Virtual inquiry practical training system for traditional Chinese medicine teaching

Abstract

The invention provides a virtual inquiry practical training system for traditional Chinese medicine teaching, and the system comprises a case and script configuration module which is used for configuring core physical sign information of cases and inquiry question and answer pairs configured for the cases based on a case library, storing the core physical sign information and the inquiry question and answer pairs in a vector database, and constructing a vector index at the same time; the inquiry module is used for calling core physical sign information of cases in a case library according to the inquiry authority of the user to carry out inquiry practical training; in the inquiry practical training process, questions input by a user are received, and answers provided by the mixed retrieval and large model secondary confirmation module are returned; the mixed retrieval and large model secondary confirmation module is used for receiving questions provided by the inquiry window module, obtaining a candidate inquiry question-answer pair set from the vector database through mixed retrieval, and carrying out secondary confirmation through a large model to obtain an optimal inquiry question-answer pair; and the inquiry scoring module is used for scoring the inquiry practical training process of the user according to a preset rule. The problems of less practical operation, poor feedback, difficult data tracing and the like of traditional Chinese medicine teaching can be solved.

Description

Technical Field

[0001] This invention relates to the technical field of integrating artificial intelligence with traditional Chinese medicine education, and in particular to a virtual consultation and training system for teaching traditional Chinese medicine. Background Technology

[0002] Currently, TCM clinical teaching generally relies on real patients or standardized patients (SPs) for consultation training. However, real patient resources are limited and their conditions are unpredictable; while SPs can be standardized, training costs are high, coverage of a limited number of diseases is limited, large-scale training is difficult, and they cannot dynamically respond to students' personalized questions. In recent years, some teaching platforms have attempted to introduce chatbots to assist in consultation practice, but these often use preset rules or limited states, lack the ability to understand complex semantics, have rigid interactions, and cannot simulate the language style, emotional changes, and disease progression of real patients.

[0003] Furthermore, existing systems rely heavily on manual scoring or simple keyword matching in the evaluation process, making it difficult to comprehensively assess students' core competencies such as dialectical thinking logic, the completeness of the consultation structure, and the accuracy of terminology usage. Therefore, there is an urgent need for a TCM virtual consultation training system that can combine the semantic understanding capabilities of large-scale models with the behavioral modeling capabilities of intelligent agents, supporting flexible script configuration, dynamic interactive responses, and intelligent scoring feedback. Summary of the Invention

[0004] This invention provides a virtual consultation and training system for traditional Chinese medicine teaching, in order to solve the problems in the background art.

[0005] Other features and advantages of the invention will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0006] According to a first aspect of the present invention, a virtual consultation training system for traditional Chinese medicine teaching is provided, comprising: The case and script configuration module is used to configure the core vital signs information of cases based on the case database and configure the corresponding question-and-answer pairs for each case, and store the question-and-answer pairs in the vector database, while building a vector index; The consultation module retrieves core vital signs information from cases in the case database based on the user's consultation permissions to conduct consultation training; during the consultation training, it receives questions input by the user and returns answers provided by the hybrid retrieval and large model secondary confirmation module; The hybrid retrieval and large model secondary confirmation module is used to receive questions provided by the consultation window module, obtain a set of candidate consultation question-answer pairs from the vector database through hybrid retrieval, obtain the optimal consultation question-answer pair through secondary confirmation by the large model, and provide the answer of the optimal consultation question-answer pair to the consultation window module. The consultation scoring module is used to score the user's consultation training process according to preset rules.

[0007] In one embodiment of this application, the consultation module includes level 1 consultation permissions, level 2 consultation permissions, and level 3 consultation permissions. Under level 1 consultation permissions, level 1 consultation training is conducted, where the consultation module assigns cases within the user's current permissions, and the user only needs to conduct consultation based on the core physical signs information of the cases. Under level 2 consultation permissions, level 2 consultation training is conducted, where the consultation module assigns cases within the user's current permissions, and the user needs to conduct consultation and dialectical analysis based on the core physical signs information of the cases. Under level 3 consultation permissions, level 3 consultation training is conducted, where the consultation module assigns cases within the user's current permissions, and the user needs to conduct consultation, dialectical analysis, and prescription based on the core physical signs information of the cases. When the score in level 1 consultation training exceeds a preset threshold, the user is upgraded to level 2 consultation permissions; when the score in level 2 consultation training exceeds a preset threshold, the user is upgraded to level 3 consultation permissions.

