Large-model campus question and answer generation method and system based on multi-source data
By integrating campus multi-source data and utilizing intention classification and large language models, the problem of insufficient coverage, real-timeness and accuracy in campus Q&A is solved, and an efficient and convenient Q&A solution is provided.
Patent Information
- Application Number
- CN202510471073.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-15
- Publication Date
- 2025-08-01
AI Technical Summary
It is difficult for the existing technology to achieve high coverage, high real-time, and high accuracy Q&A in campus scenarios, and there are problems such as misleading and insufficient real-time in large models and manual answers.
By integrating campus multi-source data, using intention classification tools and large language models to generate a question-and-answer system, real-time synchronization and accurate answers of data are achieved.
Achieve campus Q&A with high coverage, high accuracy, high real-time and convenience, reducing human resource consumption and misleading risks.
Smart Images
Figure CN120407738A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of large model Q&A, and particularly relates to a campus Q&A generation method based on the collection and integration of multi-source big data and large model Q&A. Background Art
[0002] Campus Q&A is a high-frequency scenario in colleges and universities. All kinds of campus activities, campus organizations, and campus common sense will bring heavy Q&A tasks to the relevant responsible persons. However, most of the questions are highly repetitive and require a large amount of campus resources.
[0003] In response to this situation, colleges and universities generally adopt methods such as notice pushing, document integration, and counselor publicity, which can cover most students; some colleges and universities use large models to provide Q&A for specific scenarios and can answer 24 hours a day; there are also colleges and universities that provide Q&A manually by teachers and students.
[0004] The above three types of Q&A methods are difficult to achieve high coverage, high real-time performance, and high accuracy. At the same time, they consume human resources, and the wrong answers given by the large model may cause misunderstandings among students. Their important defects are as follows:
[0005] 1. Insufficient coverage. In the campus scenario, there are a large number of various information public accounts and notice platforms, which are numerous and scattered. It is difficult for students to pay attention to all of them and quickly get accurate Q&A push. Students also need to ask each other, consult counselors, etc. to understand the relevant Q&A content.
[0006] 2. Insufficient accuracy of large model answers. The Q&A of the large model is based on the knowledge of the large model, but for a refined scenario like the campus, it is difficult to efficiently give accurate answers, and wrong answers may mislead students.
[0007] 3. Limited application scope of large model answers. The large model can only be used for common sense and subject-related Q&A, and it is difficult to provide effective answers for school-based scenarios such as campus activity arrangements.
[0008] 4. Real-time performance of manual answers cannot be guaranteed. The real-time performance of manual answers depends on people's working hours and working states, and the Q&A in the campus scenario often occupies the free time of students and counselors, and the real-time performance cannot be guaranteed.
[0009] 5. Manual answers are prone to errors. Manual answers are prone to giving wrong answers due to the unfamiliarity of the Q&A personnel with the activity content.
[0010] Therefore, how to provide a campus Q&A scenario solution with high coverage, high real-time performance, and high accuracy is an urgent problem that needs to be solved by technical personnel and management personnel in this field at present. Summary of the Invention
[0011] To solve the above problems, the present invention provides a large model campus Q&A generation method and system based on multi-source data; by exploring and configuring campus source data, using an intent classification tool to classify the user's questions by intent, and directionally screening data according to the question intent, and then using a large language model to generate answers to the questions based on the screened data.
[0012] To achieve the above object, the present invention is realized through the following technical solutions:
[0013] A large model campus Q&A generation method based on multi-source data, comprising the following steps:
[0014] Step 1: Analyze the Q&A requirements of the campus scenario, import data sources according to the campus Q&A requirements, and the data sources include campus real-time information data (data updated over time from the campus information system) and campus information text materials (text-based campus information materials published in the form of rules, documents, and articles).
[0015] Step 2: Connect the campus Q&A system with the campus information system, sort out the sources of all data, and establish a real-time information data foundation for question answering; at the same time, connect with the campus information text materials and sort out the available relevant materials.
[0016] Step 3: Train a text classifier as an intent processing tool.
[0017] Step 4: The user performs identity authentication at the system user end, and the system initially saves the user's information according to the student number or employee number information.
[0018] Step 5: The user inputs the question to be asked through the system user end.
[0019] Step 6: The system obtains the question input by the user, performs intent classification through the intent processing tool, obtains the corresponding label in the question intent, and restricts the data retrieval range according to the user's personal information, and finds relevant data information helpful for answering the question through vector search within the restricted retrieval range.
