A text auxiliary method based on LLM+RAG digital system

Abstract

This invention discloses a text assistance method for digital systems based on LLM+RAG, belonging to the field of artificial intelligence. Focusing on the text processing needs of digital project management, its core is the integration of a large language model and retrieval-enhanced generation technology to construct an assistance system comprising eight functional modules. This system can automatically generate system introductions, identify duplicate items and typos, provide intelligent question answering, extract key information, and supports multilingual processing. This method overcomes the limitations of traditional technologies, solving problems such as insufficient understanding of long texts, lagging knowledge updates, and balancing privacy and model performance. It can significantly improve document processing efficiency, reduce error rates, provide intelligent decision support, assist in international business, and comprehensively enhance the intelligence level of digital systems.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and specifically to a text-assisted method based on an LLM+RAG digitization system. Background Technology

[0002] In the field of intelligent text processing, existing technologies are mainly evolving in two directions: traditional NLP systems based on rule engines and end-to-end models based on deep learning. The former achieves structured data extraction through predefined templates, but suffers from drawbacks such as poor semantic generalization ability and high cost of multilingual adaptation; while the latter has made some improvements in semantic understanding, it faces problems such as strong dependence on training data and difficulty in domain transfer.

[0003] The document intelligent processing system based on the rule engine adopts a feature engineering + classifier architecture. Its shortcomings are: it relies on manually defined feature templates, which makes it difficult to adapt to emerging document types; it lacks context modeling capabilities, resulting in the loss of semantic associations across paragraphs; and it lacks a dynamic knowledge injection mechanism, making it impossible to update domain knowledge in real time.

[0004] While deep learning solutions have partially addressed the feature engineering problem, they suffer from limitations such as poor model interpretability and a black-box decision-making process. Furthermore, they fail to address the following industry pain points: difficulties in fusing multi-source heterogeneous data; attention dilution issues in long text processing; and the lack of privacy data leakage risk control mechanisms. These technical bottlenecks severely restrict the evolution of digital management systems towards intelligence.

[0005] In response to the limitations of long text comprehension capabilities, lagging domain knowledge updates, and the balance between privacy and model performance in the background technologies, this invention aims to comprehensively improve the text processing capabilities and intelligence level of digital systems. Summary of the Invention

[0006] This invention proposes a text-assisted method for LLM+RAG digitization systems, the method comprising the following steps:

[0007] S1. Intelligent text processing utilizes Large Language Model (LLM) to perform deep understanding and analysis of input text, achieving efficient semantic parsing and information extraction; S2. Retrieval Enhanced Generation (RAG) combines vector databases and semantic retrieval technology to quickly retrieve relevant information from massive documents, enhancing the knowledge reserves and generation capabilities of LLM; S3. Multi-module collaboration integrates multiple functional modules such as application system registration, intelligent display, and text input recognition to achieve comprehensive text-assisted processing; S4. Security Guarantee: Data encryption and access control technologies are employed to ensure the secure processing of sensitive information and the protection of privacy.

[0008] The use of large language models in step S1 specifically involves intelligent text processing technology. By fully utilizing the powerful capabilities of large language models, and through in-depth understanding and analysis of the input text, LLM can achieve efficient semantic parsing and information extraction. This advanced natural language processing method can not only accurately capture the core meaning of the text, but also identify complex language structures and contextual relationships, thus laying a solid foundation for subsequent processing tasks.

[0009] The enhanced retrieval generation in step S2 specifically involves cleverly combining vector databases and semantic retrieval techniques. This allows RAG to quickly and accurately retrieve information relevant to the current task from massive amounts of documents. This method significantly enhances the knowledge base and generation capabilities of LLM.

[0010] The multi-module collaboration in step S3 specifically includes a system introduction generation module, a similar system identification module, a misspelling identification module, an intelligent question-and-answer assistance module, a related item identification module, a difference identification module, an intelligent information extraction module, and a multilingual processing module.

[0011] The system introduction generation module utilizes a core engine to intelligently analyze user-entered information and system attachments, achieving natural language understanding and information extraction. This module automatically generates standardized system introductions, significantly improving work efficiency. To ensure the quality of the generated content, the module is equipped with a rigorous quality control mechanism and a manual review process.

