An AI-driven knowledge tree adaptive learning system and method

Abstract

The application relates to the technical field of artificial intelligence and online education, and discloses an AI-driven knowledge tree adaptive learning system and method. The system adopts a five-layer hierarchical architecture, and the core is provided with a multi-LLM Provider abstract layer, a knowledge tree intelligent generation module, an RAG enhanced knowledge question and answer module, an adaptive question generation and mastery evaluation module, and an integral account and integrity verification module; the knowledge tree is automatically generated through a large language model and an SSE streaming technology, private learning materials are fused in combination with RAG technology, and five-level mastery evaluation and adaptive question generation are realized based on a multi-dimensional weighted model. The application greatly improves the knowledge organization efficiency and the learning individualization level, effectively suppresses the large model illusion, guarantees the high availability of AI services, and can be widely applied to online education and knowledge management scenes.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and online education technology, specifically to an AI-driven knowledge tree adaptive learning system and method. Background Technology

[0002] With the rapid development of online education and knowledge management, various knowledge management and learning support systems have emerged in existing technologies. Traditional online learning platforms (such as MOOC platforms and knowledge base systems) mainly adopt static knowledge organization methods, requiring users to manually construct knowledge structures, and lack intelligent automatic knowledge graph generation capabilities.

[0003] Existing technologies suffer from the following problems: Low knowledge organization efficiency: Users need to manually create and maintain knowledge structures, and it is impossible to automatically generate structured knowledge graphs based on learning topics, resulting in time-consuming knowledge organization and an unsystematic structure. Insufficient learning personalization: Existing systems mostly rely on simple rule matching for question generation and learning recommendations, lacking the ability to adaptively generate questions based on learner profiles. Disconnect between AI question answering and knowledge base: Most existing AI question answering systems are based on general large language models, lacking deep integration with users' private learning materials. Limited learning effectiveness evaluation: Existing systems mostly use simple accuracy statistics, lacking a comprehensive mastery evaluation model with multi-dimensional indicators. Summary of the Invention

[0004] The purpose of this invention is to provide an AI-driven knowledge tree adaptive learning system and method to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an AI-driven knowledge tree adaptive learning system, comprising a front-end presentation layer, a back-end API layer, a business service layer, an AI service layer, and a data storage layer, characterized in that:

[0006] The AI ​​service layer includes a multi-LLM Provider abstraction layer, which integrates a model router, a weighted round-robin load balancing unit, a circuit breaker state machine, and a unified client interface. The model router maintains an LLM service provider registry based on a hot-load configuration file and routes service requests according to model capability tags. The weighted round-robin load balancing unit distributes requests among multiple service providers of the same model. The circuit breaker state machine counts consecutive failed requests from a single service provider and triggers circuit breaking and failover when the number of failures exceeds a threshold. The unified client interface masks the API differences between different LLM service providers.

[0007] The business service layer includes a knowledge tree intelligent generation module, a RAG enhanced knowledge question answering module, an adaptive question generation and mastery assessment module, and an points ledger and integrity verification module.

[0008] The knowledge tree intelligent generation module is used to receive the learning topic and supporting parameters input by the user, construct a structured generation prompt word template, call the multi-LLMProvider abstraction layer to match the optimal LLM service, push the knowledge node data with hierarchical relationship generated by the model to the front-end display layer for rendering in real time through the SSE streaming protocol, and persistently store the knowledge nodes and hierarchical relationships in the data storage layer.

[0009] The RAG enhanced knowledge question answering module is used to preprocess and construct vectorized indexes for private learning materials uploaded by users, complete semantic retrieval matching based on users' question-and-answer or question-generating requests, construct enhanced prompt words by combining retrieval results, knowledge node content and dialogue history, call LLM service and generate corresponding results through SSE streaming transmission;

[0010] The adaptive question generation and mastery assessment module is used to build a learner profile based on the user's answer records, generate questions that match the user's mastery level by combining the learner profile with the target knowledge node content, and calculate the user's mastery score based on a preset multi-dimensional weighted formula, which is mapped to a five-level mastery level.

[0011] The points ledger and integrity verification module adopts a hybrid billing mode that combines fixed deductions with dynamic token billing, and performs integrity verification on each billing transaction through HMAC-SHA256 chain signature.

[0012] Furthermore, the specific steps of the knowledge tree intelligent generation module are as follows:

[0013] Step D1: Receive the user's input of the learning topic and optional topic description and learning scope parameters;

[0014] Step D2: Construct prompt word templates, encapsulate the learning topics into structured generation instructions that can be understood by the large language model, and require the model to output knowledge node data containing hierarchical relationships in JSON format;

[0015] Step D3: Invoke the model router of the multiple LLM Provider abstraction layer to select the best available large language model service for inference;

[0016] Step D4: Using the SSE streaming protocol, the knowledge node data generated by the model is pushed to the front end in real time. The front end receives and renders the knowledge tree graph in real time through the asynchronous generator.

[0017] Step D5: Persistently store the generated knowledge nodes and their parent-child hierarchical relationships in the database.