[0008] In one embodiment of this application, the consultation scoring module is configured with primary scoring rules, secondary scoring rules, tertiary scoring rules, and diagnostic and prescription questions for each case in the consultation modality; wherein, the primary scoring rules are used to score the comprehensiveness and logicality of consultation in the primary consultation training; the secondary scoring rules are used to score the comprehensiveness, logicality, and diagnostic results of consultation in the secondary consultation training; and the tertiary scoring rules are used to score the comprehensiveness, logicality, diagnostic results, and prescriptions of consultation in the tertiary consultation training.

[0009] In one embodiment of this application, the construction of the vector index in the case and script configuration module specifically includes: For each question-answer pair, sparse and dense vectors are generated and stored in the vector database. An inverted index is built for the sparse vectors and an HNSW index is built for the dense vectors.

[0010] In one embodiment of this application, the sparse vector generation process includes: The TF-IDF algorithm is used to segment the question part of the question-answer pair in the consultation, filter out stop words and extract keywords to generate sparse vectors.

[0011] In one embodiment of this application, the dense vector generation process includes: The Sentence-BERT pre-trained model is used to encode the question part of the question-answer pair into a dense vector of a preset dimension.

[0012] In one embodiment of this application, the specific process of hybrid retrieval in the hybrid retrieval and large model secondary confirmation module includes: Generate sparse and dense vectors based on the user's query; The sparse vector generated based on the question is matched with the vector database to obtain the similarity between the generated sparse vector and the sparse vector of the question-and-answer pair in the vector database. The dense vector generated based on the question is matched with the vector database to obtain the similarity between the generated dense vector and the dense vector of the question-and-answer pair in the vector database; The similarity obtained from sparse vector matching and dense vector matching are weighted and calculated to obtain the comprehensive score of each consultation question and answer pair. The top 3 consultation question and answer pairs with the highest comprehensive scores are selected as candidate consultation question and answer pairs and stored in the candidate consultation question and answer pair set.

[0013] In one embodiment of this application, the specific process of secondary confirmation in the hybrid retrieval and large model secondary confirmation module is as follows: The system uses a large model to obtain user questions, candidate question-answer pairs, and the current case context, and analyzes the core intent of the questions. Compare the question intent and core intent of each question-answer pair in the candidate consultation question-answer pair set one by one, analyze the differences in expression, and evaluate the degree of intent matching; Based on intent matching degree, expression similarity, and case context relevance, the confidence level of each candidate consultation question-answer pair is calculated, and the candidate consultation question-answer pair with the highest confidence level is selected as the optimal consultation question-answer pair; if the confidence level of all candidate consultation question-answer pairs is less than the preset threshold, it means that no question-answer pair was matched, and the corresponding answer is generated by the large model based on the basic case information.

[0014] In one embodiment of this application, the hybrid retrieval and large model secondary confirmation module further includes an output optimization process: The large model is invoked to perform natural language optimization on the best question-answer pair based on the way the user asks questions and the context of the case. If the user's question is a compound question, then the answers from multiple optimal question-and-answer pairs corresponding to the compound question will be integrated to form a coherent response.

[0015] In one embodiment of this application, a question-and-answer record and data statistics module is also included, which is used to record the consultation process of each user and to collect user consultation information and provide a viewing interface; at the same time, it provides a display of user answers and a rating leaderboard for individual cases.

[0016] This application addresses the problems of insufficient practical training, poor feedback, and difficulty in data traceability in traditional Chinese medicine teaching by using large-scale patient simulation models and modular system support. It achieves large-scale practical training, precise feedback, data visualization, and balanced resource allocation, comprehensively improving teaching quality and efficiency. Specific technical effects are as follows: 1. Reduce practical training costs and effectively improve practical training efficiency. By constructing virtual patients, a 24-hour uninterrupted consultation scenario can be provided without human intervention throughout the entire process. This can significantly reduce the time teachers spend on practical training guidance and lower labor costs. At the same time, students can initiate simulated consultations independently without relying on the cooperation of teachers or standardized patients. This can meet the practical training needs of multiple students at the same time and efficiently cover high-frequency practical training scenarios across the entire school.

