Intelligent question answering system implementation method and system

Abstract

The invention relates to the technical field of intelligent questioning and answering, in particular to an intelligent questioning and answering system implementation method and system.The intelligent questioning and answering system implementation method comprises the following steps of document collection and preprocessing, document dicing, document layout analysis and table layout analysis; the method has the beneficial effects that a semantic association network and a multi-modal index system are formed through offline document collection, preprocessing, slicing, layout analysis, data extraction and knowledge graph construction; in the online part, the capabilities of Embedding vectorization, multi-index joint retrieval, tensor reordering, AI database integration and large language model generation are combined, accurate semantic understanding and rapid knowledge matching of user questions are realized, high-quality answers are generated, and interactive feedback optimization is supported.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent question answering, in particular to an intelligent question answering system implementation method and system. BACKGROUND

[0002] With the rapid development of information technology, the amount of data generated by humans is growing explosively, and text data accounts for a large proportion. Whether it is various documents, reports, knowledge bases within enterprises, or a large number of web pages, articles, forum posts on the Internet, they all contain rich knowledge and information. However, how to quickly and accurately obtain valuable knowledge from these massive text data and generate meaningful answers has become a problem to be solved.

[0003] Traditional information retrieval systems mainly rely on keyword matching to find documents related to user queries. This approach often appears to be inadequate when dealing with complex semantic queries and scenarios that require integration of multiple information sources. It is difficult to understand the true semantic intent behind user questions and cannot effectively integrate relevant information from different documents to generate coherent and comprehensive answers.

[0004] At the same time, the emergence of large language models (LLM) has brought a huge breakthrough in text generation. These models can generate fluent and natural text based on their own training data, but when faced with specific domains or the need to accurately reference given document content, they often exhibit "hallucination" phenomena, i.e., generating answers that do not match or lack basis for actual document content. This is because LLMs are not specifically designed to retrieve and reference accurate information from a specific set of documents.

[0005] In this context, RAG implementation methods have emerged, combining retrieval and generation to leverage the strengths of both. However, in practical applications, existing RAG implementation methods still have some limitations, such as inaccurate vector retrieval leading to failure to find the best relevant documents, semantic gap between questions and answers, inability to fully exploit complex structured information in documents (such as tables, charts, etc.), and difficulty in handling large-scale real-time updated document data, which limits the further improvement of their performance and effectiveness. Therefore, developing the next generation of RAG engine implementation methods has become an inevitable trend, aiming to overcome existing problems and achieve more efficient, accurate, and intelligent knowledge retrieval and generation to meet the urgent needs of knowledge acquisition and application in different fields and scenarios, such as intelligent customer service, enterprise knowledge management, intelligent medical auxiliary diagnosis, academic research assistants, and other important fields. SUMMARY

[0006] The purpose of the present application is to provide an intelligent question answering system implementation method and system to solve the problems raised in the background.

[0007] To achieve the above object, the present application provides the following technical scheme: an intelligent question and answer system implementation method, comprising the following steps:

[0008] S1: document collection and preprocessing, collecting text, PDF, Word, PPT multi-format documents, implementing coding format unification processing, removing special characters, error formats and redundant attachments, and establishing a preprocessed document library;

[0009] S2: document segmentation, text documents are segmented according to semantic paragraphs, table data is segmented according to row and column structure, multimedia documents are partitioned according to content type, and dynamic segmentation threshold is set to control the block size in the range of 512-1024 tokens;

[0010] S3: document layout analysis, analyzing the physical structure of the document, and identifying the font / typography features of the text area;

[0011] S4: table layout analysis, using a model to analyze table merged cells, and outputting structured JSON data;

[0012] S5: data extraction model application, constructing a fine-tuning model based on layout features and text features, and implementing entity relationship joint extraction;

[0013] S6: knowledge graph construction, constructing a knowledge graph using the extracted entities and relationships, taking entities as nodes and relationships as edges to construct a semantic network;