[0018] Furthermore, the model router in the multi-LLM Provider abstraction layer maintains a registry of multiple LLMProviders based on a YAML hot-reload configuration file and routes requests according to model capability tags; the weighted round-robin load balancing unit uses a weighted round-robin algorithm to distribute requests among multiple providers of the same model; the circuit breaker state machine maintains a failure counter and a circuit breaker state machine for each provider, automatically circuit breaking and transferring requests when consecutive failures exceed a threshold; the unified client interface provides three types of client interfaces: Chat, Embedding, and Vision, shielding the API differences between different providers.

[0019] Furthermore, the specific steps of the RAG enhanced knowledge question answering module are as follows:

[0020] Step R1: Data preprocessing - PDF text is extracted using pypdf, images are subjected to content recognition using a multimodal visual model, and documents are preprocessed using Chinese word segmentation;

[0021] Step R2: Vectorized Indexing – The text is segmented into blocks, and the Embedding model is called to generate vector embeddings to build a semantic vector index;

[0022] Step R3: Semantic retrieval—Encode the query into a vector and use cosine similarity to retrieve Top-K relevant paragraphs;

[0023] Step R4: Contextual Enhancement – ​​Combine search results, node content, and dialogue history to construct enhanced prompts;

[0024] Step R5: Generate answers by streaming large language models via SSE.

[0025] Furthermore, the RAG-enhanced knowledge-based question-answering module, in the data preprocessing step, for learning materials in image format, calls a multimodal visual model to extract text, formulas, and chart content from the images and converts them into structured text data; after completing the vectorized index construction, it supports incremental text segmentation and vector embedding for user-added learning materials, and synchronously updates the semantic vector index; at the same time, in the adaptive question generation scenario, it synchronously executes semantic retrieval and context enhancement steps to generate questions that fit the user's learning scope based on the content of the private learning materials.

[0026] Furthermore, the multi-dimensional weighted formula for the adaptive question generation and mastery assessment module is as follows: ,in: The preset weights are, in order, the accuracy rate, the number of attempts, the difficulty rate, and the time rate, and satisfy the following conditions: The levels are mapped to five levels: not learned, known, understood, mastered, and proficient.

[0027] Furthermore, the learner profile constructed by the adaptive question generation and mastery assessment module includes the user's historical mastery level for each knowledge node, answer accuracy, answer frequency, average answer time, and question difficulty preference data. The adaptive question generation matches questions of corresponding difficulty based on the current mastery level of the target knowledge node. Specifically, the "unlearned" or "understanding" level corresponds to basic concept questions, the "understanding" or "mastering" level corresponds to advanced application questions, and the "proficient" level corresponds to comprehensive extension questions. At the same time, the calculated mastery level is bound to the corresponding node in the knowledge tree and updated in real time in the front-end knowledge tree graph through different color indicators.

[0028] Furthermore, the points ledger and integrity verification module are configured with corresponding fixed deduction amounts and dynamic token billing rules for different function call scenarios such as knowledge tree generation, AI question answering, and adaptive question generation. Each billing transaction generates a corresponding transaction record. The HMAC-SHA256 algorithm is used to jointly calculate the transaction record, the hash signature of the previous transaction, and the user's unique identifier to generate a chain signature for the current transaction. The transaction record and the chain signature are stored synchronously, and the integrity of historical transaction data is verified by the chain signature before each transaction.

[0029] Furthermore, the system adopts a B / S architecture with a front-end and back-end separation design, and is divided into a five-layer architecture. The front-end presentation layer uses ReactFlow to implement the visual rendering of the knowledge tree graph, supporting users to drag and drop to edit the generated knowledge nodes, modify content, adjust hierarchical relationships, and bind node data. The edited node data is synchronously and persistently stored in the data storage layer. The back-end API layer is configured with JWT authentication, SSE streaming interface, RESTful API routing and request rate limiting unit, providing a data interaction channel between the front-end and the business service layer.

[0030] An AI-driven knowledge tree adaptive learning method, applied in an AI-driven knowledge tree adaptive learning system, includes the following steps:

[0031] S1. Build a multi-LLMProvider abstract scheduling framework, maintain the registry of multiple LLM service providers through hot-loading configuration files, complete the routing matching of service requests based on model capability tags, use a weighted round-robin algorithm to distribute requests among multiple service providers of the same model, count the consecutive failed requests of a single service provider, and trigger circuit breaking and failover when the number of failures exceeds the threshold. At the same time, shield the API differences of different LLM service providers through a unified interface.

[0032] S2. Receive the learning topic and supporting parameters input by the user, construct a structured prompt word template, match the optimal LLM service through the multi-LLMProvider abstract scheduling framework built in step S1, and push the hierarchical knowledge node data generated by the model to the front-end rendering in real time using the SSE streaming protocol, and persist the knowledge nodes and hierarchical relationships to complete the automatic generation of the knowledge tree.

[0033] S3. Perform text extraction, content recognition, and word segmentation preprocessing on user-uploaded private learning materials. After segmenting and cutting the preprocessed text into blocks, generate vector embeddings and build a semantic vector index library. When a user's question-and-answer or question-generating request is received, the request content is encoded into a vector. Top-K relevant paragraphs are matched through cosine similarity retrieval. Enhanced prompt words are constructed by combining the search results, corresponding knowledge node content, and dialogue history. The LLM service is called and the corresponding results are generated through SSE streaming.

