A large model-based scientific research management intelligent agent system

Abstract

The application discloses a scientific research management intelligent agent system based on a large model, and belongs to the field of scientific research management, and comprises a multi-source data integration layer, a large model intelligent agent core layer, an intelligent application layer and an infrastructure and security system. The multi-source data integration layer is connected with existing systems of an enterprise, and obtains multi-source heterogeneous data; the large model intelligent agent core layer is based on a large language model and retrieval enhancement generation technology, realizes document analysis, vectorization processing, semantic retrieval and intelligent generation; the intelligent application layer provides intelligent duplicate checking, intelligent novelty search, intelligent generation of review reports, intelligent generation of scientific and technological review, intelligent identification of uploaded achievement materials, intelligent customer service, intelligent navigation of work flow, data analysis and decision support functions. The application realizes intelligentization of the whole process of scientific research management through a layered architecture, forms a closed loop system of data collection-intelligent processing-application service-feedback iteration, and improves the efficiency and quality of scientific research management.

Description

Technical Field

[0001] This invention relates to the field of scientific research management technology, and more particularly to a scientific research management intelligent agent system based on a large model. Background Technology

[0002] With the continuous development of scientific research and the increasing number of research projects each year, the complexity of scientific research management is growing daily. Traditional scientific research management models often rely on manual processing of a large number of repetitive and administrative tasks, such as project plagiarism checks, compliance reviews, results entry, and policy consultations, which suffers from inefficiency, error-proneness, and slow response. To improve the efficiency of scientific research management, various scientific research management systems based on information technology and intelligent technology have emerged in recent years.

[0003] For example, Chinese invention patent application CN117393119A discloses an AI-based medical institution research management system, which uses keyword retrieval and the term frequency-inverse document frequency (TF-IDF) algorithm to calculate text similarity, which can reduce the duplication of research projects in medical institutions, but does not involve semantic-level deep matching. Chinese invention patent application CN121210615A mentions an intelligent deduplication method for scientific instruments based on multi-dimensional parameter decoupling and dynamic weighting. It uses enhanced index generation (RAG) as the model framework, a classification prediction model classifies the query instruments to obtain the target instruments, and a dynamic weight training model weights the similarity of the query instruments in each parameter dimension to generate the overall similarity between the query instruments and the target instruments, and determines the deduplication result. This method can be used in the field of instrument deduplication technology.

[0004] In recent years, with the rise of large-scale model technology, researchers have attempted to combine large language models with knowledge graphs and other technologies for application in the field of scientific research management. For example, Chinese invention patent application CN121766409A discloses a method for constructing a knowledge graph for enterprise scientific research projects. This method is based on large language models for knowledge extraction and uses text similarity and nearest neighbor propagation algorithms for entity disambiguation. It mainly solves the problem of structured representation of entity relationships in scientific research project data, focusing on the construction and reasoning of knowledge graphs. Chinese invention patent application CN117609436A discloses a question-and-answer system for university scientific research management that combines knowledge graphs and large language models. By introducing large language models to integrate the question-and-answer process, which is mainly based on structured queries from knowledge graphs, it improves the quality and effectiveness of question-and-answer services. Its application scenario is question-and-answer services in the field of university scientific research management.

[0005] In summary, existing technologies are mostly designed for the specific needs of universities or a particular research field, focusing on the design or improvement of independent functions in the scientific research management process. There is still room for improvement in the integration and in-depth application of large-scale models in the field of enterprise scientific research management. Summary of the Invention

[0006] To address the problems of existing scientific research management systems, such as limited functionality, narrow coverage, lack of effective collaboration between modules, low data reuse rate, inability to share knowledge, and inability to self-evolve, this invention provides a large-scale model-based intelligent scientific research management system. Designed for enterprise scientific research managers and researchers, it offers comprehensive scientific research management services, including intelligent plagiarism detection, intelligent novelty search, intelligent generation of review reports, intelligent generation of scientific and technological innovation reviews, intelligent identification and uploading of research materials, intelligent customer service, intelligent workflow navigation, data analysis, and decision support. This intelligent support for key aspects of scientific research management improves its efficiency and quality.

[0007] To achieve the above objectives, this application provides a large-scale model-based intelligent agent system for scientific research management, comprising: a multi-source data integration layer, a large-scale model intelligent agent core layer, an intelligent application layer, and an infrastructure and security system; The multi-source data integration layer, the large model intelligent agent core layer, and the intelligent application layer are sequentially connected in communication, and the infrastructure and security system provide underlying architectural support and security for each layer. The multi-source data integration layer is used to obtain multi-source heterogeneous data required for scientific research management by interfacing with the enterprise's existing systems, and to provide data support for upper-layer intelligent services; The core layer of the large model intelligent agent is built on a large language model and retrieval enhancement generation technology, including: file management and parsing services, core service clusters and large model application platform, which are used to realize the complete link of document parsing, vectorization processing, semantic retrieval, task orchestration and intelligent generation; The file management and parsing service is mainly responsible for receiving and storing the original scientific research documents uploaded by users to the object storage database, writing the metadata into the structured database, parsing the documents and converting them into semantic vectors through a vector embedding model, and storing them in the vector database. The core service cluster includes: an intelligent service module, a project review module, an intelligent summary module, and a project compliance module; The large model application platform works in collaboration with the core service cluster, including a Search Enhancement Generation (RAG) orchestration module, a guided template management module, a workflow scheduling module, a vector database, a full-text search engine, and a structured database.

[0008] The intelligent application layer is used to provide intelligent agent functions covering the entire scientific research management process for enterprise scientific research managers and scientific researchers, including intelligent plagiarism detection, intelligent novelty search, intelligent generation of review reports, intelligent generation of scientific and technological innovation reviews, intelligent identification and uploading of achievement materials, intelligent customer service, intelligent workflow navigation, and data analysis and decision-making functions. The infrastructure and security system includes system deployment architecture and security system, which are used to provide computing power support, network communication, containerized deployment and security assurance for the system.

[0009] To avoid the problem of low data reuse rates caused by isolated operation of system modules, this invention constructs a shared semantic space and a closed-loop feedback evolution mechanism to create a deeply coupled and synergistic relationship between the modules. Specifically, this includes: The multi-source heterogeneous data acquired by the multi-source data integration layer simultaneously serves all modules of the core service cluster. By reusing data for multiple scenarios, it avoids data redundancy and information silos caused by each module building its own knowledge base. The vector database and full-text search engine in the core layer of the large model intelligent agent constitute a unified shared semantic space. The core service group ensures the consistency of knowledge representation and the comparability of search results by operating within the same semantic space, thereby realizing knowledge sharing and semantic interoperability. User interaction data and feedback data generated by various functions in the intelligent application layer are written back to the multi-source data integration layer through the feedback channel. After data cleaning and labeling, an incremental training dataset is formed, which is used to continuously fine-tune and optimize the large language model and guidance template. This closed-loop feedback mechanism enables the system to have self-evolution capabilities. As usage time increases, the accuracy of plagiarism detection, the coverage of novelty searches, the quality of review reports, and the accuracy of customer service responses continue to improve, forming a complete closed loop of "data collection - intelligent processing - application service - feedback iteration - model evolution".