[0014] S7: document Embedding vectorization, generating document semantic vectors, combining layout feature encoders to generate structure feature vectors, and outputting representation vectors through a feature fusion layer;

[0015] S8: index establishment, establishing a three-index index of full-text index, vector index and graph index, supporting multi-condition joint query, and reducing response time;

[0016] S9: question input and preprocessing, after the user inputs a question, preprocessing the question, including word segmentation, part-of-speech tagging, and removing stop words, extracting key words and core semantic information in the question;

[0017] S10: question Embedding vectorization and query rewriting, inputting the preprocessed question into the Embedding model to generate the vector representation of the question, and simultaneously using the query rewriting model to rewrite the original question;

[0018] S11: multi-index joint retrieval, according to the vector representation and key words of the question, performing joint retrieval in multiple indexes to locate relevant document blocks and related nodes in the knowledge graph;

[0019] S12: Retrieval result sorting, input the preliminary retrieved information into the TensorReranker model, and re-sort the retrieval results by considering various factors;

[0020] S13: AINativeDatabase query and integration, according to the sorted retrieval results, query in AINativeDatabase, and integrate multi-source data;

[0021] S14: LLM generates answers, inputs the integrated data into a large language model to generate natural, accurate and complete answers, and presents them according to the user's required format and style;

[0022] S15: Answer output and interactive feedback, output the generated answers to the user and provide an interactive interface to collect user feedback information and adjust and optimize system parameters.

[0023] Preferably, in step S1, after collecting multi-format documents, the coding format unification processing is implemented through pre-set rules and algorithms to remove special characters, incorrect formats and redundant attachments, ensure the consistency and accuracy of the documents in the subsequent processing process, and establish a pre-processing document library for subsequent steps.

[0024] Preferably, in step S5, the fine-tuning model is based on natural language processing and machine learning technology, which realizes the joint extraction of key entities, relationships and event information in the document by fusing the layout features and text features of the document, using specific algorithms and model architectures, and provides accurate data basis for knowledge graph construction.

[0025] Preferably, in step S14, the prompt engineering is used to require the LLM to explicitly cite knowledge graph nodes and document fragments during the answer generation process, to enhance the credibility and authority of the answers, and to ensure that the generated answers have specific data sources support.

[0026] Preferably, in step S15, a feedback analysis module is constructed to collect user satisfaction scores and modification records, dynamically adjust retrieval weights and answer generation parameters through reinforcement learning algorithm, continuously optimize system performance and user experience according to user feedback, realize system continuous optimization, and form a closed-loop intelligent question and answer interaction process.

[0027] A system for implementing an intelligent question and answer system method, comprising:

[0028] The document processing module collects documents in multiple formats, including text, PDF, Word, and PPT, implements unified encoding format processing, removes special characters, incorrect formats, and redundant attachments, and establishes a pre-processed document library. It uses appropriate segmentation strategies based on document type and content characteristics to ensure that the segments are of appropriate size. It uses the document layout model and table layout model to analyze the physical structure of documents and tables, respectively, parses the table structure, and outputs structured JSON data.

[0029] The data extraction module inputs the segmented documents into a fine-tuned model based on natural language processing and machine learning technologies, integrates layout features with text features, and performs entity-relationship joint extraction to extract key entity, relationship, and event information from the document.

[0030] The knowledge graph construction module uses the extracted entities and relationships to construct a knowledge graph, using entities as nodes and relationships as edges to build a semantic network;

[0031] The document vectorization module converts the segmented documents and their related structural information into fixed-length vectors through the Embedding model, generates a document semantic vector, combines it with the layout feature encoder to generate a structural feature vector, and outputs the representation vector through the feature fusion layer;

[0032] The index building module builds multiple index structures including full-text index, vector index, sparse vector index, and graph index based on document vectors and keywords, supporting multi-condition joint queries.

[0033] The question processing module receives user input questions and preprocesses them, including word segmentation, part-of-speech tagging, and stop word removal, extracting keywords and core semantic information from the questions. The preprocessed questions are then fed into the Embedding model to generate a vector representation of the questions, while the query rewriting model is used to rewrite the original questions.