[0034] S4. Continuously collect users' answer records, build and dynamically update learner profiles, and generate questions that match the user's current mastery level based on the learner profile and the target knowledge node content. After the user completes the answer, calculate the mastery score through a preset multi-dimensional weighted formula, map the score to a five-level mastery level, and update the answer results and mastery level to the learner profile.

[0035] S5 adopts a hybrid billing model that combines fixed deductions with dynamic token billing to process billing for users' system function calls, and verifies the integrity of each billing transaction through HMAC-SHA256 chain signature.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] By automatically generating multi-level structured knowledge trees through a large language model, users do not need to manually build and maintain the knowledge structure. Combined with SSE streaming technology, knowledge nodes are rendered in real time and progressively. Compared with the traditional manual construction and whole package request-response mode, the efficiency of knowledge organization is improved. Users can view the generation progress in real time, which completely solves the pain points of low efficiency and unsystematic structure of existing technologies.

[0038] By leveraging RAG technology to deeply integrate users' private learning materials with AI question-answering and question-generating systems, and combining multimodal data recognition, incremental vectorized indexing, and cosine similarity semantic retrieval, the existing problems of AI question answering being disconnected from users' private knowledge bases and generalized but untargeted answers are solved.

[0039] A multi-dimensional, five-level mastery assessment model is constructed to overcome the limitations of traditional single accuracy statistics. It combines multiple dimensions such as answer accuracy, frequency, question difficulty, and answer time to accurately assess the user's mastery. Based on learner profiles, adaptive question generation is achieved to dynamically match the user's learning progress, solving the shortcomings of existing systems in terms of insufficient learning personalization and single assessment dimensions.

[0040] By employing model routing across multiple LLM Provider abstraction layers, weighted round-robin load balancing, and circuit breaker mechanisms, automatic failover for multiple service providers is achieved, preventing system unavailability caused by the failure of a single service provider. Simultaneously, HMAC-SHA256 chain signatures are used to implement end-to-end integrity verification of points transaction data, effectively preventing data tampering and ensuring the security of billing data. Attached Figure Description

[0041] Figure 1 This is a diagram of the overall system architecture of the present invention;

[0042] Figure 2 This is a flowchart of the intelligent knowledge tree generation process of the present invention;

[0043] Figure 3 This is a diagram of the multi-LLM Provider abstraction layer architecture of the present invention;

[0044] Figure 4 This is a flowchart of the RAG-enhanced knowledge question-and-answer process of the present invention;

[0045] Figure 5 This is a flowchart of the adaptive question generation and mastery assessment process of the present invention;

[0046] Figure 6 This is the homepage of the system interface for this invention.

[0047] Figure 7 This is the user login interface for the system operation of the present invention;

[0048] Figure 8 This is the user registration interface for the system operation of the present invention;

[0049] Figure 9 This is the main interface for the system operation of the present invention;

[0050] Figure 10 This invention provides a knowledge tree list and sidebar for the system operation.

[0051] Figure 11 The knowledge graph creation dialog box for the system operation of this invention;

[0052] Figure 12 This refers to the AI-powered intelligent dialogue interface for the system operation of this invention.

[0053] Figure 13 This is the community square page where the system of this invention operates;

[0054] Figure 14 This is the community post details page for the system operation of this invention;

[0055] Figure 15 This is the personal center page for the system operation of this invention. Detailed Implementation

[0056] Please see Figure 1 —15, This invention provides a technical solution: an AI-driven knowledge tree adaptive learning system, comprising a front-end presentation layer, a back-end API layer, a business service layer, an AI service layer, and a data storage layer, characterized in that:

[0057] The AI ​​service layer includes a multi-LLM Provider abstraction layer. This layer incorporates a model router, a weighted round-robin load balancer, a circuit breaker state machine, and a unified client interface. The model router maintains an LLM service provider registry based on hot-loaded configuration files and routes service requests according to model capability tags. The weighted round-robin load balancer distributes requests among multiple service providers of the same model. The circuit breaker state machine counts consecutive failed requests from a single service provider and triggers circuit breaking and failover when the number of failures exceeds a threshold. The unified client interface masks the API differences between different LLM service providers.

[0058] The business service layer includes a knowledge tree intelligent generation module, a RAG enhanced knowledge question answering module, an adaptive question generation and mastery assessment module, and an points ledger and integrity verification module;

[0059] The knowledge tree intelligent generation module is used to receive the learning topic and supporting parameters input by the user, build a structured generation prompt word template, call multiple LLMProvider abstraction layers to match the optimal LLM service, push the knowledge node data with hierarchical relationship generated by the model to the front-end display layer for rendering in real time through the SSE streaming protocol, and persistently store the knowledge nodes and hierarchical relationships in the data storage layer.

