Multi-agent-based college entrance examination consultation method and system

Abstract

The invention discloses a multi-agent-based college entrance examination consultation method and system, and belongs to the technical field of large language model application. The problems that in the prior art, a traditional college entrance examination consultation system based on a large language model is limited in complex problem processing capacity, insufficient in personalized service capacity and lack of expert knowledge are solved. The method comprises the following steps: S1, constructing a local knowledge base containing objective data and expert knowledge extracted through a specific process; s2, designing a knowledge retrieval tool for accessing a network and a local knowledge base; s3, introducing a long-term memory mechanism to store, apply and update personalized information of the user in a cross-session manner; and S4, constructing a multi-agent collaborative architecture, analyzing user intentions according to user personalized information, scheduling expert agents, and integrating results to realize college entrance examination consultation. According to the method, the accuracy and comprehensiveness of answering are effectively improved, continuous personalized services are provided, and the method can be applied to online college entrance examination consultation.

Description

Technical Field

[0001] This invention relates to a college entrance examination consultation method and system, and more particularly to a college entrance examination consultation method and system based on multi-agent systems, belonging to the field of large language model application technology. Background Technology

[0002] College application choices not only directly affect whether students can enter their ideal university and major, but also have a profound and even decisive impact on their future higher education path and even career development. However, faced with a massive amount of university information, complex majors, ever-changing admission policies, and fierce competition, candidates and parents often feel overwhelmed, lost, and even experience significant anxiety during the application process. Therefore, accurate, comprehensive, and personalized college application counseling services have extremely high social demand and practical significance.

[0003] Traditional college application guidance relied primarily on school teachers' guidance, family and friends' experiences, or paid consultants. However, with the rapid development of the internet and big data technologies, online college application service platforms have emerged. These platforms systematically collect, organize, and initially integrate massive amounts of publicly available data, such as admission policies, historical university admission scores, and program details. They typically provide users with convenient data queries, simple condition matching based on historical admission data, and preliminary filtering, thus improving the efficiency of students and parents in obtaining college information to some extent. However, these early systems also have significant limitations: First, they mainly serve as information aggregation and display platforms; users still need to conduct in-depth analysis and judgment of massive amounts of data to extract truly valuable application decision-making criteria, requiring a high level of analytical ability and time investment from users. Second, the application suggestions generated by these systems based on historical data and preset rules are often broad and formulaic, lacking in-depth insight into individual student differences. For example, they fail to fully consider diverse factors such as students' personal interests, strengths, personality traits, future career aspirations, and family backgrounds, making it difficult to provide truly personalized guidance. Their generalized suggestions may not accurately match the unique needs and actual circumstances of each student.

[0004] In recent years, Large Language Models (LLMs) have made groundbreaking progress, significantly improving natural language understanding and generation capabilities, leading to their widespread application in numerous industries and scenarios. In the education sector, specifically in college entrance examination (Gaokao) consultation services, some forward-thinking educational service platforms have begun exploring and integrating LLM technology to build next-generation Gaokao consultation systems. These systems typically allow users to interact with the system through a natural language dialogue interface, raising various questions about the Gaokao. The built-in LLM is responsible for interpreting the user's intent, then retrieving relevant knowledge from the platform's massive information (knowledge base), and generating responses and suggestions based on the search results. Thanks to the efficient processing and understanding of massive amounts of information by LLMs, the next-generation Gaokao consultation system can theoretically further improve the convenience and efficiency of Gaokao consultation services. Furthermore, LLMs possess the ability to analyze information by combining students' personalized information (such as interests and family circumstances), potentially providing more targeted and personalized guidance. However, the current technical solutions for college entrance examination consultation systems based on large language models are still in the early stages of development, and they face many problems that need to be solved in practical applications: (1) Limited ability to handle complex problems: Existing technical solutions usually rely on a single large language model to perform preliminary analysis of user questions and then perform knowledge retrieval. Due to the limitations of the capabilities of the large language model itself at the current stage, the system is unable to effectively complete complex tasks that require complex logical reasoning, multi-source information integration, or multi-angle analysis, resulting in the inability to comprehensively and accurately retrieve and synthesize relevant information, and the depth and robustness of its answers are limited; (2) Insufficient personalized service capabilities: The existing technical solutions have information isolation between different sessions, and only save the history of each session (short-term memory), but lack an effective memory and in-depth analysis mechanism for the user's full session history. When the user starts a new conversation, the system cannot effectively utilize the user's personalized information accumulated in the previous interaction, resulting in the lack of accurate personalized adaptation in the subsequent answers, and even repeated questioning of known information. This not only reduces the personalization of the consultation, but also damages the smoothness of the interaction and the user experience; (3) Lack of expert knowledge: The existing technical solutions mainly rely on publicly available structured or semi-structured objective data (such as college information and score lines over the years) for knowledge retrieval. However, for the experiential knowledge and professional judgment contained in the minds of senior college entrance examination consultants, there is a lack of effective acquisition, integration and application mechanisms. Therefore, it is difficult to fully meet the key needs of some users for professional and in-depth consultation services.

[0005] In conclusion, a college entrance examination consultation method and system based on multi-agent systems is needed. Summary of the Invention

[0006] A brief overview of the invention is given below to provide a basic understanding of certain aspects of it. It should be understood that this overview is not an exhaustive summary of the invention. It is not intended to identify key or essential parts of the invention, nor is it intended to limit the scope of the invention. Its purpose is merely to present certain concepts in a simplified form as a prelude to the more detailed description that follows.

[0007] In view of this, in order to solve the problems of limited ability to handle complex problems, insufficient personalized service capabilities, and lack of expert knowledge in the traditional college entrance examination consultation system based on large language models in the prior art, the present invention provides a college entrance examination consultation method and system based on multi-agent.

[0008] Technical Solution 1 is as follows: A college entrance examination consultation method based on multi-agent systems, comprising the following steps:

[0009] S1. Construct a local knowledge base that includes objective data and expert knowledge extracted through specific processes;

[0010] S2. Design a knowledge retrieval tool for accessing online and local knowledge bases;

[0011] S3. Introduce a long-term memory mechanism to store, apply, and update user personalized information across sessions;

[0012] S4. Construct a multi-agent collaborative architecture, analyze user intent based on personalized user information, schedule expert agents, and integrate results to achieve college entrance examination consultation.

[0013] Furthermore, step S1 includes the following steps:

[0014] S11. Collect and store objective data;

[0015] In S11, Python web crawling technology is used to collect college entrance examination related data from publicly available Internet channels, including the enrollment plans of major universities, the admission scores of previous years, professional information, and university information. After cleaning and organizing, the data is uniformly stored in a locally deployed MySQL structured database.

[0016] S12. Extract and construct expert knowledge;

[0017] S12 includes the following steps:

[0018] S121. Obtain and preprocess the raw data, that is, manually collect and screen high-quality videos of college entrance examination consulting experts as raw materials, use Python programming library to extract audio tracks from video files, and save them as standard audio file format MP3.