[0017] 2. Enrich case studies and enhance disease management capabilities. The case and script configuration module can associate a massive number of cases and edit core signs such as tongue appearance and pulse, restoring the process of "gradual presentation of symptoms" in real consultations, allowing students to be exposed to rare and difficult diseases, and significantly improving their ability to make multidimensional judgments about diseases.

[0018] 3. Precise feedback to improve learning efficiency By customizing scoring criteria and displaying consultation scores and details, the accuracy of feedback is greatly improved, effectively enhancing learning efficiency.

[0019] 4. Data traceability to facilitate personalized teaching By retaining complete consultation data and outputting information such as the accuracy rate of diagnosis and the compliance rate of prescriptions, it is possible to accurately identify the user's weaknesses and significantly improve the effectiveness of teaching evaluation and personalized teaching; students can also review their learning progress and summarize and improve themselves.

[0020] 5. Break down geographical limitations and balance teaching resources The system establishes a standardized training platform, allowing students from grassroots colleges and universities to participate in high-quality virtual training through the consultation module. This provides them with teaching feedback consistent with that of colleges and universities in core cities, effectively narrowing the gap in training conditions between different regions and promoting the balanced development of talent cultivation quality.

[0021] 6. Improve the accuracy and reliability of question-and-answer matching. By employing a dual mechanism of vector semantic matching and large-scale model secondary confirmation, the system can quickly filter out candidate question-answer pairs with similar semantics through vector matching, overcoming the limitations of single keyword matching. At the same time, it avoids false matching through deep intent analysis of the large-scale model, ensuring that the hit results are highly consistent with the student's question intent, thus significantly improving the accuracy and reliability of question-answer matching. Attached Figure Description

[0022] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0023] Figure 1This is a schematic diagram of the virtual consultation and training system for traditional Chinese medicine teaching proposed in the embodiments of this application. Detailed Implementation

[0024] The embodiments of this application are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar modules or modules having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application. Rather, the embodiments of this application include all variations, modifications, and equivalents falling within the spirit and scope of the appended claims.

[0025] Addressing core issues in Traditional Chinese Medicine (TCM) teaching, such as scarce clinical training resources, weak feedback mechanisms, and unquantifiable learning processes, this application's embodiments construct a virtual consultation and training system for TCM students. This system, oriented towards skill advancement and covering the entire diagnostic and treatment process, deeply integrates Large Language Model (LLM) and intelligent agent technology. Unlike existing general medical consultation simulation systems, this application does not simply simulate patient dialogue but designs an assessable, traceable, and evolving teaching loop based on the TCM logic of "four diagnostic methods—syndrome differentiation—treatment." Specifically, it innovatively constructs a dual mechanism of "vector semantic matching + large model secondary confirmation" using a "vector database + large model" architecture to solve the problem of unstable large model output, while simultaneously building a standardized training process and evaluation system.

[0026] Please refer to Figure 1 The virtual consultation training system for TCM teaching mainly includes a case and script configuration module, a consultation module, a hybrid retrieval and large model secondary confirmation module, and a consultation scoring module.

[0027] The system comprises several modules: a case and script configuration module, a case and script configuration module, and a consultation module. The case and script configuration module is used to configure the core vital signs information of cases based on the case database and to configure corresponding consultation question-and-answer pairs for each case. These pairs are stored in a vector database, and a vector index is constructed. The consultation module is used to retrieve the core vital signs information of cases from the case database based on the user's consultation permissions to conduct consultation training. During consultation training, it receives user-input questions and returns answers provided by the hybrid retrieval and large-scale model secondary confirmation module. The hybrid retrieval and large-scale model secondary confirmation module receives questions from the consultation window module, retrieves a set of candidate consultation question-and-answer pairs from the vector database through hybrid retrieval, performs secondary confirmation through the large-scale model to obtain the optimal consultation question-and-answer pair, and provides the answer to the optimal consultation question-and-answer pair to the consultation window module. The consultation scoring module is used to score the user's consultation training process according to preset rules.

