An ai-assisted learning method and system for standardized qualification certification examinations

Abstract

This invention discloses an AI-assisted learning method and system for standardized qualification certification examinations, relating to the field of computer software service technology. The method includes collecting and verifying examination question data for a specific professional field, constructing a structured knowledge base supporting semantic retrieval; receiving user queries and calling a large language model to identify the user's intent; if the identified intent is related to learning in a specific professional field, retrieving relevant knowledge content from the structured knowledge base; generating personalized learning content based on the retrieval results using retrieval enhancement generation technology and feeding it back to the user's terminal; this method supports natural language interactive retrieval and can accurately match students' semantic needs.

Description

Technical Field

[0001] This invention relates to the field of computer software service technology, and specifically to an AI-assisted learning method and system for standardized qualification certification examinations. Background Technology

[0002] With the rapid development of information technology and artificial intelligence, various online education platforms and intelligent learning tools are emerging, aiming to improve learning efficiency and optimize the learning experience. The National Medical Licensing Examination (hereinafter referred to as the "Medical Licensing Examination") is a crucial link in the training and qualification of medical personnel in my country. Its preparation process is characterized by a large number of knowledge points, rapid content updates, and strong clinical practicality. However, current preparation services for the Medical Licensing Examination still have many shortcomings, failing to meet the efficient, accurate, and personalized learning needs of medical students.

[0003] Currently, online services for preparing for the National Medical Licensing Examination mainly focus on online learning and practice. These platforms collect exam questions and complete mock exams covering various subject categories and topics, providing students with specialized practice or mock exams to help them familiarize themselves with the exam format and solidify their knowledge. Some platforms also provide answer keys.

[0004] However, existing learning and practice services only achieve resource retrieval through a "catalog classification index," which categorizes questions by subject (such as internal medicine and surgery) and exam module (such as basic medicine and clinical medicine). Students need to click through the catalog layer by layer to find the corresponding content, which makes it impossible for students to directly or accurately locate specific knowledge points and may mislead them. Summary of the Invention

[0005] To address the shortcomings of existing technologies that rely solely on "catalog classification indexes" for resource retrieval, which prevents students from directly or accurately locating specific knowledge points and may mislead them, this invention proposes an AI-assisted learning method and system for standardized qualification certification exams, thereby resolving the problems existing in the prior art.

[0006] An AI-assisted learning method for standardized qualification certification exams includes the following steps: Collect and validate exam question data for specific professional fields; The validated exam question data is transformed into vectors through a word embedding model to build a vector knowledge base that supports semantic retrieval. It receives natural language learning requests from users and identifies exam-related intentions based on a large language model; it retrieves relevant knowledge fragments from a vector knowledge base based on the identified intentions; and it synthesizes the knowledge fragments into structured auxiliary learning content that matches the learning request based on retrieval enhancement generation technology. Analyze user interaction data with structured learning content to build user profiles; Based on user profiles, we provide adaptive learning suggestions or resource recommendations; and output structured auxiliary learning content and adaptive learning suggestions or resource recommendations to user terminals.

[0007] Furthermore, it also includes preprocessing the exam question data after collecting it from a specific professional field, specifically including the following steps: For image-based materials, an optical character recognition model is used to identify the text content in the image, convert it into editable text, and save the image file to a specified path; for document-based materials, a document parsing tool is used to read the text and convert it into UTF-8 encoded plain text. Identify and segment the question stem, options, answers, and explanations from the exam question data; for questions containing images or special format content, extract their text description information and store it in association with the original file; Based on the semantic similarity of the question content, duplicate question data is identified, merged, or removed.

[0008] Furthermore, the verification of exam question data in a specific professional field specifically includes cross-validation, external source validation, and manual validation. Cross-validation involves performing vector retrieval of the question stem in a pre-constructed temporary knowledge base to verify the uniqueness of the corresponding answer, and marking questions with conflicting answers as questionable questions. External source validation involves comparing and verifying the question stem and answer of the questionable questions with data from external authoritative knowledge sources. Manual validation involves domain professionals conducting a final review of questions that remain questionable after the cross-validation and external source validation.

[0009] Furthermore, the construction of a vector knowledge base supporting semantic retrieval specifically includes the following steps: Each question is divided into multiple text sub-blocks, consisting of the question stem, the standard answer, and the answer explanation. The Qwen3-Embedding model is used to generate corresponding word embedding vectors for each text sub-block. The text sub-blocks are associated and stored with their corresponding word embedding vectors and the metadata of the questions, and a hybrid retrieval index supporting vector retrieval and full-text retrieval is constructed.

[0010] Furthermore, after retrieving relevant knowledge fragments from the vector knowledge base based on the identified intent, the method also includes calling a re-ranking model to rank the knowledge fragments; wherein the re-ranking model is used to re-rank the data based on the multi-dimensional matching degree between the query intent and the retrieval results.