[0019] S122. The speech of standard audio file format MP3 is transcribed using a locally deployed Whisper-large-v3-turbo speech recognition model and converted into raw text data;

[0020] S123. By calling the cloud-based large language model service GLM-4-Flash, the original text data is processed to complete punctuation marks and standardize paragraph division, resulting in the processed text.

[0021] S124. Use the RecursiveCharacterTextSplitter utility class to divide the processed text into blocks to obtain text blocks;

[0022] S125. For each text block, call the high-performance large language model DeepSeek-v3 to perform deep semantic understanding, automatically extract and generate multiple sets of question-and-answer pairs closely related to college entrance examination consultation;

[0023] S126. Based on the question-answer pairs, construct knowledge storage and index, and perform vectorized storage and keyword storage;

[0024] In S126, the vectorization storage process is as follows: for each extracted question-answer pair, the "question" part in the question-answer pair is converted into a high-dimensional question vector using the BGE-M3 embedding model, and then the complete question-answer pair is stored in the Milvus vector database using the high-dimensional question vector as an index.

[0025] The keyword storage process is as follows: the "question" part of the question-answer pair is processed by the Jieba word segmentation tool to extract keywords, and the complete question-answer pair containing the word segmentation results is stored in the MySQL database.

[0026] Furthermore, step S2 includes the following steps:

[0027] S21. Construct a web knowledge retrieval tool, which includes a web search tool and a web page content reading tool, and outputs a list of search results and the main HTML in Markdown format, respectively;

[0028] In S21, the process of building a web search tool is as follows: input the query string and the number of search results, call the external large model network search agent Tavily Search API to perform the actual web search operation, obtain the search results, and finally output a structured list of search results, wherein each independent search result contains the Uniform Resource Locator URL of the target webpage, the page title, and the content summary of the webpage content;

[0029] The process of building a webpage content reading tool is as follows: input the URL of the webpage's Uniform Resource Locator, then use the requests library to send an HTTP request to the URL to obtain the original Hypertext Markup Language (HTML) code of the webpage, then use the trafilatura library to perform deep parsing on the obtained HTML code, intelligently identify and extract the main part of the HTML code, and finally, the extracted main HTML is uniformly converted into Markdown format and output.

[0030] S22. Design retrieval tools for objective data in the local knowledge base, and use the objective data retrieval tools to retrieve and extract college entrance examination related information from the MySQL relational database;

[0031] In S22, the input for the admission score and enrollment plan retrieval tool is the university name, province name, and year; the input for the college information retrieval tool is the college name; and the input for the major information retrieval tool is the major name and major level. After receiving their respective preset input parameters, the above retrieval tools dynamically construct and execute corresponding SQL query statements to efficiently retrieve and extract relevant information from the locally deployed MySQL relational database. Finally, the objective data retrieval tool outputs the queried information records.

[0032] S23. Construct an expert knowledge retrieval tool;

[0033] S23 includes the following steps:

[0034] S231. Perform query decomposition, using a query decomposer based on a large language model to decompose the original query string. Decomposed into three subqueries from different perspectives ;

[0035] S232. For each subquery Where i = 1, 2 or 3, the expert knowledge retrieval tool adopts a hybrid retrieval strategy, using vector retrieval and keyword retrieval to recall candidate question-answer pairs from different data storage and indexing mechanisms;

[0036] In step S232, the vector retrieval device uses the BGE-M3 embedding model to vectorize the current subquery text, converting it into a high-dimensional semantic vector. Subsequently, using high-dimensional semantic vectors Based on this, a similarity search is performed on a pre-built Milvus vector database, which stores vectorized representations of the question portions of all expert knowledge question-answer pairs. The retrieval goal of similarity search is to find the top 10 question-answer pairs with the highest cosine similarity between the question vector and the subquery vector;

[0037] Cosine similarity The calculation formula is expressed as:

[0038]

[0039] in, Indicates the problem;

[0040] The keyword search engine uses the BM25 algorithm to perform keyword matching and retrieval in a MySQL database that stores expert question-and-answer pairs, targeting the current subquery. The calculation yields the question portion of each question-answer pair in the database. The BM25 score is used to return the top 10 question-answer pairs with the highest scores.

[0041] The formula for calculating the BM25 score is as follows:

[0042]

[0043] in, Indicates the total number of questions. Indicates included terms Number of questions Indicates terms In the question Frequency of occurrence in Indicates terms In subquery Frequency of occurrence in Hyperparameters representing the control of word frequency saturation and the impact of document length;

[0044] S233. Result Fusion: In the hybrid retrieval stage, each subquery obtains a set of candidate question-answer pairs from the vector retrieval tool and the keyword retrieval tool respectively. The inverse ranking fusion algorithm is used to calculate the comprehensive score of each question-answer pair based on its ranking in the result list of different retrieval tools. The higher the ranking, the greater the contribution to the final score, and the ranking result is obtained.

[0045] In S233, the overall score The calculation formula is expressed as:

[0046]

[0047] in, Represents the set of retrievers. Indicate the problem In the searcher The ranking in Indicates the regulating factor;

[0048] S234. A reranker is introduced to perform a secondary relevance ranking of the ranking results, that is, to concatenate the subquery text with the question part of each question-answer pair, input it into the bge-reranker-large model, and calculate a relevance score, which represents the degree of semantic relevance between the model and the current subquery.

[0049] S235. After obtaining the relevance score of each candidate question-answer pair calculated by the rearranger, filter out question-answer pairs with a score lower than 0.5. Finally, remove duplicates from the sub-retrieval results corresponding to each sub-query and obtain the final retrieval results.

[0050] Furthermore, step S3 includes the following steps:

[0051] S31. Long-term memory storage: This involves creating and maintaining a long-term memory unit for each user. The long-term memory unit is stored in the system's file system in JSON format and is indexed and retrieved using the user identifier (User ID). The data stored in the long-term memory unit is predefined key user information, specifically including: college entrance examination score, provincial ranking, province, gender, subject selection combination, hobbies, career planning, intention to study abroad, family economic situation, and preferred university location.

[0052] S32. Apply the long-term memory stored in the application to perform memory retrieval, serialize it into a plain text string, and embed it into the system prompt words of the large language model inside the multi-agent collaborative architecture;

[0053] S32 includes the following steps:

[0054] S321. When a user engages in any consultation session, retrieve the corresponding long-term memory unit based on the user ID;

[0055] S322. Serialize the retrieved JSON-formatted long-term memory data into a plain text string, and add a descriptive prefix "The following is information about the user's long-term memory:" before the text string;

[0056] S323. Embed the processed long-term memory text string into the system prompt words of the large language model within the multi-agent collaborative architecture;

[0057] S33. When a user ends a session, if the session generates new valid interaction information, i.e. the session state changes, the memory update process is triggered to update the long-term memory.

