A document filtering system and method for enhancing retrieval by incorporating verification mechanisms

Abstract

This invention discloses a retrieval enhancement document filtering system and method with integrated verification mechanisms, belonging to the field of document filtering technology. The system includes: a user input and query parsing module, which extracts key information through natural language processing and converts it into high-dimensional semantic vectors, meta-tags, and keyword sets; a multi-source document retrieval module, which obtains documents from multiple data sources using hybrid retrieval and generates a candidate set after initial screening and sorting; a credibility assessment and security verification module, which generates scores and tags after multi-dimensional evaluation and filters qualified documents; a document consistency detection module, which detects document conflicts and handles sorting to ensure logical consistency; a document acquisition and generation module, which inputs qualified documents into a generation model to generate answers with citations; and a result output and traceability module, which outputs the answers and records the entire process data to ensure traceability. This invention aims to ensure the accuracy and credibility of generated content through a multi-dimensional verification mechanism.

Description

Technical Field

[0001] This invention relates to the field of document filtering technology, specifically to a retrieval enhancement document filtering system and method that integrates verification mechanisms. Background Technology

[0002] With the widespread application of large language models and retrieval-enhanced generation (RAG) techniques, an increasing number of intelligent systems are relying on external knowledge documents to generate answers. However, current RAG methods have several problems, particularly in high-risk domains, where using untrusted or erroneous documents can lead to serious consequences. Traditional RAG systems typically rely on a one-time retrieval and directly pass the recalled documents to the generation module, lacking sufficient verification of document quality, resulting in potentially risky generated results. Furthermore, current document screening methods often fail to effectively address consistency checks and security issues between documents, thus necessitating a technology capable of automatically screening and verifying document credibility. Summary of the Invention

[0003] This invention addresses the needs and shortcomings of current technological development by providing a retrieval enhancement document filtering system and method that integrates verification mechanisms. The aim is to ensure the accuracy and credibility of the generated content through a multi-dimensional verification mechanism.

[0004] Firstly, the present invention provides a retrieval enhancement document filtering system that integrates verification mechanisms, and the technical solution adopted to solve the above-mentioned technical problems is as follows:

[0005] A document filtering system for enhanced retrieval based on integrated verification mechanisms, comprising:

[0006] The user input and query parsing module is used to parse user questions using natural language processing technology, extract key information and convert it into high-dimensional semantic vectors, meta tags and keyword sets, providing semantic support and contextual basis for subsequent modules;

[0007] The multi-source document retrieval module is used to retrieve relevant documents from multiple data sources using a hybrid retrieval method. It performs preliminary screening and weighted sorting based on factors such as timeliness and credibility, and generates a set of candidate documents for further processing by subsequent modules.

[0008] The credibility assessment and security verification module is used to evaluate the reliability of document sources and detect content risks through multiple dimensions, generate credibility scores and security labels, and filter out qualified documents that meet the standards to pass to subsequent modules.

[0009] The document consistency detection module is used to detect content conflicts between qualified documents through semantic comparison technology, process conflicts and sort them according to preset rules, and ensure that the documents that pass the detection are logically consistent and reliable.

[0010] The document acquisition and generation module is used to take the qualified documents output by the document consistency detection module as context input to generate the model and generate answers with accompanying reference information.

[0011] The results output and traceability module is used to output answers with accompanying reference information, and records the entire process data through a log system to ensure traceability. The log data is also used for strategy learning and distillation to improve the system's adaptability and reliability.

[0012] Optionally, the user input and query parsing modules involved specifically include:

[0013] The Natural Language Processing (NLP) unit receives user queries, performs preliminary analysis using NLP techniques, extracts core entities, keywords, and grammatical structures from the queries through word segmentation, named entity recognition, and dependency parsing, and further identifies potential intents and contexts to gain a deeper understanding of the query semantics.

[0014] The query transformation unit is used to convert the initially parsed query into a high-dimensional semantic vector using the pre-trained large language model BERT, ensuring that the semantics of the query are accurately expressed.