[0020] Step 7: Integrate the searched data and information with the question content, and use a large language model to generate an answer to the question based on the data information and feedback it to the user end for answering questions.
[0021] Further, the specific operation of step 2 is as follows:
[0022] 2-1: The described campus information system includes an academic affairs system, a student affairs system, and a unified identity authentication system for the network center. By docking with the campus information system, it synchronizes student class schedules, student lists, teacher lists, counselors, and class information in real time, and updates the notices and academic situation policies issued by the school in real time. It sorts out the sources of all data, that is, the specific departments, affiliated colleges, and grades that release the data, integrates data from different sources, tags different data usage scopes, and restricts the data usage scope through tags. For example, the "Training Plan for the Class of 2022 in the School of Computer Science" should be tagged with "School of Computer Science", "Class of 2022", "Computer Science and Technology", "Software Engineering", etc.
[0023] 2-2: Provide campus real-time information data through the api interface, pass the tag as a parameter of the api interface, and at the same time configure an authentication service for the api to ensure data security.
[0024] 2-3: Segment and vectorize campus information text materials, use Moka Massive Mixed Embedding to convert the text into vectors, and implement text information retrieval through the retrieval module. The retrieval module calculates the similarity of the vectors corresponding to the text, sorts them in descending order of similarity, and selects the text with high similarity as the retrieval result.
[0025] 2-4: Individually encapsulate the api interfaces and retrieval modules in 2-2 and 2-3 into call methods, that is, functions, so that the corresponding campus real-time information data or campus information text materials can be obtained through the call of the corresponding functions.
[0026] Functions are mainly divided into two types. One type does not require passing parameters and directly calls the service interface according to the user's login information to obtain campus real-time information data; the other type requires the user's question as a parameter, and also converts the question into a vector through the M3E (Moka Massive Mixed Embedding) model. By calculating the cosine similarity between the question vector and the text vector, the text corresponding to the maximum value of the result is selected as the retrieval result. The specific calculation method is as follows:
[0027] Cosine Similarity(A,B)={A·B} / {|A||B|}
[0028] Among them, A and B represent two text vectors for which the similarity needs to be compared.
[0029] 2-5: Construct a hash index structure, use the tag set of campus real-time information data or campus information text materials as the key of the hash, and use the function to obtain the corresponding data as the value to obtain a multi-source data support set, as follows:
[0030] {(label11, label12, …): function1,
[0031] (label21, label22, …): function2, …}
[0032] Among them, the labels included in the set are the labels of campus real-time information data or campus information text materials, and the function is a retrieval execution function that can retrieve data or information within the corresponding label constraint range through API calls or vector knowledge bases.
[0033] 2 - 6: Merge the key sets of the multi-source data support sets, that is, insert the keys of all hash objects into a new set to obtain the total set of all labels of the data set, denoted as InfoLabelSet.
[0034] Furthermore, the specific implementation of step 3 is as follows:
[0035] Using InfoLabelSet and the data of each data source, with the help of the RBT3 Chinese language classification model based on BERT, use the data source data and its corresponding labels as the data set to train a text classifier as an intent processing tool.
[0036] Furthermore, the specific operations of step 4 are as follows:
[0037] The user inputs the student ID or staff number, password, and conducts unified identity authentication verification with the network center at the system user end. After passing, the user enters the system, and the system initially saves the user's information according to the student ID or staff number information (student information such as student ID, name, grade, major, class, counselor, etc.; teacher information such as staff number, name, teaching information, class schedule, etc.).
[0038] 4 - 1: After the user logs in, the user end submits the student ID or staff number information of the user to the system processing end and applies for the user's personal information.
[0039] 4 - 2: After receiving the application, the processing end queries the campus database using the student ID or staff number, returns the user information, and saves it as IdentityLabelSet at the same time.
[0040] Furthermore, the specific operations of step 6 are as follows:
[0041] 6 - 1: Perform intent classification through the model trained in step 3, select the 3 categories with the highest confidence as the question intents, obtain the corresponding labels in each question intent, and construct the label set of the question intent, denoted as QuestionLabelSet.
[0042] 6-2: Take the union of the IdentityLabelSet extracted from the user login information and the QuestionLabelSet to obtain the retrieval set SearchLabelSet. The calculation method is as follows:
[0043] SearchLabelSet = IdentityLabelSet U QuestionLabelSet
[0044] 6-3: Based on the matching comparison between SearchLabelSet and InfoLabelSet, confirm the specific functions to be called through the multi-source data support set, and obtain the set of all functions to be called FunctionSet.