[0058] S33 includes the following steps:

[0059] S331. Extract the complete message history list of the session, filter the message list, remove tool messages, and keep only user messages and assistant messages to obtain the filtered message list;

[0060] S332. Convert the filtered message list into a "user-assistant" dialogue text format;

[0061] S333. Input the user's current long-term memory and the generated dialogue text retrieved in step S321 into the large language model. The large language model analyzes the relationship between the dialogue text and the existing long-term memory, automatically infers and generates JSON Patch operation instructions for incrementally updating the long-term memory, and finally the system applies the operation instructions to update the long-term memory and saves it.

[0062] Furthermore, step S4 includes the following steps:

[0063] S41. Define the role and functions of the supervisory agent;

[0064] In step S41, the supervisory agent, as the direct interaction object between the user and this college entrance examination consultation system and the core coordinator of the entire collaborative architecture, has the following responsibilities:

[0065] 1) Direct answers: For simple user questions, answer the user directly;

[0066] 2) Understanding the problem and scheduling expert agents: Deeply understand the user's complex or multifaceted consultation problem, accurately identify the user's intent, and intelligently decompose the problem into a series of sub-tasks to be handled by specific expert agents. Then, through the "Handoffs" mechanism, control is handed over to the appropriate expert agents to execute the tasks in sequence.

[0067] 3) Integrating Results and Generating the Final Answer: After each expert agent completes its task and relinquishes control, the lead agent is responsible for integrating the execution results from different expert agents and generating the final answer.

[0068] S42. The "Handoffs" mechanism is used to realize the transfer of control and information transmission between agents. The agent that initiates the handover is configured with a tool in the form of transfer_to_[TargetAgentName](**params), where [TargetAgentName] specifies the name of the target agent, and **params represents a parameter dictionary containing the information required by the target agent to perform its subsequent tasks.

[0069] S43. Design an expert agent that includes a knowledge retrieval agent, a volunteer recommendation agent, an admission probability analysis agent, and a major recommendation agent, and complete the construction of a multi-agent collaborative architecture;

[0070] S431. Design a knowledge retrieval intelligent agent;

[0071] In S431, the design of the knowledge retrieval agent is as follows:

[0072] Responsibilities: Responsible for independently retrieving and querying relevant knowledge from online and local knowledge bases;

[0073] Required information: The specific query content is transmitted by the supervisory agent through the Handoffs mechanism;

[0074] Core capabilities and tools: The knowledge retrieval agent is a ReAct-style agent with autonomous reasoning and action capabilities, and is equipped with all the knowledge retrieval tools described in step S2;

[0075] Workflow: Autonomously iterates through multiple steps. In each step, the knowledge retrieval agent determines the next action based on the current progress and long-term memory content, and through reasoning using an internal large language model, namely, selecting and calling knowledge retrieval tools, and outputting all the retrieved knowledge upon task completion.

[0076] S432. Design a volunteer recommendation agent;

[0077] In S432, the design of the volunteer recommendation agent is as follows:

[0078] Responsibilities: Based on the applicant's information, recommend suitable university and major combinations;

[0079] Required information: the candidate's ranking, province, and subject selection;

[0080] Core capabilities and tools: The volunteer recommendation agent is a WorkFlow-style agent, equipped with predefined process reasoning and actions, but without any tools.

[0081] Workflow: 1) Preliminary screening: Based on the candidate's ranking, province, and subject selection, records whose admission ranking and candidate ranking are within the preset fluctuation range are screened from the professional admission score database;

[0082] 2) Timeliness verification and updating: The preliminary screening results are compared with the college enrollment plan for that year, and majors that have stopped enrolling students that year are eliminated to form a preliminary list of candidate choices;

[0083] 3) Personalized Refinement: Using a large language model, combined with the user's explicit requirements, user information in long-term memory, and detailed information about majors in the enrollment plan for that year, the preliminary list of choices is further filtered to generate a highly personalized recommended choice form;

[0084] S433. Design an intelligent agent for admission probability analysis;

[0085] In S433, the design of the admission probability analysis agent is as follows:

[0086] Responsibilities: Analyze the probability of admission to a specific target university specified by the applicant and provide corresponding application advice;

[0087] Required information: the candidate's ranking, province, subject selection, and the name of the target university;

[0088] Core capabilities and tools: The admission probability analysis agent is a Workflow-style agent, which is equipped with predefined process reasoning and actions, but no tools are provided.

[0089] Workflow: 1) Data collection: Query the admission scores, enrollment plan details, and the latest enrollment plan for the current year of the target university for the past two years;

[0090] 2) Trend Analysis and Comprehensive Judgment: Using a large language model, combined with the dynamic changes in the enrollment plan, the fluctuations in the admission scores of previous years, and referring to the user information stored in the long-term memory as described in step S3, a comprehensive analysis and evaluation is conducted to generate probability analysis results on the candidate's admission to the target university and to provide targeted application suggestions.

[0091] S434. Intelligent agents are recommended for design majors;

[0092] In S434, the design of the admission probability analysis agent is as follows:

[0093] Responsibilities: Based on the candidate's individual needs, interests, and long-term development plans, recommend a major and provide reasons for the recommendation;

[0094] Required information: None;

[0095] Core capabilities and tools: The professional recommendation agent is a ReAct-style agent with autonomous reasoning and action capabilities, equipped with professional information retrieval tools, and has the ability to directly engage in multi-round deep interactions with users;

[0096] Workflow: Autonomous multi-step iteration. In each step, the professional recommendation agent determines the next action based on the current progress and long-term memory content, and through internal large language model reasoning. The actions include actively asking the user, selecting and calling professional knowledge retrieval tools, and outputting results upon task completion.

[0097] Technical Solution 2 is as follows: A college entrance examination consultation system based on multi-agent intelligence, used to implement the college entrance examination consultation method based on multi-agent intelligence as described in Technical Solution 1, including a multi-agent collaborative architecture, retrieval tools, a knowledge base and a memory bank;

[0098] The multi-agent collaborative architecture is connected to the retrieval tool and the memory, respectively, and the retrieval tool is connected to the knowledge base.

[0099] Furthermore, the multi-agent collaborative architecture includes a supervisor agent and an expert agent;

[0100] The expert intelligent agent includes a knowledge retrieval intelligent agent, a college application recommendation intelligent agent, an admission probability analysis intelligent agent, and a major recommendation intelligent agent;

[0101] The knowledge base includes a local knowledge base and a network knowledge base. The local knowledge base includes expert knowledge and objective data.

[0102] The memory bank includes long-term memory and short-term memory.