[0060] The RAG-enhanced knowledge question-answering module is used to preprocess and construct vectorized indexes of private learning materials uploaded by users, complete semantic retrieval matching based on users' question-answering or question-generating requests, construct enhanced prompt words by combining retrieval results, knowledge node content and dialogue history, call LLM service and generate corresponding results through SSE streaming transmission;

[0061] The adaptive question generation and mastery assessment module is used to build learner profiles based on user answer records, combine learner profiles with target knowledge node content to generate questions that match the user's mastery level, and calculate the user's mastery score based on a preset multi-dimensional weighted formula, which is mapped to a five-level mastery level.

[0062] The points ledger and integrity verification module adopts a hybrid billing model that combines fixed deductions with dynamic token billing, and performs integrity verification on each billing transaction through HMAC-SHA256 chain signature.

[0063] The specific steps of the knowledge tree intelligent generation module are as follows:

[0064] Step D1: Receive the user's input of the learning topic and optional topic description and learning scope parameters;

[0065] Step D2: Construct prompt word templates, encapsulate the learning topics into structured generation instructions that can be understood by the large language model, and require the model to output knowledge node data containing hierarchical relationships in JSON format;

[0066] Step D3: Invoke the model router of the multiple LLM Provider abstraction layer to select the best available large language model service for inference;

[0067] Step D4: Using the SSE streaming protocol, the knowledge node data generated by the model is pushed to the front end in real time. The front end receives and renders the knowledge tree graph in real time through the asynchronous generator.

[0068] Step D5: Persistently store the generated knowledge nodes and their parent-child hierarchical relationships in the database.

[0069] The model router in the multi-LLM Provider abstraction layer is based on a YAML hot-reload configuration file, maintains a registry of multiple LLMProviders, and routes requests according to model capability tags; the weighted round-robin load balancing unit uses a weighted round-robin algorithm to distribute requests among multiple providers of the same model; the circuit breaker state machine maintains a failure counter and a circuit breaker state machine for each provider, automatically circuit breaking and transferring requests when consecutive failures exceed a threshold; the unified client interface provides three types of client interfaces: Chat, Embedding, and Vision, shielding the API differences between different providers.

[0070] The specific steps for the RAG-enhanced knowledge-based question-and-answer module are as follows:

[0071] Step R1: Data preprocessing - PDF text is extracted using pypdf, images are subjected to content recognition using a multimodal visual model, and documents are preprocessed using Chinese word segmentation;

[0072] Step R2: Vectorized Indexing – The text is segmented into blocks, and the Embedding model is called to generate vector embeddings to build a semantic vector index;

[0073] Step R3: Semantic retrieval—Encode the query into a vector and use cosine similarity to retrieve Top-K relevant paragraphs;

[0074] Step R4: Contextual Enhancement – ​​Combine search results, node content, and dialogue history to construct enhanced prompts;

[0075] Step R5: Generate answers by streaming large language models via SSE.

[0076] The RAG-enhanced knowledge-based question-answering module, in the data preprocessing step, uses a multimodal visual model to extract text, formulas, and charts from image-based learning materials and converts them into structured text data. After completing the vectorized index construction, it supports incremental text segmentation and vector embedding for user-added learning materials, and synchronously updates the semantic vector index. At the same time, in adaptive question generation scenarios, it synchronously performs semantic retrieval and context enhancement steps to generate questions that fit the user's learning scope based on the content of private learning materials.

[0077] The multi-dimensional weighted formula for the adaptive question generation and mastery assessment module is as follows: ,in: The preset weights are, in order, the accuracy rate, the number of attempts, the difficulty rate, and the time rate, and satisfy the following conditions: The levels are mapped to five levels: not learned, known, understood, mastered, and proficient.

[0078] The learner profile built by the adaptive question generation and mastery assessment module includes the user's historical mastery level for each knowledge node, answer accuracy, answer frequency, average answer time, and question difficulty preference data. Adaptive question generation matches questions of corresponding difficulty based on the current mastery level of the target knowledge node. Basic concept questions are generated for the "not learned" or "understand" level, advanced application questions are generated for the "understand" or "master" level, and comprehensive extension questions are generated for the "proficient" level. At the same time, the calculated mastery level is bound to the corresponding node in the knowledge tree and updated in real time in the front-end knowledge tree graph through different color indicators.

[0079] The points ledger and integrity verification module is configured with corresponding fixed deduction amounts and dynamic token billing rules for different function call scenarios such as knowledge tree generation, AI question answering, and adaptive question generation. Each billing transaction generates a corresponding transaction record. The HMAC-SHA256 algorithm is used to jointly calculate the transaction record, the hash signature of the previous transaction, and the user's unique identifier to generate the chain signature of the current transaction. The transaction record and the chain signature are stored synchronously, and the integrity of historical transaction data is verified by the chain signature before each transaction.

[0080] The system adopts a B / S architecture with a front-end and back-end separation design, and is divided into five layers. The front-end presentation layer uses ReactFlow to realize the visualization rendering of the knowledge tree graph, and supports users to drag and drop to edit the generated knowledge nodes, modify the content, adjust the hierarchical relationship, and bind node data. The edited node data is synchronously and persistently stored in the data storage layer. The back-end API layer is configured with JWT authentication, SSE streaming interface, RESTful API routing and request rate limiting unit, providing a data interaction channel between the front-end and the business service layer.

