Intelligent RAG enhanced large model auxiliary writing system

Abstract

The invention discloses an intelligent RAG enhanced large model auxiliary writing system, and relates to the technical field of artificial intelligence auxiliary writing. In order to overcome the defects of an existing writing auxiliary tool, the adopted scheme comprises an interface and user layer for receiving a writing demand and redisplaying a generation result; the service layer is used for deploying three types of intelligent agents including writing, interactive writing enhancement and checking; the platform layer is used for carrying out industry adaptation analysis, mixed retrieval and RAG enhancement on writing demands and then sending the writing demands to the large model layer; three intelligent agents of writing, interactive writing enhancement and checking are called in sequence based on the output of the large model layer, and then a text is redisplayed to the interface and user layer after fact accuracy verification is completed through a tense knowledge graph; the data layer is used for providing a multi-type data storage system; the large model layer outputs a paragraph-level text and a reference index based on a writing large model, the paragraph-level text is returned to the platform layer, and reference index information is fed back to the interpretability and security layer for tracing; and the infrastructure layer is used for providing bottom layer resource support for each layer.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence-assisted writing technology, specifically to a large model-assisted writing system with intelligent RAG enhancement. Background Technology

[0002] Breakthroughs in large language models (LLMs) have spurred a new generation of AI-assisted writing systems, but they still face structural bottlenecks in scenarios with high professional barriers and strong compliance requirements, specifically manifested in the following ways:

[0003] 1. Knowledge blind spots and illusions: Static training corpora cannot cover in-depth fields and lack fine-grained representation of professional knowledge such as archives, law, and government affairs, resulting in vague, inaccurate, and even frequently illusory content generated.

[0004] 2. Single-model bottleneck: A single end-to-end model undertakes the entire chain of tasks from "understanding → retrieval → generation → proofreading". Indiscriminate processing leads to logical drift in long documents, fragmented chapter styles, and large quality variance.

[0005] 3. Zero verification mechanism: There is no fact-checking or logical consistency verification process within the system. Errors require secondary manual review, and the verification cost accounts for more than 30% of the project cycle.

[0006] 4. Knowledge lag: The parameterized knowledge update cycle is measured in "months", which cannot keep up with hourly changes in policies, regulations, cases, etc., creating a "publishing and becoming outdated" dilemma.

[0007] 5. Scene insensitivity: The prompt templates and hyperparameters are fixed, making it difficult to automatically switch between heterogeneous scenarios such as archival research, government documents, and legal documents, resulting in frequent misalignment of output format, terminology, and narrative framework.

[0008] 6. Black box and untraceable: The generation process lacks chain evidence, users cannot locate the source of information and the reasoning path, auditing and accountability have no basis, and credibility is limited.

[0009] 7. Security and compliance risks: The lack of sensitive entity identification, privacy desensitization and compliance rule engine makes it easy to trigger cross-border data leakage, privacy abuse and industry regulatory penalties.

[0010] 8. Industry Adaptability Gap: The lack of built-in hot-swappable industry terminology, template library, and rule package results in a continuous misalignment between terminology translation, citation format, approval elements, and industry standards, making it difficult to directly implement in the production environment.

[0011] In summary, existing large language model (LLM) writing solutions have systemic defects in eight dimensions: knowledge depth, collaboration granularity, validation strength, update speed, scenario detail, explanatory power, security level, and industry relevance. There is an urgent need for a new writing enhancement framework that is dynamic, reliable, collaborative, and scalable. Summary of the Invention

[0012] This invention addresses the shortcomings of existing writing aids in terms of knowledge timeliness, factual accuracy, collaboration on complex tasks, and industry adaptability by providing an intelligent RAG-enhanced large-model-assisted writing system.

[0013] The present invention provides a large model-assisted writing system with intelligent RAG enhancement, and the technical solution adopted to solve the above-mentioned technical problems is as follows:

[0014] A large-model-assisted writing system with intelligent RAG enhancement, employing a seven-layer fusion architecture, including:

[0015] The interface and user layer serve as the entry point for interaction between the system and the user, and are used to: receive writing requests submitted by users in natural language and display the corresponding generated results;

[0016] The service layer deploys writing agents, interactive writing augmentation agents, and proofreading agents. Its underlying layer uses a microservice cluster architecture to encapsulate two core atomic capabilities: text generation and proofreading. The text generation atomic capability includes five types of services: content generation, sentence expansion, sentence abbreviation, sentence polishing, and sentence rewriting. The proofreading atomic capability includes five types of services: punctuation standardization, sentence fluency, text accuracy, format compliance, and sensitive word screening, providing basic functional support for the upper-layer agents.

[0017] The platform layer, as the core scheduling hub, specifically implements the following operations: It receives writing requests uploaded by the interface and user layer. First, the industry adaptation module calls various libraries in the data layer to complete industry adaptation parsing. Then, the hybrid retrieval module obtains high-confidence fragments from the structured knowledge base and vector knowledge base. The RAG module concatenates the high-confidence fragments with the writing template to form an enhanced prompt, simultaneously generating a citation package, which is then sent to the large model layer. It receives paragraph-level text and citation indexes output by the large model layer. Through the multi-agent scheduling module, it triggers collaborative optimization, first scheduling the writing agent in the service layer to generate standardized text, then calling the interactive writing enhancement agent in the service layer for text enhancement, and finally calling the proofreading agent in the service layer for text correction. The corrected text is then verified for factual accuracy by the temporal knowledge graph verification module combined with the structured knowledge base and terminology database in the data layer. Finally, the verified text is displayed back to the interface and user layer.

[0018] The large model layer is equipped with a writing large model that has been continuously pre-trained, supervised fine-tuned and aligned with reinforcement learning. Specifically, it performs the following operations: receiving the enhanced Prompt and citation package transmitted by the RAG module of the platform layer, generating a draft text that meets the requirements through the writing large model, outputting paragraph-level text and citation index, and recording key inference nodes in the generation process to provide support for interpretability traceability.

[0019] The data layer provides end-to-end data assurance for the platform layer and the large model layer: First, it constructs a multi-type data storage system, covering industry feature library, structured knowledge base, industry template library, vector knowledge base, terminology library, and compliance rule library; second, it ensures data quality through metadata governance; and third, it provides standardized data access interfaces to support the data call needs of each layer.

[0020] The interpretability and security layer runs through the entire chain from writing requirement input to result delivery and is deeply linked with each layer. It is used to: receive key inference nodes from the large model layer and reference packages from the platform layer to build a knowledge tracing chain and inference process graph; generate confidence assessment by combining the fact verification results from the platform layer; perform sensitive information desensitization on the text to be displayed output from the platform layer; add digital watermarks to the final display results of the interface and user layer; and track the data of the entire chain through unified logs to form chain evidence tracing to ensure the transparency and security of the generation process.

[0021] The infrastructure layer is used to provide underlying resource support for the aforementioned layers. Through computing, networking, storage, and container orchestration resources, it ensures the horizontal elastic scaling of the service layer, the efficient computing of the large model layer, and the stable storage of the data layer, while supporting the high-availability operation of a unified gateway and unified authentication.

[0022] All the aforementioned layers are connected through a unified gateway to achieve data flow, unified authentication to ensure access security, and unified log tracking of link status, forming a closed-loop chain of "interface and user layer receiving requirements - platform layer parsing and enhancing and sending to the big model - big model layer core generation - platform layer scheduling service layer collaborative optimization and verification - interpretability and security layer assessment and security hardening - interface and user layer delivering results". The data layer and infrastructure layer provide continuous data and resource support for the entire process.

[0023] Optionally, the interfaces involved and the user layer support access to the Web platform, management backend and WPS plugin through standardized interfaces. The received writing requirements are structured, parsed, initially verified and formatted, and then synchronized to the platform layer through a unified gateway. At the same time, the text output by the platform layer is received and processed by the interpretability and security layers, and then displayed back to the user.

[0024] Optionally, after the platform layer receives the writing requirements and the user layer synchronizes them through a unified gateway, the industry adaptation module sequentially executes three steps: "industry identification—configuration loading—terminology adaptation." Finally, it calls the industry feature library, structured knowledge base, industry template library, terminology library, and compliance rule library from the data layer to complete industry adaptation parsing, mapping the task to a domain ontology event graph, and outputting a structured writing task description package. This process specifically includes:

[0025] (1.1) Industry identification: The requirement description transmitted between the interface and the user layer is vectorized and then the cosine similarity is calculated with the industry feature library pre-set in the data layer, and the scene label is output in seconds;

[0026] (1.2) Configuration loading: Based on the structured knowledge base and industry template library of the data layer hot-pull data layer of the scene tag, combined with the domain ontology event graph generated after industry identification in step (1.1), and the pre-set constraint rules in the template, a structured writing framework is generated immediately.

[0027] (1.3) Terminology adaptation: Based on the aforementioned structured writing framework, the terminology library and compliance rule library of the data layer are called to carry out industry adaptation parsing work. The general expressions, non-standard terms, unit names and citation formats in the writing framework are automatically replaced with industry standard expressions, and the compliance of the expressions is verified at the same time. Finally, a writing task description package that can directly drive subsequent hybrid retrieval and RAG enhancement is formed.