[0034] The retrieval module performs a joint search across multiple indexes based on the question's vector representation and keywords, comprehensively considering keyword matching and semantic similarity to quickly locate document blocks related to the question and related nodes in the knowledge graph;

[0035] The retrieval result ranking module inputs the initially retrieved document blocks, knowledge graph nodes, and related relationship information into the TensorReranker model, and re-ranks the retrieval results based on tensor operations and comprehensive consideration of multiple factors;

[0036] The AINativeDatabase query and integration module queries the AINativeDatabase based on the sorted search results, obtains and integrates multi-source data to form a coherent and complete semantic representation;

[0037] The answer generation module inputs the integrated data into a large language model (LLM), requires the LLM to explicitly cite knowledge graph nodes and document fragments through prompt engineering, generates natural, accurate and complete answers, and presents them in the format and style required by the user;

[0038] The answer output and interactive feedback module outputs the generated answer to the user and provides an interactive interface to collect user satisfaction scores and modification records, builds a feedback analysis module, dynamically adjusts the retrieval weight and answer generation parameters through reinforcement learning, and realizes continuous optimization of the system.

[0039] Preferably, in the document processing module, text documents are divided by semantic paragraphs, table data is divided by row and column structure, multimedia documents are divided by content type, and dynamic block threshold is set to control the block size in the range of 512-1024 tokens, to ensure that the semantic information is complete and the processing efficiency is not affected.

[0040] Preferably, the fine-tuning model in the data extraction module deeply fuses the layout features and text features of the document through specific algorithms and model architectures, accurately extracts key information in the document, and provides accurate data basis for knowledge graph construction.

[0041] Preferably, the index establishment module establishes a ternary index structure of full-text index, vector index, sparse vector index and graph index, which can quickly locate related information according to different query requirements, reduce response time and improve the retrieval efficiency of the system.

[0042] Preferably, the feedback analysis module in the answer output and interactive feedback module collects user feedback information and dynamically adjusts the retrieval weight and answer generation parameters using reinforcement learning algorithm, so that the system can continuously optimize performance and user experience according to the actual needs of the user, forming a closed-loop intelligent question and answer interaction process.

[0043] Compared with the prior art, the beneficial effects of the present application are:

[0044] The intelligent question and answer system implementation method and system provided by the application form a semantic correlation network and a multi-modal index system through offline document collection, preprocessing, cutting, layout analysis, data extraction and knowledge graph construction; the online part combines Embedding vectorization, multi-index joint retrieval, tensor reordering, AI database integration and large language model generation capability to realize accurate semantic understanding and rapid knowledge matching of user questions, generate high-quality answers and support interactive feedback optimization. The overall implementation method is committed to solving the problems of insufficient semantic understanding and difficult multi-source data fusion of traditional retrieval systems, improving the efficiency and accuracy of question and answer in complex document scenarios, and providing an intelligent and accurate information acquisition and knowledge discovery platform for users. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 The method flowchart of the application. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solution of the application clear and complete, and the advantages more clear and obvious, the embodiments of the application will be further described in detail below with reference to the drawings. It should be understood that the specific embodiments described herein are part of the embodiments of the application, not all embodiments, and are only used to explain the embodiments of the application, and do not limit the embodiments of the application. All other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0047] Embodiment one, please refer to Figure 1 The application provides a technical solution: an intelligent question and answer system implementation method, comprising the following steps:

[0048] S1: document collection and preprocessing

[0049] Collect various documents such as text files, PDF, Word documents, PPT, etc. Perform preliminary cleaning on the documents to remove irrelevant characters, formatting errors and unnecessary attachments, and unify the coding format to prevent garbled characters and prepare for S2 processing.