[0010] The various functions of the intelligent application layer form a chain of collaboration through the workflow scheduling module: During the project application stage, after parsing the target document uploaded by the user, intelligent deduplication and novelty searches are performed sequentially. The deduplication and novelty search results can be jointly input into the intelligent review report generation module to generate comprehensive review opinions; the key information extracted by the intelligent identification function of uploaded results materials is automatically filled into the collaborative information management platform, while triggering the intelligent customer service function to push a results entry confirmation notification to the user and triggering the data analysis function to update the results statistics view; the compliance judgment results of the intelligent workflow navigation are fed back to the intelligent customer service module in real time to update the executable rules in the rule base. The above-mentioned chain of collaboration creates a dynamic coupling relationship of "input-processing-output-re-input" between the functional modules, realizing synergistic efficiency of the system workflow.

[0011] This invention addresses the high demands on semantic understanding accuracy, retrieval recall, and the professionalism of generated results in tasks such as plagiarism detection, novelty search, and review report generation in scientific research management scenarios. The invention optimizes the RAG orchestration module in the following ways: In the intelligent plagiarism detection function, this invention adopts a three-dimensional similarity calculation algorithm that combines semantic vector similarity, text fingerprint similarity, and structural feature similarity. Through a vector embedding model, the target document submitted by the user and the project data in the collaborative information management platform are converted into semantic vectors, and the cosine similarity is calculated to obtain the semantic level matching result. Text fingerprints are generated by a text fingerprinting algorithm, making it easier to quickly compare, deduplicatize, and cluster similar and identical texts without comparing the entire text word by word. The Hamming distance is calculated to obtain literal matching results, which are used to identify directly copied or highly similar text fragments. Extract chapter structure features from the target document and calculate structural feature similarity to identify structural plagiarism; This algorithm employs an adaptive weighted fusion mechanism to weight and fuse three-dimensional similarity scores, outputting a comprehensive similarity score and a detailed segmented comparison report. This hybrid algorithm overcomes the limitations of single semantic similarity in identifying literal plagiarism and single text fingerprints in identifying semantic rewriting, significantly improving the accuracy and interpretability of plagiarism detection results.

[0012] In the intelligent novelty search function, this invention adopts a multi-path recall strategy that combines vector retrieval, full-text retrieval and keyword expansion retrieval. Vector retrieval recalls semantically relevant literature from the collaborative information management platform project library and external academic databases based on semantic vector similarity. Full-text search uses keyword matching and Boolean logic retrieval based on search engines to recall documents containing specific technical terms; Keyword expansion retrieval utilizes a large language model to expand the user-input research content and innovative points using synonyms, hyponyms, and domain terms, generating an expanded keyword set for supplementary retrieval. The three-way recall results are then deduplicated and merged before being input into a cross-encoder-based re-ranking model for fine-grained sorting, ultimately outputting relevant literature. This multi-path recall strategy overcomes the problems of incomplete or inaccurate recall associated with single retrieval methods, ensuring the literature coverage and relevance of the novelty search report.

[0013] In the intelligent generation modules for review reports and scientific and technological innovation summaries, this invention employs a guided template combined with a dynamic context injection generation control mechanism. The guided template management module pre-sets suitable prompt word formats for different generation tasks. During the generation process, the RAG orchestration module retrieves relevant contextual information from vector and structured databases based on the task type, dynamically concatenates it, and injects it into the input prompts of the large language model. Simultaneously, an output validation mechanism performs structured validation on the generated results; results that fail validation trigger regeneration or manual intervention. This mechanism ensures the professionalism, consistency, and controllability of the large language model's generated results.

[0014] The vector embedding model adopts a domain-adaptive model pre-trained on corpus of scientific research management. Based on the general pre-trained model, it uses domain-specific corpus such as enterprise scientific research project materials, scientific and technological achievement data, and policy library documents for further pre-training and comparative learning fine-tuning. This enables the semantic vectors generated by the model to more accurately capture professional terms, technical concepts and semantic relationships in the field of scientific research management, thereby improving the accuracy and recall of vector retrieval.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This application provides a large-scale model-based intelligent agent system for scientific research management. By systematically applying large-scale model intelligent agent technology to the scientific research management of enterprises, and by constructing a layered architecture of multi-source data integration layer, large-scale model intelligent agent core layer, and intelligent application layer, it achieves multi-functional intelligent coverage of the entire scientific research management process, from project plagiarism checking, novelty checking, review report generation, scientific and technological innovation summary generation, automatic achievement entry, intelligent customer service, workflow navigation to data analysis and decision-making. This application addresses the high demands of tasks in research management scenarios by optimizing the RAG orchestration module, including a three-dimensional hybrid similarity calculation algorithm, multi-way recall and re-ranking strategies, a generation control mechanism with guided template constraints, and a domain-adaptive vector embedding model. This significantly improves the semantic understanding accuracy, retrieval recall rate, and professionalism of the generated results in research management, such as plagiarism detection, novelty search, and review report generation. This application constructs a shared semantic space and a closed-loop feedback evolution mechanism to reuse the collected multi-source heterogeneous data multiple times, realize cross-module data sharing within the system, form a deeply coupled synergistic relationship between various functional modules, and fine-tune the model through data feedback from each module, so as to realize the continuous self-optimization of the intelligent agent system and continuously improve the efficiency and quality of scientific research management during use. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of a research management intelligent agent system structure based on a large model, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the deployment architecture of a scientific research management intelligent agent system based on a large model, according to an embodiment of the present invention. Figure 3This is a schematic diagram of the overall workflow of a large-model-based intelligent agent system for scientific research management, according to an embodiment of the present invention. Figure 4 This is a schematic diagram of the plagiarism detection function of a scientific research management intelligent agent system based on a large model, according to an embodiment of the present invention. Figure 5 This is a schematic diagram of the novelty search function of a scientific research management intelligent agent system based on a large model, according to an embodiment of the present invention. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] The large-scale model-based intelligent agent system for scientific research management refers to an intelligent agent system built for enterprise scientific research management business, which uses a large language model (LLM) as the core reasoning engine and combines retrieval augmented generation (RAG) technology. It uses natural language interaction to complete intelligent functions such as deduplication, novelty search, intelligent generation of review reports, intelligent generation of scientific and technological innovation reviews, intelligent identification and uploading of achievement materials, intelligent customer service, intelligent workflow navigation, data analysis and decision-making.

[0020] The specific implementation process of this invention will be described in detail below.

[0021] This embodiment provides a scientific research management intelligent agent system based on a large model, such as... Figure 1 As shown, it includes a multi-source data integration layer, a large-model intelligent agent core layer, and an intelligent application layer that are connected in sequence, as well as an infrastructure and security system that provides underlying architectural support and security for each layer of the system.

[0022] The multi-source data integration layer deploys a data integration interface to interface with the enterprise's existing systems.

[0023] Specifically, it interfaces with the enterprise's unified user system: a data integration interface is built to synchronize organizational data (including department codes, department names, and hierarchical relationships) and user account data (including user ID, name, employee number, department, and role permissions) daily. The synchronized data is processed through data cleaning, formatting, merging, deduplication, verification, and business rule calculation before being stored in the user information table and organizational structure table of the structured database, providing unified user identity and permission data support for upper-level intelligent services. Integration with the unified portal: A single sign-on (SSO) integration is built based on the OAuth 2.0 protocol and JWT token mechanism. After a user logs in to the unified portal, the system automatically obtains the user's identity information through token verification, eliminating the need to log in again in the agent system. At the same time, the agent application entry is embedded in the unified portal to achieve seamless navigation. Integration with the Collaborative Information Management Platform: A two-way data integration interface is constructed. On the one hand, basic project information (project number, project name, undertaking unit, start and end time, budget), process data (milestone nodes, progress reports, change records), scientific and technological achievement data (achievement number, achievement name, achievement type, personnel involved, completion time), and policy database documents (scientific and technological project management methods, project establishment and change guidelines, funding management methods) are synchronized from the collaborative information management platform. The synchronized data is processed and stored in a structured database and an object storage database. On the other hand, the processing results of the intelligent agent system (such as duplicate reports, novelty reports, review reports, and achievement entry information) are written back to the corresponding modules of the collaborative information management platform to achieve two-way data exchange.