[0011] Furthermore, the user profile includes real-time recording of incorrect answers from users' responses, a question collection folder where users can customize their collection of questions, the addition of tags and notes, and an assessment of the mastery of each knowledge point.

[0012] Furthermore, the adaptive learning recommendations include a similar question reinforcement practice package automatically generated based on the incorrect question information, or a personalized review plan generated based on the mastery assessment.

[0013] This invention also proposes an AI-assisted learning system for standardized qualification certification examinations, comprising: The data collection unit is used to collect exam question data in a specific professional field and to verify it. The knowledge base construction unit is used to convert the verified exam question data into vectors through a word embedding model in order to build a vector knowledge base that supports semantic retrieval. The intelligent retrieval and content generation unit is used to receive natural language learning requests input by users and to identify exam-related intentions based on a large language model; to retrieve relevant knowledge fragments from the vector knowledge base according to the identified intentions; and to synthesize the knowledge fragments into structured auxiliary learning content that matches the learning request based on retrieval enhancement generation technology. The user profile building unit is used to analyze user interaction data with structured learning content in order to build user profiles. The output unit is used to provide adaptive learning suggestions or resource recommendations based on user profiles; and to output structured auxiliary learning content and adaptive learning suggestions or resource recommendations to the user terminal.

[0014] This invention provides an AI-assisted learning method for standardized qualification certification exams, which has the following beneficial effects: This invention employs a vector knowledge base that supports semantic retrieval and a technique for intent recognition of learning requests based on a large language model. It transforms users' fuzzy natural language queries into precise semantic vectors and performs deep matching within the structured vector knowledge base. This not only overcomes the limitations of traditional coarse-grained classification retrieval, allowing students to directly locate specific knowledge points through natural language questions, but also significantly reduces ineffective filtering time, fundamentally improving the accuracy and efficiency of information retrieval. Furthermore, by analyzing user interaction data, constructing user profiles, and providing adaptive learning suggestions, it dynamically generates similar question reinforcement practice packages or personalized review plans based on real-time recorded incorrect answers, notes, and knowledge point mastery. This fundamentally solves the drawbacks of indiscriminate push notifications in traditional technologies, significantly improving the targeting and effectiveness of learning. Attached Figure Description

[0015] Figure 1This is a schematic diagram illustrating the collection of original materials in an embodiment of the present invention; Figure 2 This is a schematic diagram of the knowledge base construction process in an embodiment of the present invention; Figure 3 This is a schematic diagram of the user demand response process in an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0017] Online services for preparing for the National Medical Licensing Examination on the Internet are mainly divided into two categories: (1) Resource services: The core function is to collect past exam questions and authoritative mock exam questions, and provide students with resource download channels. However, high-quality resources often require payment to obtain, and only have the resource download function, without core preparation functions such as practice and mock exams, which cannot meet students' one-stop preparation needs. (2) Learning and practice services: Supports special practice by subject category and knowledge topic, and can also conduct complete mock exams, but there are obvious shortcomings - it does not support in-depth questioning of medical knowledge, and the interactive experience is poor. The platform is full of various marketing content, and the core purpose is to guide candidates to purchase paid courses.

[0018] The existing learning and practice services only achieve resource retrieval through "catalog classification index", that is, the questions are classified by subject (such as internal medicine, surgery) and examination module (such as basic medicine, clinical medicine) in a coarse granularity. Students need to click through the catalog to find the corresponding content. At the same time, it only relies on manual screening of question sources and does not establish a systematic knowledge base verification mechanism. Some question answers lack authoritative verification. Its defects and causes include: (1) Low retrieval accuracy: Due to the use of coarse catalog index mode, without the introduction of word embedding, vector matching and other technologies, it is impossible to identify the natural language semantics of students' input, and it is also impossible to directly locate specific knowledge points (such as "electrocardiographic manifestations of acute myocardial infarction"). This causes students to spend a lot of time screening resources and cannot quickly match weak points; (2) Risk of content accuracy: Due to the lack of a content guarantee mechanism of "multi-round verification (cross-validation + external verification + manual verification) + search enhancement generation (RAG)", it only relies on basic manual review, which makes it difficult to avoid the problem of incorrect question answers or incomplete analysis, which may mislead students.

[0019] The existing learning and practice services only provide the "unified question push" function, that is, push the same questions to all students according to the preset subject or topic. There is no differentiated design for individual student learning situation. Only a few platforms support "wrong question record", but it is not possible to customize the wrong questions, classify them by tags or recommend similar questions. The defects and causes are: (1) Insufficient personalization: Because the customized function module of "wrong question bank + collection question bank + custom tags + personal notes" has not been developed, and the user learning behavior analysis mechanism has not been introduced, it is impossible to record students' wrong question frequency, collection preferences, weak knowledge points and other personalized data, which makes it impossible to generate targeted practice content, and students find it difficult to make up for their weak links efficiently. (2) Poor learning continuity: Due to the lack of a recommendation algorithm for "wrong questions associated with similar questions", students' review of wrong questions is limited to a single question, and it is impossible to form a learning loop of "wrong questions-knowledge points-similar questions", making it difficult to systematically consolidate weak knowledge.