[0050] S2: document cutting

[0051] According to the document type and content characteristics extracted in S1, adopt appropriate cutting strategy. For example, for text files, cutting can be performed according to paragraphs, chapters or fixed length; for tables, cutting can be performed according to rows or columns. Ensure that each cutting size is moderate, which can completely contain semantic information and will not affect the processing efficiency too much, and prepare for S3 and S4.

[0052] S3: document layout analysis (document layout model)

[0053] The document layout model is used to analyze the physical structure of the document, identify text regions, image regions, table regions, etc. Determine the layout order, direction, font size, etc. of the text to better understand the organization of the document content, provide assistance for subsequent information extraction, and apply the model to S5.

[0054] S4: Table layout analysis (Table layout model)

[0055] For table documents, the table layout model is used to analyze the structure of the table, including the boundaries of rows, columns, and cells, merged cells, etc. Accurately extract the data content in the table and convert it into a structured data format, such as a two-dimensional array, data frame, or specified json structure, to facilitate S5 processing and analysis.

[0056] S5: Data extraction model application

[0057] The data extraction model is used to input the cut document data from S2 (S3, S4). This model is a fine-tuned model for natural language processing and machine learning techniques, which identifies and extracts key entities, relationships, events, etc. in the document, and is used for S6 knowledge graph construction.

[0058] S6: Knowledge graph construction

[0059] The knowledge graph is constructed using the entities and relationships extracted by S5. The entity is used as the node, and the relationship is used as the edge to construct a semantic network. Through the knowledge graph, information from different documents can be integrated, and potential associations between different entities can be discovered, providing rich semantic knowledge for subsequent queries and reasoning, and serving as a query for S13.

[0060] S7: Document Embedding vectorization

[0061] The cut document and its related structural information (such as document layout, table structure, etc.) from S2 (S3, S4) are converted into fixed-length vectors through the Embedding model. The Embedding model can capture the semantic information of the document, so that semantically similar documents are closer in the vector space, providing a basis for subsequent similarity calculation and retrieval, and serving as a query for S13.

[0062] S8: Index establishment

[0063] Based on the S7 document vectors and keywords, multiple index structures are established, including full-text index, vector index, sparse vector index and graph index, etc., for S13 query. Full-text index is used to quickly locate documents containing specific keywords; vector index facilitates vector similarity search to find the most similar document blocks to the query vector; sparse vector index can improve retrieval efficiency, especially when dealing with high-dimensional sparse data; graph index indexes the nodes and edges in the knowledge graph, facilitating graph query operations.

[0064] S9: Question input and preprocessing

[0065] After the user inputs the question, the question is preprocessed, including word segmentation, part-of-speech tagging, removing stop words, etc. The question is converted into a form suitable for model processing, and the keywords and core semantic information in the question are extracted to S10.

[0066] S10: Question Embedding vectorization and query rewriting

[0067] The question after S9 preprocessing is input into the Embedding model to generate the vector representation of the question. At the same time, the original question is rewritten using the query rewriting model to expand keywords, adjust sentence structure, etc. to more accurately match the information in the document, improve the recall rate and accuracy of retrieval, and then perform S11 index query.

[0068] S11: Multi-index joint retrieval

[0069] According to the vector representation and keywords of the question in S10, joint retrieval is performed in the full-text index, vector index, sparse vector index and graph index established in the S8 offline part. Considering keyword matching and semantic similarity, quickly locate the relevant document blocks and related nodes in the S6 knowledge graph.

[0070] S12: Retrieval result sorting (Tensor Reranker) Input the document blocks and knowledge graph nodes and related information retrieved in S11 into the Tensor Reranker model. Based on tensor operations, the model considers multiple factors such as the semantic similarity between the question and the document block, the quality of the document block, the correlation strength in the knowledge graph, etc. to reorder the retrieval results, and put the most relevant and valuable information in the front, which is used for S13 information integration.

[0071] S13: AI Native Database query and integration According to the sorted search results in S12, query in the AI Native Database to obtain more detailed and comprehensive related information. AI Native Database can efficiently store and manage various structured and unstructured data, including document content, knowledge graph, search results, etc., and support complex query operations. The multi-source data obtained by querying is integrated to form a coherent and complete semantic representation.