[0024] The core layer of the large model intelligent agent includes file management and parsing services, a core service cluster, and a large model application platform. It is built based on large language models and retrieval augmented generation (RAG) technology and deployed on the vLLM high-performance inference framework. This layer supports flexible selection and mixed invocation of multiple large language models, including but not limited to: DeepSeek series (suitable for complex inference and code generation), GLM series (suitable for dialogue and instruction following), Doubao series (suitable for efficient real-time interaction), and Qwen series (suitable for multilingual and general scenarios).

[0025] The file management and parsing service is primarily responsible for receiving and storing raw research documents uploaded by users, storing the binary streams of the documents in an object storage database, and simultaneously writing metadata into the document information table of a structured database. It parses documents in formats such as PDF and Word, and for scanned or image documents, it automatically calls an Optical Character Recognition (OCR) model to perform text recognition, outputting structured text with coordinate information. After cleaning and segmentation, a standard text dataset is formed. This standard text dataset is then converted into 768-dimensional semantic vectors using a domain-adaptive vector embedding model and stored in a vector database. Simultaneously, the standard text dataset is synchronously indexed to an open-source search engine (Elasticsearch, ES) to build an inverted index, supporting keyword retrieval and Boolean logic retrieval.

[0026] In this embodiment, the domain-adaptive vector embedding model uses domain-specific corpora such as enterprise research project materials, scientific and technological achievement data, and policy library documents for further pre-training and comparative learning fine-tuning.

[0027] Specifically, during the pre-training phase, approximately 28,000 historical research project applications and feasibility study reports from enterprises, approximately 12,000 technology achievement registration and award materials, and approximately 6,000 science and technology management regulations and policy documents at various levels were collected and cleaned. After document parsing, paragraph segmentation, and deduplication, approximately 550,000 standard text paragraphs (each paragraph's length controlled between 256 and 512 tokens) were constructed as the corpus for further pre-training within the domain. In the contrastive learning fine-tuning phase, based on the above corpus, approximately 120,000 text pairs (including positive and negative sample pairs) were constructed using a combination of rule-based and manual verification methods as the fine-tuning training dataset.

[0028] Based on a general pre-trained model (such as BGE-large or an open-source model of similar size), intra-domain adaptive pre-training is performed. The training epochs are set to 3, and a linear warm-up learning rate scheduling strategy is adopted, with the warm-up ratio set to 10% of the total steps. In the contrastive learning fine-tuning stage, the training epochs are set to 5, and an early stopping method is adopted. When the average retrieval accuracy on the validation set no longer improves for 2 consecutive epochs, the training is terminated early to prevent overfitting.

[0029] The positive sample pair construction methods for contrastive learning include: in the same document, the title and its corresponding text abstract constitute a positive sample pair; in the same research project, the project name and the abstract of the results constitute a positive sample pair; in the same technical field, synonymous sentences describing the same technical concept in different documents (confirmed through keyword overlap rate and manual review) constitute a positive sample pair.

[0030] Contrastive learning negative sample pair construction methods include: In-batch negative samples, which means that in the same training batch, paragraphs from other documents are used as negative samples for the current query sample, thereby reusing batch information to improve the diversity of negative samples; and Hard negative samples, which select text pairs that have similar keywords to the positive samples (such as TF-IDF vector cosine similarity greater than 0.6) but actually belong to different technical directions or different project categories, and mark them as hard negative samples to enhance the model's ability to distinguish easily confused semantics.

[0031] Contrastive learning employs the InfoNCE loss function with a temperature coefficient of 0.07 to adjust the model's penalty for negative samples. The training batch size is set to 64, the AdamW optimizer is used, and the initial learning rate is 3 × 10⁻⁶. -5 The weight decay coefficient is set to 0.01. After training, the output of the last hidden layer is aggregated into a 768-dimensional fixed-length vector using average pooling, and then L2 normalized before being stored in the vector database.

[0032] The core service cluster includes: an intelligent service module, a project review module, an intelligent summary module, and a project compliance module. Each module communicates asynchronously through a message queue (RabbitMQ) and uses a Redis cluster to build a multi-level cache.

[0033] The intelligent service module builds a knowledge base based on the system documents and project materials in the enterprise collaborative information management platform and is deployed on the large model application platform; it prioritizes the use of the Doubao-pro-32k model or the Qwen-Max model, configures streaming output, and controls the first character delay to within 500ms; it supports multi-turn dialogue context management, with a maximum context window of 32K tokens; The project review module is equipped with a three-dimensional hybrid similarity calculation engine that uses semantic vectors, text fingerprints, and structural features. Semantic vector similarity is calculated using cosine similarity, text fingerprint similarity is calculated using the SimHash algorithm, 64-bit fingerprints are generated after word segmentation and Hamming distance is calculated, structural feature similarity is calculated using a tree structure matching algorithm based on chapter title and paragraph position, and three-dimensional similarity is fused through adaptive weights to output a comprehensive similarity score. Configure a multi-path recall engine: the vector retrieval path uses ANN approximate nearest neighbor retrieval based on a vector database, the full-text retrieval path uses keyword matching retrieval based on an open-source search engine, and the keyword expansion retrieval path uses synonym expansion based on a large language model. Synonyms dictionary and domain terminology database can be used to supplement the retrieval. After deduplication and merging of the three recall results, they are input into a re-ranking model based on cross-encoder for fine ranking and output relevant documents.

[0034] In this embodiment, the cross-encoder adopts a Transformer-based encoder architecture (specifically, BERT-base-chinese is selected as the backbone network). Its input layer concatenates the target document query to be ranked with candidate documents to form a single input sequence. The model jointly encodes the concatenated sequence through a multi-layer bidirectional self-attention mechanism, ensuring that the query interacts with each token in the document at each layer, thereby capturing fine-grained relevance signals at the word and phrase levels. Finally, the model maps the 768-dimensional output vector corresponding to the position to a real-valued relevance score between 0 and 1 through a regression head consisting of two fully connected layers (dimensions 768→256→1).

[0035] The training data consists of three parts: historical novelty search data from enterprises, including 3,000 feasibility study reports and their corresponding comparative documents deemed "highly relevant" (relevance score ≥ 0.8) by experts, and interfering documents deemed "irrelevant" (relevance score ≤ 0.2). After cleaning, approximately 24,000 manually annotated query-document pairs were formed; abstracts and titles were obtained from public academic databases (such as CNKI and Wanfang), and remote supervision was used. Using a supervision approach, approximately 100,000 weakly supervised samples are automatically generated, with the keyword overlap rate (TF-IDF cosine similarity) and the similarity of the Chinese Library Classification number among the documents serving as weak labels. Among these, samples with an overlap rate higher than 0.7 and the same classification number are labeled as positive samples, while samples with an overlap rate lower than 0.3 and different classification numbers are labeled as negative samples. For hard negative samples that are prone to misjudgment in scientific research management scenarios, approximately 20,000 enhanced hard negative samples are constructed by using a large language model (DeepSeek-V3) to paraphrase or replace core technical terms in the query text, and then using interference text that has high keyword overlap with the original query but is technically different.