[0020] The existing learning and practice services adopt the interaction mode of "button click + form input". Students need to enter the question page through a fixed entry point (such as "Specialized Practice - Internal Medicine - Cardiovascular System") and cannot directly make requests through natural language (such as "Help me generate 10 A2 type questions about hypertension"); and only the answers to the questions are provided, without supporting follow-up questions on knowledge points (such as when a student asks "Why should hypertensive patients use nifedipine tablets with caution", the platform has no relevant analysis or extended content). Some platforms will also insert marketing pop-ups in the interface to guide the purchase of paid courses. Its defects and causes are: (1) Low interaction efficiency: Because it does not integrate large language models (such as Qwen3) and intelligent workflow orchestration tools (such as Dify), it cannot understand natural language semantics and can only rely on fixed operation paths, resulting in cumbersome interaction steps, which does not conform to students' exam preparation habits of "quickly making requests and quickly obtaining content". (2) Lack of knowledge extension ability: Because the knowledge extension mechanism of "knowledge base - large language model - question and answer generation" has not been built, only basic questions and answers are stored, without being associated with authoritative textbooks, clinical guidelines and other extended content. It is unable to respond to students' follow-up questions on knowledge points and it is difficult to help students build a complete knowledge system. (3) Strong interference: Because the platform's profit model relies on the promotion of paid courses, the development concept of "simple interface design + no marketing interference" has not been adopted, resulting in too much marketing content in the interface, which distracts students' attention and affects their focus on exam preparation.

[0021] The existing learning and practice services adopt a deployment mode of "general server + fixed resource allocation", without introducing GPU intelligent scheduling and model quantization technology. It only relies on a single or a small number of general servers to carry user requests, and the model runs in FP16 format, resulting in high memory usage and computational complexity. Its defects and causes are as follows: (1) Weak high concurrency carrying capacity: Because the hardware support system of "GPU cluster (Huawei 910B graphics card) + intelligent scheduling (dynamic allocation of memory and computing cores according to concurrency) + load balancing" is not deployed, it only relies on general servers. When the number of concurrent users increases (such as hundreds of people online at the same time), page lag and request timeout problems are likely to occur. (2) Slow response speed: Because int4 quantization technology is not used, the memory usage of the model is not reduced (the memory usage of FP16 format is 4 times that of int4), the inference speed cannot be improved, resulting in students waiting too long for questions to load and answers to be generated, which affects the efficiency of exam preparation.

[0022] Based on this, the present invention proposes a method for establishing an integrated AI learning platform for the physician licensing examination, which includes an offline preparation stage (knowledge base construction process) and an online service stage (user demand response process).

[0023] Offline preparation phase: S1. Collection of raw materials: Collection channels: A dual approach of "online web crawling + offline data processing" is adopted.

[0024] Online: A parallel crawler program was developed based on the Scrapy framework to analyze the page structure of medical licensing exam resource websites (such as medical exam websites and training institution question banks), accurately extract core content such as question stems, options, answers, and explanations, and avoid redundant information (such as advertisements and irrelevant links). Offline: Collect internal tutoring materials from medical colleges and universities, as well as authoritative publicly published tutoring books (such as "Analysis of Past Exam Questions for the Physician Qualification Examination"). Organize the questions into Word and PDF formats, and also collect questions containing medical images (electrocardiograms, pathological slides) (JPG / PNG format) to ensure coverage of both "text + image" question scenarios.

[0025] Results collected: Over 230,000 questions have been gathered, covering all disciplines of the medical licensing examination (internal medicine, surgery, obstetrics and gynecology, pediatrics, etc.), such as... Figure 1 The "multi-type example questions" shown include past exam questions, mock exam questions from well-known institutions, and test questions from tutoring materials, laying a data foundation for the subsequent construction of the knowledge base.

[0026] S2. Material Preprocessing: (1) Format conversion: For image-based materials (JPG / PNG): Use an OCR (Optical Character Recognition) model to recognize the text content in the image (such as image descriptions and option text in test questions), and convert it into editable text; at the same time, save the image file to a specified path, and replace the original image with the path identifier in the text (such as "[image path: / data / ecg_001.jpg]") to ensure that the image question can be searched and recognized later.

[0027] For document-type materials (Word / PDF): Use document parsing tools to read the text and convert it into UTF-8 encoded plain text to solve the problems of "garbled text and missing formulas" caused by different formats, and ensure data format consistency.