[0072] S14: LLM generates answers Input the integrated data in S13 into a large language model (LLM). LLM generates natural, accurate, and complete answers based on the input information, and can present them in the format and style required by the user. During the generation process, LLM can also reference specific content in the original document and knowledge graph to enhance the credibility and authority of the answer.

[0073] S15: Answer output and interactive feedback Output the answer generated by LLM in S14 to the user and provide an interactive interface to allow the user to evaluate the answer, ask further questions, or modify the query conditions in S9. Based on user feedback information, the parameters of the entire RAG engine can be adjusted and optimized to continuously improve system performance and user experience, so that the system can better meet user needs and achieve a closed-loop intelligent question and answer interaction process.

[0074] In example two, on the basis of example one, a system for implementing an intelligent question and answer system is proposed, which includes:

[0075] The document processing module is used to collect text, PDF, Word, PPT multi-format documents, implement coding format unification processing, remove special characters, error formats and redundant attachments, and establish a pre-processing document library; according to the document type and content characteristics, adopt appropriate cutting strategy to cut the document, ensure the cutting size is moderate; use document layout model and table layout model to analyze the physical structure of document and table respectively, parse table structure and output structured JSON data; for text documents, cut according to semantic paragraphs, table data cut according to row and column structure, multimedia documents partition according to content type, and set dynamic block threshold to control block size in the range of 512-1024 tokens, to ensure that it can contain semantic information completely, and will not be too large to affect processing efficiency.

[0076] The data extraction module inputs the segmented document into a fine-tuning model based on natural language processing and machine learning technology, fuses layout features and text features, and implements joint extraction of entities and relationships to extract key entities, relationships, and event information in the document. The fine-tuning model in the data extraction module fuses layout features and text features in depth through a specific algorithm and model architecture to accurately extract key information in the document and provide accurate data basis for knowledge graph construction.

[0077] The knowledge graph construction module constructs a knowledge graph using the extracted entities and relationships, takes entities as nodes and relationships as edges, and constructs a semantic network.

[0078] The document vectorization module converts the segmented document and its related structural information into a fixed-length vector through an Embedding model to generate a document semantic vector, combines with a layout feature encoder to generate a structural feature vector, and outputs a feature vector through a feature fusion layer.

[0079] The index establishment module establishes full-text index, vector index, sparse vector index, and graph index structures based on document vectors and keywords to support multi-condition joint query. The ternary index structure of the full-text index, vector index, sparse vector index, and graph index established by the index establishment module can quickly locate related information according to different query requirements, reduce response time, and improve the retrieval efficiency of the system.

[0080] The question processing module is used to receive user input questions, pre-process the questions, including word segmentation, part-of-speech tagging, and stop word removal, extract key words and core semantic information in the questions, input the pre-processed questions into an Embedding model to generate vector representations of the questions, and use a query rewriting model to rewrite the original questions.

[0081] The retrieval module performs joint retrieval in multiple indexes according to the vector representations and keywords of the questions, comprehensively considers keyword matching and semantic similarity, and quickly locates related document blocks and related nodes in the knowledge graph.

[0082] The retrieval result sorting module inputs the initially retrieved document blocks, knowledge graph nodes, and related relationship information into a TensorReranker model, reorders the retrieval results based on tensor operations, and comprehensively considers multiple factors.

[0083] The AINativeDatabase query and integration module queries in the AINativeDatabase according to the sorted retrieval results, obtains and integrates multi-source data to form a coherent and complete semantic representation.