[0036] Training is conducted in two phases. The first phase is general domain adaptation, which uses pre-trained cross-encoder weights on a general ranking dataset (such as MSMARCO Passage Ranking) as initialization. Preliminary domain adaptation is performed using the aforementioned remotely supervised weakly labeled data. The training consists of two epochs with a learning rate of 2×10⁻⁶. -5 The batch size was set to 32. The second stage was fine-tuning, which used manually labeled data retained from the company's historical business to fine-tune the model. The training epochs were set to 5, employing an early stopping strategy: training was terminated early if the average regress ranking on the validation set failed to improve for two consecutive epochs. The optimizer used was AdamW, with a learning rate of 1×10⁻⁶. -5 The weight decay coefficient was set to 0.01, and the batch size was set to 16. The loss function used was the hinge loss. After training, the re-ranking model was fixed in ONNX format and independently deployed on the GPU inference nodes of the core service cluster as a sub-component of the project review module, accepting real-time calls from the intelligent novelty search function, with a single re-ranking inference latency controlled within 50 milliseconds.

[0037] The intelligent review module is equipped with a timed task scheduler, which can be set to automatically query data that meets specific conditions from designated scientific research channels at regular intervals. After data cleaning and summarization, it calls the DeepSeek-V3 model or GLM-4 model to generate scientific and technological review documents, supports exporting to Word format, and can be configured to send to a designated email address at regular intervals.

[0038] The project compliance module builds a rule base based on corporate policy documents, and the rules in the rule base are stored in a structured database; it deploys a rule engine that supports forward chain reasoning and reverse chain reasoning; and it uses different large language models for reasoning according to the complexity of the rule chain reasoning: the Qwen-Max model is used for simple rule matching, and the DeepSeek-R1 model is used for complex multi-condition reasoning.

[0039] Specifically, in this embodiment, the implementation process of the forward chain reasoning and the reverse chain reasoning is as follows: Each rule in the rule base is stored in a production rule format. The data structure includes a rule number, priority, condition part, and conclusion action part. Facts are represented in the form of triples (entity, attribute, value), such as (current user, role, project leader), (current document, total funding, 500,000), (current project, domain, artificial intelligence). All facts are maintained in the working memory of the rule engine and support dynamic addition, deletion, modification, and querying of facts.

[0040] The forward chain inference process works as follows: When the system receives a compliance judgment request, it first converts the current operation's environmental parameters (user role, operation type, operation object type, operation time, amount involved, etc.) into an initial fact set and loads it into the working memory. After the inference engine starts, it repeatedly executes the following "matching-selection-execution" loop until the target state is reached or no rules can be triggered: Matching Phase: The RETE algorithm is used to perform pattern matching on all rules in the rule base. The LHS of each rule is compared with the current fact set in the working memory to find rule instances that satisfy all conditions (i.e., the "conflict set"). The RETE algorithm achieves efficient incremental matching by constructing a top-down α-node (handling single-condition matching) and β-node (handling variable binding and connection between multiple conditions) network, avoiding traversing all rules in each loop and improving inference efficiency.

[0041] Selection Phase: Rules to be executed are selected from the conflict set based on a comprehensive decision considering both priority and the Least Recently Used time. Specifically, rules are first sorted in descending order of priority, and those with the same priority are then sorted in ascending order of the timestamp of their most recent execution (earlier timestamps take precedence) to ensure fair scheduling of rules.

[0042] Execution Phase: Executes the RHS portion of the selected rule. Operations include: adding new facts to the working memory (e.g., determining "compliant" or "non-compliant"), deleting or modifying existing facts, and calling external services (e.g., triggering alert pop-ups, logging violations, or calling a large language model to generate correction suggestions). After execution, the addition of new facts may satisfy the LHS of more rules, and the engine automatically enters the next iteration until the fact set in the working memory no longer triggers any new rules. The inference terminates, and the final determination result and inference path (i.e., the sequence of executed rule chains) are output.

[0043] The reverse chain reasoning implementation process is used when the system needs to verify whether a specific compliance objective is met (e.g., "Objective: Is the current operation compliant?"). The engine first represents the objective as a fact to be verified (e.g., (current operation, compliance status, True)). The reasoning process is as follows: Target stack initialization: Push the main target onto the target stack; Rule retrieval: Search the rule base for all rules that can derive the top-of-stack objective (i.e., the conclusion matches the objective fact) from the RHS. If multiple rules are found, try them sequentially according to priority and rule confidence. Sub-goal generation: For a selected rule, all conditional clauses in its LHS are transformed into sub-goals to be verified. For example, if the rule LHS is "User role is project leader AND total funding ≤ 500,000 AND funding category is within budget", then three sub-goals are generated: "Verify if user role is project leader", "Verify if total funding is ≤ 500,000", and "Verify if funding category is within budget". These sub-goals are pushed onto the goal stack in sequence. Recursive Verification: A new sub-target is popped from the target stack, and rule retrieval and sub-target generation are recursively repeated. If a sub-target can be directly determined by existing facts in the working memory (successful match), the sub-target is marked as "verified" and backtracks to the previous level; if a sub-target cannot be determined by existing facts, and there is no RHS in the rule base that can derive the sub-target, the engine triggers an active query mechanism, calling the large model application platform (for conditions requiring semantic understanding, such as "whether the use of funding items complies with the definition of travel expenses in Article X, Paragraph Y of the Management Measures"), inputting the conditional text into the large language model (DeepSeek-R1) for qualitative judgment, and injecting the boolean value returned by the model as external facts into the working memory to continue reasoning. If all sub-targets are verified successfully, the main target is established, and the compliance and reasoning basis chain (rule chain backtracking path) is output; if any sub-target fails verification, the main target fails, and the non-compliance result and specific failure node are output.

[0044] In this embodiment, different large language models are used based on the complexity of the rule chain reasoning. Specifically, in the reverse chain reasoning, when the sub-target involves the judgment of unstructured text (such as the semantic interpretation of institutional clauses, and the judgment of business rationality), the rule engine dynamically selects the model based on the context length and logical complexity of the condition to be judged. For simple conditions (such as text comparison length < 200 characters, requiring judgment of "whether they belong to the same technical field"), the Qwen-Max model is used to balance latency and cost; for complex conditions (such as the need to compare multiple institutional clauses simultaneously, and the need to comprehensively consider the project's historical execution and industry practices), the DeepSeek-R1 model is used to obtain stronger reasoning capabilities. The structured judgment results (including confidence scores) returned by the model are parsed and injected into the working memory as new facts to drive the next round of rule chain reasoning, forming a hybrid reasoning mode of "rule logic derivation as the main method and large model semantic judgment as the auxiliary method". The entire reasoning process (including the number of all triggered rules, trigger time, and intermediate judgment results) is recorded in the operation audit log in JSON format for subsequent compliance review and rule base optimization traceability.

[0045] The large model application platform serves as the collaborative working platform for the core service cluster. In a preferred embodiment of this invention, the open-source Dify large model development platform is selected, and its configuration is as follows: RAG orchestration module: Configures multiple retrieval strategies, including pure vector retrieval suitable for semantic question answering scenarios, pure full-text retrieval suitable for precise keyword matching scenarios, and multi-way recall retrieval suitable for deduplication and novelty detection scenarios; The retrieval results and user questions or task instructions are dynamically concatenated with the context and then input into the large language model, and the concatenation template is pre-placed in the guidance template management module.