[0028] (2) Structure and paragraph marks: The text is split into three parts: "question stem, standard answer, and answer explanation". By identifying keywords (such as "answer:", "explanation:", "correct option:") or format markers (such as question stems starting with "1." and answers in bold), the three core modules of each question are accurately marked.

[0029] To handle special scenarios where the question stem and answer are not continuous (such as answers being concentrated at the end of a book): match the corresponding answer and explanation by using keywords in the question stem (such as "acute myocardial infarction") to ensure that the question and text correspond; at the same time, insert a custom separator "---" between two questions to provide an identifier for subsequent knowledge base segmentation.

[0030] (3) Data cleaning (content deduplication): Using the combination of "question stem + correct answer" as the identifier, a semantic similarity algorithm (based on the Embedding model to calculate the similarity of text vectors) is used to identify duplicate questions (those with a similarity of ≥95% are considered duplicates).

[0031] Retain questions with the most authoritative sources and the most complete explanations (e.g., prioritize past exam questions, followed by mock exams from well-known institutions) to reduce information redundancy and avoid repetitive practice for students.

[0032] (4) Data tagging (content validation): Cross-validation: The question stem is used as the query condition to search for the answer in a temporary knowledge base. If a unique match is returned, the question is considered accurate. If there are conflicting answers (such as different answers corresponding to the same question stem), the question is marked as "questionable".

[0033] External verification: The question stem and answer of the question are concatenated and then verified using the medical knowledge Q&A service of the school's teaching platform (which integrates authoritative textbooks and clinical guideline data) to determine whether they are consistent with the textbook knowledge.

[0034] Manual verification: Medical professionals reviewed the questions that remained questionable in the first two rounds, corrected incorrect answers, and supplemented missing explanations; finally, a random check was conducted (1000 questions were randomly selected) to ensure that the accuracy rate of the question bank reached over 99.9%.

[0035] S3. Knowledge base creation, such as Figure 2 As shown: (1) Creation of a knowledge base without images: ① Segmentation settings: Select the "Parent-Child Segmentation" mode. The parent block is a complete question (separated by "---"), and the child block is split into 3 fields (separated by "\n" newline character) according to "Question Stem, Standard Answer, and Answer Explanation". Also check "Replace consecutive spaces, newlines, and tabs" to avoid formatting interference with retrieval.

[0036] ②Word embedding and indexing settings: The Qwen3-Embedding model (a series of text embedding models developed based on the Qwen3 basic model) is used to convert each sub-block text into a numerical vector, compressing the text information into a computable vector form, ensuring that the retrieval can match "semantic association" (such as matching related questions such as "treatment principles of acute myocardial infarction" when a user asks "treatment of myocardial infarction"), rather than just matching keywords.

[0037] ③ Search rule configuration: Enable "hybrid search" (vector search + full-text search): Vector search matches deep semantics by calculating the similarity between the user query vector and the question vector; full-text search matches keywords (such as "hypertension" and "diuretics") through the inverted index; after the search, the Qwen3-rerank model is called to re-rank the results and output the Top 50 results according to "semantic relevance and knowledge point authority" to ensure search accuracy.

[0038] The Qwen3-rerank model is a high-precision re-ranking model developed by Tongyi Labs based on the Qwen3 series of basic models. It is divided into two main branches: plain text (Qwen3-Reranker) and multimodal (Qwen3-VL-Reranker). Its core purpose is to refine the ranking of initial recall results in the retrieval system, solving the problem of "dispersed relevance of initial retrieval results". It works in conjunction with the Qwen3-Embedding model to form a two-stage retrieval process of "efficient recall + accurate ranking", which significantly improves the final retrieval accuracy.

[0039] (2) An image knowledge base has been created: The core configuration is the same as the image-free knowledge base, with the only addition being the "Delete all URLs and email addresses" setting: Since the questions contain image storage paths (such as " / data / ecg_001.jpg"), this setting can avoid path characters interfering with the search results, ensuring that semantic matching is based solely on text content (question stem, answer, and explanation), while retaining image path identifiers for easy viewing of image-based questions.

[0040] During the online service phase, such as Figure 3 As shown: S1: User input received Users input their needs through the "conversational input box" (without marketing interference and a simple interface) in the scenario application layer, such as "generate 10 A2 type questions about diabetes" or "explain the typical symptoms of acute appendicitis".

[0041] The system receives input content through the "start node" in the diagram and stores it in the global variable sys.query, providing a unified data entry point for subsequent nodes such as "problem classification and retrieval" to ensure the continuity of process data.

[0042] S2: Question Classification and Intent Recognition Model and parameter configuration: In the "Question Classification Node", select the Qwen3 large language model as the intent recognition model, and the input variable is sys.query (user input content).

[0043] Category tag settings: Define two branch tags - "Questions related to everyday language" (such as "What day of the week is it today?" "How to register an account") and "Content related to the medical practitioner examination" (such as question generation and knowledge point consultation).