[0015] The annotation generation unit is used to annotate relevant meta tags based on the domain information of the problem, and at the same time generate a keyword set to support the domain tag constraints, access control and Boolean search of the subsequent multi-source document retrieval module;

[0016] The parsing output unit is used to output the parsed high-dimensional semantic vector, meta-tags, and keyword set, providing semantic support and contextual basis for subsequent modules.

[0017] Optionally, the multi-source document retrieval module involved specifically includes:

[0018] The preliminary retrieval unit is used to retrieve documents based on the high-dimensional semantic vectors, meta tags, and keyword sets generated by the user input and query parsing module. It adopts a hybrid strategy that combines vector retrieval, Boolean keyword retrieval, and the classic retrieval algorithm based on BM25. It calls three data sources in parallel: the enterprise's private database, the public API, and the academic index. At the same time, it selects suitable data sources to prioritize the retrieval of documents by combining the context of the query and domain tag constraints.

[0019] The preliminary screening unit is used to evaluate the source authority, timeliness, historical update frequency and review system of each document retrieved from different data sources according to the preset credibility rules, calculate the source quality vector of the document, and combine the matching degree between the document and the query and the matching of domain tags to assign a preliminary credibility score to the document and filter out documents that match the query more than a preset threshold.

[0020] The weighted sorting unit is used to sort documents by weight based on the source quality vector and the preliminary confidence score, and to prioritize retaining documents whose matching degree with the query is higher than a preset threshold.

[0021] The candidate set generation unit is used to eliminate documents from unreliable sources, ultimately forming a candidate document set. It also performs a preliminary evaluation of the relevance of each document in the candidate document set and outputs the results to subsequent modules.

[0022] Optionally, the credibility assessment and security verification modules involved specifically include:

[0023] The credibility assessment unit is used to obtain a set of candidate documents and assign an initial credibility score to each document based on its source, publication history, and peer review system.

[0024] The content security detection unit is used to scan document content, identify sensitive content, and mark potential risks by using keyword filtering, text fingerprint comparison, and entity concatenation detection.

[0025] The scoring and tagging unit is used to assign a credibility score to each document and mark it as usable, requiring review, or unusable based on the credibility score and security test results.

[0026] The qualified document screening unit is used to filter out documents with a security tag of "available" and a credibility score exceeding a set value, and pass them to subsequent modules. Documents with a security tag of "requires review" or "unavailable" are marked or removed for subsequent manual review.

[0027] Optionally, the document consistency detection module involved specifically includes:

[0028] The consistency detection unit is used to perform pairing of documents on the same topic for qualified documents selected by the credibility assessment and security verification module. It uses semantic alignment network and natural language inference model to calculate the semantic consistency and implication relationship between documents and identify whether there are semantic contradictions or logical conflicts.

[0029] The document conflict handling unit is used to filter documents according to preset rules when serious conflicts are detected between documents;

[0030] The document sorting and output unit sorts documents according to consistency scores, prioritizing documents that are logically consistent and whose conflicts do not meet the requirements, and passing the checks to subsequent modules to ensure that the content generated later is based on a consistent and reliable information source.

[0031] Optionally, the document acquisition and generation modules involved specifically include:

[0032] The document acquisition unit is used to acquire qualified documents output by the document consistency detection module, ensure that the security label of the acquired qualified documents is usable, and then pass them to the answer generation unit;

[0033] The answer generation unit is used to input qualified documents labeled as safe as available as contextual information into the generation model, while building a Prompt with reference placeholders to ensure that the generated content can accurately point to the source document;

[0034] The answer output and recording unit is used to evaluate the confidence level of the answer generated by the generative model and output the final answer based on the evaluation results, including the answer generated by the generative model, cited references, relevant document numbers, and confidence evaluation results.

[0035] Optionally, the generative model involved may employ a general large language model or a domain-optimized large language model to generate accurate, coherent, and traceable natural language output based on the security tags output by the user question and document consistency detection module as usable qualified documents.