[0028] Optionally, after the industry-adaptive module outputs the structured writing task description package, the hybrid retrieval module, driven by the writing task description package, sequentially completes slot extraction, BM25+ vector hybrid recall, and reordering to obtain high-confidence fragments from the structured knowledge base and the vector knowledge base. The specific process includes:

[0029] (2.1) The slot extraction unit uses the data layer event-entity dictionary as a basis and adopts the "BIO + pointer network" joint model to annotate the specified core fields in the structured writing task description package word by word, accurately extract the core requirements of the task and output the structured query vector, which also serves as the unified input for sparse and dense retrieval.

[0030] (2.2) The retrieval engine unit launches the BM25 sparse channel, the semantically dense vector channel, and the keyword completion channel in parallel: the BM25 sparse channel achieves accurate term matching based on the inverted index of the structured knowledge base, the semantically dense vector channel captures synonyms and implied expressions in the full embedding space of the vector knowledge base, and the keyword completion channel expands and recalls synonyms or abbreviations based on the data layer terminology database. The three results are deduplicated and merged to form an initial candidate set and each carries a basic similarity score.

[0031] (2.3) The reordering unit introduces a contrastive learning pre-trained model, which integrates four-dimensional features such as relevance, timeliness, source authority, and paragraph completeness in a "query-segment" pairing manner. The model calculates and outputs the confidence score in the 0-1 interval, extracts the Top-K high-confidence segments in descending order of score, and generates a citation package containing segment ID, source URL, publication / revision date, confidence score, and core related terms. This package is then sent to the RAG module to complete the supply of traceable and highly confident knowledge.

[0032] Optionally, the RAG module involved, under the principle of "dynamic-trustworthy-traceable," consists of four sub-pipelines connected in series and parallel: knowledge retrieval, data fusion, context awareness, and quality assessment. This enables the accurate injection of high-confidence fragments into the large model layer. The specific process includes:

[0033] (3.1) Dynamic Knowledge Retrieval: The input consists of the structured writing task description package output by the industry adaptation module and the high-confidence fragments and citation packages output by the hybrid retrieval module. First, the event graph and paragraph-level keywords are extracted from the structured writing task description package. Then, the core semantics of the high-confidence fragments and the knowledge anchors associated with the citation packages are integrated. Based on the graph nodes, the core semantics of the high-confidence fragments and the knowledge anchors of the citation packages, a three-state query of "keyword-semantics-vector" is generated. These queries are then sent to the inverted index of the structured knowledge base, the dense index of the vector knowledge base and the sparse index for parallel retrieval. At the same time, incremental knowledge patches from the data layer are merged into the index at the minute level through an incremental message queue to ensure that the candidate set is zero time difference with the latest facts. The output is a set of candidate fragments with initial similarity classification.

[0034] (3.2) Multi-source data fusion: For the candidate fragment set, the corresponding parser is automatically called according to the MIME type. Structured data is directly mapped to attribute triples, unstructured text is generated into semantic blocks after paragraph segmentation, referential resolution and entity linking, and images / tables are converted into structured text descriptions after OCR and table structure recognition. The three types of data units processed by the corresponding parser are projected into the same semantic space through a unified vectorized encoder. The bidirectional alignment loss function is used to eliminate modal differences, form a consistent knowledge representation and bind a globally unique ID, providing a basis for traceability.

[0035] (3.3) Context awareness: The fused knowledge representation and the core requirement context in the structured writing task description package are fed into the lightweight re-ranking model. The lightweight re-ranking model uses the paragraph topic vector as the query, calculates the coupling degree between the knowledge representation and the writing requirement context in real time, dynamically adjusts the keyword weight and vector distance threshold, improves the ranking of the paragraph that is strongly related to the current writing topic, and outputs high-confidence re-ranking results.

[0036] (3.4) Knowledge quality assessment: First, extract the basic information of the high-confidence rearrangement results and related reference packages, and quantify the scores from four dimensions: authority, timeliness, relevance and completeness. The scores of each dimension are weighted according to the preset weights to calculate the comprehensive score. Segments with a comprehensive score lower than the set threshold are automatically discarded. The remaining segments are arranged in descending order of comprehensive score. The overall average quality score of the output high-confidence segment set is ≥0.92.

[0037] (3.5) The RAG module calls the corresponding template from the industry template library retrieved by the industry adaptation module during the configuration loading stage, and concatenates the high-confidence fragments into an enhanced Prompt according to the template structure, context logic and writing requirements. It reuses the citation package generated by the hybrid retrieval module and supplements the globally unique ID information, and sends it to the large model layer. The large model layer performs the generation operation, and the writing large model generates a text draft that meets the requirements, outputs paragraph-level text and citation index, and records the key reasoning nodes in the generation process to provide support for subsequent interpretability traceability and for subsequent platform layer scheduling service layer collaborative optimization.

[0038] Alternatively, the large model layer involved adopts the core concept of "corpus is knowledge, knowledge is parameters" and uses a domestic computing power cluster to carry out a three-level progressive reshaping of the base model, forming a reliable writing engine for high-barrier scenarios in multiple industries, while completing the core generation task. The specific process steps are as follows:

[0039] Step 1: Continuous pre-training of the large-scale model: Build a high-quality hybrid corpus covering fifteen industries, inject high-frequency terms from each industry into the data layer to expand the vocabulary, and provide rich domain corpus support for the large model; adopt a course learning strategy from easy to difficult, first train the general domain corpus to retain basic language capabilities, and then gradually introduce domain-specific corpus, so that the large model can accurately master the segmentation and semantic representation capabilities of domain-specific expressions. Through the elastic weight decay mechanism, catastrophic forgetting is suppressed, providing a large model base with domain-based capabilities for subsequent fine-tuning and reinforcement learning;

[0040] Step 2: Large Model Supervised Fine-Tuning: Building upon the domain-specific capabilities established in Step 1's pre-training, heterogeneous writing tasks are encapsulated into a single paradigm using the "instruction-knowledge-output" triplet as the core. High-fidelity alignment training is conducted using QA instructions with <slot> slots. By amplifying gradients at error-prone locations, introducing paragraph-level attention masks, and relative position encoding, the logical order of "chapter-section-article" in the text is ensured to remain intact, achieving precise calibration of the pre-trained large model's generation logic. Simultaneously, national cryptographic algorithms are used to encrypt model parameters, and blockchain fingerprint digest technology is combined to record parameter change trajectories, enabling rollback and auditability of model parameter versions. This lays a solid foundation for security and compliance for subsequent reinforcement learning alignment and core task execution, completing high-quality policy initialization.

[0041] Step 3: Large Model Reinforcement Learning Alignment: Based on the strategy initialization completed in Step 2, a dual-track preference optimization mechanism is built. The factual accuracy track uses industry review logic supported by the structured knowledge base and terminology base of the data layer to score the risk of candidate manuscripts. The format compliance track automatically compares the layout elements according to the industry standards of the data layer. A dual-tower reward model is adopted to calculate the difference between factual and format similarity. In DPO training, the weight of factual reward is set to twice that of format reward. At the same time, a length penalty factor is introduced to suppress redundant generation. The basic large model parameters after the pre-training in Step 1 and the fine-tuning in Step 2 are kept frozen. Only the policy network is optimized. Finally, the factual accuracy of the large model is ≥95%, the one-time format compliance rate is ≥90%, and the general language ability is improved simultaneously, providing a key guarantee for the output of high-quality results for the core generation task.

[0042] Step 4: Core Generation Task Execution: After receiving the enhanced Prompt and citation package from the platform layer's RAG module, the large model layer, relying on the writing model's capabilities aligned with continuous pre-training, supervised fine-tuning, and reinforcement learning, and combined with dynamic prompting optimization strategies, transforms the domain knowledge learned during pre-training, the logical norms of fine-tuning and calibration, and the quality standards of reinforcement optimization into actual generation capabilities. It generates a first draft of text that meets the requirements, outputs paragraph-level text and citation indexes, and records key inference nodes in the generation process to support interpretability traceability. Subsequently, the output paragraph-level text and citation indexes are synchronized to the platform layer's multi-agent scheduling module for collaborative optimization, and the platform layer synchronizes the citation indexes and key inference nodes to the interpretability and security layer for source tracing.

[0043] Optionally, the platform layer receives paragraph-level text and citation indexes output from the large model layer. Through a multi-agent scheduling module, collaborative optimization is triggered. First, the writing agent in the service layer is scheduled to generate standardized text. Then, the interactive writing enhancement agent in the service layer is invoked for text enhancement. Finally, the proofreading agent in the service layer is invoked for text correction. The corrected text is then verified for factual accuracy by a temporal knowledge graph verification module combined with a structured knowledge base and terminology database in the data layer. Finally, the verified text is displayed back to the interface and user layer. This process specifically includes:

[0044] The writing agent calls the content generation atomic service encapsulated in the service layer, related resources in the data layer, and its own standardization processing capabilities to generate standardized text based on the paragraph-level text and citation index output by the large model layer.