[0046] The guidance template management module pre-sets guidance templates for various scenarios, including plagiarism report generation templates (including similarity threshold settings, comparison dimension descriptions, and conclusion format requirements), novelty search report generation templates (including novelty assessment dimensions, literature citation formats, and conclusion structure requirements), review report generation templates (including a review dimension list, scoring criteria, and suggested format requirements), and review article generation templates (including abstract structure, body text framework, and reference format). Each template can be preset with prompt word formats adapted to different large language models, such as the DeepSeek series using a system prompt + user prompt format, and the GLM series using a role setting + task description format.

[0047] The workflow scheduling module is configured with a visual workflow orchestration interface, supporting drag-and-drop workflow design and pre-set standard workflow templates, including a deduplication workflow from document upload, parsing, OCR recognition, vectorization, retrieval, model inference to result formatting; a novelty search workflow from document upload, parsing, keyword expansion, multi-path recall, re-sorting to report generation; and a review workflow from data crawling, cleaning, summarizing, model generation to review export.

[0048] In this embodiment, the workflow scheduling module implements chain-like collaboration between multiple functional modules in the intelligent application layer based on a directed acyclic graph (DAG) orchestration engine. The core mechanism of this orchestration engine includes: Node definition and dependency declaration: Each intelligent function, such as intelligent plagiarism detection, intelligent novelty detection, and intelligent review report generation, is encapsulated as an atomic task node. Each node must declare its input parameter type (e.g., document ID, report ID) and output data structure (e.g., plagiarism detection result JSON, novelty detection result JSON) when registering.

[0049] Dependencies between nodes (e.g., the "duplicate check" and "novelty check" nodes can be executed in parallel, but the "generate review report" node must depend on the successful completion of both the "duplicate check" and "novelty check" nodes) are declared through YAML configuration files to form a complete DAG definition.

[0050] Triggering Conditions and Execution Strategy: The workflow is triggered using an event-driven model. When the "Intelligent Recognition and Upload of Deliverables" function completes document parsing and vectorization, and writes the document ID into a specific status field in the structured database, this event serves as the workflow's starting trigger. After capturing the event, the scheduler begins scheduling node execution according to the pre-defined "Project Application Full-Process Workflow" DAG. The scheduling strategy supports serial, parallel, and conditional branching. For nodes without dependencies (such as duplicate and novelty checks), the scheduler distributes them to different worker nodes for parallel execution to improve processing efficiency. For nodes with dependencies, the scheduler continuously monitors the execution status of its upstream nodes, and only pushes the downstream node task into the execution queue when all upstream nodes are in a "successful" state.

[0051] State Management and Fault Tolerance: The scheduler maintains a state machine for the global workflow instance, with states including: pending execution, executing, successful, failed, and waiting for retry. The execution results of each node (success / failure, output data, and exception logs) are persisted to the workflow instance table in a structured database. When any node fails, the scheduler automatically retryes according to a preset retry strategy (maximum 3 retries with exponential backoff intervals). If the retry still fails, the entire workflow instance is marked as "failed" and an alarm is triggered. Simultaneously, a manual intervention interface provides information about the failed node and its context, supporting continuation from the failed node to avoid reprocessing completed tasks.

[0052] Data Transfer and Context Sharing: The workflow scheduling module is deeply integrated with the shared semantic space. Metadata such as the report storage path and similarity score generated by upstream nodes (e.g., the plagiarism detection module) are temporarily stored in a Redis cache by the scheduler as part of the workflow context. When a downstream node (e.g., the review report generation module) is scheduled for execution, the scheduler automatically injects the required upstream output data from Redis as its input parameters, thereby achieving data transfer without directly calling the database, reducing system coupling and improving workflow efficiency.

[0053] The vector database can be a distributed vector database, configured to store 768-dimensional vectors of scientific research project data, scientific and technological achievements, and institutional documents. It adopts a hierarchical navigable small world graph (HNSW) index, which supports efficient similarity retrieval and vector distance calculation.

[0054] The full-text search engine is configured to store the full text of project materials, scientific and technological achievements, and institutional documents in its index, and supports keyword search, fuzzy search, Boolean logic search, and highlighting.

[0055] Structured Database: Utilizes a MySQL database to store structured data such as project metadata, user information, report records, rule base, and operation logs, supporting read / write separation and high-availability switching.

[0056] The vector database and the full-text search engine together constitute the system's shared semantic space. The vector database stores 768-dimensional semantic vectors of research documents, project materials, and scientific achievements, supporting semantic retrieval based on vector similarity. The full-text search engine builds an inverted index on the full text of documents, supporting keyword retrieval and Boolean logic retrieval. This shared semantic space provides a unified knowledge access interface for all modules in the core service cluster. Updates to the knowledge base by any module can be perceived and utilized by other modules in real time, thereby achieving cross-module consistency and semantic interoperability of knowledge representation.

[0057] The intelligent application layer provides research management personnel and researchers with intelligent functions for research management, including intelligent plagiarism detection, intelligent novelty search, intelligent generation of review reports, intelligent generation of scientific and technological innovation reviews, intelligent identification and uploading of research materials, intelligent customer service, intelligent workflow navigation, data analysis, and decision-making.

[0058] The infrastructure and security system includes system deployment architecture and security system, providing computing power support, network communication, containerized deployment and all-dimensional security protection for the system.

[0059] The system deployment architecture is built on a containerized deployment system based on the container orchestration and cluster scheduling (Kubernetes, K8s) platform. It adopts a cloud-native microservice architecture and is divided into a gateway layer, an application cluster layer, and a middleware layer. Please refer to the system deployment architecture section for details. Figure 2 As shown.

[0060] The gateway layer serves as a unified entry point for access within and outside the system. It connects to external resources through external service ports and to internal systems through data integration ports, achieving isolation and security control of internal and external network traffic. It can also be combined with SLB load balancing and NGINX clusters to complete traffic distribution and reverse proxy, integrate unified authentication and authorization capabilities, and connect to the enterprise's unified user system and unified portal to ensure access compliance and security. The application cluster layer, built on the Kubernetes platform, serves as a microservice runtime environment and the carrier for the system's core business and AI capabilities. The front-end machine cluster handles user traffic forwarded by the gateway layer, providing multi-terminal access points and unified request forwarding. The back-end machine cluster deploys basic business microservices such as process event services, scheduled task services, email services, and framework main services in a containerized manner, while also carrying the core AI capability hub services, covering intelligent business needs across all scenarios. The system achieves unified scheduling of all cluster resources through the Kubernetes platform's master control node, dynamically scaling service instances using HPA horizontal auto-scaling capabilities. Supporting monitoring services provide end-to-end performance monitoring, log analysis, and link tracing. NFS services provide shared storage support, and each microservice is managed uniformly through the Nacos service registration and configuration center, ensuring stable inter-service calls and flexible scaling.

[0061] The middleware layer provides underlying support for upper-layer business operations, including data storage, asynchronous processing, and performance acceleration. It adopts a structured data storage architecture to achieve high-availability storage and read / write separation of structured data. It undertakes asynchronous processing of message queues from the core service cluster, decoupling business processes and improving system throughput. At the same time, it connects to object storage to store unstructured files and vector databases to store text vectors. Combined with the large model application platform, it provides underlying support for RAG retrieval enhancement generation and large model call orchestration for core services. The overall architecture achieves full-link layered decoupling from gateway access and business operation to data storage, providing a solid guarantee for high-concurrency processing, high-reliability operation, and continuous evolution of the system.