[0103] The beneficial effects of the present invention are as follows: (1) Constructing a multi-agent collaborative architecture to significantly improve the ability to handle complex problems: The present invention does not simply rely on a single large language model to handle all types of user problems, but constructs a collaborative architecture composed of a supervisor agent and multiple expert agents. Each expert agent is carefully designed to be proficient in specific problem types (volunteer recommendation, knowledge retrieval, professional recommendation, etc.), so that it can set its toolset, optimize its reasoning logic and response strategy in a targeted manner, and show processing efficiency and depth far exceeding that of the general model in its respective areas of expertise. The supervisor agent, as the central hub of the architecture, is responsible for intelligently parsing user intentions and coordinating expert agents to execute tasks in sequence. Finally, the execution results are efficiently integrated and refined to answer the user's questions. The architecture of the present invention enables the system to handle problems that require multi-source information and complex logical reasoning more effectively, thereby improving the accuracy and comprehensiveness of the answers.

[0104] (2) A long-term memory mechanism is introduced to achieve continuous personalized service: This invention breaks through the limitation of existing systems that rely solely on short-term memory and innovatively introduces a long-term memory mechanism. It actively identifies, extracts and structures key user information from the dialogue and continuously updates the system's long-term memory. This mechanism enables the system to maintain "cognition" of the user's identity and needs across conversations, achieving truly continuous personalized service. In addition, since the system remembers some information, the user does not need to provide it repeatedly, which greatly enhances the coherence of the consultation process and the user experience.

[0105] (3) Expert knowledge construction and application, providing in-depth guidance: Based on the integration of objective data, this invention further constructs expert knowledge containing the experience and judgment of senior college entrance examination consultants, and designs an efficient knowledge retrieval tool to enable the system to provide analysis and suggestions that go beyond simple data query, containing expert experience and in-depth insights, and meet the core demands of some users for professional and high-level consulting services. Attached Figure Description

[0106] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0107] Figure 1 This is a flowchart illustrating a multi-agent-based college entrance examination consultation method.

[0108] Figure 2 This is a schematic diagram of the structure of a college entrance examination consultation system based on multi-agent systems;

[0109] Figure 3 This is a schematic diagram illustrating an embodiment of a college entrance examination consultation system based on multi-agent technology.

[0110] Figure 4 A flowchart illustrating the process of building expert knowledge;

[0111] Figure 5 This is a flowchart illustrating the process of expert knowledge retrieval.

[0112] Figure 6 This is a flowchart illustrating the long-term memory mechanism.

[0113] Figure descriptions: 1. Multi-agent collaborative architecture; 2. Retrieval tool; 3. Knowledge base; 4. Memory bank. Detailed Implementation

[0114] To make the technical solutions and advantages of the embodiments of the present invention clearer, the exemplary embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0115] Example 1: Reference Figures 1-6 This embodiment describes a college entrance examination consultation method based on multi-agent systems, which specifically includes the following steps:

[0116] S1. Construct a local knowledge base that includes objective data and expert knowledge extracted through specific processes;

[0117] S2. Design a knowledge retrieval tool for accessing online and local knowledge bases;

[0118] S3. Introduce a long-term memory mechanism to store, apply, and update user personalized information across sessions;

[0119] S4. Construct a multi-agent collaborative architecture, analyze user intent based on personalized user information, schedule expert agents, and integrate results to achieve college entrance examination consultation.

[0120] Furthermore, step S1 includes the following steps:

[0121] S11. Collect and store objective data;

[0122] In S11, regarding the construction of objective data, Python web crawling technology is used to systematically collect authoritative data related to the college entrance examination from publicly available Internet channels. This includes the enrollment plans, historical admission scores, major information, and university information of various universities. Given that such data usually has good structured characteristics, it is cleaned, organized, and uniformly stored in a locally deployed MySQL structured database for efficient querying and management.

[0123] S12. Extract and construct expert knowledge;

[0124] In S12, regarding the construction of expert knowledge, refer to Figure 4 This invention innovatively proposes a method for extracting and organizing knowledge from unstructured expert experience, comprising the following steps:

[0125] S121. Obtain and preprocess the raw data, that is, manually collect and screen high-quality videos of college entrance examination consulting experts as raw materials, use Python programming library to extract audio tracks from video files, and save them as standard audio file format MP3.

[0126] S122. The locally deployed Whisper-large-v3-turbo speech recognition model is used to transcribe the speech of the standard audio file format MP3, and convert it into raw text data efficiently and accurately;

[0127] S123. To improve the readability of the text and the convenience of subsequent processing, the original text data is processed by calling the cloud-based large language model service GLM-4-Flash to perform standardization processing such as punctuation completion and paragraph division, forming a clearly structured text content, and thus obtaining the processed text.

[0128] S124. In order to adapt to the input length limit of the large language model and ensure semantic integrity, the RecursiveCharacterTextSplitter utility class is used to divide the processed text into blocks, and the expected length of each text block is set to 5000 characters. A 500-character overlap area is set between adjacent text blocks to ensure the continuity of key information and the integrity of the context.

[0129] S125. For each text block, call the high-performance large language model DeepSeek-v3 to perform deep semantic understanding, automatically extract and generate multiple sets of question-and-answer pairs closely related to college entrance examination consultation;

[0130] S126. Based on the question-answer pairs, a knowledge storage and index are constructed, and vectorized storage and keyword storage are performed. In order to achieve efficient retrieval of expert knowledge, this invention adopts a dual indexing and storage mechanism.

[0131] In S126, the vectorization storage process is as follows: for each extracted question-answer pair, the "question" part in the question-answer pair is converted into a high-dimensional question vector using the BGE-M3 embedding model. Then, the high-dimensional question vector is used as an index to store the complete question-answer pair in the Milvus vector database to support fast retrieval based on semantic similarity.

[0132] The keyword storage process is as follows: the "question" part of the question-answer pair is processed by the Jieba word segmentation tool to extract keywords. The complete question-answer pair containing the word segmentation results is stored in the MySQL database so that it can be supplemented by traditional keyword-based search engines.

[0133] Specifically, in step S1, a comprehensive local knowledge base is constructed to provide robust data and knowledge support for the intelligent agent, referencing... Figure 3 The local knowledge base consists of two parts: objective data and expert knowledge. Through the construction process of objective data and expert knowledge, this invention forms a local knowledge base that is rich in content, diverse in structure, and easy to search, providing a strong knowledge foundation for the college entrance examination consultation system.

