Multi-chain collaborative retrieval enhancement system, method and equipment based on large model and storage medium

Abstract

The invention relates to the technical field of artificial intelligence and large models, in particular to a multi-chain collaborative retrieval enhancement system, method and device based on a large model and a storage medium. The data extraction module is used for analysis; the mixed retrieval knowledge base construction module is used for cutting the document and constructing a mixed retrieval knowledge base by using the obtained document fragments; the query processing module is used for acquiring a query request of a user, performing mixed retrieval in the mixed retrieval knowledge base according to the query request, and outputting a document fragment corresponding to a retrieval result; the fusion sorting module is used for weighted fusion sorting; the answer generation module is used for inputting the document fragments and the query request into a large language model to generate answers; and the result output module is used for outputting an answer, wherein the answer comprises a traceability mark. According to the method, unified analysis and deep fusion can be carried out on the multi-modal heterogeneous data, multi-dimensional retrieval reasoning can be carried out, and the retrieval efficiency and the retrieval accuracy are improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and large model technology, specifically to a multi-chain collaborative retrieval enhancement system, method, device, and storage medium based on a large model. Background Technology

[0002] With the widespread application of large language models in natural language processing tasks, several core shortcomings in real-world applications have gradually become apparent, constituting a key bottleneck for further technological implementation. First, because the model's knowledge representation relies entirely on static parameterized memory from the training phase, it lacks the ability to continuously absorb new knowledge. Consequently, it often fails to provide accurate and effective answers to time-sensitive queries, severely limiting its practical value in rapidly changing environments. Second, publicly available models cannot access internal enterprise knowledge bases (such as product design documents, customer agreements, and other proprietary materials), creating private data blind spots. This makes it difficult to effectively apply the model in business scenarios that rely on internal knowledge, such as professional consulting and corporate decision-making. Furthermore, when faced with queries outside the training distribution or semantically ambiguous inputs, the model is prone to generating seemingly plausible or even completely fictitious content, such as incorrectly citing legal clauses or misinterpreting professional terminology. Such illusions and semantic biases pose a serious challenge to its reliability and security.

[0003] To overcome these shortcomings, existing technical solutions mainly include two mainstream paths: fine-tuning and Retrieval Augmentation (RAG). While fine-tuning techniques can leverage domain-labeled data for targeted model optimization, this process typically relies on a large number of high-quality labeled samples, resulting in high implementation costs, long iteration cycles, and a high risk of declining model generalization ability, leading to significant performance degradation when facing unknown tasks. Traditional RAG methods enhance the accuracy of generated content by introducing external knowledge retrieval mechanisms, alleviating some of the pressure on model knowledge updates. However, their retrieval modules usually rely solely on similarity matching between semantic vectors, exhibiting several inherent limitations: mechanical fixed-length slicing of input text often disrupts the logical integrity of structured content such as tables and formulas; at the semantic understanding level, it struggles to effectively distinguish between synonyms and polysemous words, resulting in unsatisfactory recall and precision; furthermore, this method generally ignores the hierarchical structure and logical connections within documents, such as citation relationships between legal clauses or process dependencies in technical documents, leading to a lack of overall consistency and interpretability in the retrieval results.

[0004] The core of the current technological bottleneck lies in two aspects. On the one hand, existing methods lack the ability to uniformly parse and deeply integrate multimodal heterogeneous data (such as mixed text and image layouts in scanned documents, cross-references between tables and text, etc.). On the other hand, their retrieval mechanisms remain fragmented, failing to collaboratively integrate signals from multiple dimensions, such as semantic similarity, syntactic structure information, and knowledge graph associations, into a coherent and efficient retrieval reasoning framework. Therefore, constructing a new generation of language model enhancement mechanisms that can deeply integrate multi-source heterogeneous knowledge, possess dynamic update capabilities, and support complex semantic understanding has become a key technical problem that urgently needs to be addressed in this field. Summary of the Invention

[0005] One of the objectives of this invention is to provide a multi-chain collaborative retrieval enhancement system based on a large model, which performs unified analysis and deep fusion of multimodal heterogeneous data, and can collaboratively integrate signals from multiple dimensions such as semantic similarity, syntactic structure information, and knowledge graph association into a coherent and efficient retrieval reasoning framework, thereby improving retrieval efficiency and accuracy.