[0062] The security system is designed according to the Level 2 Information Security Protection Standard, covering all aspects of system security protection, including network security, application security, data security, large model security, and interface security.

[0063] For network security, unified authentication, access rate limiting, and routing control are achieved through the API gateway, while traffic filtering and access control are implemented at the access layer, strictly adhering to the enterprise's internal and external network isolation strategy. For application security, based on unified identity authentication, functional permissions and data access scope are assigned according to roles to prevent unauthorized input and unauthorized operations. Regarding data security, for important data such as research project materials, results materials, funding information, institutional documents, and user information, we strictly adhere to the enterprise's data security and data classification management requirements. Data access is strictly controlled according to permissions. During the system's operation, the operation audit log records every intelligent question and answer, navigation request, data analysis, and export behavior, providing a basis for compliance checks and security incident tracing.

[0064] To ensure the security of large models, a model output verification mechanism and security constraints on prompt words are established to ensure that intelligent question answering, plagiarism detection, novelty checking, and review report generation are compliant, accurate, and free from misleading information, and to prevent malicious inducement and injection attacks.

[0065] For interface security, the system connects with the unified user system, unified portal, and collaborative information management platform through internal interfaces. All interfaces implement identity authentication, authorization verification, and access control to ensure that data is not stolen or tampered with during the data exchange process.

[0066] Please refer to the overall workflow of the system of this invention. Figure 3 This embodiment uses the project application process initiated by enterprise researchers as a typical scenario to illustrate in detail the working process of the system of the present invention to realize the functions of intelligent plagiarism detection, intelligent novelty search, intelligent generation of review reports, and intelligent identification and uploading of achievement materials, as follows: When a user uploads a feasibility study report (PDF or Word document format) through the front-end interface, the system triggers the following processing flow: Step 1: The file management and parsing service receives the target document, stores it in the object storage database, performs text extraction and format parsing on the target document, outputs structured text with coordinate information, calls the large language model to extract core keywords and technical terms from the research content and innovation points, cleans and segments the parsed text, converts it into semantic vectors through a domain-adaptive vector embedding model, and stores it in the vector database of the shared semantic space.

[0067] Step 2: Trigger the project review module to perform a plagiarism check on the target document. For the workflow, please refer to [link / reference needed]. Figure 4 As shown, it specifically includes: S21. Similarity calculation is performed on the target document. The project review module calls the three-dimensional hybrid similarity calculation engine to compare the semantic vector, text fingerprint, and structural features of the user-uploaded document with the existing project data in the collaborative information management platform. The comprehensive similarity score is calculated through adaptive weighted fusion. The specific calculation method is as follows: S211, Semantic vector similarity calculation: Using the domain-adaptive vector embedding model, the user-uploaded target document Dt and the source document Ds in the database are converted into 768-dimensional semantic vectors Vt and Vs, respectively. Then, the similarity score Sim_semantic between the two is calculated using the cosine similarity formula:

[0068] The value range is [0,1]. The larger the value, the more similar the semantics. It is used to capture deep semantic similarities such as synonym rewriting and word order transformation. S212, Text Fingerprint Similarity Calculation: The SimHash algorithm is used to calculate the text fingerprint of a document. The specific process is as follows: First, the document is segmented into words, and a traditional hash function is applied to each word to generate a 64-bit hash value. Then, the hash values ​​of all words are weighted and summed according to their weights (e.g., TF-IDF values) (adding weight if the hash bit is 1, subtracting weight if it's 0), resulting in a 64-bit weighted sum vector. Finally, each bit of this vector is converted to 1 (>0) or 0 (≤0) to generate the final 64-bit text fingerprint. For the target document Dt and the source document Ds, the Hamming distance H(Dt, Ds) between their fingerprints is calculated. The formula for calculating the text fingerprint similarity score Sim_fingerprint is:

[0069] This score is used to quickly identify highly similar content at the literal level, such as direct text copying and partial rewriting.

[0070] S213, Structural Feature Similarity Calculation: Extracting the chapter structure features of the target and source documents. Specifically, parsing the chapter titles (e.g., "1.", "1.1", "Chapter 1") and their hierarchical relationships to construct a structure tree T. For each structure tree, extracting its path set P, where each path represents a hierarchical sequence from the root node to a leaf chapter. The structural similarity Sim_structure between two documents is measured by calculating the intersection-union ratio of their path sets, using the formula:

[0071] This score is used to identify structural plagiarism that involves using the same or similar chapter frameworks.

[0072] S214, Adaptive Weighted Fusion: The calculated 3D similarity scores are weighted and fused to obtain the final comprehensive similarity score, Sim_total. The fusion formula is:

[0073] Wherein, α, β, and γ are adaptive weights, satisfying α+β+γ=1. In this embodiment, considering that scientific research plagiarism detection has the highest requirements for semantic understanding, the default weights are set as semantic weight α=0.5, text fingerprint weight β=0.3, and structural feature weight γ=0.2.

[0074] It should be noted that the weights in this embodiment are not fixed, but are dynamically adjusted based on the characteristics of the target document itself. The specific strategy is as follows: S2141, Adjusting γ based on document structure richness: The system first analyzes the chapter level depth and paragraph structure completeness of the target document, and calculates the structure richness index ρ (value range 0~1, derived by normalizing the number of chapter titles, level depth, and standardized paragraphs in the document). When ρ≥0.6, it indicates that the document structure features are significant and have strong discriminative power, and γ is increased (adjustment coefficient is ρ×0.15); when ρ<0.3, it indicates that the document is a typical continuous narrative text with weak structural feature discriminative power, and γ is decreased to 0.05, and the weights of α and β are increased accordingly to keep the sum at 1.

[0075] S2142, Adjusting β based on document length: Calculate the text length factor λ of the target document (normalized in thousands of characters). When the document length is short (λ < 0.5 thousand characters), the Hamming distance discriminative power of the SimHash fingerprint decreases due to feature sparsity, and the system automatically multiplies β by a decay coefficient of 0.6; when the document length is long (λ > 3 thousand characters), the fingerprint stability is enhanced, and β remains at its default value or is increased to 0.35.

[0076] S2143, Weight Normalization Output: The final weight value is calculated by normalization using the above adaptive adjustment coefficients combined with the default baseline weight of the scene.

[0077] This dynamic adjustment strategy ensures that the overall similarity score maintains a reasonable dimensional contribution ratio across different document types and plagiarism detection scenarios.

[0078] S22. For documents with a similarity exceeding the set threshold of 0.7, the system extracts similar segments, labels the sources of similarity, calculates the similarity distribution, and generates a structured plagiarism report. The report includes an overall similarity score, segmental similarity details, a list of similar sources, and modification suggestions.

[0079] S23. The plagiarism report is stored in a structured database and simultaneously written back to the corresponding project's database on the collaborative information management platform for project managers to access.