[0134] Furthermore, step S2 includes the following steps:

[0135] S21. Construct a web knowledge retrieval tool, which includes a web search tool and a web page content reading tool, and outputs a list of search results and the main HTML in Markdown format, respectively;

[0136] In step S21, to ensure that the agent can acquire and utilize the latest information and dynamic knowledge on the Internet in real time, this invention designs and integrates two collaborative network knowledge retrieval tools. The process of constructing the network search tool is as follows: input the query string and the number of search results, call the external large model network search agent Tavily SearchAPI to perform the actual network search operation, obtain the search results, and finally output a structured list of search results, wherein each independent search result contains the Uniform Resource Locator URL of the target webpage, the page title, and the content summary of the webpage content;

[0137] Given that the content summaries returned by web search tools are usually quite brief, in order to obtain more complete and detailed information about the target webpage, this invention further designs a webpage content reading tool. The process of constructing the webpage content reading tool is as follows: input the Uniform Resource Locator (URL) of the webpage, then use the requests library to send an HTTP request to the URL to obtain the original Hypertext Markup Language (HTML) code of the webpage, then use the trafilatura library to perform deep parsing on the obtained HTML code, intelligently identify and extract the main part of the HTML code. This process can effectively filter out elements with low relevance to the core information, such as advertisements, navigation bars, and footers, thereby ensuring the purity and relevance of the extracted content. Finally, in order to facilitate the understanding, processing and analysis of the subsequent large language model and improve the efficiency of information utilization, the extracted main HTML is uniformly converted into Markdown format and output.

[0138] Through the coordinated use of the aforementioned web search tools and web page content reading tools, the intelligent agent constructed by this invention can efficiently and accurately acquire the required knowledge from the Internet, providing strong external information support for subsequent complex problem-solving and personalized consulting services.

[0139] S22. Design retrieval tools for objective data in the local knowledge base, and use the objective data retrieval tools to retrieve and extract college entrance examination related information from the MySQL relational database;

[0140] In step S22, for the objective data (admission scores, enrollment plans, university information, and major information) in the local knowledge base, the present invention designs corresponding retrieval tools. The input for the admission scores and enrollment plans retrieval tools is the university name, province name, and year; the input for the university information retrieval tool is the university name; and the input for the major information retrieval tool is the major name and major level (undergraduate or associate degree). After receiving their respective preset input parameters, the above retrieval tools dynamically construct and execute corresponding SQL query statements to efficiently retrieve and extract relevant information directly from the locally deployed MySQL relational database. Finally, the objective data retrieval tool outputs the retrieved information records as output, which can be directly called and referenced by the corresponding agents in the multi-agent system when making decisions, integrating information, or answering user inquiries, thereby ensuring the accuracy and reliability of the objective data sources on which the system relies.

[0141] S23. Construct an expert knowledge retrieval tool;

[0142] In step S23, for unstructured expert question-and-answer pairs stored in the local knowledge base, this invention designs and implements an efficient expert knowledge retrieval tool, referencing... Figure 5The expert knowledge retrieval tool takes a query string as input and aims to accurately filter and output knowledge items highly relevant to the query intent from a massive number of question-answer pairs. This includes the following steps:

[0143] S231. Query decomposition is performed. To improve the comprehensiveness and accuracy of the retrieval, a query decomposer based on a large language model is first used to decompose the original query string. Decomposed into three subqueries from different perspectives This multi-perspective segmentation can cover different aspects of user intent and enhance the diversity of subsequent searches;

[0144] S232. For each subquery Where i = 1, 2 or 3, the expert knowledge retrieval tool adopts a hybrid retrieval strategy, using vector retrieval and keyword retrieval to recall candidate question-answer pairs from different data storage and indexing mechanisms in order to improve the comprehensiveness and accuracy of the recall.

[0145] In step S232, the vector retrieval device uses the BGE-M3 embedding model to vectorize the current subquery text, converting it into a high-dimensional semantic vector. Subsequently, using high-dimensional semantic vectors Based on this, a similarity search is performed on a pre-built Milvus vector database, which stores vectorized representations of the question portions of all expert knowledge question-answer pairs. The retrieval goal of similarity search is to find the top 10 question-answer pairs with the highest cosine similarity between the question vector and the subquery vector;

[0146] Cosine similarity The calculation formula is expressed as:

[0147]

[0148] in, Indicates the problem;

[0149] The keyword search engine uses the BM25 (Best Match 25) algorithm to perform keyword matching and retrieval in a MySQL database storing expert question-and-answer pairs, targeting the current subquery. The calculation yields the question portion of each question-answer pair in the database. The BM25 score is used to return the top 10 question-answer pairs with the highest scores.

[0150] The formula for calculating the BM25 score is as follows:

[0151]

[0152] in, Indicates the total number of questions. Indicates included terms Number of questions Indicates terms In the question Frequency of occurrence in Indicates terms In subquery Frequency of occurrence in Hyperparameters representing the control of word frequency saturation and the impact of document length;

[0153] S233. Result Fusion: In the hybrid retrieval stage, each subquery obtains a set of candidate question-answer pairs from the vector retrieval machine and the keyword retrieval machine respectively. In order to integrate these two sets of results and obtain a unified and better ranked list, this invention adopts the Reciprocal Rank Fusion (RRF) algorithm. The comprehensive score of each question-answer pair is calculated according to its ranking in the result lists of different retrieval machines. The higher the ranking, the greater the contribution to the final score, and the ranking result is obtained.

[0154] In S233, the overall score The calculation formula is expressed as:

[0155]

[0156] in, Represents the set of retrievers. Indicate the problem In the searcher The ranking in Indicates the regulating factor;

[0157] S234. Although the question-answer pair list after the inverse ranking fusion algorithm has been optimized at the recall level, its ranking is mainly based on the similarity calculation of the retrieval itself. In order to further improve the true relevance between the retrieval results and the subqueries, this invention introduces a reranker to perform a secondary relevance ranking of the ranking results. That is, the subquery text is concatenated with the question part of each question-answer pair and input into the bge-reranker-large model to calculate a fine relevance score, which represents the degree of semantic relevance between the model and the current subquery.

[0158] S235. After obtaining the relevance score of each candidate question-answer pair calculated by the rearranger, filter out question-answer pairs with a score lower than 0.5 to further ensure the quality and relevance of the final output results. Finally, remove duplicates from the sub-retrieval results corresponding to each sub-query and obtain the final retrieval results.

[0159] For details, please refer to Figure 3Step S2 aims to design a series of knowledge retrieval tools to enable intelligent agents to autonomously access and utilize knowledge. The knowledge required by the system covers two major categories: network knowledge and local knowledge. To ensure that intelligent agents can acquire knowledge comprehensively and effectively, this invention designs customized retrieval tools for different types of knowledge.

[0160] Furthermore, step S3 includes the following steps:

[0161] S31. Long-term memory storage: This involves creating and maintaining an independent long-term memory unit for each user. The long-term memory unit is stored in the system's file system in JSON format and is indexed and retrieved using a unique user identifier (User ID). The data stored in the long-term memory unit is predefined key user information, specifically including: college entrance examination score, provincial ranking, province, gender, subject selection combination, hobbies, career planning, intention to study abroad, family economic situation, and preferred university location.