[0045] 6-4: Call the functions in FunctionSet in sequence to obtain data and information materials.
[0046] Furthermore, retain the question history. Users can follow up through the user terminal, and the system combines the question history through the large language model to continue generating targeted answers.
[0047] A large model campus Q&A system based on multi-source data, including a configuration terminal, a user terminal, and a processing terminal:
[0048] Configuration terminal: Import data from different fields by docking with the campus information system, sort out the sources of all data and configure label information for the data, import campus information text materials through the method of text material upload, and provide campus real-time information data through the api interface.
[0049] User terminal: Authenticate the user. After successful verification, initially save the user's information according to the student number or work number information, including the question feedback function and the function of viewing reference materials and documents. The user terminal provides a question input port for the user. After sending the question input by the user to the processing terminal for processing, display the output answer of the processing terminal.
[0050] Processing terminal: Includes a large model call module and a data storage module.
[0051] Model call module, through calling the model, perform auxiliary intent recognition and answer generation work. Classify the intent by calling the intent processing tool, obtain the corresponding label in the question intent, and constrain the data retrieval range according to the user's personal information. Search for relevant data information helpful for question answering in the data storage module within the constrained retrieval range through vector search; call the large language model to generate an answer to the question based on the data information and feedback it to the user terminal.
[0052] The data storage module stores the imported data information, and also stores the API call link, user identity information, and user Q&A logs.
[0053] Advantages of the present invention
[0054] For various Q&A scenarios, the following can be achieved:
[0055] 1. High coverage rate. All users can obtain all the required information through a single dialogue window and can directly ask questions when they have information needs, highly covering all users on campus.
[0056] 2. High accuracy. The answers corresponding to the questions are generated strictly referring to text materials or campus data APIs. The generated answers can be refined to all the detailed questions of campus life, and the answers are accurate. The generated answers will combine the background information of the user's question and the content of the retrieved data to ensure the accuracy and reliability of the answers.
[0057] 3. High real-time performance. Answers can be obtained immediately after a question is asked.
[0058] 4. High convenience. Asking questions is not restricted by time and space and can be done anytime and anywhere.
[0059] 5. Easy maintenance. Only need to upload and delete data, and the multi-source data can be maintained through the campus Q&A system, ensuring the reliability of the system's Q&A and easy maintenance.
[0060] 6. Available on multiple channels. The deployment of the system can be deeply integrated with various social platforms, such as QQ, WeChat official accounts, DingTalk, etc., to expand the user's question-asking channels. Description of the drawings
[0061] Figure 1 It is a schematic diagram of the advantages of the present invention over the prior art.
[0062] Figure 2 It is a schematic diagram of the process of the campus Q&A generation method of the present invention. Detailed implementation manners
[0063] The technical solution of the present invention will be further described below in conjunction with the drawings and embodiments.
[0064] A large model campus Q&A generation method based on multi-source data includes the following steps:
[0065] Step 1: First, it is necessary to connect the campus Q&A system with the campus information system, including the academic affairs system, the student affairs system, the unified identity authentication system of the network center, etc. By synchronizing student information (such as student ID, name, grade, major, class, counselor, etc.), teacher information (such as staff number, name, teaching information, class schedule, etc.) and relevant campus information (such as exam arrangements, course arrangements, activity notices) in real time, a basic multi-source data support set is constructed.
[0066] Step 2: Integrate policy documents (such as credit requirements, training programs, etc.), manage all data with tagging. For example, add tags such as "Computer Science College" and "Class of 2022" to the "Training Program for the Class of 2022 in the Computer Science College", and access it through a standardized API interface to ensure data timeliness and security. At the same time, vectorize unstructured text materials using the Moka MassiveMixed Embedding method for semantic retrieval.
[0067] Step 3: The system supports teachers and students to log in through unified identity authentication. When students log in at the user end, they need to enter their student ID and password. After passing the verification with the unified identity authentication of the network center, they enter the system, and the system automatically loads the basic information of the students, including grade, major, class, etc. When teachers log in at the user end, they need to enter their staff number and password. After passing the identity authentication, they enter the system, and the system loads their teaching courses and class information according to the teacher's identity information to ensure that each user can access the campus data related to themselves.