[0080] Step 3: Perform a novelty search analysis on the target document. For the workflow, please refer to [link / reference]. Figure 5 As shown, it specifically includes: S31, invoke the multi-path recall engine. This engine executes the following three retrieval strategies in parallel to achieve a high recall rate, as follows: S311, Vector retrieval: Based on the vector database, perform an approximate nearest neighbor (ANN) retrieval on the semantic vector Vt of the target document. Using the HNSW index, quickly recall the Top-N documents (N=50 in this embodiment) most semantically relevant to the target document from the collaborative information management platform project repository and indexed external academic databases. This recall method focuses on capturing documents with related content but different expressions. S312, Full-text Search: The full-text search engine builds an inverted index on the full text of the feasibility study report uploaded by the user and the core keywords extracted from it. Boolean queries (such as research content AND innovation points NOT existing achievements) are executed to recall documents that contain the same or similar terms in specific fields (such as title, abstract, and keywords). The recall quantity for this approach is also set to Top-N (N=50 in this embodiment) to ensure high-precision terminology matching. S313, Keyword-based Expanded Retrieval: First, a large language model (such as DeepSeek-V3) is invoked to semantically understand the research content and innovative points input by the user, generating synonyms, hyponyms, and domain-specific terms. Simultaneously, a pre-built dictionary of scientific research management terminology is used for supplementation. Then, the expanded keyword set is used to perform a supplementary search in the full-text search engine within the shared semantic space, recalling the Top-M documents (M=30 in this embodiment). This recall approach is used to overcome the semantic gap problem caused by relying solely on original keywords.

[0081] S32, the search results from the three recall methods are merged and deduplicated based on the unique identifier of each document (such as DOI or internal ID). The deduplicated candidate document set is then input into a re-ranking model based on a cross-encoder. This model concatenates the target document with each candidate document into an input sequence, performs joint encoding through a deep Transformer network, and directly outputs a relevance score within the range [0,1]. Compared to the dual encoder (Bi-Encoder) used for vector retrieval, the cross-encoder model has higher accuracy and can capture more granular interaction information. The system sorts all candidate documents in descending order based on this score, and finally selects the top K (K=20 in this embodiment) most relevant documents as the novelty search results.

[0082] S33, the system performs multi-dimensional comparisons between the research content of the target document uploaded by the user and the most relevant literature, including dimensions such as similarity of technical solutions, difference of innovation points, comparability of research methods, and overlap of expected results. It generates comparison analysis results and calls a large language model to generate a structured novelty search report. The report includes technical novelty assessment, comparison of domestic and foreign research progress, overlap index of similar projects, uniqueness analysis of innovation points, and novelty search conclusions.

[0083] S34. The novelty search report is stored in a structured database and simultaneously written back to the collaborative information management platform for reference by project review experts.

[0084] Step 4: Generate a review report based on the plagiarism and novelty search reports, which specifically includes: S41, the RAG orchestration module retrieves contextual information related to the project, such as policy documents, industry standards, and historical review comments, from the vector database and structured database of the shared semantic space.

[0085] S42 guides the template management module to load the review report and generate a template. The template defines the review dimensions (compliance with industry development plans, significance and necessity of project research, project research content, feasibility of implementation plan, output, budget, and guarantee conditions) and the scoring criteria and content requirements for each dimension.

[0086] S43, combine the user report, deduplication and novelty search report, retrieved context information and guidance template into a complete prompt word, input it into the large language model (DeepSeek-V3 or GLM-4 is preferred in this embodiment), and the model generates a draft review report.

[0087] S44. The system performs structured validation on the generated results, checking the completeness of chapters, data consistency, and accuracy of citations. Results that fail the validation are regenerated.

[0088] S44: The approved review report is output in a formatted manner, and supports online preview and export to Word format.

[0089] In the daily management scenarios of enterprise scientific research managers, this invention system also includes the implementation of functions such as intelligent generation of scientific and technological innovation reviews, intelligent customer service, intelligent workflow navigation, and data analysis and decision support. Please refer to [link / reference]. Figure 2 The specific steps are as follows: Step 5: The system configures the scheduled task to trigger the intelligent summary module, specifically as follows: S51, the intelligent review module initiates queries to multiple designated research channels in parallel according to preset query conditions. The research channels include, but are not limited to: CNKI, Wanfang Data, and the results database of the enterprise collaborative information management platform.

[0090] The system is configured by default to automatically execute tasks every Monday at 2:00 AM. The default query criteria are publications from the past week and the main areas the company is involved in. Users can also define their own time and preset query criteria.

[0091] S52: The captured data is deduplicated, formatted, and missing values ​​are handled before being categorized and summarized by technical field.

[0092] S53, call the DeepSeek-V3 or GLM-4 model, load the review generation template, and generate a scientific and technological innovation review document containing research background, technological progress, hot spot analysis, trend prediction, and references. During this process, the RAG orchestration module retrieves relevant field background knowledge from the shared semantic space as context supplement.

[0093] S54, the system will export the generated science and technology innovation review document as a Word document and send it to the designated research management personnel's email address on a scheduled basis according to the preset configuration, while storing it in the object storage database for future reference.

[0094] Step 6: The user inputs a natural language question through the chat window on the front-end interface. The system triggers the intelligent service module, which calls the large language model to identify the user's intent and determine the question type. The question types include, but are not limited to: process consultation, policy inquiry, operation guidance, and professional knowledge Q&A. Simultaneously, the project compliance module is triggered for monitoring, as follows: S61, the project compliance module loads an executable rule base converted from corporate policy documents from a structured database. The system obtains the environmental parameters of the current operation (user role, operation type, operation object, operation time, etc.). The rule engine performs rule chain reasoning based on the environmental parameters to determine the compliance of the current operation.

[0095] S611, for compliant operations, the system generates operation suggestions and next steps based on a large language model and pushes them to the user in the form of a non-intrusive floating window; S612: For non-compliant operations, the system triggers an early warning mechanism, pops up a window to inform the user of the specific reason for the violation, and provides suggestions for compliant correction. S613: After the compliance determination results and the user's subsequent corrective actions are reviewed and confirmed, they are fed back to the rule base as incremental data, realizing the semi-automatic continuous updating of the rule base.

[0096] S62, In this embodiment, the RAG orchestration module adopts a vector retrieval hybrid full-text retrieval strategy to balance recall and precision. Based on the identified intent type, it retrieves the most relevant institutional documents, project materials or policy clauses from the vector database.

[0097] S63, the template management module loads the question and answer template, and the RAG orchestration module dynamically concatenates the user's original question, the retrieved context information and the question and answer template into complete prompt words, which are then input into the large language model (Qwen-Max is preferred in this embodiment). The model generates a natural language answer that conforms to the enterprise's specifications.

[0098] S64: The answer is streamed to the front-end interface via a WebSocket channel. Upon receiving the first token from the large language model, the system immediately pushes that token to the front-end interface via WebSocket. Subsequent tokens generated by the model are pushed in real-time until the complete answer is sent. The first-character delay is controlled within 500ms, improving the user experience.

[0099] In S65, after each question-and-answer session, a satisfaction rating control pops up in the chat interface, allowing users to rate the quality of the response from 1 to 5 stars. The rating data, along with the question-and-answer records, is stored in a structured database as a closed-loop feedback input for subsequent model fine-tuning and template optimization.

[0100] Step 7: The user inputs a natural language command (e.g., "Analyze the distribution of project funding in various technical fields over the past three years") through the front-end interface. The intelligent service module calls a large language model (DeepSeek-V3 is preferred in this embodiment) to perform semantic parsing of the command and output a structured data analysis task. This task should include at least the following fields: target data source (e.g., project information table, funding details table), analysis dimensions (e.g., time dimension "past three years", classification dimension "technical fields"), statistical methods (e.g., summation, mean, percentage), and expected output format (e.g., bar chart, pie chart). Details are as follows: S71, based on the data analysis task requirements, automatically generate structured query statements (SQL) to retrieve the required data from the relevant tables in the structured database (MySQL).