[0045] The interactive writing enhancement agent subscribes to the standard text and citation index in real time, parses the rewrite / expansion / abbreviation / polishing instructions synchronized between the interface and the user layer in real time, dynamically calls the four types of atomic services encapsulated in the service layer, namely sentence expansion, sentence abbreviation, sentence polishing and sentence rewriting, and makes fine-grained adjustments to the chapter structure, terminology standardization and sentence fluency of the standard text. The adjustment results are written back to the system message bus in real time for subsequent review agents to share and reuse.

[0046] The proofreading agent initiates five atomic services in parallel: punctuation standardization, sentence fluency, text accuracy, format compliance, and sensitive word screening. Simultaneously, it links with the data layer compliance rule library and terminology library, reuses the platform layer's basic verification logic, and performs full-dimensional verification character by character, paragraph by paragraph, and style by style, returning difference markers and precise correction suggestions. It also generates quantitative confidence scores from five dimensions: semantic consistency, character accuracy, format compliance, sensitive information risk, and content compliance. If the overall confidence score is lower than a preset threshold, the proofreading agent initiates a second-iteration optimization request to the interactive writing enhancement agent through a "game-cooperation" mechanism. The writing agent simultaneously provides standardized basic parameters, and the three work together to execute a cyclical "standardization adaptation - enhancement adjustment - verification evaluation" process based on the game-cooperation mechanism until global convergence.

[0047] The temporal knowledge graph verification module performs final factual accuracy verification on the reviewed text. It constructs a temporal association verification link based on authoritative facts in the data layer structured knowledge base and standardized expressions in the terminology base to check for issues such as timeline contradictions, factual deviations, and misuse of terminology. Only text that passes the verification is allowed to pass.

[0048] Finally, the platform layer will synchronize the optimized results after the closed-loop processing of "standardization-enhancement-proofreading-compliance-fact verification" and the full-process traceability records to the interface and user layers to complete the result delivery.

[0049] Optionally, the temporal knowledge graph verification module involved uses a three-level overlay graph of "general-industry-scenario" as its core, linking the structured knowledge base, terminology base, and compliance rule base of the data layer. It completes fact verification according to the process of "graph alignment-incremental update-multidimensional verification-anomaly feedback," achieving multidimensional, fine-grained, and real-time verification of the generated content. The specific process includes:

[0050] (1) Graph alignment: First, the domain ontology event graph output by the platform layer industry adaptation module is matched with the "general-industry-scenario" three-level superimposed graph built into the temporal knowledge graph verification module. The general graph provides basic entities and higher-level concepts, the industry graph reuses the terminology library loaded in the industry adaptation stage, and the scenario graph accurately matches the pre-set operation-level elements in the template of the writing task configuration loading stage. The cross-graph alignment is achieved through the dual mechanism of "unique entity ID matching + semantic vector similarity verification" to ensure that the verification granularity corresponds to the writing task scenario requirements at the same scale.

[0051] (2) Incremental update: Link the incremental message queue of the data layer to capture newly released / revised authoritative knowledge in the structured knowledge base in real time, complete the online injection of graph nodes, attributes and relationships within 3 minutes, and timestamp and archive the old version knowledge and associate the timestamp information in the reference package to achieve "zero lag" hot update; the update record is synchronously associated with the source URL in the reference package and the knowledge release certificate of the data layer, forming a "knowledge update-source tracing" closed loop to prevent policy reference and terminology usage deviations caused by knowledge lag;

[0052] (3) Multidimensional verification execution: The temporal knowledge graph verification module simultaneously initiates dual-dimensional verification of logical consistency and factual accuracy: On the logical consistency verification side, it calls the pre-set temporal constraints and causal rule base to automatically detect anomalies such as contradictions in relationships between entities, inversion of event time sequence, and data range conflicts. The positioning accuracy reaches the clause level within the paragraph. Based on the severity of the anomaly, it triggers confidence penalty or initiates paragraph knowledge re-retrieval request with the associated RAG module; On the factual accuracy verification side, it combines the fragment ID, source authority weight, graph entity attributes, and relation edges in the reference package to perform bidirectional reasoning. Forward reasoning verifies the consistency between the text statement and the graph nodes and relationships, and reverse reasoning completes the missing or implicit key facts in the text. The reasoning process forms a cross-verification with the quality assessment dimension of the RAG module to improve the credibility of the verification.

[0053] (4) Anomaly Handling and Feedback: If an anomaly is detected, the temporal knowledge graph verification module marks the anomaly location in real time and associates it with the corresponding node in the domain ontology event graph. At the same time, it generates a traceable alternative expression based on authoritative evidence in the data layer terminology library, compliance rule library, and structured knowledge library, with source links, confidence scores, timestamps, and verification logic descriptions. It also outputs a structured anomaly report and correction suggestions. It is recommended to support two modes: "one-click user location and review" or "automatic triggering of the service layer interactive writing enhancement agent for paragraph optimization" to ensure closed-loop processing of anomaly issues.

[0054] Optionally, the interpretability and security layer adopts a three-layer in-depth architecture of "knowledge tracing—reasoning chain—confidence assessment," running through the entire process from input of writing requirements to delivery of results, deeply linking with each layer to achieve transparency in the generation process, credibility of results, and controllable risks; among which:

[0055] As the foundational layer of the architecture, the knowledge tracing module receives key inference nodes from the large model layer and reference packages from the platform layer. It records the original source, retrieval path, processing timestamp, and verification status of each knowledge fragment using a Merkle tree structure. It ensures data integrity through SHA-256 hashing and supports second-level backtracking to the original policy text or database records, providing reliable source support for subsequent inference and evaluation.

[0056] The reasoning chain module is based on the credible source of the knowledge tracing module. It uses the Do-Calculus algorithm to construct a step-by-step logical path from input requirements to output content. Combined with the causal reasoning graph, it forms a complete reasoning link, constructs a reasoning process graph, and provides interactive visualization functions through D3.js, allowing users to expand and view the decision basis at any stage, so as to make the reasoning process explainable.

[0057] The confidence assessment module combines the deduction results of the reasoning chain module with the fact verification results of the platform layer, and integrates four-dimensional indicators: source credibility, logical consistency, statistical confidence and historical accuracy. It uses a weighted scoring model to calculate a standardized score of 0-100. The score is refreshed in real time after each paragraph is generated or modified, and a detailed scoring explanation is attached, allowing users to quickly judge the reliability of the current content.

[0058] Optionally, the interpretability and security layers involved also achieve triple capabilities of privacy protection, security traceability, and compliance auditing:

[0059] (a) Privacy protection: Sensitive information is desensitized on the text to be displayed at the platform layer, ε-differential privacy technology is introduced to ensure data privacy and security, and sensitive content such as personal information, trade secrets and classified data is automatically identified by combining named entity recognition and rule-based hybrid methods, and precise desensitization processing is performed;

[0060] (b) Security traceability: Add invisible digital watermarks to the results displayed by the interface and user layer, and track the data of the whole link through a unified log system to form an immutable chain evidence traceability system;

[0061] (c) Compliance audit: The integrated compliance audit engine module serves as a supplement to security capabilities, constructing a regulatory knowledge graph covering 15 industries. It contains over 500,000 relevant articles, industry standards, and regulatory rules, and employs a hybrid reasoning mechanism of rule engine and machine learning to perform real-time compliance scanning and risk warnings on the generated text. Violation paragraphs are highlighted instantly, and specific modification suggestions are provided simultaneously.

[0062] The intelligent RAG-enhanced large model-assisted writing system of the present invention has the following advantages compared with the prior art:

[0063] This invention addresses the challenges of large-model-driven assisted writing scenarios by systematically integrating industry-adaptive parsing, hybrid retrieval, RAG enhancement, large-model generation, multi-agent scheduling, temporal knowledge graph verification, and interpretable security mechanisms to construct an integrated writing enhancement framework. This framework enables end-to-end content generation, polishing, and proofreading for complex writing tasks, solving the problem that existing intelligent writing systems often encounter difficulties in writing high-barrier medical texts due to industry context mismatch, terminology drift, and insufficient rigor in wording, leading to easily rejected or compliance-risk published texts. Attached Figure Description

[0064] Appendix Figure 1 This is a system architecture diagram of the present invention;

[0065] Appendix Figure 2 This is a flowchart of the system operation of the present invention. Detailed Implementation

[0066] 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.

[0067] Example 1:

[0068] Combined with appendix Figure 1 This embodiment proposes a large model-assisted writing system with intelligent RAG enhancement, characterized by its adoption of a seven-layer fusion architecture, including:

[0069] (i) Interface and user layer, as the entry point for system-user interaction, is used to: receive writing requests submitted by users in natural language and display the corresponding generated results.

[0070] In this embodiment, the interface and user layer support access to the Web platform, management backend, and WPS plugin through a standardized interface. The received writing requirements are structured, parsed, preliminarily validated, and formatted before being synchronized to the platform layer through a unified gateway. At the same time, the text output by the platform layer is received and processed by the interpretability and security layers, and then displayed back to the user.

[0071] (II) Service layer, which is equipped with writing agent, interactive writing enhancement agent and proofreading agent. Its underlying layer is a microservice cluster architecture that encapsulates two core atomic capabilities: text generation and proofreading. The text generation atomic capability includes five types of services: content generation, sentence expansion, sentence abbreviation, sentence polishing and sentence rewriting. The proofreading atomic capability includes five types of services: punctuation standardization, sentence fluency, text accuracy, format compliance and sensitive word screening, which provide basic functional support for the upper-layer agents.