[0012] The approximate system identification module employs advanced similarity algorithms to efficiently identify and display similar or duplicate systems. This module not only visualizes the relationships between systems but also generates intuitive system relationship graphs. It incorporates a duplicate registration early warning mechanism to effectively prevent redundant work and improve resource utilization efficiency. Furthermore, the module supports customizable similarity thresholds, allowing users to adjust the identification accuracy according to specific needs.

[0013] The misspelling detection module, based on its core engine, implements intelligent character recognition technology, providing context-aware automatic error correction. It supports multilingual misspelling databases and can identify and correct common errors in various languages. A key feature of this module is its user feedback learning mechanism, which continuously improves the misspelling database based on user corrections. Furthermore, the module supports custom dictionaries, allowing users to add specialized terminology from specific fields to further enhance recognition accuracy.

[0014] The intelligent question-answering assistance module is built upon a core engine to create a powerful question-answering system. It supports contextual understanding and multi-turn dialogue, accurately understanding user intent and providing corresponding answers. The module also features personalized recommendation capabilities, offering customized information based on user query history and preferences. Furthermore, it maintains a real-time updated FAQ database, continuously learning and accumulating new questions and answers to continuously improve service quality.

[0015] The related project identification module, based on the core engine, implements multi-dimensional data analysis and correlation, achieving accurate project similarity scoring. This module can visually display historical processes and similar projects, helping users quickly understand the connections between projects. It also features intelligent recommendation capabilities, automatically suggesting relevant reference materials based on the characteristics of the current project. Furthermore, the module supports custom correlation rules, allowing users to adjust the identification criteria for related projects according to specific needs.

[0016] The discrepancy identification module uses a precise text comparison algorithm to effectively track differences between multiple document versions. This module not only highlights changed content but also automatically generates change summaries, helping users quickly grasp the key points of document modifications. The module supports multiple document formats and can automatically generate detailed discrepancy reports. Furthermore, the discrepancy identification module provides version management functionality for discrepancy tracking, allowing users to easily trace the document's historical changes.

[0017] The intelligent information extraction module, based on a core engine, achieves highly efficient document understanding capabilities. It can quickly locate and extract key information from documents and supports user-defined extraction templates. The module also features batch processing capabilities, enabling it to process a large number of documents simultaneously and intelligently aggregate the extracted information. Furthermore, it supports multiple document formats, including structured and unstructured text, greatly improving the flexibility and applicability of information extraction.

[0018] The multilingual processing module, based on the core engine, implements translation functionality, supporting high-quality translation between multiple languages. It employs intelligent text layout preservation technology to ensure that the translated document format remains consistent with the original. The module integrates a specialized terminology database and industry-specific dictionaries, guaranteeing the accuracy of translations within specialized fields. Furthermore, it supports the construction and maintenance of multilingual parallel corpora, continuously improving translation quality. The multilingual support capabilities of the module greatly facilitate cross-language communication and information sharing.

[0019] In the security assurance of step S4, the system manages user input text, conversation content and intermediate results in a controlled manner to avoid long-term exposure of plaintext in unnecessary stages, and reduces the risk of cross-task and cross-user data misuse through isolation mechanisms.

[0020] At the data transmission and interface call level, TLS and other transmission encryption protocols are used to protect link security and prevent text from being eavesdropped on or tampered with during interaction. Text content and log fragments that need to be written to disk or cached are encrypted using symmetric encryption methods such as AES, and key management and regular rotation are implemented to reduce the risk of key leakage.

[0021] In the RAG / retrieval enhancement stage, in addition to encrypting and storing the original text, logical isolation and access control are implemented for semantic features, vector indexes, and their mapping relationships to ensure that the retrieval only returns information within the authorized scope. In the output stage, sensitive information identification and de-identification strategies are combined to perform security filtering and auditing of the generated results, preventing sensitive fields from being carried out by the model or improperly disseminated.

[0022] These modules, through close integration and collaborative work, collectively build a highly efficient and intelligent digital system text-assisted platform, significantly improving the efficiency and quality of project management. The platform's modular design not only ensures the independence and professionalism of each function but also enables seamless collaboration between modules, providing users with a comprehensive text processing solution.