[0084] An answer generation module inputs the integrated data into a large language model (LLM), requires the LLM to explicitly cite knowledge graph nodes and document fragments using prompt engineering, generates natural, accurate, and complete answers, and presents them in the format and style required by the user;

[0085] An answer output and interactive feedback module outputs the generated answers to the user and provides an interactive interface to collect user satisfaction ratings and modification records, builds a feedback analysis module, dynamically adjusts the retrieval weights and answer generation parameters through reinforcement learning, and realizes continuous optimization of the system; The feedback analysis module in the answer output and interactive feedback module collects user feedback information, dynamically adjusts the retrieval weights and answer generation parameters using reinforcement learning algorithms, so that the system can continuously optimize performance and user experience according to the actual needs of the user, forming a closed-loop intelligent question and answer interaction process

[0086] Although embodiments of the present application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for implementing an intelligent question-answering system, characterized by: The following steps are involved: S1: Document collection and preprocessing: collect documents in multiple formats, including text, PDF, Word, and PPT, implement unified encoding format processing, remove special characters, incorrect formats, and redundant attachments, and establish a preprocessed document library; S2: Document segmentation: Text documents are segmented by semantic paragraphs, table data is segmented by row and column structure, and multimedia documents are partitioned by content type. Dynamic segmentation thresholds are set to control the segment size within the range of 512-1024 tokens. S3: Document layout analysis, parsing the document's physical structure and identifying font / typography features in text areas; S4: Table layout analysis, using the model to parse table merged cells and output structured JSON data; S5: Data extraction model application, building a fine-tuning model, integrating layout features and text features, and implementing entity relationship joint extraction; S6: Knowledge graph construction, using the extracted entities and relationships to build a knowledge graph, using entities as nodes and relationships as edges to build a semantic network; S7: Document Embedding Vectorization, which generates a document semantic vector and combines it with a layout feature encoder to generate a structural feature vector, and then outputs a representation vector through a feature fusion layer; S8: Index creation: Create a ternary index consisting of full-text index, vector index, and graph index to support multi-condition joint queries and reduce response time. S9: Question input and preprocessing. After the user enters a question, the question is preprocessed, including word segmentation, part-of-speech tagging, and stop word removal to extract keywords and core semantic information from the question. S10: Question Embedding Vectorization and Query Rewriting: Input the preprocessed question into the Embedding model to generate a vector representation of the question, and use the query rewriting model to rewrite the original question. S11: Multi-index joint retrieval: Based on the vector representation and keywords of the question, a joint search is performed in multiple indexes to locate the document blocks related to the question and the relevant nodes in the knowledge graph; S12: Ranking of search results: Input the initially retrieved information into the TensorReranker model and re-rank the search results by considering multiple factors. S13: AINativeDatabase query and integration: Based on the sorted search results, query the AINativeDatabase to obtain and integrate multi-source data; S14: LLM generates answers by inputting the integrated data into a large language model to generate natural, accurate, and complete answers, and presents them in the format and style required by the user; S15: Answer output and interactive feedback: output the generated answer to the user, provide an interactive interface, collect user feedback information, and adjust and optimize system parameters.

2. The method for implementing an intelligent question-answering system according to claim 1, wherein: In step S1, after collecting multi-format documents, the encoding format is unified through preset rules and algorithms to remove special characters, incorrect formats and redundant attachments, ensure the consistency and accuracy of the documents in subsequent processing, and establish a pre-processed document library for subsequent steps.

3. The method for implementing an intelligent question-answering system according to claim 2, wherein: In step S5, the fine-tuning model constructed is based on natural language processing and machine learning technology. By integrating the layout features and text features of the document, and adopting specific algorithms and model architectures, it realizes the joint extraction of key entities, relationships, and event information in the document, providing an accurate data foundation for the construction of the knowledge graph.

4. The method for implementing an intelligent question-answering system according to claim 3, wherein: In step S14, prompt engineering is used to require LLM to explicitly reference knowledge graph nodes and document fragments in the process of generating answers, so as to enhance the credibility and authority of the answers and ensure that the generated answers are supported by specific data sources.