[0072] (III) The platform layer, as the core scheduling hub, specifically implements the following operations: It receives writing requests uploaded by the interface and user layer. First, the industry adaptation module calls various libraries in the data layer to complete industry adaptation parsing. Then, the hybrid retrieval module obtains high-confidence fragments from the structured knowledge base and vector knowledge base. The RAG module concatenates the high-confidence fragments with the writing template to form an enhanced Prompt, while simultaneously generating a citation package, which is then sent to the large model layer. It receives paragraph-level text and citation indexes output by the large model layer. Through the multi-agent scheduling module, it triggers collaborative optimization, first scheduling the writing agent in the service layer to generate standardized text, then calling the interactive writing enhancement agent in the service layer to perform text enhancement, and finally calling the proofreading agent in the service layer to perform text correction. The corrected text is then verified for factual accuracy by the temporal knowledge graph verification module combined with the structured knowledge base and terminology database in the data layer. Finally, the verified text is echoed back to the interface and user layer.

[0073] 1. After the platform layer receives the writing requirements synchronized with the user layer through a unified gateway, the industry adaptation module sequentially executes three steps: "industry identification—configuration loading—terminology adaptation." Finally, it calls the industry feature library, structured knowledge base, industry template library, terminology library, and compliance rule library from the data layer to complete industry adaptation parsing, mapping the task to a domain ontology event graph, and outputting a structured writing task description package. This process specifically includes:

[0074] (1.1) Industry identification: The requirement description transmitted between the interface and the user layer is vectorized and then the cosine similarity is calculated with the industry feature library pre-set in the data layer. The scene labels such as archives, regulations, and finance are output in seconds.

[0075] (1.2) Configuration loading: Based on the structured knowledge base and industry template library of the hot-pull data layer of the scene tag, combined with the domain ontology event graph generated after industry identification in step (1.1), the template has pre-set constraints such as title level, citation format and approval elements, and immediately generates a structured writing framework.

[0076] (1.3) Terminology adaptation: Based on the aforementioned structured writing framework, the terminology library and compliance rule library of the data layer are called to carry out industry adaptation parsing work. The general expressions, non-standard terms, unit names and citation formats in the writing framework are automatically replaced with industry standard expressions, and the compliance of the expressions is verified at the same time. Finally, a writing task description package that can directly drive subsequent hybrid retrieval and RAG enhancement is formed.

[0077] 2. After the industry adaptation module outputs the structured writing task description package, the hybrid retrieval module, driven by the writing task description package, sequentially completes slot extraction, BM25+ vector hybrid recall, and re-ranking to obtain high-confidence fragments from the structured knowledge base and vector knowledge base. The specific process includes:

[0078] (2.1) The slot extraction unit uses the data layer event-entity dictionary as a basis and adopts the "BIO + pointer network" joint model to annotate the specified core fields in the structured writing task description package word by word, accurately extract the core requirements of the task (such as industry attributes, writing topics, key entities, etc.) and output the structured query vector, which also serves as a unified input for sparse and dense retrieval.

[0079] (2.2) The retrieval engine unit launches the BM25 sparse channel, the semantically dense vector channel, and the keyword completion channel in parallel: the BM25 sparse channel achieves accurate term matching based on the inverted index of the structured knowledge base, the semantically dense vector channel captures synonyms and implied expressions in the full embedding space of the vector knowledge base, and the keyword completion channel expands and recalls synonyms or abbreviations based on the data layer terminology database. The three results are deduplicated and merged to form an initial candidate set and each carries a basic similarity score.

[0080] (2.3) The reordering unit introduces a contrastive learning pre-trained model, which integrates four-dimensional features such as relevance (the degree of matching with the writing task), timeliness (the publication / update time of the knowledge fragment), source authority (the qualification level of the knowledge source), and paragraph integrity (the completeness of the fragment information) in a "query-fragment" pairing manner. The model calculates and outputs a confidence score in the range of 0–1, and extracts the Top-K (the K value can be dynamically configured according to the task complexity) high-confidence fragments in descending order of score. At the same time, it generates a citation package containing fragment ID, source URL, publication / revision date, confidence score, and core related terms, and sends it to the RAG module to complete the supply of traceable and highly confident knowledge.

[0081] 3. Under the principles of "dynamic-reliable-traceable," the RAG module consists of four sub-pipelines connected in series and parallel: knowledge retrieval, data fusion, context awareness, and quality assessment. This enables the precise injection of high-confidence fragments into the large model layer. The specific process includes:

[0082] (3.1) Dynamic Knowledge Retrieval: The input consists of the structured writing task description package output by the industry adaptation module and the high-confidence fragments and citation packages output by the hybrid retrieval module. First, the event graph and paragraph-level keywords are extracted from the structured writing task description package. Then, the core semantics of the high-confidence fragments and the knowledge anchors associated with the citation packages are integrated. Based on the graph nodes, the core semantics of the high-confidence fragments and the knowledge anchors of the citation packages, a three-state query of "keyword-semantics-vector" is generated. These queries are then sent to the inverted index of the structured knowledge base, the dense index of the vector knowledge base and the sparse index for parallel retrieval. At the same time, incremental knowledge patches from the data layer are merged into the index at the minute level through an incremental message queue to ensure that the candidate set is zero time difference with the latest facts. The output is a set of candidate fragments with initial similarity classification.

[0083] (3.2) Multi-source data fusion: For the candidate fragment set, the corresponding parser is automatically called according to the MIME type. Structured data is directly mapped to attribute triples, unstructured text is generated into semantic blocks after paragraph segmentation, referential resolution and entity linking, and images / tables are converted into structured text descriptions after OCR and table structure recognition. The three types of data units processed by the corresponding parser are projected into the same semantic space through a unified vectorized encoder. The bidirectional alignment loss function is used to eliminate modal differences, form a consistent knowledge representation and bind a globally unique ID (fragment ID in the associated reference package), providing a basis for traceability.

[0084] (3.3) Context awareness: The fused knowledge representation and the core requirement context in the structured writing task description package are fed into the lightweight re-ranking model. The lightweight re-ranking model uses the paragraph topic vector as the query, calculates the coupling degree between the knowledge representation and the writing requirement context in real time, dynamically adjusts the keyword weight and vector distance threshold, improves the ranking of the paragraph that is strongly related to the current writing topic, and outputs high-confidence re-ranking results.

[0085] (3.4) Knowledge quality assessment: First, extract the basic information of the high-confidence rearrangement results and associated citation packages, and quantify them from four dimensions: authority (source qualification, source URL qualification in associated citation description files), timeliness (publication / revision date, time information in associated citation description files), relevance (matching degree with writing task), and completeness (fragment information coverage). The scores of each dimension are weighted according to preset weights to calculate the comprehensive score. Fragments with a comprehensive score lower than the set threshold are automatically discarded, and the remaining fragments are sorted in descending order of comprehensive score. The overall average quality score of the output high-confidence fragment set is ≥0.92.

[0086] (3.5) The RAG module calls the corresponding template from the industry template library retrieved by the industry adaptation module during the configuration loading stage, and concatenates the high-confidence fragments into an enhanced Prompt according to the template structure, context logic and writing requirements. It reuses the citation package generated by the hybrid retrieval module and supplements the globally unique ID information, and sends it to the large model layer. The large model layer performs the generation operation, and the writing large model generates a text draft that meets the requirements, outputs paragraph-level text and citation index, and records the key reasoning nodes in the generation process to provide support for subsequent interpretability traceability and for subsequent platform layer scheduling service layer collaborative optimization.

[0087] 4. The platform layer receives paragraph-level text and citation indexes output from the large model layer. Through the multi-agent scheduling module, it triggers collaborative optimization. First, it schedules the writing agent in the service layer to generate standardized text. Then, it calls the interactive writing enhancement agent in the service layer to perform text enhancement. Finally, it calls the proofreading agent in the service layer to perform text correction. The corrected text is then verified for factual accuracy by the temporal knowledge graph verification module, combined with the structured knowledge base and terminology database in the data layer. Finally, the verified text is displayed back to the interface and user layer. This process specifically includes:

[0088] The writing agent invokes content generation atomic services encapsulated in the service layer, data-related resources (such as industry template libraries and terminology libraries), and its own standardization processing capabilities to generate standardized text based on paragraph-level text and citation indexes output from the large model layer. The system provides four types of atomic services: content generation, sentence expansion, sentence abbreviation, sentence polishing, and sentence rewriting.

[0089] The interactive writing enhancement agent subscribes to the standard text and citation index in real time, parses the rewrite / expansion / abbreviation / polishing instructions synchronized between the interface and the user layer in real time, dynamically calls the four types of atomic services encapsulated in the service layer, namely sentence expansion, sentence abbreviation, sentence polishing and sentence rewriting, and makes fine-grained adjustments to the chapter structure, terminology standardization and sentence fluency of the standard text. The adjustment results are written back to the system message bus in real time for subsequent review agents to share and reuse.