[0044] Path diversion: The Qwen3 model determines the type of request based on semantic understanding: if it is "everyday language", it jumps to the "everyday answer node"; if it is "medical related", it jumps to the "preliminary search node" and starts the knowledge base search process.

[0045] S3: Handling Everyday Language Needs If the user inputs "everyday phrases" (such as "how to modify personal information"), the system enters the "everyday answer node" and calls the Qwen3 language model.

[0046] Set the system prompt ("As a friendly platform assistant, provide concise and appropriate answers to users' daily questions, without involving medical professional content"), pass the user input (sys.query) as the prompt to the model, generate an answer, jump to the "response node", and feed back to the user through streaming output (displaying content in real time without waiting for complete generation).

[0047] S4: Medical-related needs – Dual knowledge base retrieval Preliminary search (with image knowledge base): The "Preliminary Search Node" is associated with the "Medical Licensing Examination Knowledge Base with Images," and the query variable is sys.query.

[0048] Retrieve the Top 100 results according to the "hybrid retrieval" rule (vector retrieval + full-text retrieval) and store them in the variable retrieval_result_img; Jump to the "Reflection Node" to standardize the format of the results (such as removing image path interference) in preparation for subsequent merging.

[0049] Search again (without image knowledge base): The "Re-search Node" is associated with "Medical Licensing Examination Knowledge Base without Images", and the query variable is still sys.query.

[0050] Enable the "metadata filtering component" (which can filter by "question type" and "subject". For example, if the user's requirement is "A2 type questions", only A2 type questions will be retrieved). After retrieving the Top 100 results, store them in the variable retrieval_result_noimg.

[0051] Jump to the "Summary Node" and merge retrieval_result_img and retrieval_result_noimg into a unified list retrieval_result_all to ensure coverage of all relevant questions in the "text + image" category.

[0052] S5: Reordering and Summarizing Search Results The "Summary Node" calls the Qwen3-rerank model, with the input parameters being "user requirements (sys.query) + merged results (retrieval_result_all)".

[0053] The model scores based on three dimensions: ① deep semantic matching (e.g., if a user asks "treatment", it prioritizes matching analyses containing "treatment principles" and "medication regimen"); ② authority of knowledge points (it prioritizes matching relevant content from past exam questions and textbook texts); ③ clinical applicability (it prioritizes matching questions that conform to the latest clinical guidelines).

[0054] Sort by score from highest to lowest, retain the top 50 results as the final candidate set final_result, and remove low-relevance content (such as removing questions related to "hypertension" if the user asks "diabetes"), thus solving the problem of "disorganized search results requiring manual filtering" in existing technologies.

[0055] S6: Professional Response Generation Model selection and prompt word configuration: For the "Thinking and Reasoning Node", the GPT model with a faster response speed is selected (15% faster than Qwen3 in actual tests). The system prompt words are set as follows: "As a medical licensing exam tutor, based on the knowledge base results, generate the response in the following format: ① Question generation requirements: question stem + options + answer + explanation; ② Knowledge point consultation: point-by-point explanation (core definition, clinical characteristics, precautions), combined with examples when necessary."

[0056] Response generation: Pass the "user request (sys.query) + final candidate set (final_result)" into the model to generate structured content. For example, if the user request is "Generate 2 A2 type questions about myocardial infarction", the model outputs: "1. Question: Male, 65 years old, sudden chest pain for 2 hours... Options: A.... B.... Answer: A Explanation: Typical manifestations of acute myocardial infarction are... 2. Question: Female, 58 years old, chest pain accompanied by profuse sweating... Options: A.... B.... Answer: B Explanation:...", ensuring professional content and clear format.

[0057] S7: Service Output and Personalized Data Synchronization Service output: The "Reply Node" adopts "streaming output" technology to display the reply content in real time (e.g., when 10 questions are generated, each question is displayed in real time without waiting for all questions to be generated); it also supports "copying the reply" and "downloading the questions" (PDF format) to facilitate offline review for users.

[0058] Personalized data synchronization: Error sync: If a user answers a question incorrectly during practice, the system automatically records the "question content, incorrect answer, number of errors (+1), and most recent error time" to the "error database".

[0059] Synchronize favorites: When a user clicks "Favorite Question", the system will add the question to a custom favorites folder (such as "Cardiovascular System - Emergency").

[0060] Tags and notes are synchronized: Users can tag questions (such as "high-frequency test points") or take notes (such as "thrombolysis time window for myocardial infarction: within 12 hours of onset"), and the system will link the questions to the stored data, supporting quick location of content by "tag filtering" and "note search".