[0028] To address the issue of content illusion or semantic drift that can easily occur in traditional Chinese medicine (TCM) professional contexts using large-scale models, this embodiment proposes a RAG enhancement architecture. This architecture involves pre-setting standardized question-and-answer pairs and storing them in a vector database. When a user asks a question, a semantically similar candidate set is first obtained through vector retrieval. Then, the large-scale model performs intent verification and context optimization on the candidate set, ultimately outputting an accurate and traceable answer. This mechanism ensures both the professional accuracy of the answer and the flexibility of natural language interaction, effectively avoiding mismatches and missed matches, and is a core technology supporting the reliability of teaching. In this embodiment, it is mainly implemented through two parts: a case and script configuration module and a hybrid retrieval and large-scale model secondary confirmation module.

[0029] In this embodiment, the case and script configuration module is mainly divided into a case basic information editing module and a consultation script. Users can link to the case database through the case basic information editing module and edit the core physical signs information of each case in the database. The core physical signs information includes at least tongue appearance and pulse characteristics. For each case, a complete consultation question-and-answer pair can be edited and scoring points can be set through the consultation script. In practical applications, consultation question-and-answer pairs can be imported in batches to complete the editing. It should be noted that the edited consultation question-and-answer pair should cover the core symptoms, physical signs, medical history, and other key information of the case. For example, for a case of "wind-heat common cold," the preset consultation question is "Do you have fever symptoms?" The corresponding answer is "Significant fever, body temperature approximately 38.5℃."

[0030] After editing the question-and-answer pairs of the cases, they are stored in a vector database, and a vector index is built to provide data support for dual matching.

[0031] Furthermore, this embodiment provides a specific method for constructing a vector index. For each question-answer pair, two types of vectors, sparse and dense, are generated and stored in a vector database (Milvus is preferred). An inverted index is constructed for the sparse vectors to improve keyword matching efficiency. An HNSW index is constructed for the dense vectors, with parameters configured as M=16 and efConstruction=200 to ensure semantic retrieval speed, thus forming a hybrid retrieval foundation supported by a "dual vector index".

[0032] Specifically, sparse vectors represent keyword features. In this embodiment, the TF-IDF algorithm is used to segment the question portion of the consultation question-answer pair, filter stop words, and extract keywords (such as "fever," "sore throat," and "pulse") to generate sparse vectors. (The dimension is the size of the system vocabulary, which is 100,000 dimensions by default. Only the dimension corresponding to the keyword has a non-zero weight value.)

[0033] Dense vectors represent semantic features. In this embodiment, the Sentence-BERT pre-trained model (chinese-roberta-wwm-ext) is used to encode the question portion of the question-answer pair into a 768-dimensional dense vector. The core semantics of the problem are captured through the mean pooling strategy.

[0034] The hybrid retrieval and large model secondary confirmation module combines the precise matching capability of hybrid retrieval from the vector library with the intent understanding capability of the large model to retrieve the corresponding answers from the vector database based on the user's question, thus simulating a real consultation scenario.

[0035] Specifically, after a user submits a question, the hybrid retrieval and large-model secondary confirmation module uses a hybrid retrieval algorithm to filter out a set of candidate question-answer pairs with similar semantics. The specific process is as follows: First, sparse and dense vectors are generated based on the user's question text (such as "Are you hot? Do you have a sore throat?").

[0036] Then, a dual-path retrieval is performed in the vector database, including the keyword path and the semantic path.

[0037] Keyword Path: The inverted index is used to match question-and-answer pairs containing core keywords in the vector database. Specifically, sparse vectors generated from the question are matched against the vector database to obtain the similarity between the sparse vector and the sparse vectors of the question-and-answer pairs in the database. .

[0038]

[0039] in Represents the similarity of sparse vectors for precise keyword matching. Ask a question to the user sparse vectors, Let be a sparse vector of the question-and-answer pairs. For the vocabulary dimension, These represent the TF-IDF weights for the corresponding dimensions. In this embodiment, the similarity of sparse vectors is represented by the dot product similarity. The magnitude of the dot product similarity can be used to determine which question-answer pairs in the vector database contain the core keywords from the user's question.