[0081] An AI-driven knowledge tree adaptive learning method, applied in an AI-driven knowledge tree adaptive learning system, includes the following steps:

[0082] S1. Build a multi-LLMProvider abstract scheduling framework, maintain the registry of multiple LLM service providers through hot-loading configuration files, complete the routing matching of service requests based on model capability tags, use a weighted round-robin algorithm to distribute requests among multiple service providers of the same model, count the consecutive failed requests of a single service provider, and trigger circuit breaking and failover when the number of failures exceeds the threshold. At the same time, shield the API differences of different LLM service providers through a unified interface.

[0083] S2. Receive the learning topic and supporting parameters input by the user, construct a structured prompt word template, match the optimal LLM service through the multi-LLMProvider abstract scheduling framework built in step S1, and push the hierarchical knowledge node data generated by the model to the front-end rendering in real time using the SSE streaming protocol, and persist the knowledge nodes and hierarchical relationships to complete the automatic generation of the knowledge tree.

[0084] S3. Perform text extraction, content recognition, and word segmentation preprocessing on user-uploaded private learning materials. After segmenting and cutting the preprocessed text into blocks, generate vector embeddings and build a semantic vector index library. When a user's question-and-answer or question-generating request is received, the request content is encoded into a vector. Top-K relevant paragraphs are matched through cosine similarity retrieval. Enhanced prompt words are constructed by combining the search results, corresponding knowledge node content, and dialogue history. The LLM service is called and the corresponding results are generated through SSE streaming.

[0085] S4. Continuously collect users' answer records, build and dynamically update learner profiles, and generate questions that match the user's current mastery level based on the learner profile and the target knowledge node content. After the user completes the answer, calculate the mastery score through a preset multi-dimensional weighted formula, map the score to a five-level mastery level, and update the answer results and mastery level to the learner profile.

[0086] S5 adopts a hybrid billing model that combines fixed deductions with dynamic token billing to process billing for users' system function calls, and verifies the integrity of each billing transaction through HMAC-SHA256 chain signature.

[0087] The implementation environment for the above system is as follows;

[0088] Server-side: Linux Ubuntu 22.04, Python 3.12, FastAPI 0.104, MySQL 8.0, Nginx 1.24.

[0089] Client-side: Chrome 120+ browser; front-end based on Next.js 16 and React 19.

[0090] AI services: DeepSeek as the main provider, Tongyi Qianwen DashScope as an alternative, BAAI / bge-m3Embedding model, and Qwen2-VL-72B visual model.

[0091] Example 1: AI-Driven Automatic Knowledge Tree Generation Implementation

[0092] This embodiment is based on the above system and method, and automatically generates scenarios for knowledge trees of user learning topics. The specific implementation process is as follows:

[0093] In the knowledge graph creation dialog box on the front end, users select the "AI generation" mode, enter the learning topic as "Machine Learning Fundamentals", and optionally fill in the topic description as "a core knowledge system of machine learning for beginners with zero foundation, covering core concepts, mainstream algorithms and basic applications", and limit the learning scope to "undergraduate introductory level content", and submit the generation request.

[0094] The knowledge tree intelligent generation module executes step D1, receiving the above learning topic, topic description, and learning scope parameters;

[0095] Perform step D2 to construct a standardized prompt template. The template content is: "You are a professional expert in building an educational knowledge system. Based on the learning topic input by the user, you need to generate a structured, hierarchical knowledge tree graph. The requirements are as follows: 1. Generate three levels of knowledge nodes: the first level is the core knowledge domain, the second level is the subdivided knowledge module, and the third level is the specific knowledge point; 2. The total number of nodes should be controlled between 20 and 30, and the node content should be aligned with the learning scope and learning objectives specified by the user; 3. Strictly output in JSON format, with each node containing a unique node ID, node name, node description, and parent node ID fields to ensure a clear parent-child hierarchical relationship."

[0096] The output should only contain JSON data and should not include any additional explanatory text. The user input should be: "Learning Topic: {Learning Topic}, Topic Description: {Topic Description}, Learning Scope: {Learning Scope}". The user input parameters should be populated into the template and encapsulated into structured generation instructions that can be understood by the large language model.

[0097] Execute step D3, call the model router of the multi-LLMProvider abstraction layer. The model router matches the optimal DeepSeek model for the "knowledge system generation" scenario based on the capability tags in the YAML hot-reload configuration file, and distributes the request to the available DeepSeek service nodes through the weighted round-robin load balancing unit. If the current DeepSeek service fails more than 3 times in a row, the circuit breaker state machine automatically triggers the circuit breaker and transfers the request to the alternative Tongyi Qianwen DashScope service.

[0098] In step D4, a long connection is established with the large language model through the SSE streaming protocol, and the knowledge node JSON data generated by the model is pushed to the front end in real time in chunks. The front end continuously receives the data stream through the JavaScript asynchronous generator. Each time a complete node data is received, the corresponding node and hierarchical connection are immediately rendered on the canvas in real time through the ReactFlow component, realizing the progressive visualization of the knowledge tree without waiting for all content to be generated.