[0061] This invention constructs a high-quality structured knowledge base, introducing word embedding, vector retrieval, Retrieval Enhancement Generation (RAG), and multi-round validation (cross-validation + external validation + manual validation) technologies. This allows students to directly retrieve corresponding knowledge point questions using natural language (e.g., "treatment principles of acute left ventricular failure"). Simultaneously, it ensures a high level of accuracy in the question bank content, avoiding misleading information. This addresses the problem of existing learning and practice services using a "coarse directory index + basic manual review" approach, which lacks a systematic semantic retrieval and content verification mechanism. This results in students being unable to accurately locate knowledge points using natural language, and questions often contain incorrect answers. This addresses the problem by developing a wrong question bank (recording incorrect answers and frequency), a collection question bank (supporting categorized management), custom tags (labeling questions by organ / disease), and a similar question recommendation algorithm. This automatically links student learning behavior data to generate practice content tailored to individual weaknesses, forming a personalized learning loop of "wrong questions - knowledge points - similar questions." This solves the problem that existing learning and practice services lack a "personal learning data tracking + customized content generation" module, only providing standardized question pushes and failing to record students' wrong questions, collection preferences, and weak knowledge points, making it difficult for students to specifically address their weaknesses. Based on large language models (such as Qwen3) and Dify... This workflow orchestration tool allows students to directly describe their exam preparation needs using natural language (e.g., "Generate 10 A3-type questions about diabetes") and ask follow-up questions on specific knowledge points (e.g., "Why do diabetic patients need to monitor glycated hemoglobin?"). It also features a simple, conversational interface, eliminating marketing content and improving interaction efficiency and learning depth. This addresses the problems of existing learning and practice services that lack integration of large language models and intelligent workflow orchestration tools, employ a fixed interaction mode of "button clicks + form input," lack connection to authoritative knowledge extensions, and suffer from significant marketing interference, resulting in cumbersome student interactions, difficulty in in-depth questioning, and reduced focus. This is achieved through deployment... The Huawei 910B GPU cluster (long-running + dedicated inference GPUs) and private cloud virtual machines support the front-end application. Combined with intelligent GPU scheduling (dynamically allocating GPU memory / computing cores according to concurrency and task type) and int4 quantization technology (reducing GPU memory usage and improving inference speed), it ensures that high-speed response (approximately 3000 tokens per second) can be maintained even when thousands of students are online at the same time. This avoids lag and timeouts in high-load scenarios, thus solving the problem that existing learning and practice services use a "general server + fixed resource allocation" solution without introducing GPU intelligent scheduling and model quantization technology, which leads to lag when the number of concurrent users increases and slow model inference speed.

[0062] The beneficial effects achievable by this invention are as follows: Through a four-layer architecture of "computing power resource layer + material resource layer + artificial intelligence platform layer + scenario application layer", combined with the core working principles of "RAG retrieval enhanced generation + large language model understanding + intelligent GPU scheduling + personalized question bank management", this invention achieves multi-dimensional beneficial effects against the closest existing technology for "learning and practice services", as detailed below: (i) Knowledge retrieval and content accuracy: From "extensive matching" to "precise and reliable", it achieves precise retrieval at the natural language semantic level, and the accuracy rate of the question bank is over 99.9%, completely solving the problems of "time-consuming retrieval and content error" in existing technologies.

[0063] The resource layer integrates over 270,000 questions through a combination of web scraping (Scrapy framework) and offline collection. A high-quality knowledge base is constructed using a combination of format conversion (OCR recognition of image text), structural tagging (splitting the question stem / answer / analysis), and multi-round verification (cross-validation: internal knowledge base retrieval verification; external verification: calling the medical Q&A service of the school's teaching platform; manual verification: random sampling and review), ensuring content accuracy from the source. The AI ​​platform layer introduces an embedding word model to convert questions and user natural language queries (such as "diagnostic criteria for chronic obstructive pulmonary disease") into vectors. Through a hybrid retrieval system (vector retrieval matching semantics + full-text retrieval matching keywords) and the Qwen3-rerank model, knowledge-point-level resources are accurately located without needing to click through directories layer by layer. The workflow utilizes a question classification node (Qwen3 model identifying medical / non-medical questions) and a RAG mechanism (retrieval results as input to the large model) to ensure answers are based on an authoritative knowledge base, avoiding the "illusion" of the large model.

[0064] (II) Personalized Exam Preparation: From "Unified Push" to "Precise Adaptation": Realizing "personal learning data-driven customized services" to help students focus on their weak points and improve exam preparation efficiency by more than 50% (based on trial feedback).

[0065] The application layer develops a "mistake database (recording incorrect answers, number of errors, and the time of the most recent error in real time, with support for manual modification) + collection database (customizable category folders) + custom tags (labeling questions by system / organ / disease) + personal notes (related questions, fuzzy search)" functional module to fully record the individual's learning trajectory; the AI ​​platform analyzes weak knowledge points through user behavior data (frequency of mistakes, collection type, tag preferences), calls the question database association algorithm to recommend similar questions, forming a learning closed loop of "mistakes - knowledge points - similar questions", avoiding ineffective practice caused by the "indiscriminate push" of existing technologies.