[0023] Compared with the prior art, the beneficial effects of the present invention include: (1) By combining retrieval and generation capabilities, this invention enables the system to quickly find relevant information in a vast document library and automatically generate high-quality answers or content. This not only improves the speed of document processing but also reduces reliance on manual operations, thereby saving a significant amount of human resources, especially in large-scale data processing scenarios.

[0024] (2) Traditional document processing methods often rely on manual input and judgment, which is prone to errors or information omissions. This invention, through automated processes, not only reduces human intervention but also retrieves and generates content based on the latest information, thereby reducing error rates and information inconsistencies. This is of great value for business scenarios that require accurate and consistent information, such as the legal and medical fields.

[0025] (3) This invention can combine with external knowledge bases to provide users with real-time decision support, helping project managers and decision-makers to quickly obtain relevant information and suggestions during project management. Through intelligent question answering and content generation, this invention can improve the scientific nature of decision-making, reduce the risks of relying on experience and intuition, and enhance the decision-making efficiency of enterprises.

[0026] (4) With the advancement of globalization, enterprises increasingly need to handle multilingual document content. This invention supports the retrieval and generation of multilingual documents, making the conversion and processing between different languages ​​more efficient, thereby supporting the international business expansion of enterprises. Whether in customer service, product document management, or cross-border cooperation, this invention can provide strong support.

[0027] (5) The combined retrieval and generation mechanism of the architecture of this invention can not only process documents in actual business, but also learn and integrate new knowledge in real time, enhancing the system's ability to generate content in different fields and on different topics. This makes this invention not only a "static" system, but also a "dynamic" knowledge base, which can continuously optimize and improve its processing capabilities as time goes by and information is updated, becoming a powerful knowledge management tool. Attached Figure Description

[0028] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0029] Figure 1 This is an overall flowchart of the method of the present invention.

[0030] Figure 2 This is a schematic diagram illustrating the functions of each module in the method of the present invention.

[0031] Figure 3 This is a diagram showing the text analysis function interface of the method system of the present invention. Detailed Implementation

[0032] Example 1 The overall flowchart of the text-assisted method based on the LLM+RAG digitization system described in this invention is as follows: Figure 1 As shown, the specific steps are as follows: S1. Intelligent text processing utilizes a large language model (LLM) to perform deep understanding and analysis of the input text, achieving efficient semantic parsing and information extraction; S2. Enhanced RAG generation: Combining vector databases and semantic retrieval technology, it quickly retrieves relevant information from massive documents, enhancing the knowledge reserves and generation capabilities of LLM. S3. Multi-module collaboration integrates multiple functional modules such as application system registration, intelligent display, and text input recognition to achieve comprehensive text-assisted processing; S4. Security Guarantee: Data encryption and access control technologies are employed to ensure the secure processing of sensitive information and the protection of privacy.

[0033] Example 2 A text-aiding system based on the LLM+RAG digitization system text-aiding method, with the following specific functions: S1. Intelligent text processing utilizes Large Language Models (LLM) for deep understanding and analysis of input text, achieving efficient semantic parsing and information extraction. Its page display is as follows: Figure 3 As shown; This invention employs intelligent text processing technology, fully leveraging the powerful capabilities of large language models. Through deep understanding and analysis of the input text, LLM enables efficient semantic parsing and information extraction. This advanced natural language processing method not only accurately captures the core meaning of the text but also identifies complex language structures and contextual relationships, thus laying a solid foundation for subsequent processing tasks.

[0034] The main component characteristics of LLM include: It adopts a hybrid expert architecture with a total of 671 billion parameters; Each token can activate 37 billion parameters; Dynamic gated routing technology is used to achieve load balancing without auxiliary loss; Integrated RoPE rotational position encoding, supporting context windows with 128k tokens; Introducing the multi-head latent attention (MLA) mechanism improves inference efficiency.