[0068] Step 4: Users enter questions through the system user end. After the questions are forwarded to the processing end, the processing end calls the model to perform semantic analysis on the questions and identify the intent of the questions. For example, question scenarios such as "credit policy", "exam arrangement", or "activity notice". Match the identified intent with the data tags, and constrain the retrieval scope according to the user's identity information (such as major, grade, etc.), and call the API interface or perform vector retrieval to locate relevant data. After the data location is completed, the processing end combines the retrieved content with the question intent, and calls the large language model to generate targeted answers. The generated answers combine the background information of the user's question and the content of the retrieved data to ensure the accuracy and reliability of the answers.
[0069] Step 5: The system pushes the generated answers to the user end. The answer content includes a clear reply and an attached reference material link for the user to further understand the detailed information.
[0070] Furthermore, the system also supports users to follow up based on the current question. For example, after viewing the specific time of an activity, further inquire about the registration process, etc. The system will continue to generate targeted answers through the large language model according to the context history of the questions, and archive the entire Q&A record for users to view at any time.
[0071] Step 6: The system has scalability and maintainability. By regularly importing and updating data such as policy files and course schedules, the accuracy of the answer content is ensured. The system can be integrated with social platforms such as WeChat official accounts and DingTalk to expand the application scenarios of Q&A. In terms of security, by adding an authentication service to the API interface, it is ensured that users can only access the data within their authorized scope. If the large model fails to generate an accurate answer, the system will mark the question and submit it to manual processing. In addition, the system supports model optimization. By analyzing user questions and feedback, the accuracy and coverage of the answers are continuously improved to provide users with a convenient Q&A experience.
[0072] The large model campus Q&A system based on multi-source data mainly includes a configuration end, a user end, and a processing end:
[0073] Configuration end: Import data from different fields by docking with the campus information system, sort out the sources of all data and configure tag information for the data, import campus information text materials through the method of text material upload, and provide campus real-time information data through the api interface. The specific operations are as follows:
[0074] (1): The campus information system includes a teaching affairs system, a student affairs system, and a unified identity authentication system for the network center. By docking with the campus information system, the student class schedule, student list, teacher list, counselor, and class information are synchronized in real time, and the notices and academic situation policies issued by the school are updated in real time; sort out the sources of all data, that is, the specific departments, affiliated colleges, and grades that publish the data, integrate data from different sources, and label different data usage scopes. The usage scope of the data is restricted by the label. For example, the "Cultivation Plan for the Class of 2022 in the School of Computer Science" should be labeled with "School of Computer Science", "Class of 2022", "Computer Science and Technology", "Software Engineering", etc.
[0075] (2): Provide the campus real-time information data through the api interface, pass the label as an api interface parameter, and at the same time configure an authentication service for the api to ensure data security.
[0076] (3): Segment and vectorize the campus information text materials, use Moka Massive Mixed Embedding to convert the text into vectors, and realize the retrieval of text information through the retrieval module. The retrieval module calculates the similarity of the vectors corresponding to the text, sorts them according to the similarity level, and selects the text with high similarity as the retrieval result.
[0077] (4): Package the api interface and retrieval module in (2) and (3) into call methods respectively, that is, functions, so that the corresponding campus real-time information data or campus information text materials can be obtained by calling the corresponding functions.
[0078] The function is mainly divided into two types. One type does not require passing in parameters and directly calls the service interface based on the user's login information to obtain real-time campus information data. The other type requires taking the question raised by the user as a parameter, and through the M3E (Moka Massive Mixed Embedding) model, the question is also transformed into a vector. By calculating the cosine similarity with the text vector, the text corresponding to the maximum value of the result is taken as the retrieval result. The specific calculation method is as follows:
[0079] Cosine Similarity(A,B}={A·B} / {||A||B|}
[0080] Among them, A and B represent the text vectors corresponding to the two texts whose similarity needs to be compared.
[0081] (5): Construct a hash index structure, use the label set of real-time campus information data or campus information text materials as the key of the hash, and use the function to obtain the corresponding data as the value, to obtain a multi-source data support set, specifically as follows:
[0082] {(label11,label12,…):function1,
[0083] (label21,label22,…):function2,…}
[0084] Among them, the label included in the set is the label of real-time campus information data or campus information text materials, and the function is the retrieval execution function, which can obtain the data or information within the corresponding label constraint range through api call or vector knowledge base retrieval.