[0040] Semantic Path: Retrieve semantically similar question-and-answer pairs using the HNSW index. Specifically, the dense vector generated based on the question is matched against a vector database to obtain the similarity between this dense vector and the dense vectors of the question-and-answer pairs in the vector database. :

[0041] in, Cosine similarity for dense vectors captures semantic associations; Ask a question to the user Dense vectors, For dense vectors of question-and-answer pairs in a medical consultation, For vector dot product, , These are the L2 norms of the vectors. In this embodiment, cosine similarity is used to characterize the similarity of dense vectors. The magnitude of the cosine similarity can be used to determine which question-answer pairs in the vector database are semantically similar to the user's question.

[0042] Finally, the similarity scores obtained from sparse vector matching and dense vector matching are weighted and calculated to obtain the comprehensive score for each question-answer pair. Three question-and-answer pairs with the highest to lowest overall scores were selected as candidate question-and-answer pairs and stored in the candidate question-and-answer pair set. In this embodiment, to ensure coverage and control subsequent confirmation costs, three question-answer pairs are selected as candidate question-answer pairs. In practical applications, the number can be adjusted. Overall Score The calculation formula is as follows:

[0043] in, Ask a question to the user With the A consultation Q&A The overall score (range 0-1). This is the semantic similarity weight coefficient (0≤β≤1, default value 0.6, can be customized).

[0044] To avoid erroneous search results due to mixed retrieval matching, this embodiment also utilizes a large model to perform a secondary confirmation process through a mixed retrieval and a large model secondary confirmation module. That is, the large model is used to verify the candidate question-answer pair set. Perform deep intent verification to ensure the results match the user's question. The secondary confirmation process is as follows: (1) Precise intent parsing: The large model receives user questions, candidate question-answer pairs, and the current case context, and parses the core intent of the question. For example, "Do you feel hot?" is parsed as "asking about fever symptoms"; "Do you have a sore throat?" is parsed as "asking about sore throat symptoms".

[0045] (2) Intent comparison: By comparing the question intent and core intent of each question and answer pair in the candidate question and answer pair set, and analyzing the differences in expression (such as synonym substitution, sentence structure change, colloquial omission, etc.), the degree of intent matching is evaluated.

[0046] (3) Confidence calculation and result judgment: Based on the degree of intent matching, expression similarity and case context relevance, the confidence of each candidate consultation question-answer pair is calculated. The candidate question-answer pair with the highest confidence level is selected as the optimal question-answer pair.

[0047] In one embodiment, a reliability threshold needs to be preset. If the confidence level of any candidate question-and-answer pair is greater than the preset threshold, the candidate question-and-answer pair with the highest confidence level is selected as the optimal pair, indicating a successful match. If the confidence level of all candidate question-and-answer pairs is less than the preset threshold, it indicates a failed match. In this case, a reasonable answer is generated by the large model based on the basic information of the case. For example, if a failed match is not found, the large model generates a reasonable response based on the core information of the case, while avoiding speculative content (e.g., in a case of "wind-heat cold" without dizziness, if the user asks "Are you dizzy?", the generated answer is "I don't feel dizzy; my main symptoms are fever and sore throat"), and prompts the user to ask questions about the core symptoms related to the case.

[0048] To ensure the naturalness of the output, in this embodiment, after the hybrid retrieval and large-model secondary confirmation module determines the answer, the output is further optimized by calling the large model. Specifically, the large model performs natural language optimization based on the way the user asks the question (e.g., colloquial, complex questions) and the case context, combined with the determined optimal question-answer pair or a reasonable answer generated based on basic case information. Natural language optimization mainly includes sentence structure adaptation and multi-intent integration.

[0049] Sentence adaptation adjusts the preset written answers to match the tone of the user's question (e.g., when a user asks in a conversational tone, "Are you hot?", the preset answer "I have a noticeable fever, and my body temperature is about 38.5℃" is optimized to "I have a noticeable fever, and my body temperature is about 38.5℃").