[0006] To achieve the above objectives, a multi-chain collaborative retrieval enhancement system based on a large model is provided, including:

[0007] Document receiving module: Used to receive multi-format documents uploaded by users;

[0008] Data extraction module: used to parse the received documents; the parsing includes document format judgment, text conversion and structured parsing;

[0009] Hybrid retrieval knowledge base construction module: used to segment the parsed document into several document fragments; construct a hybrid retrieval knowledge base based on the document fragments, the hybrid retrieval knowledge base including a vector database, a keyword database, and a knowledge graph; Vector database: generates and stores semantic vectors of document fragments; Keyword database: extracts and stores keywords based on document fragments; Knowledge graph: extracts entity-relation triples based on document fragments and updates the knowledge graph;

[0010] Query processing module: This module acquires user query requests, performs a hybrid search within the hybrid retrieval knowledge base based on these requests, and outputs the corresponding document fragments. The hybrid search includes vector similarity retrieval, keyword matching, and knowledge graph reasoning. Vector similarity retrieval calculates the cosine similarity between the query vector and the document fragment vector. Keyword matching retrieval extracts technical terms from the query requirements for keyword retrieval. Knowledge graph reasoning retrieval locates entities and associated paths in the knowledge graph database based on the query request.

[0011] The fusion and sorting module is used to perform weighted fusion and sorting of the document fragments corresponding to the search results;

[0012] Answer generation module: Used to input the document fragments ranked at the front and the query request into the large language model to generate answers;

[0013] Results output module: Used to output the answer, which includes a source tag.

[0014] Furthermore, in the hybrid retrieval knowledge base construction module, documents are segmented according to semantic segmentation rules, wherein the semantic segmentation rules are as follows:

[0015] The semantic segmentation rules are as follows:

[0016] The smallest unit of segmentation is a complete sentence;

[0017] The maximum length of a fragment does not exceed a predetermined length threshold.

[0018] The table in the document is protected as a whole, and each identified complete table is treated as an independent fragment and not merged with other text content.

[0019] Furthermore, the fusion sorting module performs weighted fusion sorting using a fusion sorting formula, which is as follows:

[0020] Score=α·cos(vec_Q,vec_S)+β·KeyMatch(S)+γ·GraphRel(S)

[0021] Where α+β+γ=1, vec_Q and vec_S are the vector of the question and the shard vector, respectively, keyMatch is the matching score between the key and the shard key, and GraphRel(S) is the matching score between the entity and the entity in the graph database.

[0022] Furthermore, the generation of the traceability marker in the result output module includes the following steps:

[0023] Based on the sorting results and document fragment source information provided by the fusion sorting step, locate its source, including {document ID, page number, starting line number};

[0024] The marked content is determined according to a predefined priority rule, in which fragments originating from vector retrieval are given the highest priority;

[0025] When answer fragments originate from multiple sources, the information from those sources will be merged.

[0026] Output a traceability tag containing the source information or the merged source information, which is attached to the generated answer in a predetermined format.

[0027] The second objective of this invention is to provide a multi-chain collaborative retrieval enhancement method based on a large model, comprising the following steps:

[0028] Document receiving steps: Receive multi-format documents uploaded by users;

[0029] Data extraction steps: Parsing the received document; the parsing includes document format judgment, text conversion and structured parsing;

[0030] The steps for constructing a hybrid retrieval knowledge base are as follows: The parsed document is segmented into several document fragments; a hybrid retrieval knowledge base is constructed based on these document fragments, comprising a vector database, a keyword database, and a knowledge graph; the vector database generates and stores semantic vectors for the document fragments; the keyword database extracts and stores keywords based on the document fragments; and the knowledge graph extracts entity-relation triples based on the document fragments and updates the knowledge graph.

[0031] Query processing steps: Obtain the user's query request, perform a hybrid search in the hybrid retrieval knowledge base according to the query request, and output the document fragments corresponding to the search results; Hybrid retrieval includes vector similarity retrieval, keyword matching, and knowledge graph reasoning; Vector similarity retrieval: Calculate the cosine similarity between the query vector and the document fragment vector; Keyword matching retrieval: Extract professional terms from the query requirements and perform keyword retrieval; Knowledge graph reasoning retrieval: Locate entities and related paths in the knowledge graph database based on the query request;

[0032] The fusion and sorting step involves weighted fusion and sorting of the document fragments corresponding to the search results.

[0033] Answer generation steps: Input the document fragments ranked at the front end and the query request into the large language model to generate the answer;

[0034] Output steps: Output the answer, which includes a source tag.

[0035] Furthermore, in the hybrid retrieval knowledge base construction step, documents are segmented according to semantic segmentation rules, wherein the semantic segmentation rules are as follows:

[0036] The smallest unit of segmentation is a complete sentence;

[0037] The maximum length of a fragment does not exceed a predetermined length threshold.

[0038] The table in the document is protected as a whole, and each identified complete table is treated as an independent fragment and not merged with other text content.