[0035] S2. Retrieval Enhanced Generation (RAG) combines vector databases and semantic retrieval technology to quickly retrieve relevant information from massive documents, enhancing the knowledge reserves and generation capabilities of LLM; Retrieval-Enhanced Generation (RAG) technology is another core component of this invention. By cleverly combining vector databases and semantic retrieval techniques, RAG can quickly and accurately retrieve information relevant to the current task from massive amounts of documents. This method significantly enhances the knowledge base and generation capabilities of LLM.

[0036] Key features of RAG components include: The integrated retrieval and generation model combines the advantages of retrieval enhancement and generation models, enabling the generation model to be enhanced with the help of external knowledge bases when generating answers; The generative architecture based on the BART model utilizes a pre-trained BART model for text generation, which can effectively generate high-quality natural language output. Efficient retrieval using vector search: By integrating vector databases such as FAISS, relevant information can be quickly retrieved from large-scale documents. The two-stage retrieval and generation process first uses a retrieval tool to find the most relevant documents from an external document collection. Then, these documents are fed into the generative model along with the input question to generate a more accurate and informative answer. End-to-end training: The RAG model trains the retrieval and generation models together, making the retrieval and generation processes complement each other and enhancing the overall performance of the model. The efficient training and inference process, through the adoption of streamlined retrieval strategies and generation optimizations, enables RAG to handle large-scale training data while reducing computational overhead during inference. It supports large-scale context windows, and the model can process longer context information by combining efficient retrieval mechanisms, breaking through the limitation of traditional generative models in processing shorter contexts. Flexible model extensibility supports different types of search engines, allowing the model to select the most suitable search method according to different application scenarios; By combining retrieval and generation, the RAG architecture demonstrates significant performance improvements in this system, especially in tasks such as question answering, information extraction, and dialogue generation that require a large amount of external knowledge.

[0037] The system adopts a layered processing architecture to achieve efficient and intelligent text processing: Input parsing: Use regular expressions to accurately extract key elements; Parallel retrieval: Triggers a hybrid retrieval engine to acquire relevant knowledge from multiple dimensions; Knowledge fusion: Deduplicating and synthesizing search results; Hint building: Utilize the integrated background knowledge to generate structured hint templates; Dynamic reasoning: generating initial responses through expert-blended LLM; Output validation: Ensures the accuracy and compliance of the generated content; These two core components work together to provide the invention with superior natural language processing and information retrieval capabilities, thereby enabling efficient and accurate text assistance functions.

[0038] The large language model component of this invention overcomes the limitations of traditional natural language processing systems through the comprehensive application of the aforementioned advanced technologies. Its superior parameter scale and innovative architecture design enable the model to deeply understand complex language structures and semantic relationships, laying a solid foundation for subsequent text processing tasks. Particularly in long text processing, the large language model component of this invention performs exceptionally well, effectively capturing long-range dependencies and providing users with more accurate and coherent language generation services.

[0039] S3. Multi-module collaboration integrates multiple functional modules such as application system registration, intelligent display, and text input recognition to achieve comprehensive text-assisted processing; This embodiment sets up eight core modules, mainly including a system introduction generation module, a similar system identification module, a misspelling identification module, an intelligent question-and-answer assistance module, a related item identification module, a difference identification module, an intelligent information extraction module, and a multilingual processing module.

[0040] The system introduction generation module utilizes a core engine to intelligently analyze user-entered information and system attachments, achieving natural language understanding and information extraction. This module automatically generates standardized system introductions, significantly improving work efficiency. To ensure the quality of the generated content, the module is equipped with a rigorous quality control mechanism and a manual review process.

[0041] The approximate system identification module employs advanced similarity algorithms to efficiently identify and display similar or duplicate systems. This module not only visualizes the relationships between systems but also generates intuitive system relationship graphs. It includes a built-in duplicate registration warning mechanism to effectively prevent redundant work and improve resource utilization efficiency. Furthermore, the approximate system identification module supports customizable similarity thresholds, allowing users to adjust the identification accuracy according to specific needs.

[0042] The misspelling detection module, based on its core engine, implements intelligent character recognition technology, providing context-aware automatic error correction. It supports multilingual misspelling databases and can identify and correct common errors in various languages. A key feature of this module is its user feedback learning mechanism, which continuously improves the misspelling database based on user corrections. Furthermore, the module supports custom dictionaries, allowing users to add specialized terminology from specific fields to further enhance recognition accuracy.