[0090] The proofreading agent initiates five atomic services in parallel: punctuation standardization, sentence fluency, text accuracy, format compliance, and sensitive word screening. Simultaneously, it links with the data layer compliance rule base and terminology base, reuses the platform layer's basic verification logic, and performs full-dimensional verification character by character, paragraph by paragraph, and style by style, returning difference markers and precise correction suggestions. Simultaneously, it generates quantitative confidence scores from five dimensions: semantic consistency, character accuracy, format compliance, sensitive information risk, and content compliance. If the overall confidence score is lower than a preset threshold, the proofreading agent initiates a second-iteration optimization request to the interactive writing enhancement agent through a "game-cooperation" mechanism. The writing agent simultaneously provides standardized basic parameters, and the three work together in a cyclical process of "standardization adaptation (writing agent) - enhancement adjustment (interactive writing enhancement agent) - verification evaluation (proofreading agent)" based on the game-cooperation mechanism until global convergence (meeting the confidence threshold requirement).

[0091] 5. The temporal knowledge graph verification module performs final factual accuracy verification on the reviewed text. Based on authoritative facts in the data layer structured knowledge base and standardized expressions in the terminology base, it constructs a temporal association verification link to check for issues such as timeline contradictions, factual deviations, and misuse of terminology. Only text that passes the verification is allowed to pass.

[0092] Specifically, the temporal knowledge graph verification module uses a three-level overlay graph of "general-industry-scenario" as its core, linking the structured knowledge base, terminology base, and compliance rule base of the data layer. It completes fact verification according to the process of "graph alignment-incremental update-multidimensional verification-anomaly feedback," achieving multidimensional, fine-grained, and real-time verification of the generated content. The specific process includes:

[0093] (5.1) Graph Alignment: First, the domain ontology event graph output by the platform layer industry adaptation module is matched with the "general-industry-scenario" three-level superimposed graph built into the temporal knowledge graph verification module. The general graph provides basic entities and higher-level concepts, the industry graph reuses the terminology library loaded in the industry adaptation stage, and the scenario graph accurately matches the pre-set operation-level elements in the template of the writing task configuration loading stage. The cross-graph alignment is achieved through the dual mechanism of "unique entity ID matching + semantic vector similarity verification" to ensure that the verification granularity corresponds to the writing task scenario requirements at the same scale.

[0094] (5.2) Incremental Update: Linking the incremental message queue of the data layer, the system captures newly released / revised authoritative knowledge (such as industry standards, policy documents, and standard terminology) in the structured knowledge base in real time. The system completes the online injection of graph nodes, attributes, and relationships within 3 minutes. At the same time, it timestamps and archives the old version of knowledge and associates the timestamp information in the reference package to achieve "zero-latency" hot updates. The update record is synchronously associated with the source URL in the reference package and the knowledge release certificate of the data layer, forming a "knowledge update-source tracing" closed loop to prevent policy citation and terminology usage deviations caused by knowledge lag.

[0095] (5.3) Multidimensional verification execution: The temporal knowledge graph verification module simultaneously initiates dual-dimensional verification of logical consistency and factual accuracy: On the logical consistency verification side, it calls the pre-set temporal constraints and causal rule base (associated with the compliance rule base of the data layer) to automatically detect anomalies such as contradictions in relationships between entities, inversion of event time sequence, and data range conflicts. The positioning accuracy reaches the clause level within the paragraph. Based on the severity of the anomaly, it triggers confidence penalty (such as deducting the corresponding dimension score for low confidence anomalies) or initiates a paragraph knowledge re-retrieval request in association with the RAG module; On the factual accuracy verification side, it combines the fragment ID, source authority weight, graph entity attributes, and relation edges in the reference package to perform bidirectional reasoning. Forward reasoning verifies the consistency between the text statement and the graph nodes and relationships, and backward reasoning completes the missing or implicit key facts in the text. The reasoning process forms a cross-verification with the quality assessment dimensions (authority, timeliness, etc.) of the RAG module to improve the credibility of the verification;

[0096] (5.4) Anomaly Handling and Feedback: If an anomaly is detected, the temporal knowledge graph verification module marks the anomaly location in real time (accurate to the sentence level) and associates it with the corresponding node in the domain ontology event graph. At the same time, it generates a traceable alternative expression based on authoritative evidence in the data layer terminology library, compliance rule library, and structured knowledge base, with source links, confidence scores, timestamps, and verification logic explanations. It also outputs a structured anomaly report and correction suggestions. It is recommended to support two modes: "one-click user location and review" or "automatic triggering of the service layer interactive writing enhancement agent for paragraph optimization" to ensure closed-loop processing of anomaly issues.

[0097] Finally, the platform layer will synchronize the optimized results after the closed-loop processing of "standardization-enhancement-proofreading-compliance-fact verification" and the full-process traceability records to the interface and user layers to complete the result delivery.

[0098] (iv) Large model layer, which is equipped with a writing large model that has been continuously pre-trained, supervised fine-tuned and aligned with reinforcement learning. Specifically, it performs the following operations: receives the enhanced Prompt and citation package transmitted by the RAG module of the platform layer, generates a text draft that meets the requirements through the writing large model, outputs paragraph-level text and citation index, and records key inference nodes in the generation process to provide support for interpretability traceability.

[0099] In this embodiment, the large model layer adopts the core concept of "corpus is knowledge, knowledge is parameters" and uses a domestic computing power cluster to carry out a three-level progressive reshaping of the base large model, forming a reliable writing engine for high-barrier scenarios in multiple industries, while completing the core generation task. The specific process steps are as follows:

[0100] Step 1: Continuous pre-training of the large-scale foundation model: Build a high-quality hybrid corpus covering fifteen industries (including structured specifications, practical cases, authoritative literature, etc.), inject high-frequency terms from various industries into the data layer to expand the vocabulary, and provide rich domain corpus support for the large-scale model; adopt a course learning strategy from easy to difficult, first train the general domain corpus to retain basic language capabilities, and then gradually introduce domain-specific corpus, so that the large-scale model can accurately master the segmentation and semantic representation capabilities of domain-specific expressions. Through the elastic weight decay mechanism, catastrophic forgetting is suppressed, providing a large-scale model foundation with domain-based capabilities for subsequent fine-tuning and reinforcement learning;

[0101] Step 2: Large Model Supervised Fine-Tuning: Building upon the domain-specific capabilities established in Step 1's pre-training, heterogeneous writing tasks (such as report writing and text drafting) are encapsulated into a single paradigm using the "instruction-knowledge-output" triplet as the core. High-fidelity alignment training is conducted using QA instructions with <slot> slots. By amplifying gradients at error-prone locations, introducing paragraph-level attention masks, and relative position encoding, the logical order of "chapter-section-article" in the text is ensured to remain intact, achieving precise calibration of the pre-trained large model's generation logic. Simultaneously, national cryptographic algorithms are used to encrypt model parameters, and blockchain fingerprint digest technology is combined to record parameter change trajectories, enabling rollback and auditability of model parameter versions. This lays a solid foundation for security and compliance for subsequent reinforcement learning alignment and core task execution, completing high-quality policy initialization.

[0102] Step 3: Large Model Reinforcement Learning Alignment: Based on the strategy initialization completed in Step 2, a dual-track preference optimization mechanism is built. The factual accuracy track uses industry review logic supported by the structured knowledge base and terminology base of the data layer to score the risk of candidate manuscripts. The format compliance track automatically compares the layout elements according to the industry standards of the data layer. A dual-tower reward model is adopted to calculate the difference between factual and format similarity. In DPO training, the weight of factual reward is set to twice that of format reward. At the same time, a length penalty factor is introduced to suppress redundant generation. The basic large model parameters after the pre-training in Step 1 and the fine-tuning in Step 2 are kept frozen. Only the policy network is optimized. Finally, the factual accuracy of the large model is ≥95%, the one-time format compliance rate is ≥90%, and the general language ability is improved simultaneously, providing a key guarantee for the output of high-quality results for the core generation task.

[0103] Step 4: Core Generation Task Execution: After receiving the enhanced Prompt and citation package from the platform layer's RAG module, the large model layer, relying on the writing model's capabilities aligned with continuous pre-training, supervised fine-tuning, and reinforcement learning, and combined with dynamic prompting optimization strategies, transforms the domain knowledge learned during pre-training, the logical norms of fine-tuning and calibration, and the quality standards of reinforcement optimization into actual generation capabilities. It generates a first draft of text that meets the requirements, outputs paragraph-level text and citation indexes, and records key inference nodes in the generation process to support interpretability traceability. Subsequently, the output paragraph-level text and citation indexes are synchronized to the platform layer's multi-agent scheduling module for collaborative optimization, and the platform layer synchronizes the citation indexes and key inference nodes to the interpretability and security layer for source tracing.

[0104] (V) Data Layer: Providing end-to-end data assurance for the platform layer and the large model layer. First, it constructs a multi-type data storage system, covering industry feature libraries, structured knowledge bases, industry template libraries (adapting to different genres), vector knowledge bases (for efficient retrieval), terminology libraries (domain-specific vocabulary), and compliance rule libraries (compliance requirements for various domains). Second, it ensures data quality through metadata governance (including data cleaning, updates, and permission grading). Third, it provides standardized data access interfaces to support data access needs at each layer in areas such as requirement analysis, industry adaptation, retrieval enhancement, and fact verification.