[0099] Execute step D5. After all nodes are generated, the 24 generated knowledge nodes (including 3 first-level nodes for "supervised learning", "unsupervised learning" and "reinforcement learning", 10 second-level nodes and 11 third-level nodes) and their parent-child hierarchical relationships and node description information are synchronously and persistently stored in the MySQL database and bound to the current user account.

[0100] The knowledge tree generation process in this embodiment takes 11 seconds, which is 12 times more efficient than the traditional manual knowledge structure building and whole package request-response generation mode. Users can view the generation progress in real time, greatly improving the user experience.

[0101] Example 2: Implementation of Private Data Augmented Knowledge Question Answering Based on RAG Technology

[0102] This embodiment is based on the above system and method, and specifically addresses the AI ​​question-answering scenario for user-owned learning materials. The implementation process is as follows:

[0103] Users can upload the "Statistical Learning Methods" textbook in PDF format to the data management interface corresponding to the "Machine Learning Fundamentals" knowledge tree on the front end. The system will then store the PDF file in Alibaba Cloud OSS file storage.

[0104] The RAG enhanced knowledge question answering module executes step R1, which uses the pypdf library to extract the full text content of the PDF, and performs cleaning, deduplication, and Chinese word segmentation preprocessing on the extracted text; for image pages in the PDF containing formulas and charts, the Qwen2-VL-72B multimodal visual model is called to extract the text, mathematical formulas, and chart data from the images, convert them into LaTeX format and structured text data, and merge them with the full text content;

[0105] In step R2, a fixed window segmentation strategy is adopted to divide the merged text into segments of 500 characters each, with adjacent segments overlapping by 50 characters to avoid semantic breaks. The BAAI / bge-m3Embedding model is called to generate a 768-dimensional vector embedding for each text block. The vector embedding data is stored in the vector index library to build a semantic vector index. At the same time, the text blocks are associated and bound with the "Machine Learning Basics" node of the corresponding knowledge tree.

[0106] In the AI ​​dialogue interface, users can select the "Machine Learning Fundamentals" knowledge tree that is linked to the above materials, enter the question "What are the kernel functions of SVM, and what scenarios are they suitable for?", and submit a question-and-answer request.

[0107] In step R3, the user's question is encoded into a 768-dimensional query vector using the BAAI / bge-m3 model. The top 5 relevant text paragraphs with the highest matching degree with the query content are retrieved from the vector index library by calculating cosine similarity.

[0108] In step R4, an enhanced prompt word template is constructed by combining the top 5 relevant paragraphs retrieved, the content of the "SVM" node corresponding to the current knowledge tree, and the user's historical dialogue records with the system, and filling them into the prompt word template. The large language model is required to answer only based on the provided context and must not fabricate information outside the context, thus effectively reducing model illusions.

[0109] In step R5, the large language model is called via SSE streaming, and an accurate answer is generated based on the enhanced prompt words. The answer content includes the core definitions, applicable scenarios and advantages and disadvantages of linear kernel, multinomial kernel, Gaussian kernel and Sigmoid kernel, which are completely matched with the content of the textbook "Statistical Learning Methods" uploaded by the user. The answer is pushed to the front end for rendering and display in real time via Markdown+KaTeX format.

[0110] In this embodiment, the AI ​​question-answering system based on RAG enhancement achieves an 85% improvement in accuracy and a 90% reduction in hallucination rate compared to the native question-answering system of a general large language model, realizing a deep integration of user-owned learning materials and the AI ​​question-answering system. Simultaneously, when a user initiates a question request for a specific node, the system concurrently executes the aforementioned semantic retrieval and context enhancement steps, generating questions tailored to the user's learning scope based on the textbook content, thus achieving full-scenario adaptation of private materials.

[0111] Example 3: Implementation of Adaptive Question Generation and Multi-Dimensional Five-Level Mastery Assessment

[0112] This embodiment is based on the above system and method, and is implemented for the scenario of adaptive question generation and mastery assessment of knowledge nodes. The specific implementation process is as follows:

[0113] Prerequisites: The user has completed 5 quizzes in the "Linear Regression" node of the "Machine Learning Fundamentals" knowledge tree, with a total of 10 questions answered, 6 of which were answered correctly, resulting in a historical accuracy rate of 60%. The system has built a learner profile for the user based on the historical quiz record, and the historical mastery level of this node in the profile is "understanding".

[0114] On the front-end "Linear Regression" node page, users can click the "Generate Question" button to submit a question request;

[0115] The adaptive question generation and mastery assessment module reads the user's learner profile, obtains the historical answer records, accuracy rate, historical mastery level, and average answer time data for this node, and combines the core content of the "linear regression" node and the bound private learning materials to construct a differentiated question prompt word template;

[0116] Based on the user's current level of "understanding", the system generates advanced questions at the "comprehension" level, specifically two advanced multiple-choice questions and one basic fill-in-the-blank question. The difficulty of the questions is appropriate for the user's current level of understanding, covering basic concepts as well as simple application scenarios and calculations, avoiding questions that are too difficult or too easy.