[0039] Furthermore, in the fusion sorting step, a weighted fusion sort is performed using a fusion sorting formula, which is as follows:

[0040] Score=α·cos(vec_Q,vec_S)+β·KeyMatch(S)+γ·GraphRel(S)

[0041] Where α+β+γ=1, vec_Q and vec_S are the vector of the question and the shard vector, respectively, keyMatch is the matching score between the key and the shard key, and GraphRel(S) is the matching score between the entity and the entity in the graph database.

[0042] Furthermore, the generation of the traceability marker in the result output step includes the following steps:

[0043] Based on the sorting results and document fragment source information provided by the fusion sorting step, locate its source, including {document ID, page number, starting line number};

[0044] The marked content is determined according to a predefined priority rule, in which fragments originating from vector retrieval are given the highest priority;

[0045] When answer fragments originate from multiple sources, the information from those sources will be merged.

[0046] Output a traceability tag containing the source information or the merged source information, which is attached to the generated answer in a predetermined format.

[0047] A third objective of this invention is to provide a computer-readable storage medium storing a computer program that, when executed, implements the steps of the multi-chain collaborative retrieval enhancement method based on a large model as described above.

[0048] The fourth objective of this invention is to provide an electronic device, including a memory, a processor, and a computer program stored in the memory, characterized in that: when the processor executes the program, it implements the steps of the multi-chain collaborative retrieval enhancement method based on a large model as described above.

[0049] Principles and advantages:

[0050] 1. Multimodal Hybrid Retrieval and Dynamic Knowledge Update: This invention effectively overcomes the shortcomings of large-scale model knowledge solidification by constructing a hybrid retrieval knowledge base that integrates vectors, keywords, and knowledge graphs, and establishing an automated parsing and incremental update pipeline after new document uploads. The system can integrate external multi-format documents (including internal enterprise private data) in real time, ensuring that search results are always based on the latest and most relevant knowledge, significantly improving the accuracy of responses in time-sensitive scenarios (such as financial market analysis and policy and regulatory inquiries).

[0051] 2. Deep Analysis and Integrity Protection of Structured Content: By employing a multimodal parsing engine that combines natural language processing, layout analysis, and specialized table conversion technologies, this invention can accurately identify and retain complex structured content in documents (such as merged cells and formula calculation logic). Its semantic segmentation rules ensure that tables, charts, etc., are processed as a whole unit, avoiding information fragmentation and semantic damage caused by traditional fixed-length slicing, thereby maintaining the logical integrity and accuracy of knowledge during retrieval and generation.

[0052] 3. Multi-chain Collaborative Retrieval and Semantic Accuracy Enhancement: The system employs a three-way parallel retrieval mechanism (vector similarity, keyword Boolean matching, and knowledge graph reasoning) combined with an adaptive fusion ranking algorithm, comprehensively considering semantic relevance, terminology accuracy, and entity association. This method effectively suppresses model illusions and reduces misinterpretations and incorrect citations of technical terms. Particularly in high-risk fields such as law and medicine, the reliability and security of the generated answers are fundamentally enhanced.

[0053] 4. End-to-end controllable output and traceability mechanism: By outputting answers with precise traceability markers (document ID, page number, line number) and assigning priorities to different search sources, the system not only improves the interpretability of the results but also allows users to verify the source of the answers. This mechanism enhances user trust and provides a solid technical foundation for enterprise compliance auditing and quality control. Attached Figure Description

[0054] Figure 1 This is a logical block diagram of a multi-chain collaborative retrieval enhancement system based on a large model, according to an embodiment of the present invention.

[0055] Figure 2 This is a flowchart of a multi-chain collaborative retrieval enhancement method based on a large model, according to an embodiment of the present invention. Detailed Implementation

[0056] The following detailed description illustrates the specific implementation method:

[0057] Example 1

[0058] A multi-chain collaborative retrieval enhancement system based on a large model, basically as follows: Figure 1 As shown, it includes a server and a user terminal, which are connected via remote communication. The server includes the following modules:

[0059] Document receiving module: used to receive multi-format documents uploaded by users through the user terminal; the multi-format documents include DOCX, PDF, IMAGE, EXCEL, PPT, TXT, MD, JSON and HTML;

[0060] Data extraction module: used to parse the received documents; the parsing includes document format judgment, text conversion and structured parsing; including document format judgment unit, text conversion unit and structured parsing unit;

[0061] The document format determination unit is used to identify the document format and layout structure through layout recognition technology.