[0066] (III) Interaction and Knowledge Extension: From “Tedious Interference” to “High Efficiency and Depth”: Natural language interaction efficiency is improved by 3 times (no need for multiple operations), supports in-depth questioning of knowledge points, and is free from marketing interference, significantly improving focus.

[0067] The AI ​​platform integrates the Qwen3 large language model and the Dify workflow orchestration tool (Chatflow mode supports multi-turn dialogue). Students directly input their natural language needs (e.g., "Generate 5 A2-type questions about acute appendicitis"). The system responds quickly through an automated process: "Start node receives input → Question classification node judges intent → Retrieval node matches resources → Summary node generates results," replacing the cumbersome operation of "button clicks + directory filtering" in existing technologies. The knowledge base covers "past exam questions + mock questions + tutoring material analysis + clinical guidelines," supporting follow-up questions (e.g., after a student asks "indications for acute appendicitis surgery," they can further ask "prevention and treatment of postoperative complications"). The large language model generates answers based on the retrieved extended knowledge, helping to build a complete knowledge system. The scenario application layer adopts a "conversational input box + streaming output" interface design, eliminating all marketing content and solving the problems of "pop-up interference and complex operation" in existing technologies.

[0068] (iv) High concurrency performance: From "lag and timeout" to "smooth and stable": Supports thousands of students online at the same time, with a response speed of about 3,000 tokens per second (throughput of 2,922.5 tokens / s at 64 concurrency, with an average latency of 1.647s), meeting the needs of large-scale use in universities.

[0069] The computing resource layer deploys "3 Huawei 910B graphics cards (for long-term operation) + 8 Huawei 910B graphics cards (for dedicated inference) + private cloud virtual machines (for front-end web applications)" to build a dedicated high-computing cluster, replacing the weak hardware support of the existing "general-purpose server" technology. The AI ​​platform introduces an "intelligent GPU scheduling mechanism": dynamically allocating video memory and computing cores based on concurrency (e.g., a single card can carry multiple int4 model instances under low load, and scheduling to the GPU cluster under high concurrency) and task type (knowledge question answering / simulated question generation). Combined with int4 quantization technology (reducing video memory usage by 75% and increasing inference speed by 3-4 times), it solves the stuttering problem caused by the existing technology's "fixed resource allocation and slow model inference".

[0070] The knowledge base construction and retrieval technology proposed in this invention, which combines "multi-round verification + vector retrieval + Reranking", achieves a 99.9% accuracy rate for the question bank through "cross-validation (vector matching within the knowledge base to verify the uniqueness of answers) + external verification (connecting to authoritative data from university teaching platforms) + manual sampling and review". It also employs "parent-child segmentation (parent block is the complete question, child block is split into question stem / answer / analysis) + hybrid retrieval (vector + full text) + Qwen3-rerank model for fine ranking" to ensure the accuracy and completeness of the retrieval results.

[0071] The interactive scheduling technology proposed in this invention, which combines "Chatflow workflow + large language model intent recognition", is based on the Dify platform to build a Chatflow workflow. Through an automated process of "start node to store input → question classification node (Qwen3 model) to determine intent → dual knowledge base retrieval (with or without image) → summary node to merge results → thinking and reasoning node (GPT model) to generate answer", it realizes natural language multi-turn dialogue and accurate response.

[0072] The high-concurrency optimization technology proposed in this invention, namely "int4 quantization + dynamic GPU scheduling", uses int4 quantization technology to compress large language models, combined with a GPU scheduling strategy of "concurrency awareness + task type adaptation" (low load single card multiple instances, high concurrency cluster scheduling), and is equipped with message queues and load balancing to achieve low latency and high throughput under high concurrency.

[0073] The beneficial effects of the above technologies include: (1) Upgrading the quality of knowledge services: The "multi-round verification + retrieval reordering" technology ensures that the question bank is error-free and the retrieval is bias-free. Students do not need to verify the accuracy of the content, and their learning trust and efficiency are significantly improved. The "content reliability" score in the trial feedback reached 9.2 / 10. (2) Revolutionizing the interactive experience: The "Chatflow workflow + intent recognition" technology changes the interaction from "human adapting to system" to "system adapting to human". Students do not need to learn the operation logic. They can quickly obtain the required content by inputting natural language, and the operation time is reduced by more than 60%. (3) Feasibility of large-scale application: The "quantification + dynamic scheduling" technology breaks through the hardware resource limitation. A single GPU cluster can support thousands of students to use at the same time without the need for a significant increase in hardware investment, providing technical support for large-scale deployment in universities.

[0074] Based on the same inventive concept, this invention also proposes an AI-assisted learning system for standardized qualification certification examinations, comprising: The data collection unit is used to collect and verify exam question data in a specific professional field.

[0075] The knowledge base construction unit is used to transform the validated exam question data into vectors through a word embedding model in order to build a vector knowledge base that supports semantic retrieval.

