge base is updated based on user feedback on the answer or recommended operations (such as likes, dislikes, and follow-up questions), forming a business closed loop of "question and answer - feedback - knowledge optimization."

[0050] Next, a specific case will be used to illustrate the implementation method of the large-scale question-answering system based on enterprise knowledge and business systems provided by the present invention.

[0051] 1. Low-code workflow setup: Workflow Design: Based on a low-code platform (coze platform is used in this case), a complete workflow of "user input - intent recognition - branch judgment - resource recall - response result - resource recommendation" is designed. The workflow can be quickly built and flexibly adjusted through visual configuration and modular components.

[0052] Workflow Execution and Monitoring: Responds to user input in real time and completes workflow execution quickly. Monitors workflow status in real time, records key log information, and sets up exception handling mechanisms to ensure stable system operation.

[0053] 2. Project-side content aggregation: Comprehensive data query: The results returned by the comprehensive workflow can be used to query data from both internal business systems (such as training courses and new academic affairs systems) and external business systems (such as industry databases). Through data fusion processing, data silos are eliminated.

[0054] Results integration and optimization: Weighted fusion of information from different sources is performed using weighted allocation (e.g., knowledge relevance weight 0.4, credibility weight 0.3, and answer completeness weight 0.3), followed by deduplication and result sorting to ensure information completeness and consistency.

[0055] 3. Front-end interface presentation: The front-end interface is the core of the intelligent question-answering system's direct interaction with the user. Its design goal is to provide users with a simple, intuitive, and feature-rich interactive environment, ensuring that users can efficiently ask questions and receive satisfactory answers. This front-end interface includes functions such as user inquiries, question-and-answer responses, recommended questions, feedback mechanisms, and history records. The following is a detailed description of its implementation: (1) When using the intelligent question-and-answer system, the first thing users encounter is the user question interface. The interface is designed to be simple and clear, with a prominent input box at its core, supporting users to ask questions through text input or voice input. Text input is suitable for scenarios where users need to describe the problem in detail, while voice input provides users with a more convenient way to interact.

[0056] (2) After a user submits a question, the system will enter the question-and-answer phase. At this time, the system will generate a clear, intuitive, and easy-to-understand answer based on the knowledge base search results and in-depth consideration. To meet the needs of different users, the system supports multiple content presentation methods, including but not limited to text, images, tables, and flowcharts. For complex answer content, the system will flexibly select the most suitable display format according to the answer type and content characteristics to ensure the accuracy and efficiency of information transmission. At the same time, to adapt to the display needs of different devices, the system adopts adaptive layout technology. On the PC, the system can make full use of screen space to provide richer display formats; while on the mobile device, the system will optimize the layout to ensure a good reading experience on small screens and avoid content that is too crowded or difficult to read. In addition, the system also supports multilingual answers and can automatically switch languages ​​according to the user's language preferences or system settings to meet the needs of international users.

[0057] (3) To further enhance the user experience, the system will provide a recommended question function below the answer area. This function analyzes the semantics and context of the user's question based on the user's query history and current question, and mines potential needs to recommend questions related to the current question. The recommended questions are displayed in a list format, with each recommended question accompanied by a brief description or keywords to help users quickly understand the question content. The number of recommended questions will be flexibly adjusted according to screen space and user needs to ensure the cleanliness and usability of the interface. More importantly, the list of recommended questions is not static, but will be dynamically updated according to the user's real-time interaction to ensure that the recommended content is always closely related to the user's current needs and guides the user to further explore knowledge.

[0058] (4) To facilitate users' review and tracing of the problem-solving process, the system also features a history function. The system automatically records users' query history and interaction records, including detailed information such as the time and content of user questions, system responses, and user feedback. These records are not only of significant value to individual users but also provide valuable data support for system optimization. The system provides a history query entry in the interface, allowing users to quickly find historical questions and answers through various methods such as timelines and keywords. History records can be categorized and displayed according to time sequence or question type, making it easy for users to quickly locate the content they need. Furthermore, to protect user privacy, the system allows users to choose to delete specific history records. In addition, the system regularly cleans up expired or irrelevant historical data to ensure the efficient operation of the history function while protecting user privacy.