[0050] Multi-intent integration is mainly for users who ask complex questions (such as "Are you hot? Do you have a sore throat?"). By integrating the answers to the optimal question-answer pairs corresponding to each sub-question, a coherent response is formed (such as "I have a significant fever, with a body temperature of about 38.5℃, and a severe sore throat").

[0051] In practical applications, the system can also automatically record information such as mixed search scores, confidence levels, and answers before and after optimization, providing quantitative basis for subsequent scoring and review.

[0052] The hybrid retrieval and large-model secondary confirmation module in this embodiment solves the limitations of single keyword matching or semantic matching through hybrid retrieval, and ensures matching accuracy through large-model secondary confirmation. At the same time, it ensures the naturalness of the output through preset answer optimization, completely avoiding the "content illusion" of the large model, and realizing the traceability and quantification of the consultation process.

[0053] The consultation module serves as the training entry point for the virtual consultation training system for Traditional Chinese Medicine (TCM) teaching in this embodiment. It is primarily designed for training users, such as students. After logging into the system through the consultation module, users can access the core physical signs information of cases in the case database to conduct consultation training, based on their consultation permissions. Specifically, users conduct the consultation process based on the core physical signs information of the case. They input questions in the display window provided by the consultation module. After the questions are retrieved, confirmed, and optimized by the hybrid retrieval and large-model secondary confirmation module, the answers are displayed in the display window. Users can then continue to ask questions based on the answers until the consultation of the current case is completed.

[0054] In this embodiment, the consultation module provides users with three levels of permissions, including level one consultation permission, level two consultation permission, and level three consultation permission.

[0055] Under Level 1 consultation access, users can conduct Level 1 consultation training, where they only need to conduct consultations based on the core vital signs information of the case. The system will assign cases that match the current level for questioning and answering. Notably, cases at each level can be pre-configured.

[0056] Under Level 2 consultation access, users can conduct Level 2 consultation training, where they need to conduct consultations and dialectical analysis based on the core physical signs of the case. In this Level 2 consultation training, in addition to assigning cases that match the current level for questioning and answering, the system also provides dialectical questions, requiring users to provide a dialectical conclusion after the consultation.

[0057] Under the three-level consultation permission, users can conduct three-level consultation training, where they need to conduct consultations, dialectical analysis, and prescription based on the core physical signs and symptoms of a case. In implementing the three-level consultation, building upon the two-level consultation training, the system also provides prescription questions, which users need to write after the consultation.

[0058] In this embodiment, when a user logs in for the first time, the user is given Level 1 consultation privileges by default. When the user's score in the Level 1 consultation training is greater than a preset threshold (e.g., 80 points), the user is upgraded to Level 2 consultation privileges. When the user's score in the Level 2 consultation training is greater than the preset threshold (80 points), the user is upgraded to Level 3 consultation privileges.

[0059] Corresponding to the consultation module, the consultation scoring module in this embodiment is also configured with three levels of rules, including first-level scoring rules, second-level scoring rules, and third-level scoring rules, as well as diagnostic questions and prescription questions for each case in the consultation modality; wherein, the first-level scoring rules are used to score the comprehensiveness and logicality of consultation in the first-level consultation training; the second-level scoring rules are used to score the comprehensiveness, logicality, and diagnostic results of consultation in the second-level consultation training; and the third-level scoring rules are used to score the comprehensiveness, logicality, diagnostic results, and prescriptions of consultation in the third-level consultation training.

[0060] In this embodiment, the diagnosis and prescription questions for each case can be pre-edited in the consultation and scoring module, along with corresponding scoring rules. Generally, the scoring rules for the diagnosis and prescription questions are to determine whether the diagnosis conclusion, prescription, and herbal compatibility are correct; if correct, points are awarded.

[0061] Specifically, during Level 1 consultation training, the system defaults to a hybrid retrieval + large-model secondary confirmation mode, focusing on training students' ability to ask questions about core symptoms and their standardized expression habits. The consultation scoring module focuses on the success rate of hybrid retrieval, helping users identify problems such as "inaccurate question keywords" and "vague semantic expression."