[0062] The text conversion unit is used to convert documents into processable text content, such as converting PDF / IMAGE format documents into text. This unit has multi-mode processing capabilities:

[0063] For scanned documents, noise-reducing OCR processing is performed, specifically including:

[0064] Adaptive thresholding is used to eliminate image noise.

[0065] High-precision character recognition is achieved through the Tesseract OCR engine;

[0066] The BERT error correction model is used to correct the recognition results, with an error correction rate of no less than 98%.

[0067] For raw electronic documents, the text content is directly extracted and the metadata is parsed, including:

[0068] DOCX format: Parsing style hierarchy and document embedding objects;

[0069] PDF format: Extract font embedding information and vector graphics data.

[0070] The structured parsing unit is used for deep parsing of the extracted text content, specifically including:

[0071] Text semantics are extracted using a large language model based on natural language processing technology; the large language model is accessed through a large model interface: connecting to models such as DeepSeek and Qwen through the openai component.

[0072] The table is extracted using table conversion technology, and the identified table (table recognition: running the DeepDoc table recognition tool) is converted into HTML or Markdown structured format. The table conversion technology includes:

[0073] (1) OCR-based table region detection model;

[0074] (2) Identify cell coordinates using the Tabula algorithm;

[0075] (3) Convert the table structure according to the formula, for example: <td rowspan＝"2"> Merge cells Standard cell ;

[0076] (4) Retain the calculation logic of the formulas in the table to ensure semantic integrity.

[0077] A hybrid retrieval knowledge base construction module is used to segment the parsed documents into several document fragments. This module segments documents according to semantic segmentation rules, where the smallest unit of a fragment is a complete sentence; the maximum length of a fragment does not exceed a predetermined length threshold; and tables in the document are protected as a whole, with each identified complete table treated as an independent fragment and not merged with other text content. A hybrid retrieval knowledge base is constructed based on the document fragments. This knowledge base includes a vector database, a keyword database, and a knowledge graph. The vector database generates and stores semantic vectors for the document fragments; the keyword database extracts and stores keywords based on the document fragments; and the knowledge graph extracts entity-relation triples based on the document fragments and updates the knowledge graph. The construction of the knowledge graph includes:

[0078] (1) Entity extraction: The BiLSTM-CRF model is used to identify domain entities;

[0079] (2) Relation extraction: based on a pre-trained relation classifier;

[0080] (3) Map update mechanism: Incremental update is triggered after a new document is uploaded.

[0081] The query processing module is used to obtain query requests (i.e., questions or requirements) sent by users through the client, perform mixed retrieval in the hybrid retrieval knowledge base according to the query request, and output the document fragments corresponding to the search results. The hybrid retrieval includes vector similarity retrieval, keyword matching, and knowledge graph reasoning. Vector similarity retrieval: calculates the cosine similarity between the query vector and the document fragment vector; keyword matching retrieval: extracts professional terms from the query requirements for keyword retrieval; knowledge graph reasoning retrieval: locates entities and related paths in the knowledge graph database based on the query request.

[0082] The fusion and sorting module is used to perform weighted fusion and sorting of document fragments corresponding to the search results. The fusion and sorting module uses a fusion and sorting formula as follows:

[0083] Score=α·cos(vec_Q,vec_S)+β·KeyMatch(S)+γ·GraphRel(S)

[0084] Where α+β+γ=1, vec_Q and vec_S are the vector of the question raised by the user and the shard vector, respectively, keyMatch is the matching score between the keyword and the shard keyword, and GraphRel(S) is the matching score between the entity and the entity in the graph database.

[0085] Answer generation module: used to input the document fragments ranked at the top of the list and the query request into the large language model to generate answers; the process of hybrid retrieval and answer generation is as follows:

[0086] (1) Phase 1: Three-way parallel retrieval: Vector retrieval segmentation recalls the Top 20 document fragments; Keyword matching segmentation recalls the Top 15 document fragments; Graph reasoning segmentation recalls 10 related paths; Result deduplication;

[0087] (2) Phase 2: The fusion sorting step calculates the Score value according to the formula to achieve recall and re-sorting;

[0088] (3) Stage 3: Return to the Top-3 document fragments for the answer generation step. Generate prompts using the relevant slice knowledge (i.e., document fragments and query requests) through the LLM big model template, and then generate the answer through the LLM big model.

[0089] Result output module: Used to output the answer, which includes a source tracing marker. The generation of the source tracing marker in the result output module includes the following steps:

[0090] Based on the sorting results and document fragment source information provided by the fusion sorting step, locate its source, including {document ID, page number, starting line number};

[0091] The marked content is determined according to a predefined priority rule, in which fragments originating from vector retrieval are given the highest priority;

[0092] When answer fragments originate from multiple sources, the information from those sources will be merged.