[0162] S32. Apply the long-term memory stored in the application to perform memory retrieval, serialize it into a plain text string, and embed it into the system prompt words of the large language model inside the multi-agent collaborative architecture;

[0163] S32 includes the following steps:

[0164] S321. When a user engages in any consultation session, retrieve the corresponding long-term memory unit based on the user ID;

[0165] S322. Serialize the retrieved JSON format long-term memory data into a plain text string and add a descriptive prefix "The following is information about the user's long-term memory:" to the text string to improve the efficiency of the large language model in recognizing and utilizing memory content;

[0166] S323. Embed the processed long-term memory text string into the system prompt words of the large language model inside the multi-agent collaborative architecture to ensure that the model can perceive and integrate the user's key information when processing any user request;

[0167] S33. When a user ends a session, if the session generates new valid interaction information, i.e. the session state changes, the memory update process is triggered to update the long-term memory.

[0168] S33 includes the following steps:

[0169] S331. Extract the complete message history list of the session, and filter the message list to remove tool messages and keep only user messages and assistant messages to obtain the filtered message list;

[0170] S332. Convert the filtered message list into a clear, continuous "user-assistant" dialogue text format;

[0171] S333. Input the user's current long-term memory and the generated dialogue text retrieved in step S321 into the large language model. The large language model analyzes the relationship between the dialogue text and the existing long-term memory, automatically infers and generates JSON Patch operation instructions for incrementally updating the long-term memory, and finally the system applies the operation instructions to update the long-term memory and saves it.

[0172] For details, please refer to Figure 5 The existing college entrance examination consultation system saves the messages generated in each consultation session in real time and inserts them into the message context of the large language model to realize multi-turn dialogue service under the current consultation session, which is a short-term memory mechanism. However, relying solely on short-term memory cannot achieve continuous personalized service across sessions. Step S3 aims to introduce a long-term memory mechanism to effectively solve this problem.

[0173] Furthermore, step S4 includes the following steps:

[0174] S41. Define the role and functions of the supervisory agent;

[0175] In step S41, the supervisory agent, as the direct interaction object between the user and this college entrance examination consultation system and the core coordinator of the entire collaborative architecture, has the following responsibilities:

[0176] 1) Direct answers: For simple user questions, including but not limited to small talk, greetings, and common sense inquiries, answer the user directly;

[0177] 2) Understanding the problem and scheduling expert agents: Deeply understand the user's complex or multifaceted consultation problem, accurately identify the user's intent, and intelligently decompose the problem into a series of sub-tasks to be handled by specific expert agents. Then, through the "Handoffs" mechanism, control is handed over to the appropriate expert agents to execute the tasks in sequence.

[0178] 3) Integrating Results and Generating the Final Answer: After each expert agent completes its task and relinquishes control, the lead agent is responsible for integrating the execution results from different expert agents to generate a comprehensive and accurate final answer.

[0179] S42. The "Handoffs" mechanism is used to realize the transfer of control and information transmission between agents. The agent that initiates the handover is configured with a tool in the form of transfer_to_[TargetAgentName](**params), where [TargetAgentName] specifies the name of the target agent, and **params represents a parameter dictionary containing the information required by the target agent to perform its subsequent tasks.

[0180] S43. Design an expert agent that includes a knowledge retrieval agent, a volunteer recommendation agent, an admission probability analysis agent, and a major recommendation agent, and complete the construction of a multi-agent collaborative architecture;

[0181] S431. Design a knowledge retrieval intelligent agent;

[0182] In S431, the design of the knowledge retrieval agent is as follows:

[0183] Responsibilities: Responsible for independently retrieving and querying relevant knowledge from online and local knowledge bases;

[0184] Required information: The specific query content is transmitted by the supervisory agent through the Handoffs mechanism;

[0185] Core capabilities and tools: The knowledge retrieval agent is a ReAct-style agent (Reasoning and Action), possessing autonomous reasoning and action capabilities, and equipped with all the knowledge retrieval tools described in step S2;

[0186] Workflow: Autonomously iterates through multiple steps. In each step, the knowledge retrieval agent determines the next action based on the current progress and long-term memory content, and through reasoning using an internal large language model, namely, selecting and calling knowledge retrieval tools, and outputting all the retrieved knowledge upon task completion.

[0187] S432. Design a volunteer recommendation agent;

[0188] In S432, the design of the volunteer recommendation agent is as follows:

[0189] Responsibilities: Based on the applicant's information, recommend suitable university and major combinations;

[0190] Required information: the candidate's ranking, province, and subject selection;

[0191] Core capabilities and tools: The volunteer recommendation agent is a WorkFlow-style agent, equipped with predefined process reasoning and actions, but without any tools.

[0192] Workflow: 1) Preliminary screening: Based on the candidate's ranking, province, and subject selection, records whose admission ranking and candidate ranking are within the preset fluctuation range are screened from last year's professional admission score database;

[0193] 2) Timeliness verification and updating: The preliminary screening results are compared with the college enrollment plan for that year, and majors that have stopped enrolling students that year are eliminated to form a preliminary list of candidate choices;

[0194] 3) Personalized Refinement: Using a large language model, combined with the user's explicit requirements, user information in long-term memory (family situation, regional preferences, hobbies, etc.), and detailed information about the majors in the enrollment plan for that year (tuition fees, major descriptions), the preliminary list of choices is further filtered to generate a highly personalized recommended choice form;

[0195] S433. Design an intelligent agent for admission probability analysis;

[0196] In S433, the design of the admission probability analysis agent is as follows:

[0197] Responsibilities: Analyze the probability of admission to a specific target university specified by the applicant and provide corresponding application advice;

[0198] Required information: the candidate's ranking, province, subject selection, and the name of the target university;

[0199] Core capabilities and tools: The admission probability analysis agent is a Workflow-style agent, which is equipped with predefined process reasoning and actions, but no tools are provided.

[0200] Workflow: 1) Data collection: Query the admission scores, enrollment plan details, and the latest enrollment plan for the current year of the target university for the past two years;

[0201] 2) Trend Analysis and Comprehensive Judgment: Using a large language model, combined with the dynamic changes in the enrollment plan (increase or decrease in the number of students enrolled, adjustment of majors), fluctuations in the admission scores of previous years, and referring to the user information stored in the long-term memory as described in step S3, a comprehensive analysis and evaluation is conducted to generate probability analysis results on the candidate's admission to the target university and to provide targeted application suggestions.

[0202] S434. Intelligent agents are recommended for design majors;

[0203] In S434, the design of the admission probability analysis agent is as follows:

[0204] Responsibilities: Based on the candidate's individual needs, interests, and long-term development plans, recommend suitable majors and provide detailed reasons for the recommendations;

[0205] Required information: None;

[0206] Core Capabilities and Tools: The professional recommendation agent is a ReAct-style agent with autonomous reasoning and action capabilities. It is equipped with professional information retrieval tools, enabling it to query information on specific professions as needed during the interaction process to assist its recommendation decision and enrich the recommendation reasons. In addition, considering the complexity of professional selection and the need for deep personalization, the professional recommendation agent has the ability to directly engage in multiple rounds of in-depth interaction with users to fully explore and understand users' potential needs and concerns.