[0105] (vi) The interpretability and security layer runs through the entire chain from writing requirement input to result delivery and is deeply linked with each layer. It is used to: receive key reasoning nodes of the large model layer and reference packages of the platform layer to build knowledge traceability links and reasoning process graphs, and generate confidence assessments by combining the fact verification results of the platform layer; perform sensitive information desensitization on the text to be displayed output by the platform layer, and add digital watermarks to the final display results of the interface and user layer; and track the data of the entire chain through unified logs to form chain evidence tracing to ensure the transparency and security of the generation process.

[0106] Specifically, the interpretability and security layer adopts a three-layer in-depth architecture of "knowledge tracing—reasoning chain—confidence assessment," running through the entire process from input of writing needs to delivery of results, and deeply linking with each layer to achieve transparency in the generation process, credibility of results, and controllable risks; among which:

[0107] As the foundational layer of the architecture, the knowledge tracing module receives key inference nodes from the large model layer and reference packages from the platform layer. It records the original source, retrieval path, processing timestamp, and verification status of each knowledge fragment using a Merkle tree structure. It ensures data integrity through SHA-256 hashing and supports second-level backtracking to the original policy text or database records, providing reliable source support for subsequent inference and evaluation.

[0108] The reasoning chain module is based on the credible source of the knowledge tracing module. It uses the Do-Calculus algorithm to construct a step-by-step logical path from input requirements to output content. Combined with the causal reasoning graph, it forms a complete reasoning link, constructs a reasoning process graph, and provides interactive visualization functions through D3.js, allowing users to expand and view the decision basis at any stage, so as to make the reasoning process explainable.

[0109] The confidence assessment module combines the deduction results of the reasoning chain module with the fact verification results of the platform layer, and integrates four-dimensional indicators: source credibility, logical consistency, statistical confidence, and historical accuracy. It uses a weighted scoring model to calculate a standardized score of 0-100. The score is updated in real time after each paragraph is generated or modified, and a detailed scoring explanation (including the score ratio of each dimension and the explanation of deduction items) is provided to help users quickly judge the reliability of the current content.

[0110] The explainability and security layer also enables triple capabilities: privacy protection, security traceability, and compliance auditing.

[0111] (a) Privacy protection: Sensitive information is desensitized on the text to be displayed at the platform layer, ε-differential privacy technology is introduced to ensure data privacy and security, and the Named Entity Recognition (NER) and rule-based hybrid method are combined to automatically identify sensitive content such as personal information, trade secrets, and classified data, and perform precise desensitization processing;

[0112] (b) Security traceability: Add invisible digital watermarks to the results displayed by the interface and user layer (supporting the traceability of associated user operation information), and at the same time track the data of the whole chain through a unified log system (including operation records of all stages such as requirement input, industry adaptation, search enhancement, model generation, and verification optimization), forming an immutable chain evidence traceability system;

[0113] (c) Compliance audit: The integrated compliance audit engine module serves as a supplement to security capabilities, constructing a legal knowledge graph covering 15 industries. It contains over 500,000 relevant articles, industry standards, and regulatory rules, and employs a hybrid reasoning mechanism combining rule engines and machine learning to perform real-time compliance scanning and risk warnings on generated text. Violations are highlighted instantly, and specific modification suggestions (based on relevant legal provisions) are provided simultaneously.

[0114] (vii) Infrastructure layer, used to: provide underlying resource support for the aforementioned layers (interface and user layer, service layer, platform layer, large model layer, data layer and interpretability and security layer), and ensure horizontal elastic scaling of the service layer, efficient computing of the large model layer and stable storage of the data layer through computing, network, storage and container orchestration resources, while supporting high availability operation of unified gateway and unified authentication.

[0115] All the aforementioned layers (interface and user layer, service layer, platform layer, large model layer, data layer, interpretability and security layer, and infrastructure layer) are interconnected through a unified gateway to achieve data flow, unified authentication to ensure access security, and unified log tracking of link status. This forms a closed-loop end-to-end process: "interface and user layer receives requirements — platform layer parses and enhances and sends to the large model — large model layer generates core components — platform layer schedules service layer to collaboratively optimize and verify — interpretability and security layer evaluates and strengthens security — interface and user layer delivers results." The data layer and infrastructure layer provide continuous data and resource support for the entire process.

[0116] Reference Appendix Figure 2 The specific workflow of the large model-assisted writing system based on this embodiment is as follows:

[0117] S1. The interface and user layer receive writing requests submitted by users in natural language. After structured parsing, preliminary verification, and format standardization, the requests are synchronized to the platform layer through a unified gateway.

[0118] S2. The platform layer receives writing requests from the user layer via the interface and performs the following operations:

[0119] First, the industry adaptation module calls the various libraries in the data layer to complete the industry adaptation parsing. Then, the hybrid retrieval module obtains high-confidence fragments from the structured knowledge base and vector knowledge base. The RAG module concatenates the high-confidence fragments with the writing template to form an enhanced Prompt, and generates a reference package, which is then sent to the large model layer.

[0120] S3, the large model layer receives the enhanced Prompt and citation package transmitted by the RAG module of the platform layer, generates a draft text that meets the requirements through the writing large model, outputs paragraph-level text and citation index, and records key reasoning nodes in the generation process to support interpretability traceability.

[0121] S4. The platform layer receives paragraph-level text and citation indexes output from the large model layer. It triggers collaborative optimization through the multi-agent scheduling module: first, it schedules the writing agent of the service layer to generate standard text, then calls the interactive writing enhancement agent of the service layer to enhance the text, and finally calls the proofreading agent of the service layer to correct the text. The corrected text is verified by the temporal knowledge graph verification module in combination with the structured knowledge base and terminology base of the data layer to verify the factual accuracy. Finally, the text that passes the verification is echoed back to the interface and user layer.

[0122] It should be added that, in the entire process from input of writing requirements to delivery of results, the interpretability and security layer receives key inference nodes from the large model layer and reference packages from the platform layer to construct a knowledge tracing link and inference process graph, and generates a confidence assessment by combining the fact verification results of the platform layer; performs sensitive information desensitization on the text to be displayed output by the platform layer, adds digital watermarks to the final display results of the interface and user layer; and tracks the data of the entire link through unified logs to form a chain of evidence tracing to ensure the transparency and security of the generation process.

[0123] In summary, the intelligent RAG-enhanced large-model assisted writing system of this invention can construct an integrated "dynamic-intelligent-trustworthy" writing enhancement framework and complete the content generation, polishing and proofreading of complex writing tasks end-to-end. This solves the problem that existing intelligent writing systems are prone to rejection or compliance risks when writing texts in high-barrier fields such as medical fields, due to industry context mismatch, terminology drift and insufficient wording rigor.

[0124] 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. An intelligent RAG-enhanced large model assisted writing system, characterized in that, It adopts a seven-layer fusion architecture, including: Interface and user layer, as the entrance for system interaction with users, used for: receiving the writing requirements submitted by users in natural language form and displaying the corresponding generation results; Service layer, deploying writing agents, interactive writing enhancement agents and proofreading agents, which are encapsulated by micro-service cluster architecture at the bottom, encapsulating two core atomic capabilities of text generation and proofreading, where the text generation atomic capability includes five types of services: content generation, sentence expansion, sentence abbreviation, sentence polishing and sentence rewriting, and the proofreading atomic capability includes five types of services: punctuation specification, sentence fluency, word accuracy, format compliance and sensitive word screening, providing basic function support for upper agents; Platform layer, as the core scheduling center, specifically implements the following operations: receiving the writing requirements uploaded by the interface and user layer, first calling the industry adaptation module to complete the industry adaptation analysis of the data layer library, then obtaining high-confidence fragments from the structured knowledge base and vector knowledge base through the hybrid retrieval module, and through the RAG module, the high-confidence fragments are spliced into enhanced prompts with writing templates, and reference packages are generated, which are sent to the large model layer; receiving the paragraph-level text and reference index output by the large model layer, triggering collaborative optimization through the multi-agent scheduling module, first scheduling the writing agent of the service layer to generate standard text, then calling the interactive writing enhancement agent of the service layer for text enhancement, and finally calling the proofreading agent of the service layer for text correction, the corrected text is verified by the temporal knowledge graph verification module combined with the structured knowledge base and term library of the data layer for fact accuracy verification, and the text that passes the verification is finally echoed to the interface and user layer; Large model layer, configured with writing large models aligned by continuous pre-training, supervised fine-tuning and reinforcement learning, specifically implements the following operations: receiving the enhanced Prompt and reference package transmitted by the RAG module of the platform layer, generating text drafts that meet the requirements through the writing large model, outputting paragraph-level text and reference index, and recording key reasoning nodes in the generation process to provide support for explainability traceability; Data layer, providing full-link data guarantee for platform layer and large model layer: first, building a multi-type data storage system, covering industry feature library, structured knowledge base, industry template library, vector knowledge base, term library and compliance rule library; second, ensuring data quality through metadata governance; third, providing standardized data access interface to support data calling needs of each layer; Explainability and security layer, running through the whole link from writing requirement input to result delivery and deeply linked with each layer, used for: receiving key reasoning nodes from the large model layer and constructing knowledge traceability link and reasoning process graph with the reference package from the platform layer, generating confidence evaluation combined with the fact verification result of the platform layer; performing sensitive information desensitization on the text output by the platform layer to be echoed, adding digital watermark to the final echoed result of the interface and user layer; tracking the whole link data through unified log to form chain evidence tracking, ensuring transparent and safe generation process; An infrastructure layer, configured to provide underlying resource support for the foregoing layers, and to guarantee horizontal elastic expansion of the service layer, efficient computation of the large model layer, and stable storage of the data layer through computing, network, storage, and container orchestration resources, while supporting high-availability operation of the unified gateway and unified authentication; The foregoing all layers are connected through the unified gateway to realize data flow, unified authentication to guarantee access security, and unified log tracking link status, forming a full-link closed loop of "interface and user layer receives requirements-platform layer analyzes and enhances and sends to large model-large model layer core generates-platform layer dispatches service layer to optimize and verify-explainability and security layer evaluates and secures-results are delivered to the interface and user layer", and the data layer and the infrastructure layer provide continuous data and resource support for the whole process.