[0043] The intelligent question-answering assistance module is built upon a core engine to create a powerful question-answering system. It supports contextual understanding and multi-turn dialogue, accurately understanding user intent and providing corresponding answers. The module also features personalized recommendation capabilities, offering customized information based on user query history and preferences. Furthermore, it maintains a real-time updated FAQ database, continuously learning and accumulating new questions and answers to continuously improve service quality.

[0044] The related project identification module, based on the core engine, implements multi-dimensional data analysis and correlation, achieving accurate project similarity scoring. This module can visually display historical processes and similar projects, helping users quickly understand the connections between projects. It also features intelligent recommendation capabilities, automatically suggesting relevant reference materials based on the characteristics of the current project. Furthermore, the module supports custom correlation rules, allowing users to adjust the identification criteria for related projects according to specific needs.

[0045] The discrepancy identification module uses a precise text comparison algorithm to effectively track differences between multiple document versions. It not only highlights changed content but also automatically generates change summaries, helping users quickly grasp the key points of document modifications. The module supports multiple document formats and can automatically generate detailed discrepancy reports. Furthermore, it provides version management functionality for discrepancy tracking, allowing users to easily trace historical changes to documents.

[0046] The intelligent information extraction module, based on a core engine, achieves highly efficient document understanding capabilities. It can quickly locate and extract key information from documents and supports user-defined extraction templates. The module also features batch processing capabilities, enabling it to process a large number of documents simultaneously and intelligently aggregate the extracted information. Furthermore, it supports multiple document formats, including structured and unstructured text, greatly improving the flexibility and applicability of information extraction.

[0047] The multilingual processing module, based on the core engine, implements translation functionality, supporting high-quality translation between multiple languages. It employs intelligent text layout preservation technology to ensure that the translated document format remains consistent with the original. The module integrates a specialized terminology database and industry-specific dictionaries, guaranteeing the accuracy of translations within specialized fields. Furthermore, it supports the construction and maintenance of multilingual parallel corpora, continuously improving translation quality. This multilingual support capability significantly facilitates cross-language communication and information sharing.

[0048] The functions of each module in the method of this invention are illustrated as follows: Figure 2 As shown, these modules, through close integration and collaborative work, jointly construct a highly efficient and intelligent digital system text assistance platform, significantly improving the efficiency and quality of project management. The platform's modular design not only ensures the independence and professionalism of each function but also achieves seamless collaboration between modules, providing users with a comprehensive text processing solution.

[0049] S4. Security Guarantee: Advanced data encryption and access control technologies are employed to ensure the secure processing of sensitive information and the protection of privacy. In text-assisted scenarios within digital systems, security assurance needs to cover the entire process of text "input—processing—storage—output". The system of this invention provides controlled management of user-input text, conversation content, and intermediate results during processing, preventing plaintext from being exposed in unnecessary stages for extended periods, and reducing the risk of data misuse across tasks and users through isolation mechanisms.

[0050] At the data transmission and interface call level, TLS and other transmission encryption protocols are used to protect link security and prevent text from being eavesdropped on or tampered with during interaction. Text content and log fragments that need to be written to disk or cached are encrypted using symmetric encryption methods such as AES, and key management and regular rotation are implemented to reduce the risk of key leakage.

[0051] In the RAG / retrieval enhancement stage, in addition to encrypting and storing the original text, logical isolation and access control are implemented for semantic features, vector indexes, and their mapping relationships to ensure that the retrieval only returns information within the authorized scope. In the output stage, sensitive information identification and de-identification strategies are combined to perform security filtering and auditing of the generated results, preventing sensitive fields from being carried out by the model or improperly disseminated.

Claims

1. A text-assisted method based on an LLM+RAG digitization system, characterized in that, Includes the following steps: S1. Intelligent text processing utilizes a large language model (LLM) to perform deep understanding and analysis of the input text, achieving efficient semantic parsing and information extraction; S2. Enhanced RAG generation: Combining vector databases and semantic retrieval technology, it quickly retrieves relevant information from massive documents, enhancing the knowledge reserves and generation capabilities of LLM. S3. Multi-module collaboration integrates multiple functional modules such as application system registration, intelligent display, and text input recognition to achieve comprehensive text-assisted processing; S4. Security Guarantee: Data encryption and access control technologies are employed to ensure the secure processing of sensitive information and the protection of privacy.