[0207] Workflow: Autonomous multi-step iteration. In each step, the professional recommendation agent determines the next action based on the current progress and long-term memory content, and through internal large language model reasoning. The actions include actively asking the user, selecting and calling professional knowledge retrieval tools, and outputting results upon task completion.

[0208] For details, please refer to Figure 3 To achieve efficient handling of complex college entrance examination consultation issues, this invention constructs an innovative multi-agent collaborative architecture, which consists of a supervisor agent and a group of expert agents. Through clearly defined division of responsibilities and collaborative mechanisms, each expert agent is carefully designed to be proficient in a specific type of problem, thereby enabling them to set their toolsets, optimize their reasoning logic and response strategies, and demonstrate processing efficiency and depth far exceeding that of general models in their respective areas of expertise.

[0209] Specifically, in this embodiment, the process of using the present invention for college entrance examination consultation is as follows:

[0210] 1. System Environment and Initialization:

[0211] 1.1 Configure hardware environment: Use a server with sufficient GPU and video memory;

[0212] 1.2 Software Environment Configuration: The operating system is Linux, the programming language is Python, the database uses MySQL to store structured objective data and some expert knowledge keyword indexes, and Milvus to store the vector indexes of expert knowledge;

[0213] 1.3 Building a local knowledge base:

[0214] a. Objective data: Enrollment plans, admission scores, institution information, and major information of universities nationwide for the past two years have been collected and stored in a MySQL database in advance;

[0215] b. Expert knowledge: Several video lectures by college entrance examination consulting experts have been processed, and expert Q&A pairs on various topics have been extracted and generated, and stored in MySQL and Milvus vector databases respectively;

[0216] 1.4 Design of knowledge retrieval tools: Corresponding retrieval tools have been implemented for local knowledge bases and online knowledge;

[0217] 1.5 Introducing a long-term memory mechanism: Assume that user "Xiaoming" has previously interacted with the system, and the following information has been recorded in his long-term memory unit: {"College Entrance Examination Score": 580, "Provincial Ranking": 12000, "Province": "Shandong Province", "Gender": "Male", "Subject Combination": "Physics + Chemistry + Biology", "Interests": "Computer Programming, Robotics", "Career Planning": "None", "University Location Preference": "Beijing, Shanghai"};

[0218] 1.6 Constructing a multi-agent collaboration framework: The supervisor agent, knowledge retrieval agent, volunteer recommendation agent, admission probability analysis agent, and major recommendation agent have all been configured with their respective system prompts, tools, and workflows according to the design, and collaborate through the Handoffs mechanism;

[0219] 2. User interaction and system response:

[0220] 2.1 Sending a message: User Xiaoming logs into the system and initiates a request: "Hello, please recommend some universities and majors for me."

[0221] 2.2 Long-term memory application: The system retrieves the user's long-term memory unit based on the user's ID, and embeds the JSON data into the system prompt words of the large language model inside all agents after textification, forming a partial prompt: "The following is information about the user's long-term memory: the examinee's gender is male, the college entrance examination score is 580, the provincial ranking is 12000, the province is Shandong, the subject combination is physics + chemistry + biology, the hobbies are computer programming and robotics, and the preferred university location is Beijing and Shanghai."

[0222] 2.3 The supervisor agent receives and processes the request: The supervisor agent analyzes the user's intent as "volunteer recommendation", determines that this task should be handled by the volunteer recommendation agent, and calls transfer_to_volunteer recommendation agent(**params), and passes the ranking, province, and subject selection information obtained from long memory in the form of tool parameters;

[0223] 2.4 The volunteer recommendation agent performs tasks: Upon receiving information, the volunteer recommendation agent performs a pre-set workflow:

[0224] 2.4.1 Preliminary screening: From the MySQL database, we screened universities and majors in Shandong Province last year that admitted students with a physics, chemistry and biology combination ranking that fluctuated by 15% around 12,000th place.

[0225] 2.4.2 Timeliness Verification: Compare the screening results with this year's enrollment plan and remove major A, which has been discontinued.

[0226] 2.4.3 Personalized Refinement: Utilizing the capabilities of a large language model, combined with Xiaoming's interest in computer programming, his preference for Beijing and Shanghai, and information such as program descriptions, the candidate list is sorted and further filtered to generate a recommended list: "University A - Beijing - Artificial Intelligence (Tuition 7000 RMB, some courses are taught in English), University B - Shanghai - Computer Science (Tuition 8000 RMB, including the International Artificial Intelligence Class, the New Engineering Innovation Experimental Class, and the Software Elite Class)...". The recommended list is then returned to the supervising agent.

[0227] 2.5 The supervisory agent integrates and replies to the user: After receiving the recommendation list, the supervisory agent organizes it and replies to Xiaoming: "Hello Xiaoming, based on your situation, we recommend the following majors for you: ... These majors are a good match for your interests, and their past admission statistics are close to your academic ranking."

[0228] 2.6 User follow-up question: "I want to become an artificial intelligence engineer in the future. What's this major like? What are the job prospects?"

[0229] 2.7 The supervisor agent receives and processes the request: The supervisor agent analyzes the user's intent as "to query specific professional information and employment prospects", determines that this task should be handled by the "knowledge retrieval agent", and calls transfer_to_knowledge retrieval agent(**params) to pass the query content "introduction to artificial intelligence major and employment prospects" as a tool parameter;

[0230] 2.8 Knowledge Retrieval Agent Execution: Upon receiving a query request, the knowledge retrieval agent autonomously decides to invoke knowledge retrieval tools: it invokes objective data retrieval tools to query information about the Artificial Intelligence major in MySQL, including official introductions, curriculum, employment data, etc.; it invokes expert knowledge retrieval tools to query "Employment Prospects for Artificial Intelligence Majors," obtaining relevant expert Q&A pairs such as: "What are the employment prospects for Artificial Intelligence majors? Artificial Intelligence is a future trend with broad employment prospects and a high upper limit for employment. However, it must be noted that the lower limit for employment in this major is also quite low, and the educational requirements are relatively high..." etc. If local knowledge is insufficient, the agent will also invoke online knowledge retrieval tools to search for the latest industry reports. Finally, the knowledge retrieval agent integrates the retrieved multi-source information and returns the information to the supervising agent.

[0231] 2.9 The supervisory agent integrates and replies to the user: Upon receiving the message, the supervisory agent, after organizing it, replies to Xiaoming: "Regarding the Artificial Intelligence major, it [cites the major introduction from objective data]. According to recent data, the main destinations for graduates of this major are [cites employment data from objective data], and the employment prospects are generally good. However, it is important to note [integrate in-depth guidance from expert knowledge]."