[0076] The intelligent retrieval and content generation unit receives natural language learning requests input by users and identifies exam-related intentions based on a large language model. It then retrieves relevant knowledge fragments from a vector knowledge base based on the identified intentions and synthesizes the knowledge fragments into structured auxiliary learning content that matches the learning request using retrieval enhancement generation technology.

[0077] The user profile building unit is used to analyze user interaction data with structured learning content in order to build user profiles.

[0078] The output unit is used to provide adaptive learning suggestions or resource recommendations based on user profiles; and to output structured auxiliary learning content and adaptive learning suggestions or resource recommendations to the user terminal.

[0079] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An AI-assisted learning method for standardized qualification certification exams, characterized in that, Includes the following steps: Collect and validate exam question data for specific professional fields; The validated exam question data is transformed into vectors through a word embedding model to build a vector knowledge base that supports semantic retrieval. It receives natural language learning requests from users and identifies exam-related intentions based on a large language model; it then retrieves relevant knowledge fragments from a vector knowledge base based on the identified intentions. Based on retrieval enhancement generation technology, knowledge fragments are synthesized into structured auxiliary learning content that meets the learning request; Analyze user interaction data with structured learning content to build user profiles; Based on user profiles, we provide adaptive learning suggestions or resource recommendations; and output structured auxiliary learning content and adaptive learning suggestions or resource recommendations to user terminals.

2. The AI-assisted learning method for standardized qualification certification examinations according to claim 1, characterized in that, It also includes preprocessing the exam question data after collecting it from a specific professional field, specifically including the following steps: For image-based materials, an optical character recognition model is used to identify the text content in the image, convert it into editable text, and save the image file to a specified path; for document-based materials, a document parsing tool is used to read the text and convert it into UTF-8 encoded plain text. Identify and segment the question stem, options, answers, and explanations from the exam question data; for questions containing images or special format content, extract their text description information and store it in association with the original file; Based on the semantic similarity of the question content, duplicate question data is identified, merged, or removed.

3. The AI-assisted learning method for standardized qualification certification examinations according to claim 1, characterized in that, The verification of exam question data in a specific professional field includes cross-validation, external source validation, and manual validation. Cross-validation involves performing vector retrieval of the question stem in a pre-constructed temporary knowledge base to verify the uniqueness of the corresponding answer, and marking questions with conflicting answers as questionable questions. External source validation involves comparing the question stem and answer of the questionable questions with data from authoritative external knowledge sources. Manual validation involves a final review by domain professionals on questions that remain questionable after cross-validation and external source validation.

4. The AI-assisted learning method for standardized qualification certification examinations according to claim 1, characterized in that, The construction of a vector knowledge base that supports semantic retrieval specifically includes the following steps: Each question is divided into multiple text sub-blocks, consisting of the question stem, the standard answer, and the answer explanation. The Qwen3-Embedding model is used to generate corresponding word embedding vectors for each text sub-block. The text sub-blocks are associated and stored with their corresponding word embedding vectors and the metadata of the questions, and a hybrid retrieval index supporting vector retrieval and full-text retrieval is constructed.

5. The AI-assisted learning method for standardized qualification certification examinations according to claim 1, characterized in that, It also includes, after retrieving relevant knowledge fragments from the vector knowledge base based on the identified intent, calling a re-ranking model to rank the knowledge fragments; wherein, the re-ranking model is used to re-rank based on the multi-dimensional matching degree between the query intent and the retrieval results.

6. The AI-assisted learning method for standardized qualification certification examinations according to claim 1, characterized in that, The user profile includes real-time recording of incorrect answers, a question collection folder where users can customize their question collections, tags and notes, and an assessment of mastery of each knowledge point.

7. The AI-assisted learning method for standardized qualification certification examinations according to claim 6, characterized in that, The adaptive learning recommendations include a similar question reinforcement practice package automatically generated based on the incorrect question information, or a personalized review plan generated based on the mastery assessment.

8. An AI-assisted learning system for standardized qualification certification examinations, characterized in that, include: The data collection unit is used to collect exam question data in a specific professional field and to verify it. The knowledge base construction unit is used to convert the verified exam question data into vectors through a word embedding model in order to build a vector knowledge base that supports semantic retrieval. The intelligent retrieval and content generation unit is used to receive natural language learning requests input by users, and to identify exam-related intentions based on a large language model; and to retrieve relevant knowledge fragments from the vector knowledge base according to the identified intentions. Based on retrieval enhancement generation technology, knowledge fragments are synthesized into structured auxiliary learning content that meets the learning request; The user profile building unit is used to analyze user interaction data with structured learning content in order to build user profiles. The output unit is used to provide adaptive learning suggestions or resource recommendations based on user profiles. It also outputs structured learning aids and adaptive learning suggestions or resource recommendations to the user's terminal.

Images