2. A text-aiding system based on the LLM+RAG digitization system text-aiding method described in claim 1, characterized in that, Specifically, it includes a system introduction generation module, a similar system identification module, a typo identification module, an intelligent question-and-answer assistance module, a related item identification module, a difference identification module, an intelligent information extraction module, and a multilingual processing module.

3. The text assistance system according to claim 2, characterized in that, The system introduction generation module uses a core engine to intelligently analyze user-filled information and system material attachments, achieving natural language understanding and information extraction to automatically generate a compliant system introduction. It is also equipped with a quality control mechanism and a manual review process to ensure the quality of the generated content.

4. The text-assisted system according to claim 2, characterized in that, The approximate system identification module uses a similarity algorithm to identify and display similar or repeating systems. It not only visualizes the relationships between systems but also generates an intuitive system relationship map. The system identification module has a built-in duplicate registration early warning mechanism to prevent duplicate work and improve resource utilization efficiency; The system's recognition module supports custom similarity thresholds, allowing users to adjust the recognition accuracy according to specific needs.

5. The text assistance system according to claim 2, characterized in that, The misspelling detection module, based on the core engine, implements intelligent character recognition technology and provides context-aware automatic error correction function; The misspelling recognition module supports multilingual misspelling databases and can identify and correct common errors in various languages. The misspelling recognition module has a user feedback learning mechanism, which can continuously improve the misspelling database based on user corrections. In addition, the misspelling recognition module also supports custom dictionaries, allowing users to add professional terms from specific fields to further improve recognition accuracy.

6. The text-assisted system according to claim 2, characterized in that, The intelligent question-answering assistance module is built on a core engine to create a question-answering system that supports contextual understanding and multi-turn dialogue. It can accurately understand user intent and provide corresponding answers. The intelligent question-answering assistance module has a personalized recommendation function, which can provide customized information based on the user's query history and preferences; The intelligent question-and-answer assistance module maintains a real-time updated FAQ database, continuously learning and accumulating new questions and answers to continuously improve service quality.

7. The text-assisted system according to claim 2, characterized in that, The associated project identification module, based on the core engine, realizes multi-dimensional data analysis and association, achieves accurate project similarity scoring, and can visualize historical processes and similar projects, helping users quickly understand the relationship between projects; The related project identification module also has an intelligent recommendation function, which can automatically recommend relevant reference materials based on the characteristics of the current project. In addition, the related project identification module supports custom association rules, allowing users to adjust the identification criteria of related projects according to specific needs.

8. The text assistance system according to claim 2, characterized in that, The difference recognition module uses a text comparison algorithm to track the differences between multiple versions of documents, highlight the changed content, and automatically generate a change summary to help users quickly grasp the key points of document modification. The difference recognition module supports multiple document formats and can automatically generate detailed difference reports. In addition, the typo recognition module also provides a version management function for difference tracking, making it convenient for users to trace the historical changes of documents.

9. The text assistance system according to claim 2, characterized in that, The intelligent information extraction module, based on the core engine, achieves document understanding capabilities, quickly locates and extracts key information from documents, supports user-defined extraction templates, has batch processing capabilities, can process a large number of documents simultaneously, and intelligently aggregates the extracted information. In addition, the intelligent information extraction module also supports multiple document formats, including structured and unstructured text, improving the flexibility and applicability of information extraction.

10. The text assistance system according to claim 2, characterized in that, The multilingual processing module, based on the core engine, implements translation functionality, supporting high-quality translation between multiple languages. It employs intelligent text layout preservation technology to ensure that the translated document format remains consistent with the original. The module integrates a professional terminology database and industry dictionaries, guaranteeing the accuracy of translations in specialized fields. Furthermore, it supports the construction and maintenance of multilingual parallel corpora, continuously improving translation quality. The multilingual support capabilities of the module promote cross-language communication and information sharing.

Images