[0036] Optionally, the output and traceability modules involved specifically include:

[0037] The results output unit is used to output the answer with citation information, including the answer generated by the generative model, cited references, relevant document numbers, and confidence assessment results;

[0038] The recording and storage unit is used to record all document retrieval, scoring, strategy decision-making and generation results in the log system, and write each state-action-reward in the generation process into a JSON-Ledger to provide data support for subsequent strategy optimization and model training;

[0039] The policy learning and distillation unit is used to train policies using accumulated log data, optimize retrieval policies, document filtering mechanisms and generative models, and periodically distill the policy network to generate lightweight auxiliary models to accelerate inference and reduce system response latency.

[0040] Secondly, this invention provides a retrieval enhancement method for generating and filtering documents by integrating verification mechanisms. The technical solution adopted to solve the above-mentioned technical problems is as follows:

[0041] A document filtering method for enhanced retrieval with integrated verification mechanisms, based on the system described in the first aspect, includes the following steps:

[0042] S1. The user input and query parsing module uses natural language processing technology to parse user questions, extract key information, and convert it into high-dimensional semantic vectors, meta-tags, and keyword sets.

[0043] S2. Based on the high-dimensional semantic vector, meta-tags and keyword set extracted and transformed from user questions, the multi-source document retrieval module retrieves relevant documents from multiple data sources through a hybrid retrieval method, performs preliminary screening and weighted sorting based on timeliness and credibility factors, and generates a candidate document set.

[0044] S3. For each document in the candidate document set, the credibility assessment and security verification module evaluates its source reliability and detects content risks from multiple dimensions, generates a credibility score and security label for each document, and selects qualified documents that meet the standards.

[0045] S4. For the qualified documents that meet the standards, the document consistency detection module uses semantic comparison technology to detect content conflicts between qualified documents, processes the conflicts and sorts them according to preset rules, and ensures that the documents that pass the detection are logically consistent and reliable.

[0046] S5. Obtain the qualified documents output by the document consistency detection module. The document acquisition and generation module uses these documents as context input to generate the model and generates answers with accompanying reference information.

[0047] S6. The result output and traceability module outputs the answer with reference information and records the entire process data through the log system to ensure traceability.

[0048] The document filtering system and method for enhanced retrieval based on a fusion verification mechanism of the present invention have the following advantages compared with the prior art:

[0049] 1. By introducing a multi-dimensional verification mechanism, this invention can perform rigorous credibility assessment, security verification, and consistency detection on documents before generating content, thereby effectively reducing generation errors caused by using unreliable or erroneous documents and improving the quality and reliability of the generated results.

[0050] 2. By introducing multi-dimensional credibility assessment and consistency verification into the retrieval enhancement generation process, this invention ensures that the generated content is based only on reliable and consistent documents. This process not only relies on the credibility score of the document source, but also combines semantic consistency detection between documents, effectively avoiding the introduction of erroneous or inconsistent documents. Especially in the medical and legal fields, it ensures that the generated answers are based on authoritative, up-to-date and conflict-free literature, significantly improving the accuracy and credibility of the generated content.

[0051] 3. By detecting the security of candidate documents, this invention can effectively avoid the introduction of non-compliant, sensitive or harmful content. Through methods such as keyword filtering, text fingerprint comparison and entity table association, potential risks in documents (such as illegal clauses, sensitive drug information, etc.) can be identified and blocked before generation, thereby ensuring the compliance and security of the generated content.

[0052] 4. All document screening, strategy decisions, and generation results in this invention are recorded in the system log, providing detailed process data for subsequent result auditing and error investigation; the generated answers also include document citations and confidence information, ensuring that each answer can be traced back to the source document; this design greatly enhances the interpretability of the system, and is particularly suitable for fields such as law and medicine that have high requirements for compliance and traceability.