[0117] The front end displays the generated questions, the user completes the answers and submits them, the system automatically evaluates the user's answers, and the system calculates the correctness rate of the answer to be 100% (all 3 questions are answered correctly), and records the average answer time per question as 45 seconds, while the preset standard answer time for the corresponding questions is 60 seconds;

[0118] The system calculates the total mastery score for this node using a multi-dimensional weighted formula. In this embodiment, the preset weights are: The calculation rules for each coefficient and the final score are as follows:

[0119] Accuracy: The accuracy rate for this test was 100%, combined with the historical accuracy rate of 60%, the overall accuracy rate is 0.7.

[0120] Frequency coefficient: When a user has accumulated 6 valid answers, reaching the preset baseline of 5 answers, the frequency coefficient is set to 0.6.

[0121] Difficulty level: The average difficulty level of the questions in this quiz is 0.7. Since the user answered all the questions correctly, the difficulty level is set to 0.7.

[0122] Time coefficient: Average user answering time 45 seconds / Standard answering time 60 seconds = 0.75, the time coefficient is set to 0.5;

[0123] Final mastery score;

[0124] The system's preset five-level mastery score mapping range is as follows: 0≤Score<0.1 indicates no learning (gray mark), 0.1≤Score<0.3 indicates understanding (blue mark), 0.3≤Score<0.7 indicates comprehension (green mark), 0.7≤Score<0.9 indicates mastery (yellow mark), and 0.9≤Score≤1 indicates proficiency (red mark). In this embodiment, Score=0.62 is mapped to the "comprehension" level, and the "linear regression" node in the front-end knowledge tree graph is synchronously updated to the green mark to intuitively display the user's mastery status.

Claims

1. An AI-driven knowledge tree adaptive learning system, comprising a front-end presentation layer, a back-end API layer, a business service layer, an AI service layer, and a data storage layer, characterized in that: The AI ​​service layer includes a multi-LLM Provider abstraction layer, which incorporates a model router, a weighted round-robin load balancing unit, a circuit breaker state machine, and a unified client interface. The model router maintains an LLM service provider registry based on a hot-load configuration file and routes service requests according to model capability tags. The weighted round-robin load balancing unit distributes requests among multiple service providers of the same model. The circuit breaker state machine counts consecutive failed requests from a single service provider and triggers circuit breaking and failover when the number of failures exceeds a threshold. The unified client interface is used to shield the API differences between different LLM service providers; The business service layer includes a knowledge tree intelligent generation module, a RAG enhanced knowledge question answering module, an adaptive question generation and mastery assessment module, and an points ledger and integrity verification module. The knowledge tree intelligent generation module is used to receive the learning topic and supporting parameters input by the user, construct a structured generation prompt word template, call the multi-LLMProvider abstraction layer to match the optimal LLM service, push the knowledge node data with hierarchical relationship generated by the model to the front-end display layer for rendering in real time through the SSE streaming protocol, and persistently store the knowledge nodes and hierarchical relationships in the data storage layer. The RAG enhanced knowledge question answering module is used to preprocess and construct vectorized indexes for private learning materials uploaded by users, complete semantic retrieval matching based on users' question-and-answer or question-generating requests, construct enhanced prompt words by combining retrieval results, knowledge node content and dialogue history, call LLM service and generate corresponding results through SSE streaming transmission; The adaptive question generation and mastery assessment module is used to build a learner profile based on the user's answer records, generate questions that match the user's mastery level by combining the learner profile with the target knowledge node content, and calculate the user's mastery score based on a preset multi-dimensional weighted formula, which is mapped to a five-level mastery level. The points ledger and integrity verification module adopts a hybrid billing mode that combines fixed deductions and dynamic token billing, and performs integrity verification on each billing transaction through HMAC-SHA256 chain signature.

2. An AI-driven knowledge tree adaptive learning method, applied in the AI-driven knowledge tree adaptive learning system described in claim 1, characterized in that, Includes the following steps: S1. Build a multi-LLMProvider abstract scheduling framework, maintain the registry of multiple LLM service providers through hot-loading configuration files, complete the routing matching of service requests based on model capability tags, use a weighted round-robin algorithm to distribute requests among multiple service providers of the same model, count the consecutive failed requests of a single service provider, and trigger circuit breaking and failover when the number of failures exceeds the threshold. At the same time, shield the API differences of different LLM service providers through a unified interface. S2. Receive the learning topic and supporting parameters input by the user, construct a structured prompt word template, match the optimal LLM service through the multi-LLMProvider abstract scheduling framework built in step S1, and push the hierarchical knowledge node data generated by the model to the front-end rendering in real time using the SSE streaming protocol, and persist the knowledge nodes and hierarchical relationships to complete the automatic generation of the knowledge tree. S3. Perform text extraction, content recognition, and word segmentation preprocessing on user-uploaded private learning materials. After segmenting and cutting the preprocessed text into blocks, generate vector embeddings and build a semantic vector index library. When a user's question-and-answer or question-generating request is received, the request content is encoded into a vector. Top-K relevant paragraphs are matched through cosine similarity retrieval. Enhanced prompt words are constructed by combining the search results, corresponding knowledge node content, and dialogue history. The LLM service is called and the corresponding results are generated through SSE streaming. S4. Continuously collect users' answer records, build and dynamically update learner profiles, and generate questions that match the user's current mastery level based on the learner profiles and target knowledge node content. After the user completes the answer, the mastery score is calculated using a preset multi-dimensional weighted formula. The score is then mapped to a five-level mastery level, and the answer result and mastery level are fed back to update the learner profile. S5 adopts a hybrid billing model that combines fixed deductions with dynamic token billing to process billing for users' system function calls, and verifies the integrity of each billing transaction through HMAC-SHA256 chain signature.