[0101] S72, the query results are input into the visualization engine in the format of a dataset (ECharts is used in this embodiment), and the corresponding visualization charts (such as bar charts, pie charts, line charts, heatmaps, etc.) are generated according to the preset output format.

[0102] S73 uses the data analysis results as context to call the large language model again, guiding it to generate a structured scientific research management decision analysis report and optimization suggestions. The report format is predefined by the "Decision Analysis Template" in the guidance template management module, and the content may include: data overview, anomaly indicator identification, cause attribution analysis, and improvement suggestions.

[0103] In S74, the system integrates visual charts and diagnostic reports into a unified front-end analytics dashboard. Users can export charts (PNG format) and diagnostic reports (Word format) with a single click for reporting or archiving.

[0104] The above describes the workflow steps for intelligent generation of scientific and technological innovation summaries, intelligent customer service, intelligent workflow navigation, and data analysis and decision-making functions in a daily work scenario. This section complements the aforementioned project application scenario examples, together constituting a complete implementation plan of the present invention covering the entire scientific research management process.

[0105] The above-described embodiments demonstrate the intelligent support capabilities of the system of the present invention, covering the entire scientific research management process from the project application front-end to the management decision-making back-end. Through the layered collaboration of the multi-source data integration layer, the large-model intelligent agent core layer, and the intelligent application layer, as well as the reliable guarantee of infrastructure and security system, efficient, intelligent, and closed-loop scientific research management is achieved.

[0106] The system of this invention not only provides intelligent support for the entire scientific research management process in terms of functionality, but also builds a comprehensive security system covering identity authentication, access control, data encryption, communication protection, input / output filtering, model security protection, and audit traceability, which can meet the strict security and compliance requirements of enterprise-level applications.

[0107] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A large-model-based intelligent agent system for scientific research management, comprising a multi-source data integration layer, a large-model intelligent agent core layer, and an intelligent application layer, characterized in that, The core layer of the large model intelligent agent includes file management and parsing services, a core service cluster, and a large model application platform, realizing full-process functions of scientific research management. The file management and parsing service parses and vectorizes the scientific research documents uploaded by users, and stores the generated semantic vectors in the shared semantic space. The core service cluster includes an intelligent service module, a project review module, an intelligent summary module, and a project compliance module. When performing tasks such as plagiarism checking, novelty checking, review report generation, or summary generation, each module uniformly calls the large model application platform. The large model application platform includes a retrieval enhancement generation and orchestration module, a guided template management module, and a workflow scheduling module, and maintains a shared semantic space composed of a vector database and a full-text search engine; The retrieval enhancement generation orchestration module is configured with multiple retrieval strategies, the guidance template management module is pre-set with guidance templates for multiple scenarios, and the workflow scheduling module is used to orchestrate the task execution order of multiple business modules in the core service cluster. The user interaction data and feedback data generated by each function in the intelligent application layer are written back to the multi-source data integration layer through the feedback channel. After processing, they form an incremental training dataset, which is used to continuously fine-tune and optimize the large language model and guidance template in the large model application platform, forming a closed-loop feedback. The intelligent application layer provides intelligent agent functions including: intelligent plagiarism detection, intelligent novelty search, intelligent generation of review reports, intelligent generation of scientific and technological innovation reviews, intelligent identification of uploaded achievement materials, intelligent customer service, intelligent workflow navigation, and data analysis and decision-making functions. The intelligent plagiarism detection function relies on the scientific and technological project data in the shared semantic space. Based on the feasibility study report uploaded by the user, it calls the three-dimensional hybrid similarity calculation engine of the project review module. The three-dimensional hybrid similarity calculation engine uses an adaptive weighted fusion method of semantic vector similarity, text fingerprint similarity and structural feature similarity to calculate a comprehensive similarity score and output a plagiarism detection report. The intelligent novelty search function is based on the shared semantic space and legally authorized public academic databases. It calls the multi-path recall engine of the project review module. The multi-path recall engine performs vector retrieval, full-text retrieval and keyword expansion retrieval in parallel. After deduplication and re-sorting based on cross encoder, it outputs relevant literature and automatically generates a novelty search report. The intelligent review report generation function is based on the user-submitted feasibility study report and the results of plagiarism and novelty searches. It retrieves relevant policy documents and historical review opinions through the enhanced retrieval and arrangement module, and loads the review report generation template through the guide template management module. It then automatically generates a structured review report using a large language model. The intelligent science and technology innovation summary generation function queries data from designated channels through a scheduled task scheduler. After data cleaning and classification, it calls the summary generation template, generates a science and technology innovation summary document using a large language model, and supports export and scheduled sending. The intelligent customer service function provides a natural language question-and-answer interaction service. It retrieves policy documents and project information from the shared semantic space through the retrieval enhancement generation and arrangement module, and generates answers by combining them with guidance templates. The intelligent workflow navigation function transforms enterprise policy documents into executable rules and incorporates them into the rule base. It performs rule chain reasoning through the rule engine of the project compliance module, generates operation suggestions based on a large language model, judges the compliance of operations and triggers an early warning mechanism, and feeds the judgment results back to the rule base for continuous updates.

2. The scientific research management intelligent agent system based on a large model according to claim 1, characterized in that, The multi-source data integration layer synchronizes and unifies organizational data and user account data of the user system by building a data integration interface, integrates and connects to the unified portal through single sign-on, and connects to the collaborative information management platform through the data integration interface.

3. The scientific research management intelligent agent system based on a large model according to claim 1, characterized in that, The file management and parsing service parses and performs OCR recognition on documents. After cleaning and segmentation, it forms a standard text dataset, which is then converted into semantic vectors through a domain-adaptive vector embedding model and stored in the vector database. At the same time, the standard text dataset is synchronously indexed to the full-text search engine. The various retrieval strategies include pure vector retrieval, pure full-text retrieval, and multi-path recall retrieval strategies. The guidance templates for various scenarios include guidance templates for plagiarism checking, novelty checking, review reports, reviews, and question-and-answer scenarios. The workflow scheduling module has pre-set standard workflows for plagiarism checking, novelty checking, and review generation.

4. The scientific research management intelligent agent system based on a large model according to claim 1, characterized in that, The core service cluster decouples asynchronous tasks through message queues, making document parsing, vectorization processing, and large model inference operations asynchronous, and building multi-level caches.

5. The scientific research management intelligent agent system based on a large model according to claim 1, characterized in that, The intelligent recognition function for uploaded results materials involves the file management and parsing service parsing and OCR recognizing the uploaded results attachments. The large model application platform calls the large language model to extract key information and automatically fills it into the results entry page of the user's collaborative information management platform through the workflow scheduling module.

6. The scientific research management intelligent agent system based on a large model according to claim 1, characterized in that, The system also includes an infrastructure and security system, which comprises a system deployment architecture and a security framework. The system deployment architecture is based on a cloud-native microservice architecture and is divided into a gateway layer, an application cluster layer, and a middleware layer. The gateway layer serves as a unified entry point and integrates unified authentication and authorization. The application cluster layer carries core services. The middleware layer includes a vector database, a full-text search engine, a structured database, and a message queue. The security framework includes network security, application security, data security, large model security, and interface security.

Images