[0053] 5. This invention introduces a continuous log recording and strategy learning mechanism, which can continuously optimize document filtering, retrieval strategies and generation models, and gradually improve their performance in different application scenarios; the use of strategy distillation technology enables the system to significantly reduce response time and improve real-time reasoning capabilities while ensuring accuracy, thereby further enhancing the user experience.

[0054] 6. This invention significantly improves the accuracy, compliance, and security of generated content by introducing a comprehensive document verification mechanism based on credibility scoring, semantic consistency detection, and security auditing. It is particularly suitable for high-risk fields such as law and medicine, and has broad application prospects and profound technical value. Attached Figure Description

[0055] Appendix Figure 1 This is a module connection block diagram of Embodiment 1 of the present invention;

[0056] Appendix Figure 2 This is a flowchart of the method in Embodiment 2 of the present invention. Detailed Implementation

[0057] To make the technical solution, the technical problem solved, and the technical effect of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with specific embodiments.

[0058] Example 1:

[0059] Combined with appendix Figure 1 This embodiment proposes a retrieval enhancement document filtering system with integrated verification mechanism, which includes:

[0060] The user input and query parsing module is used to parse user questions using natural language processing technology, extract key information and convert it into high-dimensional semantic vectors, meta tags and keyword sets, providing semantic support and contextual basis for subsequent modules;

[0061] The multi-source document retrieval module is used to retrieve relevant documents from multiple data sources using a hybrid retrieval method. It performs preliminary screening and weighted sorting based on factors such as timeliness and credibility, and generates a set of candidate documents for further processing by subsequent modules.

[0062] The credibility assessment and security verification module is used to evaluate the reliability of document sources and detect content risks through multiple dimensions, generate credibility scores and security labels, and filter out qualified documents that meet the standards to pass to subsequent modules.

[0063] The document consistency detection module is used to detect content conflicts between qualified documents through semantic comparison technology, process conflicts and sort them according to preset rules, and ensure that the documents that pass the detection are logically consistent and reliable.

[0064] The document acquisition and generation module is used to take the qualified documents output by the document consistency detection module as context input to generate the model and generate answers with accompanying reference information.

[0065] The results output and traceability module is used to output answers with accompanying reference information, and records the entire process data through a log system to ensure traceability. The log data is also used for strategy learning and distillation to improve the system's adaptability and reliability.

[0066] In this embodiment, the user input and query parsing module specifically includes:

[0067] The Natural Language Processing (NLP) unit receives user queries, performs initial parsing using NLP techniques, extracts core entities, keywords, and syntactic structures from the query through word segmentation, named entity recognition (NER), and dependency parsing, and further identifies potential intents and context to gain a deeper understanding of the query semantics.

[0068] The query transformation unit is used to convert the initially parsed query into a high-dimensional semantic vector using the pre-trained large language model BERT, ensuring that the semantics of the query are accurately expressed.

[0069] The annotation generation unit is used to annotate relevant meta tags (such as "medical" tag or "legal" tag) based on the domain information of the problem (such as medical, legal, etc.), and at the same time generate a keyword set K to support the domain tag constraints, access control and Boolean search of the subsequent multi-source document retrieval module;

[0070] The parsing output unit is used to output the parsed high-dimensional semantic vector, meta tags (such as "medical" or "legal" tags), and keyword set K, providing semantic support and contextual basis for subsequent modules.

[0071] In this embodiment, the multi-source document retrieval module specifically includes:

[0072] The preliminary retrieval unit is used to retrieve documents based on the high-dimensional semantic vectors, meta tags (such as "medical" or "legal"), and keyword set K generated by the user input and query parsing module. It employs a hybrid strategy that combines vector retrieval, Boolean keyword retrieval, and the classic retrieval algorithm based on BM25. It simultaneously calls three data sources—a private enterprise database, a public API, and an academic index—to retrieve documents. At the same time, it combines the context of the query and domain tags (such as "medical" or "legal") to select suitable data sources and prioritize the retrieval of documents.