[0093] Output a traceability tag containing the source information or the merged source information, which is attached to the generated answer in a predetermined format.

[0094] A multi-chain collaborative retrieval enhancement method based on a large model, such as Figure 2 As shown, it includes the following steps:

[0095] Document receiving steps: Receive multi-format documents uploaded by users; the multi-format documents include DOCX, PDF, IMAGE, EXCEL, PPT, TXT, MD, JSON and HTML;

[0096] Data extraction steps: Parsing the received document; the parsing includes document format judgment, text conversion, and structured parsing; including a document format judgment unit, a text conversion unit, and a structured parsing unit;

[0097] The document format determination unit is used to identify the document format and layout structure through layout recognition technology.

[0098] The text conversion unit is used to convert documents into processable text content, such as converting PDF / IMAGE format documents into text. This unit has multi-mode processing capabilities:

[0099] For scanned documents, noise-reducing OCR processing is performed, specifically including:

[0100] Adaptive thresholding is used to eliminate image noise.

[0101] High-precision character recognition is achieved through the Tesseract OCR engine;

[0102] The BERT error correction model is used to correct the recognition results, with an error correction rate of no less than 98%.

[0103] For raw electronic documents, the text content is directly extracted and the metadata is parsed, including:

[0104] DOCX format: Parsing style hierarchy and document embedding objects;

[0105] PDF format: Extract font embedding information and vector graphics data.

[0106] The structured parsing unit is used for deep parsing of the extracted text content, specifically including:

[0107] Text semantics are extracted using a large language model based on natural language processing technology; the large language model is accessed through a large model interface: connecting to models such as DeepSeek and Qwen through the openai component.

[0108] The table is extracted using table conversion technology, and the identified table (table recognition: running the DeepDoc table recognition tool) is converted into HTML or Markdown structured format. The table conversion technology includes:

[0109] (1) OCR-based table region detection model;

[0110] (2) Identify cell coordinates using the Tabula algorithm;

[0111] (3) Convert the table structure according to the formula, for example: <td rowspan＝"2"> Merge cells Standard cell ;

[0112] (4) Retain the calculation logic of the formulas in the table to ensure semantic integrity.

[0113] The hybrid retrieval knowledge base construction steps are as follows: The parsed document is segmented into several document fragments; the document is segmented according to semantic segmentation rules, whereby the smallest unit of a fragment is a complete sentence; the maximum length of a fragment does not exceed a predetermined length threshold; tables in the document are protected as a whole, with each identified complete table treated as an independent fragment and not merged with other text content. A hybrid retrieval knowledge base is constructed based on the document fragments. This hybrid retrieval knowledge base includes a vector database, a keyword database, and a knowledge graph; the vector database generates and stores semantic vectors for the document fragments; the keyword database extracts and stores keywords based on the document fragments; and the knowledge graph extracts entity-relation triples based on the document fragments and updates the knowledge graph.

[0114] Query processing steps: Obtain the user's query request (i.e., ask a question or submit a requirement); perform a hybrid search in the hybrid retrieval knowledge base based on the query request; and output the document fragments corresponding to the search results. The hybrid search includes vector similarity retrieval, keyword matching, and knowledge graph reasoning. Vector similarity retrieval: calculate the cosine similarity between the query vector and the document fragment vector; keyword matching retrieval: extract professional terms from the query requirements and perform keyword retrieval; knowledge graph reasoning retrieval: locate entities and related paths in the knowledge graph database based on the query request.

[0115] The fusion and ranking step involves weighted fusion and ranking of the document fragments corresponding to the search results. This fusion and ranking step uses a fusion and ranking formula, which is as follows:

[0116] Score=α·cos(vec_Q,vec_S)+β·KeyMatch(S)+γ·GraphRel(S)

[0117] Where α+β+γ=1, vec_Q and vec_S are the vector of the question raised by the user and the shard vector, respectively, keyMatch is the matching score between the keyword and the shard keyword, and GraphRel(S) is the matching score between the entity and the entity in the graph database.

[0118] Answer generation steps: Input the document fragments ranked at the front end and the query request into the large language model to generate the answer;

[0119] The process of hybrid retrieval and answer generation is as follows:

[0120] (1) Phase 1: Three-way parallel retrieval: Vector retrieval segmentation recalls the Top 20 document fragments; Keyword matching segmentation recalls the Top 15 document fragments; Graph reasoning segmentation recalls 10 related paths; Result deduplication;

[0121] (2) Phase 2: The fusion sorting step calculates the Score value according to the formula to achieve recall and re-sorting;

[0122] (3) Stage 3: Return to the Top-3 document fragments for the answer generation step. Generate prompts using the relevant slice knowledge (i.e., document fragments and query requests) through the LLM big model template, and then generate the answer through the LLM big model.