2. The intelligent RAG-augmented large model assisted writing system according to claim 1, wherein, The interface and user layer supports access of the Web platform, the management background, and the WPS plug-in through a standardized interface, receives writing requirements, performs structured analysis and preliminary verification and format regularization on the writing requirements, and synchronizes the writing requirements to the platform layer through the unified gateway; meanwhile, the interface and user layer receives text processed by the explainability and security layer and echoes the text to the user.

3. The intelligent RAG-augmented large model assisted writing system of claim 1, wherein, After receiving the writing requirements synchronized by the interface and user layer through the unified gateway, the platform layer executes, in sequence, three steps of "industry recognition-configuration loading-terminology adaptation" by an industry adaptation module, finally calls the industry feature library, the structured knowledge base, the industry template library, the terminology library, and the compliance rule library of the data layer to complete industry adaptation analysis, maps the task to a domain ontology event graph, and outputs a structured writing task description package, and the process specifically includes: (1.1) Industry recognition: after the requirement description transmitted by the interface and user layer is vectorized, the cosine similarity is calculated with the industry feature library preloaded in the data layer, and the scene label is output in seconds; (1.2) Configuration loading: according to the scene label, the structured knowledge base and the industry template library of the data layer are hot-pulled, and a structured writing framework is generated in combination with the domain ontology event graph generated after step (1.1) industry recognition, and the constraint rules preloaded in the template are generated immediately; (1.3) Terminology adaptation: based on the foregoing structured writing framework, the industry adaptation analysis work is carried out by calling the terminology library and the compliance rule library of the data layer, the general expressions, non-standard terms, unit names, and citation formats in the writing framework are automatically replaced with industry standard expressions, the compliance of the expressions is verified synchronously, and finally a writing task description package that can directly drive subsequent hybrid retrieval and RAG enhancement is formed.

4. The intelligent RAG-augmented large model assisted writing system according to claim 3, wherein, After the industry adaptation module outputs the structured writing task description package, the hybrid retrieval module completes, in sequence, slot extraction, BM25+vector hybrid recall and reordering under the driving of the writing task description package, to obtain high-confidence fragments from the structured knowledge base and the vector knowledge base, and the specific process includes: (2.1) The slot extraction unit takes the event-entity dictionary of the data layer as a benchmark, adopts a "BIO+pointer network" joint model to perform word-by-word labeling on the specified core fields in the structured writing task description package, accurately extracts the core requirements of the task, and outputs a structured query vector, which is used as the unified input of sparse and dense retrieval; (2.2) Retrieval engine unit starts BM25 sparse channel, semantic dense vector channel and keyword completion channel in parallel: BM25 sparse channel realizes accurate term hit for the inverted index of structured knowledge base, semantic dense vector channel captures synonymous and implication expression in the whole library embedding space of vector knowledge base, keyword completion channel expands and recalls synonymous or abbreviated forms based on data layer term library, three-way results are combined to form the initial screening candidate set and carry basic similarity scores respectively; (2.3) Reordering unit introduces a contrastive learning pre-training model to fuse relevance, timeliness, source authority and paragraph integrity in the "query-fragment” pairing mode, calculates and outputs confidence in the interval of 0-1 through the model, and cuts Top-K high-confidence fragments in descending order of score, and generates a reference package containing fragment ID, source URL, publication / revised date, confidence score and core associated terms, which are sent to the RAG module to complete the knowledge supply with traceability and high confidence.

5. The intelligent RAG-augmented large model assisted writing system according to claim 4, wherein, The RAG module is composed of four sub-pipelines of knowledge retrieval, data fusion, context awareness and quality evaluation in series and parallel under the principle of "dynamic-trustable-traceable”, realizes the precise injection of high-confidence fragments into the large model layer, and the specific process includes: (3.1) Dynamic knowledge retrieval: the input is the structured writing task description package output by the industry adaptation module and the high-confidence fragments and reference package output by the mixed retrieval module, the event graph and paragraph-level topic words are extracted from the structured writing task description package, the core semantics of the high-confidence fragments are fused, and the knowledge anchor points associated with the reference package are fused, then a "keyword-semantic-vector” three-state query is generated based on the graph nodes, core semantics of high-confidence fragments and knowledge anchor points of reference package, and is sent to the structured knowledge base inverted index, vector knowledge base dense index and sparse index in parallel for recall; at the same time, the incremental knowledge patches of the data layer are merged into the index through the incremental message queue every minute to ensure that the candidate set is zero-time-difference with the latest facts, and the output is a candidate fragment set with initial similarity scores; (3.2) Multi-source data fusion: for the candidate fragment set, the corresponding parser is automatically called according to the MIME type, among which the structured data is directly mapped into attribute triples, the unstructured text is segmented into semantic blocks after paragraph segmentation, anaphora resolution and entity linking, and the image / table is converted into structured text description after OCR and table structure recognition; the three types of data units processed by the corresponding parser are projected into the same semantic space through a unified vectorization encoder, the modal difference is eliminated by using a bidirectional alignment loss function, forming a consistent knowledge representation and binding a globally unique ID for tracing; (3.3) Context awareness: the fused knowledge representation and the core demand context in the structured writing task description package are sent into the lightweight rearrangement model together; the lightweight rearrangement model takes the paragraph theme vector as the query, calculates the coupling degree of the knowledge representation and the writing demand context in real time, dynamically adjusts the keyword weight and vector distance threshold, makes the fragments strongly related to the current writing theme rank up, and outputs the high-confidence rearrangement result; (3.4) Knowledge quality evaluation: First, extract high-confidence rearrangement results and associated reference package basic information, and quantify scores from four dimensions of authority, timeliness, relevance, and completeness; each dimension score is weighted by a pre-set weight to calculate a comprehensive score, automatically discarding fragments with a comprehensive score below a set threshold, and arranging the remaining fragments in descending order of comprehensive score, with the overall average quality score of the output high-confidence fragment set ≥0.92; (3.5) The RAG module calls the industry adaptation module to splice the high-confidence fragments into enhanced prompts according to the template structure, context logic, and writing requirements, reuses the reference package generated by the hybrid retrieval module and supplements the globally unique ID information, and sends them to the large model layer; the large model layer performs generation operations, generates a text draft that meets the requirements through the writing large model, outputs paragraph-level text and reference index, and records key reasoning nodes in the generation process to support subsequent traceability for explainability, and is used for subsequent platform layer scheduling service layer collaborative optimization.

6. The system of claim 5, wherein the system is configured to: The large model layer takes "corpus as knowledge, knowledge as parameter" as the core concept, uses domestic computing power clusters to implement three-level progressive remodeling of the base large model, forms a multi-industry high-barrier scenario credible writing engine, and at the same time completes the core generation task, the specific process steps are as follows: Step 1: Continuous pre-training of the base large model: build a high-quality mixed corpus covering fifteen industries, inject data layer industry high-frequency term expansion word table, provide rich field corpus support for the large model; adopt an easy-to-difficult curriculum learning strategy, first train general field corpus to retain basic language ability, then gradually introduce field professional corpus, make the large model accurately master the segmentation and semantic representation ability of field-specific expressions, and inhibit catastrophic forgetting through elastic weight decay mechanism, provide large model base with field foundation ability for subsequent fine-tuning and reinforcement learning; Step 2: Large model supervised fine-tuning: Based on the field foundation ability formed in step 1, take the "instruction-knowledge-output" triple as the core, encapsulate heterogeneous writing tasks into the same paradigm, and use QA instructions with 〈slot〉 slots for high-fidelity alignment training; amplify the gradient of easy-to-mistake positions, introduce paragraph-level attention mask and relative position encoding to ensure that the logical order of "chapter-section-article" in the text is not damaged, and realize precise calibration of the pre-trained large model generation logic; At the same time, encrypt model parameters using national encryption algorithm, combine blockchain fingerprint summary technology to record parameter change track, realize model parameter version rollback and audit, build a safe and compliant foundation for subsequent reinforcement learning alignment and core task execution, complete high-quality strategy initialization; Step 3: Large model reinforcement learning alignment: Initialize the large model based on the policy completed in step 2, build a double-track preference optimization mechanism, the fact accuracy track is supported by the industry review logic of the data layer structured knowledge base and the term library to score the candidate manuscript, the format compliance track automatically compares the layout elements according to the data layer industry standards; adopt a double-tower structure reward model to calculate the difference in fact and format similarity, set the fact reward weight to twice the format reward in DPO training, and introduce a length penalty factor to suppress redundancy generation; keep the basic large model parameters frozen after pre-training in step 1 and fine-tuning in step 2, only optimize the policy network, finally realize large model fact accuracy ≥ 95%, format one-time compliance rate ≥ 90% and synchronous improvement of general language ability, provide key protection for core generation task to output high-quality results; Step 4: Core generation task execution: After receiving the enhanced Prompt and reference package from the platform layer RAG module, the large model layer relies on the writing large model capability of continuous pre-training, supervised fine-tuning and reinforcement learning alignment, combines dynamic prompt optimization strategy, converts the domain knowledge learned by pre-training, logical specification calibrated by fine-tuning and quality standards optimized by reinforcement into actual generation ability, generates the initial draft of text that meets the requirements, outputs paragraph-level text and reference index, and records the key reasoning nodes in the generation process to provide support for traceability; synchronously output the paragraph-level text and reference index to the multi-agent scheduling module in the platform layer for collaborative optimization, and synchronously output the reference index and key reasoning nodes to the explainability and security layer for traceability.