[0073] The initial screening unit evaluates documents retrieved from different data sources based on preset credibility rules, assessing each document's source authority, timeliness, historical update frequency, and peer review process. It calculates the document's source quality vector and, combined with the document's match to the query and the matching of domain tags (e.g., "medical" or "legal"), assigns a preliminary credibility score to each document, filtering out documents whose match to the query exceeds a preset threshold. This preliminary credibility score considers both the basic quality of the source and its match to the query, aiming to narrow down the processing scope for subsequent units.

[0074] The weighted sorting unit is used to sort documents by weight based on the source quality vector and the preliminary confidence score, and to prioritize retaining documents whose matching degree with the query is higher than a preset threshold.

[0075] The candidate set generation unit is used to eliminate documents from unreliable sources, ultimately forming a candidate document set. It also performs a preliminary evaluation of the relevance of each document in the candidate document set and outputs the results to subsequent modules.

[0076] In this embodiment, the credibility assessment and security verification module specifically includes:

[0077] The credibility assessment unit is used to acquire a set of candidate documents and assign an initial credibility score to each document based on its source, publication history, and peer review process. For example, documents from authoritative institutions or academic journals are given higher credibility scores, while documents from unverified or unknown sources are given lower scores. Third-party verification databases are also used to verify some documents that cannot be directly judged to ensure the reliability and professionalism of the documents. The initial credibility score mentioned here is generated after the candidate documents enter the credibility assessment stage. It focuses only on the core credibility dimensions of the document itself and does not involve the matching degree with the query. It is the starting point for in-depth evaluation of document quality and provides a basis for subsequent modules.

[0078] The content security detection unit is used to scan document content using keyword filtering, text fingerprint comparison, and entity concatenation detection to identify sensitive content and mark potential risks. For example, prescription drug names and purchase links in medical documents may be marked as sensitive content, and illegal clauses in legal documents will also be identified and marked.

[0079] The scoring and tagging unit is used to assign a credibility score to each document and mark it as usable, requiring review, or unusable (due to containing sensitive content) based on the credibility score and security detection results.

[0080] The qualified document screening unit is used to filter out documents with a security tag of "available" and a credibility score exceeding a set value, and pass them to subsequent modules. Documents with a security tag of "requires review" or "unavailable" are marked or removed for subsequent manual review.

[0081] In this embodiment, the document consistency detection module specifically includes:

[0082] The consistency detection unit is used to perform pairwise comparisons of documents on the same topic for qualified documents selected by the credibility assessment and security verification module, using semantic alignment networks (such as ESIM) and natural language reasoning (NLI) models, to calculate the semantic consistency and implication relationships between documents, and to identify whether there are semantic contradictions or logical conflicts.

[0083] The document conflict handling unit is used to filter documents according to preset rules when serious conflicts are detected between documents; for example, if the diagnostic opinions of two documents are inconsistent, the document with the more authoritative diagnosis will be selected first.

[0084] The document sorting and output unit sorts documents according to consistency scores, prioritizing documents that are logically consistent and whose conflicts do not meet the requirements, and passing the checks to subsequent modules to ensure that the content generated later is based on a consistent and reliable information source.

[0085] In this embodiment, the document acquisition and generation module specifically includes:

[0086] The document acquisition unit is used to acquire qualified documents output by the document consistency detection module, ensure that the security label of the acquired qualified documents is usable, and then pass them to the answer generation unit;

[0087] The answer generation unit is used to input qualified documents labeled as safe as available as contextual information into the generation model, while building a Prompt with reference placeholders to ensure that the generated content can accurately point to the source document;

[0088] The answer output and recording unit is used to evaluate the confidence level of the answer generated by the generative model and output the final answer based on the evaluation results, including the answer generated by the generative model, cited references, relevant document numbers, and confidence evaluation results.