[0123] Result output step: Output the answer, which includes a source tracing marker. The generation of the source tracing marker in the result output step includes the following steps:

[0124] Based on the sorting results and document fragment source information provided by the fusion sorting step, locate its source, including {document ID, page number, starting line number};

[0125] The marked content is determined according to a predefined priority rule, in which fragments originating from vector retrieval are given the highest priority;

[0126] When answer fragments originate from multiple sources, the information from those sources will be merged.

[0127] Output a traceability tag containing the source information or the merged source information, which is attached to the generated answer in a predetermined format.

[0128] A computer-readable storage medium includes a large-model-based multi-chain collaborative retrieval enhancement program, which, when executed by a processor, implements the steps of the large-model-based multi-chain collaborative retrieval enhancement method described above.

[0129] Those skilled in the art will understand that implementing all or part of the processes in the above-described multi-chain collaborative retrieval enhancement method based on a large model can be accomplished by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium. When executed, the program can include the processes of various embodiments of the above-described multi-chain collaborative retrieval enhancement method based on a large model. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0130] A computer device includes a storage medium as described above, and a processor for executing a multi-chain collaborative retrieval enhancement program based on a large model stored in the storage medium. The computer device can be implemented as various types of devices, such as personal computers (PCs), server devices, mobile devices, etc. The processor controls all functions of the electronic device by executing a program stored in the storage medium on the electronic device.

[0131] Example 2

[0132] The difference between Embodiment 2 and Embodiment 1 is that the system further includes:

[0133] A human-machine collaborative optimization module, connected to the source tracing output module, the fusion sorting module, and the hybrid retrieval knowledge base, forms a closed-loop optimization system. This module includes:

[0134] A feedback receiving unit is used to receive user feedback on the output answer; the feedback includes:

[0135] Display positive and negative feedback: for example, a user's likes / dislikes for an answer;

[0136] Answer correction: The correct answer text provided by the user;

[0137] Source tagging: Users can verify the correctness of source segments supporting the answer or mark incorrect source segments.

[0138] The effectiveness evaluation unit is used to automatically evaluate the effectiveness of the current search-generation process based on user feedback and predefined indicators; and to generate structured evaluation results; the evaluation indicators include at least:

[0139] Answer quality scoring: A comprehensive evaluation is conducted based on user positive and negative feedback, the consistency between the answer and the source fragment, and the confidence level of the generated answer;

[0140] Search performance score: The score is based on the relevance of high-ranking segments in the search results to user feedback.

[0141] Source tracing credibility score: based on user confirmation of the correctness of the source fragment.

[0142] A strategy tuning unit is used to dynamically adjust the weight parameters (α, β, γ) of the fusion ranking module and / or optimize the retrieval strategy of the hybrid retrieval engine based on the evaluation results. The strategy tuning unit dynamically adjusts the weight parameters of the fusion ranking module using a reinforcement learning algorithm, and its reward function is constructed based on positive user feedback or answer evaluation scores. The optimization of the retrieval strategy includes, but is not limited to:

[0143] Adjust the enabling or disabling of each search channel and its priority; for example, if multiple evaluations find that keyword retrieval is consistently ineffective for a specific query, its weight can be temporarily reduced or the channel can be skipped.

[0144] Dynamically expand or tighten the search scope; for example, dynamically adjust the number of candidate segments returned by vector retrieval (Top-K value) based on query complexity;

[0145] Query rewriting: Based on historical feedback, queries that do not perform well are automatically rewritten into synonym queries that are more likely to recall relevant fragments;

[0146] A knowledge base optimization unit is used to optimize the hybrid retrieval knowledge base based on user feedback; the optimization includes:

[0147] For answers that receive negative feedback or are marked as incorrect, and their source fragments, the corresponding fragments in the knowledge base will be subject to weight decay or error marking, thereby reducing their priority when retrieved subsequently.

[0148] For correct answers or positive feedback provided by users, the corresponding segments in the knowledge base are weighted more effectively.

[0149] This solution reconstructs the traditional one-way retrieval and generation process into a continuously iterative closed-loop system by introducing a human-machine collaborative optimization module. This system can automatically evaluate performance based on user feedback, dynamically adjust ranking weights and retrieval strategies, and mark low-quality or erroneous fragments in the knowledge base. This allows the system to continuously self-optimize after deployment, becoming increasingly accurate with use, fundamentally solving the problem of traditional RAG systems being "statically fixed" and unable to learn from usage. It achieves closed-loop iterative optimization and system self-evolution capabilities.