[0062] When conducting training on secondary-level medical consultations, a diagnostic and answering step is added to the primary-level consultation. The system also adopts a hybrid retrieval + large-model secondary confirmation mode. Users need to submit a diagnostic conclusion based on the consultation results (such as "the syndrome type is wind-heat common cold"). The consultation scoring module simultaneously scores the consultation performance (including indicators related to hybrid retrieval and secondary confirmation) and the accuracy of the diagnostic diagnosis from two dimensions.

[0063] When conducting Level 3 consultation training, after users obtain Level 3 consultation access, they can independently select cases from different departments (such as internal medicine and gynecology) provided in the case database. Building upon Level 2 consultation, a new component adds prescription response and herbal compatibility response stages, supporting users to directly conduct open-ended consultations using a large-scale model (adapting to advanced training needs). The consultation scoring module, based on Level 2 consultation scoring, incorporates prescription compliance and the rationality of herbal compatibility into the scoring system, forming a complete clinical training loop of consultation-diagnosis-prescription.

[0064] The consultation and scoring modules in this application embodiment realize a multi-level progressive practical training for TCM teaching. Level 1 consultation focuses on symptom collection skills training (only the consultation stage is open); Level 2 consultation introduces syndrome differentiation and answering questions, training syndrome identification skills; Level 3 consultation extends to prescription writing and herbal compatibility, forming a complete clinical thinking loop of "consultation → syndrome differentiation → prescription". This training structure dynamically adapts to students' cognitive development patterns and automatically advances through comprehensive scoring (≥80 points), achieving a leap in ability from basic consultation to advanced clinical decision-making.

[0065] In one embodiment, the TCM teaching virtual consultation training system also includes a question-and-answer record and data statistics module. This module records each user's consultation process and compiles user consultation information, providing a viewing interface. It also displays user responses and a ranking list for individual cases. After completing the training, users can use this module to query statistical information about the consultation (such as question-and-answer accuracy and consultation duration) and personal learning statistics (such as pass rates at each level and average scores) to intuitively understand their training performance. They can also view their personal historical dialogues and scores, and obtain recommendations for excellent consultation records and related cases to specifically reinforce weak areas. This question-and-answer record and data statistics module also provides administrators (such as teachers) with the ability to view the consultation records for each case and user consultation details, as well as student ranking lists and statistical data, enabling a visualized evaluation of teaching effectiveness.

[0066] For those skilled in the art, the specific meanings of the above terms in this invention can be understood according to the specific circumstances; the accompanying drawings in the embodiments are used to clearly and completely describe the technical solutions in the embodiments of this invention. Obviously, the described embodiments are some embodiments of this invention, but not all embodiments. Generally, the components of the embodiments of this invention described and shown in the accompanying drawings can be arranged and designed in various different configurations.

[0067] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A virtual consultation and training system for traditional Chinese medicine teaching, characterized in that, include: The case and script configuration module is used to configure the core vital signs information of cases based on the case database and configure the corresponding question-and-answer pairs for each case, and store the question-and-answer pairs in the vector database, while building a vector index; The consultation module retrieves core vital signs information from cases in the case database based on the user's consultation permissions to conduct consultation training; during the consultation training, it receives questions input by the user and returns answers provided by the hybrid retrieval and large model secondary confirmation module; The hybrid retrieval and large model secondary confirmation module is used to receive questions provided by the consultation window module, obtain a set of candidate consultation question-answer pairs from the vector database through hybrid retrieval, obtain the optimal consultation question-answer pair through secondary confirmation by the large model, and provide the answer of the optimal consultation question-answer pair to the consultation window module. The consultation scoring module is used to score the user's consultation training process according to preset rules.

2. The virtual consultation and training system for traditional Chinese medicine teaching according to claim 1, characterized in that, The consultation module includes first-level consultation permissions, second-level consultation permissions, and third-level consultation permissions. Under the first-level consultation permission, first-level consultation training is carried out. The consultation module assigns cases under the current permissions to the user, and the user only needs to conduct consultation based on the core vital signs information of the case. Under Level 2 consultation privileges, Level 2 consultation training is conducted. The consultation module assigns cases within the user's current privileges, and the user needs to conduct consultation and dialectical analysis based on the core physical signs of the case. Under Level 3 consultation privileges, Level 3 consultation training is conducted. The consultation module assigns cases within the user's current privileges, and the user needs to conduct consultation, dialectical analysis, and prescribe medicine based on the core physical signs of the case. When the score of Level 1 consultation training exceeds a preset threshold, the user is upgraded to Level 2 consultation privileges. When the score of the Level 2 consultation training exceeds the preset threshold, the user is upgraded to Level 3 consultation privileges.