[0089] If the confidence level of the answer generated by the generative model is lower than a preset threshold, a hybrid mode is enabled: this returns not only the generated answer but also the cited original text excerpt to ensure the verifiability of the generated content; if the confidence level of the answer generated by the generative model is higher than the preset threshold, only the generated answer and the citation information of relevant literature are returned. Furthermore, document usage information and cited document numbers can be output to ensure users can trace the source of the answer.

[0090] Specifically, the generative models involved employ either general-purpose large language models or domain-optimized large language models. These models are used to generate accurate, coherent, and traceable natural language output based on the security tags output by the user question and document consistency detection module, ensuring the document is a usable and qualified document. General-purpose large language models include, but are not limited to, GPT-4, GPT-3.5, LLaMA 2 (70B / 13B versions), Claude 2, and PaLM 2. These are suitable for cross-domain general question-answering scenarios and possess strong contextual understanding and natural language generation capabilities. Domain-optimized large language models include, but are not limited to, Med-PaLM (medical domain), LawGPT (legal domain), and CodeLlama (technical documentation domain). These models are fine-tuned for specific domain knowledge, enabling them to more accurately understand technical terminology and logic, and generate output that conforms to domain specifications.

[0091] In this embodiment, the result output and traceability module specifically includes:

[0092] The results output unit is used to output the answer with citation information, including the answer generated by the generative model, cited references, relevant document numbers, and confidence assessment results;

[0093] The recording and storage unit is used to record all document retrieval, scoring, strategy decision-making and generation results in the log system, and write each state-action-reward in the generation process into a JSON-Ledger to provide data support for subsequent strategy optimization and model training;

[0094] The policy learning and distillation unit is used to train policies using accumulated log data, optimize retrieval policies, document filtering mechanisms and generative models, and periodically distill the policy network to generate lightweight auxiliary models to accelerate inference and reduce system response latency.

[0095] Example 2:

[0096] Combined with appendix Figure 2This embodiment proposes a retrieval enhancement document filtering method that integrates verification mechanisms. Based on the system described in the first aspect, the implementation process includes the following steps:

[0097] S1. The user input and query parsing module uses natural language processing technology to parse user questions, extract key information, and convert it into high-dimensional semantic vectors, meta-tags, and keyword sets.

[0098] S2. Based on the high-dimensional semantic vector, meta-tags and keyword set extracted and transformed from user questions, the multi-source document retrieval module retrieves relevant documents from multiple data sources through a hybrid retrieval method, performs preliminary screening and weighted sorting based on timeliness and credibility factors, and generates a candidate document set.

[0099] S3. For each document in the candidate document set, the credibility assessment and security verification module evaluates its source reliability and detects content risks from multiple dimensions, generates a credibility score and security label for each document, and selects qualified documents that meet the standards.

[0100] S4. For the qualified documents that meet the standards, the document consistency detection module uses semantic comparison technology to detect content conflicts between qualified documents, processes the conflicts and sorts them according to preset rules, and ensures that the documents that pass the detection are logically consistent and reliable.

[0101] S5. Obtain the qualified documents output by the document consistency detection module. The document acquisition and generation module uses these documents as context input to generate the model and generates answers with accompanying reference information.

[0102] S6. The result output and traceability module outputs the answer with reference information and records the entire process data through the log system to ensure traceability.

[0103] In summary, the document filtering system and method for enhanced retrieval using the fusion verification mechanism of this invention significantly improves the accuracy, compliance, and security of the generated content by introducing a comprehensive document verification mechanism based on credibility scoring, semantic consistency detection, and security auditing. It is particularly suitable for high-risk fields such as law and medicine, and has broad application prospects and profound technical value.

[0104] The above specific examples illustrate the principles and implementation methods of the present invention in detail. These embodiments are merely for the purpose of helping to understand the core technical content of the present invention. Based on the above specific embodiments of the present invention, any improvements and modifications made to the present invention by those skilled in the art without departing from the principles of the present invention should fall within the patent protection scope of the present invention.