5. The method for implementing an intelligent question-answering system according to claim 4, characterized in that: In step S15, a feedback analysis module is constructed to collect user satisfaction scores and modification records, and the retrieval weights and answer generation parameters are dynamically adjusted through the reinforcement learning algorithm. The system performance and user experience are continuously optimized based on user feedback, so as to achieve continuous optimization of the system and form a closed-loop intelligent question-answering interaction process.

6. A system for implementing the intelligent question-answering system method according to claim 5, characterized in that: include: The document processing module is used to collect documents in multiple formats, including text, PDF, Word, and PPT, implement unified coding format processing, remove special characters, incorrect formats, and redundant attachments, and establish a pre-processed document library; Based on the document type and content characteristics, the document is segmented using an appropriate segmentation strategy to ensure that the segment sizes are appropriate. The document layout model and table layout model are used to analyze the physical structures of the document and table respectively, parse the table structure, and output structured JSON data. The data extraction module inputs the segmented documents into a fine-tuned model based on natural language processing and machine learning technologies, integrates layout features with text features, and performs entity-relationship joint extraction to extract key entity, relationship, and event information from the document. The knowledge graph construction module uses the extracted entities and relationships to construct a knowledge graph, using entities as nodes and relationships as edges to build a semantic network; The document vectorization module converts the segmented documents and their related structural information into fixed-length vectors through the Embedding model, generates a document semantic vector, combines it with the layout feature encoder to generate a structural feature vector, and outputs the representation vector through the feature fusion layer; The index building module builds multiple index structures including full-text index, vector index, sparse vector index, and graph index based on document vectors and keywords, supporting multi-condition joint queries. The question processing module receives user input questions and preprocesses them, including word segmentation, part-of-speech tagging, and stop word removal, extracting keywords and core semantic information from the questions. The preprocessed questions are then fed into the Embedding model to generate a vector representation of the questions, while the query rewriting model is used to rewrite the original questions. The retrieval module performs a joint search across multiple indexes based on the question's vector representation and keywords, comprehensively considering keyword matching and semantic similarity to quickly locate document blocks related to the question and related nodes in the knowledge graph; The retrieval result ranking module inputs the initially retrieved document blocks, knowledge graph nodes, and related relationship information into the TensorReranker model, and re-ranks the retrieval results based on tensor operations and comprehensive consideration of multiple factors; The AINativeDatabase query and integration module queries the AINativeDatabase based on the sorted search results, obtains and integrates multi-source data to form a coherent and complete semantic representation; The answer generation module feeds the integrated data into a large language model (LLM). It uses prompt engineering to require the LLM to explicitly reference knowledge graph nodes and document fragments, generating natural, accurate, and complete answers and presenting them in the format and style required by the user. The answer output and interactive feedback module outputs the generated answers to the user, provides an interactive interface, collects user satisfaction scores and modification records, builds a feedback analysis module, and dynamically adjusts the retrieval weights and answer generation parameters through reinforcement learning to achieve continuous optimization of the system.

7. A system according to claim 6, characterized in that: In the document processing module, text documents are segmented into semantic paragraphs, tabular data is segmented into blocks according to row and column structures, and multimedia documents are partitioned by content type. A dynamic segmentation threshold is set to control the block size within the range of 512-1024 tokens to ensure that the semantic information is fully contained without significantly affecting processing efficiency.

8. A system according to claim 7, characterized in that: The fine-tuning model in the data extraction module deeply integrates the layout features and text features of the document through specific algorithms and model architectures, achieving accurate extraction of key information in the document and providing an accurate data foundation for knowledge graph construction.

9. A system according to claim 8, characterized in that: The ternary index structure of full-text index, vector index, sparse vector index and graph index established by the index establishment module can quickly locate relevant information according to different query requirements, reduce response time and improve the system's retrieval efficiency.

10. A system according to claim 9, characterized in that: The feedback analysis module in the answer output and interactive feedback module collects user feedback information and uses the reinforcement learning algorithm to dynamically adjust the retrieval weight and answer generation parameters, enabling the system to continuously optimize performance and user experience based on the user's actual needs, forming a closed-loop intelligent question-answering interactive process.

Images