[0150] By proactively incorporating user feedback mechanisms and cleaning and optimizing the knowledge base, the system can gradually eliminate unreliable information sources, continuously improving the overall quality and reliability of the output answers. This process is transparent and controllable, greatly enhancing users' trust in the AI ​​system and providing crucial safeguards for its application in high-risk areas.

[0151] The method further includes a human-machine collaborative optimization step, which specifically includes:

[0152] Feedback reception: This step receives user feedback on the output answer; the feedback includes: explicit positive and negative feedback, answer correction, and source annotation;

[0153] Display positive and negative feedback: for example, a user's likes / dislikes for an answer;

[0154] Answer correction: The correct answer text provided by the user;

[0155] Source tagging: Users can verify the correctness of source segments supporting the answer or mark incorrect source segments.

[0156] Effectiveness evaluation steps: The effectiveness of this search-generation process is automatically evaluated based on user feedback and predefined indicators, and a structured evaluation result is generated; the evaluation indicators include at least: answer quality score, search effectiveness score, and source traceability credibility score.

[0157] Answer quality scoring: A comprehensive evaluation is conducted based on user positive and negative feedback, the consistency between the answer and the source fragment, and the confidence level of the generated answer;

[0158] Search performance score: The score is based on the relevance of high-ranking segments in the search results to user feedback.

[0159] Source tracing credibility score: based on user confirmation of the correctness of the source fragment.

[0160] Strategy optimization steps: Based on the evaluation results, dynamically adjust the weight parameters (α, β, γ) of the fusion ranking step and / or optimize the retrieval strategy of the hybrid retrieval step; the strategy optimization unit dynamically adjusts the weight parameters of the fusion ranking module through a reinforcement learning algorithm, and its reward function is constructed based on positive user feedback or answer evaluation scores. The optimization of the retrieval strategy includes, but is not limited to:

[0161] Adjust the enabling or disabling of each search channel and its priority; for example, if multiple evaluations find that keyword retrieval is consistently ineffective for a specific query, its weight can be temporarily reduced or the channel can be skipped.

[0162] Dynamically expand or tighten the search scope; for example, dynamically adjust the number of candidate segments returned by vector retrieval (Top-K value) based on query complexity;

[0163] Query rewriting: Based on historical feedback, queries that do not perform well are automatically rewritten into synonym queries that are more likely to recall relevant fragments;

[0164] Knowledge base optimization steps: Optimize the hybrid retrieval knowledge base based on user feedback; the optimization includes:

[0165] For answers that receive negative feedback or are marked as incorrect, and their source fragments, the corresponding fragments in the knowledge base will be subject to weight decay or error marking, thereby reducing their priority when retrieved subsequently.

[0166] For correct answers or positive feedback provided by users, the corresponding segments in the knowledge base are weighted more effectively.

[0167] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, based on the guidance provided in this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A multi-chain collaborative retrieval enhancement system based on a large model, characterized in that, include: Document receiving module: Used to receive multi-format documents uploaded by users; Data extraction module: used to parse the received documents; the parsing includes document format judgment, text conversion and structured parsing; Hybrid retrieval knowledge base construction module: used to segment the parsed document into several document fragments; construct a hybrid retrieval knowledge base based on the document fragments, the hybrid retrieval knowledge base including a vector database, a keyword database, and a knowledge graph; Vector database: generates and stores semantic vectors of document fragments; Keyword database: extracts and stores keywords based on document fragments; Knowledge graph: extracts entity-relation triples based on document fragments and updates the knowledge graph; Query processing module: This module acquires user query requests, performs a hybrid search within the hybrid retrieval knowledge base based on these requests, and outputs the corresponding document fragments. The hybrid search includes vector similarity retrieval, keyword matching, and knowledge graph reasoning. Vector similarity retrieval calculates the cosine similarity between the query vector and the document fragment vector. Keyword matching retrieval extracts technical terms from the query requirements for keyword retrieval. Knowledge graph reasoning retrieval locates entities and associated paths in the knowledge graph database based on the query request. The fusion and sorting module is used to perform weighted fusion and sorting of the document fragments corresponding to the search results; Answer generation module: Used to input the document fragments ranked at the front and the query request into the large language model to generate answers; Results output module: Used to output the answer, which includes a source tag.

2. The multi-chain collaborative retrieval enhancement system based on a large model according to claim 1, characterized in that: In the hybrid retrieval knowledge base construction module, documents are segmented according to semantic segmentation rules, wherein the semantic segmentation rules are: The smallest unit of segmentation is a complete sentence; The maximum length of a fragment does not exceed a predetermined length threshold. The table in the document is protected as a whole, and each identified complete table is treated as an independent fragment and not merged with other text content.