[0059] In addition, the front-end interface not only provides users with an efficient and convenient interactive experience, but also continuously optimizes its functions and performance based on user feedback, thereby improving user satisfaction and usage frequency of the intelligent question-and-answer system.

[0060] 4. Follow-up questions regarding the results: Users can ask follow-up questions to the system-generated answers. Based on the previous dialogue context and knowledge base, the system can further analyze the user's intentions to ask follow-up questions, obtain more detailed and specific relevant information, realize multi-round dialogue interaction, and continuously provide users with a coherent knowledge service experience until the user's problem is satisfactorily resolved.

[0061] 5. Feedback optimization: By collecting user feedback and analyzing system operation data, system parameters and content are dynamically adjusted to achieve performance optimization. Specifically: (1) Collection of user feedback: Feedback mechanisms are an indispensable part of user-system interaction. The system includes "like" and "dislike" buttons below each answer, allowing users to quickly rate their satisfaction with the answer through simple clicks. This intuitive feedback method not only facilitates user operation but also allows the system to collect real-time, direct feedback on answer quality. Like and dislike data will be used for model training and optimization, helping the system better adjust its answering strategies and improve answer quality.

[0062] In addition to the like and dislike functions, the system provides a feedback portal where users can describe their opinions and suggestions on the answers in detail. Users can provide specific evaluations on the accuracy, timeliness, and comprehensibility of the answers, and can even add background information or raise new questions. Written feedback will be manually reviewed and used for updating and improving the knowledge base, ensuring that the system can continuously improve based on users' actual needs.

[0063] Meanwhile, the system automatically records various data during the question-and-answer process, such as query logs, answer generation time, and user interaction behavior. This data provides rich quantitative evidence for system optimization, helping the development team identify potential problems and optimization directions.

[0064] Through the aforementioned like / dislike function, text feedback function, and automatically recorded operational data, the system can comprehensively collect user feedback, providing a solid foundation for subsequent optimization.

[0065] (2) System optimization and adjustment: Based on user feedback and operational data, the system dynamically adjusts the parameters of the large model to optimize answer quality and better meet user needs. Simultaneously, the system optimizes prompts to reduce the illusion of a large model and improve the accuracy and relevance of answers. Based on user feedback, the system promptly updates the knowledge base, supplementing missing knowledge points and correcting errors to ensure the accuracy and completeness of knowledge. Furthermore, the system analyzes problems and bottlenecks in the workflow process, optimizing the workflow to improve operational efficiency and user experience.

[0066] In summary, the large-scale question-answering system implementation method based on enterprise knowledge and business systems provided by this invention has broad applicability and can be applied to multiple industries without industry limitations. Whether in education, manufacturing, service, finance, or any other industry, enterprises can flexibly apply this invention to build a knowledge-based question-answering system tailored to their own business characteristics and needs. In this way, enterprises can achieve efficient management and accurate retrieval of internal knowledge, improving their overall operational efficiency and competitiveness.

[0067] This invention constructs a question-and-answer system based on enterprise knowledge, enabling employees to quickly and accurately access work-related professional knowledge, avoiding the tedious process of blindly searching and filtering information across multiple business systems. Employees can solve work-related problems more efficiently, improving work efficiency and quality. Furthermore, the system can provide personalized knowledge recommendation services based on employees' search history and behavioral habits, further enhancing the efficiency of employees' knowledge acquisition and utilization, creating greater value for the enterprise.

[0068] This invention not only addresses the pain points of knowledge retrieval for enterprise employees but also helps enterprises optimize and innovate their knowledge management. By integrating and structuring internal business knowledge to form a unified knowledge base, enterprises can better manage and utilize their knowledge assets. Simultaneously, the system supports knowledge sharing and dissemination, promotes internal learning and communication, stimulates employee innovative thinking, and drives the enterprise's sustainable development.

[0069] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0070] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for implementing a large-scale question-answering system based on enterprise knowledge and business systems, characterized in that, Includes the following steps: The system parses the natural language questions input by the user to understand their business intent and generates business context information. Based on the aforementioned business context information, multi-source retrieval is performed in the pre-built enterprise knowledge base and the connected business systems to obtain multiple knowledge fragments; Based on the natural language question, the knowledge fragment, and the business context information, and according to the entity relationships in the enterprise knowledge base, a structured chain of evidence is generated. The evidence chain and preset business rules are provided as constraints to the large language model to constrain the reasoning process and generate a final answer consistent with the evidence.