3. The virtual consultation and training system for traditional Chinese medicine teaching according to claim 2, characterized in that, The consultation scoring module is configured with primary scoring rules, secondary scoring rules, and tertiary scoring rules, as well as diagnostic and prescription questions for each case in the consultation modality. The primary scoring rules are used to score the comprehensiveness and logicality of the consultation in the primary consultation training; the secondary scoring rules are used to score the comprehensiveness, logicality, and diagnostic results in the secondary consultation training; and the tertiary scoring rules are used to score the comprehensiveness, logicality, diagnostic results, and prescriptions in the tertiary consultation training.

4. The virtual consultation and training system for traditional Chinese medicine teaching according to claim 1, characterized in that, The specific steps involved in constructing the vector index within the case and script configuration module are as follows: For each question-answer pair, sparse and dense vectors are generated and stored in the vector database. An inverted index is built for the sparse vectors and an HNSW index is built for the dense vectors.

5. The virtual consultation and training system for traditional Chinese medicine teaching according to claim 4, characterized in that, The process of generating the sparse vector includes: The TF-IDF algorithm is used to segment the question part of the question-answer pair in the consultation, filter out stop words and extract keywords to generate sparse vectors.

6. The virtual consultation and training system for traditional Chinese medicine teaching according to claim 4 or 5, characterized in that, The process of generating the dense vector includes: The Sentence-BERT pre-trained model is used to encode the question part of the question-answer pair into a dense vector of a preset dimension.

7. The virtual consultation and training system for traditional Chinese medicine teaching according to claim 4, characterized in that, The specific process of hybrid retrieval in the hybrid retrieval and large model secondary confirmation module includes: Generate sparse and dense vectors based on the user's query; The sparse vector generated based on the question is matched with the vector database to obtain the similarity between the generated sparse vector and the sparse vector of the question-and-answer pair in the vector database. The dense vector generated based on the question is matched with the vector database to obtain the similarity between the generated dense vector and the dense vector of the question-and-answer pair in the vector database; The similarity obtained from sparse vector matching and dense vector matching are weighted and calculated to obtain the comprehensive score of each consultation question and answer pair. The top 3 consultation question and answer pairs with the highest comprehensive scores are selected as candidate consultation question and answer pairs and stored in the candidate consultation question and answer pair set.

8. The virtual consultation and training system for traditional Chinese medicine teaching according to claim 7, characterized in that, In the hybrid retrieval and large model secondary confirmation module, the specific process of secondary confirmation is as follows: The system uses a large model to obtain user questions, candidate question-answer pairs, and the current case context, and analyzes the core intent of the questions. Compare the question intent and core intent of each question-answer pair in the candidate consultation question-answer pair set one by one, analyze the differences in expression, and evaluate the degree of intent matching; Based on intent matching degree, expression similarity, and case context relevance, the confidence level of each candidate consultation question-answer pair is calculated, and the candidate consultation question-answer pair with the highest confidence level is selected as the optimal consultation question-answer pair; if the confidence level of all candidate consultation question-answer pairs is less than the preset threshold, it means that no question-answer pair was matched, and the corresponding answer is generated by the large model based on the basic case information.

9. The virtual consultation and training system for traditional Chinese medicine teaching according to claim 8, characterized in that, The hybrid retrieval and large model secondary confirmation module also includes an output optimization process: The large model is invoked to perform natural language optimization on the best question-answer pair based on the way the user asks questions and the context of the case. If the user's question is a compound question, then the answers from multiple optimal question-and-answer pairs corresponding to the compound question will be integrated to form a coherent response.

10. The virtual consultation and training system for traditional Chinese medicine teaching according to claim 1, characterized in that, It also includes a question and answer record and data statistics module, which records each user's consultation process and compiles user consultation information, and provides a viewing interface; it also provides a display of user answers and a rating leaderboard for individual cases.

Images