3. The multi-chain collaborative retrieval enhancement system based on a large model according to claim 2, characterized in that: The fusion sorting module performs weighted fusion sorting using a fusion sorting formula, which is as follows: Score=α·cos(vec_Q,vec_S)+β·KeyMatch(S)+γ·GraphRel(S) Where α+β+γ=1, vec_Q and vec_S are the vector of the question raised by the user and the shard vector, respectively, keyMatch is the matching score between the keyword and the shard keyword, and GraphRel(S) is the matching score between the entity and the entity in the graph database.

4. The multi-chain collaborative retrieval enhancement system based on a large model according to claim 3, characterized in that: The process of generating the source tracing marker in the result output module includes the following steps: Based on the sorting results and document fragment source information provided by the fusion sorting step, locate its source, including {document ID, page number, starting line number}; The marked content is determined according to a predefined priority rule, in which fragments originating from vector retrieval are given the highest priority; When answer fragments originate from multiple sources, the information from those sources will be merged. Output a traceability tag containing the source information or the merged source information, which is attached to the generated answer in a predetermined format.

5. A multi-chain collaborative retrieval enhancement method based on a large model, characterized in that: Includes the following steps: Document receiving steps: Receive multi-format documents uploaded by users; Data extraction steps: Parsing the received document; the parsing includes document format judgment, text conversion and structured parsing; The steps for constructing a hybrid retrieval knowledge base are as follows: The parsed document is segmented into several document fragments; a hybrid retrieval knowledge base is constructed based on these document fragments, comprising a vector database, a keyword database, and a knowledge graph; the vector database generates and stores semantic vectors for the document fragments; the keyword database extracts and stores keywords based on the document fragments; and the knowledge graph extracts entity-relation triples based on the document fragments and updates the knowledge graph. Query processing steps: Obtain the user's query request, perform a hybrid search in the hybrid retrieval knowledge base according to the query request, and output the document fragments corresponding to the search results; Hybrid retrieval includes vector similarity retrieval, keyword matching, and knowledge graph reasoning; Vector similarity retrieval: Calculate the cosine similarity between the query vector and the document fragment vector; Keyword matching retrieval: Extract professional terms from the query requirements and perform keyword retrieval; Knowledge graph reasoning retrieval: Locate entities and related paths in the knowledge graph database based on the query request; The fusion and sorting step involves weighted fusion and sorting of the document fragments corresponding to the search results. Answer generation steps: Input the document fragments ranked at the front end and the query request into the large language model to generate the answer; Output steps: Output the answer, which includes a source tag.

6. The multi-chain collaborative retrieval enhancement method based on a large model according to claim 5, characterized in that: In the hybrid retrieval knowledge base construction step, documents are segmented according to semantic segmentation rules, where the semantic segmentation rules are as follows: The smallest unit of segmentation is a complete sentence; The maximum length of a fragment does not exceed a predetermined length threshold. The table in the document is protected as a whole, and each identified complete table is treated as an independent fragment and not merged with other text content.

7. The multi-chain collaborative retrieval enhancement method based on a large model according to claim 6, characterized in that: In the fusion sorting step, a weighted fusion sort is performed using a fusion sorting formula, which is as follows: Score=α·cos(vec_Q,vec_S)+β·KeyMatch(S)+γ·GraphRel(S) Where α+β+γ=1, vec_Q and vec_S are the vector of the question raised by the user and the shard vector, respectively, keyMatch is the matching score between the keyword and the shard keyword, and GraphRel(S) is the matching score between the entity and the entity in the graph database.

8. The multi-chain collaborative retrieval enhancement method based on a large model according to claim 7, characterized in that: The process of generating the source tracing marker in the result output step includes the following steps: Based on the sorting results and document fragment source information provided by the fusion sorting step, locate its source, including {document ID, page number, starting line number}; The marked content is determined according to a predefined priority rule, in which fragments originating from vector retrieval are given the highest priority; When answer fragments originate from multiple sources, the information from those sources will be merged. Output a traceability tag containing the source information or the merged source information, which is attached to the generated answer in a predetermined format.

9. A computer-readable storage medium storing a computer program, characterized in that: When the program is executed, it implements the steps of the multi-chain collaborative retrieval enhancement method based on a large model as described in any one of claims 5-8.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that: When the processor executes the program, it implements the steps of the multi-chain collaborative retrieval enhancement method based on a large model as described in any one of claims 5-8.

Images