[0085] (6): Merge the key sets of the multi-source data support set, that is, insert the keys of all hash objects into a new set to obtain the total set of all labels of the data set, denoted as InfoLabelSet.
[0086] User side: Authenticate the user. After successful authentication, initially save the user's information according to the student number or employee number information, including the question feedback function and the function of viewing reference materials and documents. The user side provides a question input port for the user. After sending the question input by the user to the processing side for processing, the output answer of the processing side is displayed.
[0087] Processing side: Includes a large model call module and a data storage module.
[0088] The model invocation module conducts auxiliary intent recognition and answer generation by invoking models. It classifies intents by invoking the intent processing tool to obtain the corresponding label in the question intent, and restricts the data retrieval scope according to the user's personal information. It searches for relevant data information helpful for question answering in the data storage module within the restricted retrieval scope through vector search. It generates an answer to the question based on the data information by invoking the large language model and feeds it back to the user side. The specific implementation is as follows:
[0089] Through InfoLabelSet and data from each data source, with the help of the RBT3 Chinese language classification model based on BERT, the data from the data source and its corresponding labels are used as a dataset to train a text classifier as the intent processing tool.
[0090] Perform intent classification through the trained model, select the 3 categories with the highest confidence as the question intent, obtain the corresponding label in each question intent, and construct a label set for the question intent, denoted as QuestionLabelSet.
[0091] According to the IdentityLabelSet extracted from the user login information, take the union with QuestionLabelSet to obtain the retrieval set SearchLabelSet. The calculation method is as follows:
[0092] SearchLabelSet = IdentityLabelSet U QuestionLabelSet
[0093] Based on the matching comparison between SearchLabelSet and InfoLabelSet, confirm the specific functions to be invoked through the multi-source data support set, and obtain the set of all functions to be invoked, denoted as FunctionSet.
[0094] Invoke the functions in FunctionSet in sequence to obtain data and information materials.
[0095] Integrate the retrieved data and information with the question content, and through the large language model, that is, models with natural language understanding and processing capabilities such as DeepSeek-V3, chatgpt-4o, etc., generate an answer to the question based on the data information and feed it back to the user side for answering questions.
[0096] The data storage module stores the imported data information, and also stores the api call link, user identity information, and user question and answer logs.
[0097] The above content is a further detailed description of the present invention in combination with specific / preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, they can also make several substitutions or modifications to these described embodiments, and these substitution or modification methods should all be regarded as belonging to the protection scope of the present invention.
[0098] The parts not detailed in the present invention belong to the well-known technologies in the art.
Claims
1. A method for generating large model campus Q&A based on multi-source data, characterized in that, It includes the following steps: Step 1: Analyze the Q&A requirements of the campus scenario, and import data sources according to the campus Q&A requirements. The data sources include campus real-time information data and campus information text materials; Step 2: Connect the campus Q&A system with the campus information system, sort out the sources of all data, and establish a real-time information data basis for question answering; at the same time, connect with the campus information text materials and sort out the available relevant materials; Step 3: Train a text classifier as an intent processing tool; Step 4: The user conducts identity authentication at the system user end, and the system initially saves the user's information according to the student number or employee number information; Step 5: The user inputs the question to be asked through the system user end; Step 6: The system obtains the question input by the user, conducts intent classification through the intent processing tool, obtains the corresponding label in the question intent, and restricts the data retrieval scope according to the user's personal information. Then, relevant data information helpful for answering the question is found through vector search within the restricted retrieval scope; Step 7: Integrate the retrieved data and information with the question content, generate an answer to the question through a large language model, and feedback it to the user end for answering questions.