7. The system of claim 6, wherein the system is configured to: The platform layer receives the paragraph-level text and reference index output by the large model layer, triggers collaborative optimization through the multi-agent scheduling module, first schedules the writing agent in the service layer to generate standard text, then calls the interactive writing enhancement agent in the service layer to enhance the text, and finally calls the proofreading agent in the service layer to correct the text. The corrected text is verified for fact accuracy by the temporal knowledge graph verification module in combination with the data layer structured knowledge base and term library, and the verified text is finally echoed to the interface and user layer. This process specifically includes: The writing agent calls the content generation atomic service encapsulated by the service layer, the relevant resources in the data layer and its own standardized processing capability, generates standard text based on the paragraph-level text and reference index output by the large model layer; The interactive writing enhancement agent subscribes to the standard text and reference index in real time, analyzes the rewrite / expand / write / embellish instructions synchronized by the interface and user layer in real time, dynamically calls four types of atomic services of sentence expansion, sentence abbreviation, sentence embellishment and sentence rewriting encapsulated by the service layer, adjusts the chapter structure, term standard expression and sentence fluency of the standard text in fine granularity, and writes the adjustment results back to the system message bus in real time for sharing and reuse by the subsequent proofreading agent; The proofreading intelligent agent starts five types of atomic services in a parallel manner, including "punctuation specification, sentence fluency, character accuracy, format compliance, and sensitive word screening". Meanwhile, it links the data layer compliance rule library and the term library, reuses the platform layer basic verification logic, and conducts full-dimensional verification character by character, paragraph by paragraph, and style by style. It returns difference markers and accurate correction suggestions. It also generates quantitative confidence from five dimensions, including semantic consistency, character accuracy, format compliance, sensitive information risk, and content compliance. If the comprehensive confidence is lower than the preset threshold, the proofreading intelligent agent initiates a secondary iteration optimization request to the interactive writing enhancement intelligent agent through the "game-collaboration" mechanism. The writing intelligent agent synchronously provides standardized basic parameters. The three parties cyclically execute the "standardization adaptation-enhancement adjustment-verification evaluation" process based on the game-collaboration mechanism until global convergence. The temporal knowledge graph verification module conducts final factual accuracy verification on the proofread text. Based on the authoritative facts in the data layer structured knowledge base and the standard expressions in the term library, it constructs a temporal correlation verification link to check timeline contradictions, factual deviations, and term misuse problems. Only the verified text is released. Finally, the platform layer synchronizes the optimized results after the "standardization-enhancement-proofreading-compliance-factual verification" closed-loop processing and the full-process traceability record to the interface and user layer to complete the result delivery.

8. The system of claim 7, wherein the system is configured to: The temporal knowledge graph verification module takes a three-level superimposed graph of "general-industry-scenario" as the core, links the structured knowledge base, term library, and compliance rule library of the data layer, and completes factual verification according to the "graph alignment-incremental update-multi-dimensional verification-exception feedback" process to achieve multi-dimensional, fine-grained, and real-time verification of generated content. The specific process includes: (1) Graph alignment: First, the platform layer industry adaptation module outputs the domain ontology event graph and the three-level superimposed graph of "general-industry-scenario" built in the temporal knowledge graph verification module. The general graph provides basic entities and high-level concepts, the industry graph reuses the term library loaded in the industry adaptation stage, and the scenario graph accurately matches the operation-level elements preloaded in the template in the writing task configuration loading link. Through the "entity ID unique matching + semantic vector similarity verification" dual mechanism, accurate alignment across graphs is achieved to ensure that the verification granularity corresponds to the writing task scenario demand at the same scale. (2) Incremental update: Link the data layer incremental message queue to capture new released / revised authoritative knowledge in the structured knowledge base in real time. Within 3 minutes, online injection of graph nodes, attributes, and relationships is completed. At the same time, the timestamp of the old version of knowledge is archived and associated with the timestamp information in the reference package, realizing "zero lag" hot update. The update record is synchronized to the source URL and data layer knowledge release credentials in the reference package, forming a "knowledge update-traceability" closed loop to eliminate policy reference and term usage deviations caused by knowledge lag. (3) Multi-dimensional verification execution: The temporal knowledge graph verification module synchronously starts the consistency and accuracy of the logic verification of two dimensions: on the side of logical consistency verification, the preset timing constraints and causal rule library are called to automatically detect the relationship between entities, inverted event timing and data range conflict anomalies, locate the accuracy to the intra-sentence clause level, trigger confidence penalty or associated RAG module to initiate paragraph knowledge re-search request according to the severity of the anomaly; on the side of fact accuracy verification, the segment ID, source authority weight and graph entity attribute, relationship edge in the reference package are combined for bidirectional reasoning, forward reasoning verifies the consistency of the text statement and the graph node, and backward reasoning completes the missing or implicit key facts in the text, the reasoning process and the quality evaluation dimension of the RAG module form cross verification, and the verification reliability is improved; (4) Abnormal processing and feedback: If an anomaly is detected, the temporal knowledge graph verification module marks the abnormal position in real time and associates the corresponding node of the domain ontology event graph, and generates a traceable alternative representation based on the authoritative basis in the data layer terminology library, compliance rule library and structured knowledge base, with source link, confidence score, timestamp and verification logic explanation, and synchronously outputs structured abnormal report and correction suggestion; the suggestion supports two modes of "user one-key positioning review" or "automatic triggering of service layer interactive writing enhancement intelligent agent for paragraph optimization", ensuring closed-loop processing of abnormal problems.

9. The intelligent RAG-augmented large model assisted writing system of claim 8, wherein, The explainability and security layer adopts a three-layer deep architecture of "knowledge traceability-reasoning chain-confidence evaluation", which runs through the whole process from writing demand input to result delivery, and is deeply linked with each layer to realize transparent generation process, reliable results and controllable risks; among them: The knowledge traceability module, as the basic layer of the architecture, receives the key reasoning nodes of the large model layer and the reference package of the platform layer, records the original source, retrieval path, processing timestamp and verification status of each knowledge fragment in the Merkle tree structure, guarantees data integrity through SHA-256 hashing, supports second-level backward traceability to policy original text or database records, and provides reliable source support for subsequent reasoning and evaluation; The reasoning chain module, based on the reliable source of the knowledge traceability module, uses the Do-Calculus algorithm to build a step-by-step logical path from input demand to output content, forms a complete reasoning link with the causal reasoning graph, constructs a reasoning process graph, and provides interactive visualization function through D3.js, so that users can view the decision basis at any level, realize the explainability of the reasoning process; The confidence evaluation module combines the reasoning results of the reasoning chain module and the fact verification results of the platform layer, integrates the four-dimensional indicators of source reliability, logical consistency, statistical confidence and historical accuracy, uses a weighted scoring model to calculate a 0-100 standardized score, and refreshes the score and detailed scoring explanation in real time after each paragraph generation or modification, so that users can quickly judge the reliability of the current content.

10. The intelligent RAG-augmented large model assisted writing system of claim 9, wherein, The explainability and security layer also realizes the three capabilities of privacy protection, security traceability and compliance audit: (a) Privacy protection: Perform sensitive information desensitization on the echo text output by the platform layer, introduce ε-differential privacy technology to ensure data privacy security, automatically identify sensitive content by combining named entity recognition and rule-based methods, and perform precise desensitization processing; (b) Secure traceability: Add invisible digital watermark to the results echoed by the interface and user layer, and track all-link data through a unified log system to form an unalterable chain evidence tracking system; (c) Compliance audit: Integrate a compliance audit engine module as a security capability supplement, build a regulation knowledge graph covering 15 industries, and embed 500,000+ relevant provisions, industry standards, and regulatory rules. Use a rule engine and machine learning hybrid reasoning mechanism to perform real-time compliance scanning and risk warning on generated text, highlight violating paragraphs, and provide specific modification suggestions in real time.

Images