2. The method for implementing a large-scale question-answering system based on enterprise knowledge and business systems as described in claim 1, characterized in that, Also includes: The system presents the final answer to the user and recommends subsequent actions based on the evidence chain and preset business process template; at the same time, it updates the enterprise knowledge base based on the user's feedback on the answer or recommended actions.

3. The method for implementing a large-scale question-answering system based on enterprise knowledge and business systems as described in claim 1, characterized in that, The process of parsing the natural language questions input by the user to generate business context information specifically includes: The intent type of the natural language question is identified; the intent type includes knowledge query, business operation guidance, data report query, question feedback, and casual conversation. By combining user identity information with the current dialogue context, determine the business domain and business scenario to which the current intent type belongs; Based on preset structured slots and preset business rules, the business entity parameters required for the business scenario are extracted from the natural language problem; Based on the intent type, business domain, business scenario, and business entity parameters, business context information is generated.

4. The method for implementing a large-scale question-answering system based on enterprise knowledge and business systems as described in claim 1, characterized in that, The process of building the enterprise knowledge base includes: Based on the enterprise's organizational structure and pre-set core business processes, heterogeneous data from multiple business systems are topologically analyzed. Based on the position of the heterogeneous data in the business decision chain and the frequency of cross-departmental calls after topological sorting, the priority of the heterogeneous data for entering the database is determined. After the heterogeneous data is entered into the database, business semantic analysis is performed to extract business scenario metadata containing entities, relationships, and associations, and a knowledge graph is constructed as an enterprise knowledge base.

5. The method for implementing a large-scale question-answering system based on enterprise knowledge and business systems as described in claim 4, characterized in that, The topology sorting includes extracting, cleaning, and mapping the heterogeneous data based on a preset enterprise-level business terminology standard.

6. The method for implementing a large-scale question-answering system based on enterprise knowledge and business systems as described in claim 4, characterized in that, The knowledge graph has a two-layer structure, specifically including a bottom general semantic layer and an upper business constraint layer. The business constraint layer corrects and constrains the underlying semantic relationships by injecting preset business rules.

7. The method for implementing a large-scale question-answering system based on enterprise knowledge and business systems as described in claim 1, characterized in that, The dynamic maintenance process of the enterprise knowledge base includes: Record the source path of knowledge fragments referenced during the question-and-answer process; when the cumulative negative user feedback for the same knowledge fragment reaches a preset threshold, automatically trigger the verification and update process for that knowledge fragment; and generate content completion suggestions for the enterprise knowledge base based on the current dialogue context.

8. The method for implementing a large-scale question-answering system based on enterprise knowledge and business systems as described in claim 1, characterized in that, Based on the natural language question, the knowledge fragment, and the business context information, and according to the entity relationships in the enterprise knowledge base, a structured chain of evidence is generated; including: Based on the semantic relevance of each knowledge fragment to the natural language problem and the business scenario fit of each knowledge fragment to the business context information, a comprehensive score is calculated for each knowledge fragment; and the knowledge fragments are re-ranked based on the comprehensive score. Based on the entity relationships in the pre-built enterprise knowledge base, the reordered knowledge fragments are organized into a chain of evidence with logically supporting relationships; duplicate content from different sources is merged, and all source annotations are retained.

9. The method for implementing a large-scale question-answering system based on enterprise knowledge and business systems as described in claim 8, characterized in that, The overall score is expressed as follows: Score(s) = α·Relation(s, q) + β·Trust(s) + γ·Comp(s) + δ·We(Context q ,Tag ri ) Where Score(ri) represents the fit; Rel(ri, q) represents the semantic similarity between the i-th knowledge fragment ri and the natural language question q; Trust(si) represents the data source corresponding to the knowledge fragment ri; Comp(ri) represents the completeness of the knowledge fragment ri; Context q This represents the business context information corresponding to the current natural language question q; Tag ri The Biz (Context) label represents the business context attached to the knowledge fragment ri; q Tag ri ) represents the business context information. q Business Context Tags ri The degree of matching between them; α, β, γ and δ are all weight coefficients, and α+β+γ+δ=1.

Images