2. A method for generating campus Q&A of large models based on multi-source data according to claim 1, characterized in that, The specific operations of Step 2 are as follows: 2-1: The campus information system includes a teaching affairs system, a student affairs system, and a unified identity authentication system for the network center. By connecting with the campus information system, it synchronizes the student class schedule, student list, teacher list, counselor, and class information in real time, and updates the notices and academic situation policies issued by the school in real time; sort out the sources of all data, that is, the specific departments, affiliated colleges, and grades that release the data, integrate the data from different sources, and label the data for different usage scopes to restrict the usage scope of the data through the labels; 2-2: Provide the campus real-time information data through the api interface, pass the label as an api interface parameter, and at the same time configure an authentication service for the api to ensure data security; 2-3: Segment and vectorize the campus information text materials, use Moka Massive Mixed Embedding to convert the text into vectors, and implement the retrieval of text information through the retrieval module; the retrieval module calculates the similarity of the vectors corresponding to the text, sorts them according to the similarity from high to low, and selects the text with high similarity as the retrieval result; 2-4: Package the api interfaces and retrieval modules in 2-2 and 2-3 into call methods, that is, functions, so that the corresponding campus real-time information data or campus information text materials can be obtained through the call of the corresponding function; There are two types of functions. One does not require passing parameters and directly calls the service interface according to the user login information to obtain the campus real-time information data; the other requires the question raised by the user as a parameter, and also converts the question into a vector through the M3E model. By calculating the cosine similarity between the question vector and the text vector, the text corresponding to the maximum value of the result is selected as the retrieval result. The specific calculation method is: Cosine Similarity(A,B)={A·B} / {||A||B|} Among them, A and B represent two text vectors corresponding to which the similarity needs to be compared; 2-5: Construct a hash index structure, use the label set of campus real-time information data or campus information text materials as the key of the hash, and use the function of obtaining the corresponding data as the value to obtain a multi-source data support set, specifically as follows: {(label11,label12,…):function1, (label21,label22,…):function2,…} Among them, the label included in the set is the label of campus real-time information data or campus information text materials, and the function is a retrieval execution function that can retrieve data or information within the scope of the corresponding label constraint through api call or vector knowledge base; 2-6: Merge the key sets of the multi-source data support set, that is, insert the keys of all hash objects into a new set to obtain the total set of all labels of the data set, denoted as InfoLabelSet.
3. A method for generating large model campus Q&A based on multi-source data according to claim 1, characterized in that, Step 3 is specifically implemented as follows: Through InfoLabelSet and the data of each data source, with the help of the RBT3 Chinese language classification model based on BERT, use the data source data and its corresponding labels as the data set to train a text classifier as an intent processing tool.
4. A method for generating large model campus Q&A based on multi-source data according to claim 1 or 2, characterized in that, Step 4 is specifically operated as follows: The user inputs the student number or work number, password and unified identity authentication of the network center at the system user end. After passing the authentication, the user enters the system, and the system initially saves the user's information according to the student number or work number information; 4-1: After the user logs in, the user end submits the student number or work number information of the user to the system processing end and applies for the user's personal information; 4-2: After receiving the application, the processing end queries the campus database using the student number or work number, returns the user information, and saves it as IdentityLabelSet at the same time.
5. A method for generating large model campus Q&A based on multi-source data according to claim 4, characterized in that, Step 6 is specifically operated as follows: 6-1: Perform intent classification through the model trained in step 3, select the 3 categories with the highest confidence as the question intent, obtain the corresponding label in each question intent, and construct the label set of the question intent, denoted as QuestionLabelSet; 6-2: According to IdentityLabelSet extracted from the user login information, take the union with QuestionLabelSet to obtain the retrieval set SearchLabelSet; 6-3: According to the matching comparison between SearchLabelSet and InfoLabelSet, confirm the specific function to be called through the multi-source data support set, and obtain the set of all functions to be called FunctionSet; 6-4: Call the functions in FunctionSet in turn to obtain data and information materials.
6. A method for generating campus Q&A of large models based on multi-source data according to claim 1, characterized in that, Retain the question history, the user can ask follow-up questions through the user end, and the system continues to generate targeted answers by combining the question history with the large language model.
7. A large model campus Q&A system based on multi-source data, characterized in that, Including a configuration end, a user end and a processing end: Configuration side: Import data from different fields by docking with the campus information system, sort out the sources of all data and configure label information for the data, import campus information text materials by uploading text materials, and provide campus real-time information data through the api interface; User side: Authenticate the user. After successful authentication, initially save the user's information according to the student number or staff number information, including the problem feedback function and the function of viewing reference materials and documents. The user side provides a problem input port for the user. After sending the problem input by the user to the processing side for processing, the output answer of the processing side is displayed; Processing side: Includes a large model call module and a data storage module; Model call module: Through calling the model, perform auxiliary intent recognition and answer generation work; classify the intent by calling the intent processing tool to obtain the corresponding label in the problem intent, and constrain the data retrieval range according to the user's personal information. Search for relevant data information helpful for answering the question in the data storage module within the constrained retrieval range through vector search; call the large language model to generate an answer to the question based on the data information and feedback it to the user side; Data storage module: Store the imported data information, and at the same time store the api call link, user identity information, and user question and answer logs.