3. The system according to claim 1, characterized in that: The specific steps of the knowledge tree intelligent generation module are as follows: Step D1: Receive the user's input of the learning topic and optional topic description and learning scope parameters; Step D2: Construct prompt word templates, encapsulate the learning topics into structured generation instructions that can be understood by the large language model, and require the model to output knowledge node data containing hierarchical relationships in JSON format; Step D3: Invoke the model router of the multiple LLM Provider abstraction layer to select the best available large language model service for inference; Step D4: Using the SSE streaming protocol, the knowledge node data generated by the model is pushed to the front end in real time. The front end receives and renders the knowledge tree graph in real time through the asynchronous generator. Step D5: Persistently store the generated knowledge nodes and their parent-child hierarchical relationships in the database.

4. The system according to claim 1, characterized in that: The model router in the multi-LLM Provider abstraction layer is based on a YAML hot-reload configuration file, maintains a registry of multiple LLM Providers, and routes requests according to model capability tags; The weighted round-robin load balancing unit uses a weighted round-robin algorithm to distribute requests among multiple providers of the same model; the circuit breaker state machine maintains a failure counter and a circuit breaker state machine for each provider, and automatically circuit breaks and transfers requests when consecutive failures exceed the threshold; The unified client interface provides three types of client interfaces: Chat, Embedding, and Vision, shielding the API differences between different providers.

5. The system according to claim 1, characterized in that: The specific steps of the RAG enhanced knowledge question-answering module are as follows: Step R1: Data preprocessing - PDF text is extracted using pypdf, images are subjected to content recognition using a multimodal visual model, and documents are preprocessed using Chinese word segmentation; Step R2: Vectorized Indexing – The text is segmented into blocks, and the Embedding model is called to generate vector embeddings to build a semantic vector index; Step R3: Semantic retrieval—Encode the query into a vector and use cosine similarity to retrieve Top-K relevant paragraphs; Step R4: Contextual Enhancement – ​​Combine search results, node content, and dialogue history to construct enhanced prompts; Step R5: Generate answers by streaming large language models via SSE.

6. The system according to claim 1, characterized in that: The multi-dimensional weighted formula for the adaptive question generation and mastery assessment module is as follows: ,in: The preset weights are, in order, the accuracy rate, the number of attempts, the difficulty rate, and the time rate, and satisfy the following conditions: The levels are mapped to five levels: not learned, known, understood, mastered, and proficient.

7. The system according to claim 1, characterized in that: The points ledger and integrity verification module are configured with corresponding fixed deduction amounts and dynamic token billing rules for different function call scenarios such as knowledge tree generation, AI question answering, and adaptive question generation. Each billing transaction generates a corresponding transaction record. The HMAC-SHA256 algorithm is used to jointly calculate the transaction record, the hash signature of the previous transaction, and the user's unique identifier to generate the chain signature of the current transaction. The transaction record and the chain signature are stored synchronously, and the integrity of historical transaction data is verified by the chain signature before each transaction.

8. The system according to claim 6, characterized in that: The learner profile constructed by the adaptive question generation and mastery assessment module includes the user's historical mastery level for each knowledge node, answer accuracy, answer frequency, average answer time, and question difficulty preference data. The adaptive question generation matches questions of corresponding difficulty based on the current mastery level of the target knowledge node. Specifically, the "unlearned" or "understanding" level corresponds to basic concept questions, the "understanding" or "mastering" level corresponds to advanced application questions, and the "proficient" level corresponds to comprehensive extension questions. At the same time, the calculated mastery level is bound to the corresponding node in the knowledge tree and updated in real time in the front-end knowledge tree graph through different color indicators.

9. The system according to claim 5, characterized in that: The RAG-enhanced knowledge-based question-answering module, in the data preprocessing step, calls a multimodal visual model to extract text, formulas, and charts from image-formatted learning materials and converts them into structured text data. After completing the vectorized index construction, it supports incremental text segmentation and vector embedding for user-added learning materials, and synchronously updates the semantic vector index. At the same time, in the adaptive question generation scenario, it synchronously performs semantic retrieval and context enhancement steps to generate questions that fit the user's learning scope based on the content of the private learning materials.

10. The system according to claim 1, characterized in that: The system adopts a B / S architecture with a front-end and back-end separation design, and is divided into a five-layer architecture. The front-end display layer is based on ReactFlow to realize the visualization rendering of the knowledge tree graph, and supports users to drag and drop to edit the generated knowledge nodes, modify the content, adjust the hierarchical relationship, and bind node data. The edited node data is synchronously and persistently stored in the data storage layer. The backend API layer is configured with JWT authentication, SSE streaming interface, RESTful API routing and request rate limiting units, providing a data interaction channel between the frontend and the business service layer.

Images