Claims

1. A document filtering system for enhanced retrieval based on integrated verification mechanisms, characterized in that, It includes: The user input and query parsing module is used to parse user questions using natural language processing technology, extract key information and convert it into high-dimensional semantic vectors, meta tags and keyword sets, providing semantic support and contextual basis for subsequent modules; The multi-source document retrieval module is used to retrieve relevant documents from multiple data sources using a hybrid retrieval method. It performs preliminary screening and weighted sorting based on factors such as timeliness and credibility, and generates a set of candidate documents for further processing by subsequent modules. The credibility assessment and security verification module is used to evaluate the reliability of document sources and detect content risks through multiple dimensions, generate credibility scores and security labels, and filter out qualified documents that meet the standards to pass to subsequent modules. The document consistency detection module is used to detect content conflicts between qualified documents through semantic comparison technology, process conflicts and sort them according to preset rules, and ensure that the documents that pass the detection are logically consistent and reliable. The document acquisition and generation module is used to take the qualified documents output by the document consistency detection module as context input to generate the model and generate answers with accompanying reference information. The results output and traceability module is used to output answers with accompanying reference information, and records the entire process data through a log system to ensure traceability. It also uses log data for strategy learning and distillation to improve the system's adaptability and reliability. The credibility assessment and security verification module specifically includes: The credibility assessment unit is used to obtain a set of candidate documents and assign an initial credibility score to each document based on its source, publication history, and peer review system. The content security detection unit is used to scan document content, identify sensitive content, and mark potential risks by using keyword filtering, text fingerprint comparison, and entity concatenation detection. The scoring and tagging unit is used to assign a credibility score to each document and mark it as usable, requiring review, or unusable based on the credibility score and security test results. The qualified document screening unit is used to filter out documents with a security tag of "available" and a credibility score exceeding a set value, and pass them to subsequent modules. Documents with a security tag of "requires review" or "unavailable" are marked or removed for subsequent manual review.

2. The document filtering system for enhanced retrieval based on a fusion verification mechanism according to claim 1, characterized in that, The user input and query parsing module specifically includes: The Natural Language Processing (NLP) unit receives user queries, performs preliminary analysis using NLP techniques, extracts core entities, keywords, and grammatical structures from the queries through word segmentation, named entity recognition, and dependency parsing, and further identifies potential intents and contexts to gain a deeper understanding of the query semantics. The query transformation unit is used to convert the initially parsed query into a high-dimensional semantic vector using the pre-trained large language model BERT, ensuring that the semantics of the query are accurately expressed. The annotation generation unit is used to annotate relevant meta tags based on the domain information of the problem, and at the same time generate a keyword set to support the domain tag constraints, access control and Boolean search of the subsequent multi-source document retrieval module; The parsing output unit is used to output the parsed high-dimensional semantic vector, meta-tags, and keyword set, providing semantic support and contextual basis for subsequent modules.

3. The document filtering system for enhanced retrieval based on a fusion verification mechanism according to claim 2, characterized in that, The multi-source document retrieval module specifically includes: The preliminary retrieval unit is used to retrieve documents based on the high-dimensional semantic vectors, meta tags, and keyword sets generated by the user input and query parsing module. It adopts a hybrid strategy that combines vector retrieval, Boolean keyword retrieval, and the classic retrieval algorithm based on BM25. It calls three data sources in parallel: the enterprise's private database, the public API, and the academic index. At the same time, it selects suitable data sources to prioritize the retrieval of documents by combining the context of the query and domain tag constraints. The preliminary screening unit is used to evaluate the source authority, timeliness, historical update frequency and review system of each document retrieved from different data sources according to the preset credibility rules, calculate the source quality vector of the document, and combine the matching degree between the document and the query and the matching of domain tags to assign a preliminary credibility score to the document and filter out documents that match the query more than a preset threshold. The weighted sorting unit is used to sort documents by weight based on the source quality vector and the preliminary confidence score, and to prioritize retaining documents whose matching degree with the query is higher than a preset threshold. The candidate set generation unit is used to eliminate documents from unreliable sources, ultimately forming a candidate document set. It also performs a preliminary evaluation of the relevance of each document in the candidate document set and outputs the results to subsequent modules.

4. The document filtering system for enhanced retrieval based on a fusion verification mechanism according to claim 3, characterized in that, The document consistency detection module specifically includes: The consistency detection unit is used to perform pairing of documents on the same topic for qualified documents selected by the credibility assessment and security verification module. It uses semantic alignment network and natural language inference model to calculate the semantic consistency and implication relationship between documents and identify whether there are semantic contradictions or logical conflicts. The document conflict handling unit is used to filter documents according to preset rules when serious conflicts are detected between documents; The document sorting and output unit sorts documents according to consistency scores, prioritizing documents that are logically consistent and whose conflicts do not meet the requirements, and passing the checks to subsequent modules to ensure that the content generated later is based on a consistent and reliable information source.

5. The document filtering system for enhanced retrieval based on a fusion verification mechanism according to claim 4, characterized in that, The document acquisition and generation module specifically includes: The document acquisition unit is used to acquire qualified documents output by the document consistency detection module, ensure that the security label of the acquired qualified documents is usable, and then pass them to the answer generation unit; The answer generation unit is used to input qualified documents labeled as safe as available as contextual information into the generation model, while building a Prompt with reference placeholders to ensure that the generated content can accurately point to the source document; The answer output and recording unit is used to evaluate the confidence level of the answer generated by the generative model and output the final answer based on the evaluation results, including the answer generated by the generative model, cited references, relevant document numbers, and confidence evaluation results.

6. The document filtering system for enhanced retrieval based on a fusion verification mechanism according to claim 5, characterized in that, The generative model employs a general large language model or a domain-optimized large language model, which is used to generate accurate, coherent, and traceable natural language output based on the security tags output by the user question and document consistency detection module as usable and qualified documents.

7. The document filtering system for enhanced retrieval based on a fusion verification mechanism according to claim 6, characterized in that, The result output and traceability module specifically includes: The results output unit is used to output the answer with citation information, including the answer generated by the generative model, cited references, relevant document numbers, and confidence assessment results; The recording and storage unit is used to record all document retrieval, scoring, strategy decision-making and generation results in the log system, and write each state-action-reward in the generation process into a JSON-Ledger to provide data support for subsequent strategy optimization and model training; The policy learning and distillation unit is used to train policies using accumulated log data, optimize retrieval policies, document filtering mechanisms and generative models, and periodically distill the policy network to generate lightweight auxiliary models to accelerate inference and reduce system response latency.

8. A retrieval enhancement method for document filtering that integrates verification mechanisms, characterized in that, Based on the system as described in any one of claims 1-7, the implementation process includes the following steps: S1. The user input and query parsing module uses natural language processing technology to parse user questions, extract key information, and convert it into high-dimensional semantic vectors, meta-tags, and keyword sets. S2. Based on the high-dimensional semantic vector, meta-tags and keyword set extracted and transformed from user questions, the multi-source document retrieval module retrieves relevant documents from multiple data sources through a hybrid retrieval method, performs preliminary screening and weighted sorting based on timeliness and credibility factors, and generates a candidate document set. S3. For each document in the candidate document set, the credibility assessment and security verification module evaluates its source reliability and detects content risks from multiple dimensions, generates a credibility score and security label for each document, and selects qualified documents that meet the standards. S4. For the qualified documents that meet the standards, the document consistency detection module uses semantic comparison technology to detect content conflicts between qualified documents, processes the conflicts and sorts them according to preset rules, and ensures that the documents that pass the detection are logically consistent and reliable. S5. Obtain the qualified documents output by the document consistency detection module. The document acquisition and generation module uses these documents as context input to generate the model and generates answers with accompanying reference information. S6. The result output and traceability module outputs the answer with reference information and records the entire process data through the log system to ensure traceability.

Images