[0232] 3. Session End and Long-Term Memory Update: When Xiaoming ends the current consultation session, the system detects a change in the session state (a new interaction occurs). Then, the system extracts the messages between the user and the assistant in this session (tool messages have been filtered), inputs the message text and Xiaoming's current long-term memory into the large language model, and generates a JSON Patch instruction after model analysis: set the value of "career planning" to "work in the field of artificial intelligence, artificial intelligence engineer";

[0233] This embodiment demonstrates in detail how the system provides users with efficient, in-depth, and personalized college entrance examination consultation services through multi-agent collaboration, utilizing local knowledge bases and long-term memory mechanisms, effectively solving the problems of complex problem handling, personalized services, and expert knowledge application in existing systems.

[0234] Example 2: Reference Figures 2-3 This embodiment describes a college entrance examination consultation system based on multi-agent technology, used to implement the college entrance examination consultation method based on multi-agent technology described in Embodiment 1, including a multi-agent collaborative architecture 1, a retrieval tool 2, a knowledge base 3, and a memory bank 4;

[0235] The multi-agent collaborative architecture 1 is connected to the retrieval tool 2 and the memory bank 4 respectively, and the retrieval tool 2 is connected to the knowledge base 3.

[0236] Furthermore, the multi-agent collaborative architecture 1 includes a supervisor agent and an expert agent;

[0237] The expert intelligent agent includes a knowledge retrieval intelligent agent, a college application recommendation intelligent agent, an admission probability analysis intelligent agent, and a major recommendation intelligent agent;

[0238] The knowledge base 3 includes a local knowledge base and a network knowledge base. The local knowledge base includes expert knowledge and objective data.

[0239] The memory bank 4 includes long-term memory and short-term memory.

[0240] Although the invention has been described with reference to a limited number of embodiments, those skilled in the art will understand from the foregoing description that other embodiments are conceivable within the scope of the invention described herein. Furthermore, it should be noted that the language used in this specification has been chosen primarily for readability and instructional purposes, and not for the purpose of interpreting or limiting the subject matter of the invention. Therefore, many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the appended claims. The disclosure of the invention is illustrative and not restrictive, and the scope of the invention is defined by the appended claims.

Claims

1. A multi-agent based college entrance examination counseling method, characterized in that, Comprise the following steps: S1. Construct a local knowledge base containing objective data and expert knowledge extracted through specific processes; S2. Design a knowledge retrieval tool for accessing the network and the local knowledge base; S3. Introduce a long-term memory mechanism to store, apply and update user personalized information across sessions; S4. Build a multi-agent collaboration architecture to analyze user intent, dispatch expert agents and integrate results based on user personalized information, realizing college entrance examination counseling.

2. The method of claim 1, wherein, In S1, the following steps are included: S11. Collect and store objective data; In S11, Python crawler technology is used to collect data related to the college entrance examination from public channels on the Internet, including enrollment plans of major universities, historical admission scores, professional information, and university information. After cleaning and organizing, they are stored in a locally deployed MySQL structured database; S12. Extract and build expert knowledge; In S12, the following steps are included: S121. Obtain and preprocess raw data, i.e. manually collect and select high-quality college entrance examination counseling expert explanation videos as raw materials, use Python programming libraries to extract audio tracks from video files, and save them as standard audio file format MP3; S122. Use the locally deployed Whisper-large-v3-turbo speech recognition model to transcribe the speech of the standard audio file format MP3 into raw text data; S123. Call the cloud large language model service GLM-4-Flash to complete the punctuation and paragraph division of the raw text data, and get the processed text; S124. Use the RecursiveCharacterTextSplitter tool class to split the processed text into blocks, and get the text blocks; S125. For each text block, call the high-performance large language model DeepSeek-v3 for deep semantic understanding, automatically extract and generate multiple sets of question and answer pairs closely related to college entrance examination counseling; S126. According to the question and answer pairs, build knowledge storage and index, and store vectors and keywords; In S126, the process of vector storage is as follows: for each extracted question and answer pair, use the BGE-M3 embedding model to convert the "question" part of the question and answer pair into a high-dimensional question vector, then use the high-dimensional question vector as an index to store the complete question and answer pair in the Milvus vector database; The process of keyword storage is as follows: the "question" part of the question and answer pair is processed by the Jieba word segmentation tool for Chinese word segmentation to extract keywords, and the complete question and answer pair containing the word segmentation results is stored in the MySQL database.

3. The method of claim 2, wherein the method is based on multi-agent. In S2, the following steps are included: S21. Build a network knowledge retrieval tool, which includes a network search tool and a web page content reading tool, respectively outputting a search result list and a main body HTML in Markdown format; In S21, the process of constructing the network search tool is as follows: input the query string and the number of search results, call the external large model network search agent Tavily Search API to perform actual network search operation, obtain search results, and finally output a structured search result list, wherein each independent search result contains the uniform resource locator URL of the target webpage, the page title title, and the summary content of the webpage content; The process of constructing the webpage content reading tool is as follows: input the uniform resource locator URL of the webpage, then use the requests library to initiate an HTTP request to the URL to obtain the original hypertext markup language HTML code of the webpage, then use the trafilatura library to deeply parse the obtained HTML code, intelligently identify and extract the main part of the HTML code, and finally, the extracted main HTML is converted into Markdown format and output; S22. Design respective retrieval tools for objective data in the local knowledge base, retrieve and extract high school examination related information in the MySQL relational database through the objective data retrieval tool; In S22, the input of the admission score line and enrollment plan retrieval tool is the university name, province name and year, the input of the university information retrieval tool is the university name, the input of the professional information retrieval tool is the professional name and professional level, and the above retrieval tools, after receiving the respective preset input parameters, all retrieve and extract related information from the locally deployed MySQL relational database through dynamic construction and execution of corresponding SQL query statements, and finally, the objective data retrieval tool records the queried information as output; S23. Construct an expert knowledge retrieval tool; In S23, the following steps are included: S231. Perform query decomposition, using a large language model-based query decomposer to break the original query string into three different angle sub-queries ; S232. For each sub-query where i = 1, 2 or 3, the expert knowledge retrieval tool employs a hybrid retrieval strategy, utilizing a vector retriever and a keyword retriever to recall candidate question-answer pairs from different data storage and indexing mechanisms; In the S232, the vector retriever uses the BGE-M3 embedding model to vectorize the current subquery text, converting it into a high-dimensional semantic vector , and then performs a similarity search in the pre-constructed Milvus vector database, which stores the vectorized representations of the question parts in all expert knowledge Q&A pairs , with the high-dimensional semantic vector as the reference. The retrieval target of the similarity search is to find the top 10 Q&A pairs with the highest cosine similarity between the question vector and the subquery vector. Cosine similarity The formula for the calculation of the cosine similarity is given by: wherein indicates a problem; The keyword search engine uses the BM25 algorithm to perform keyword matching and retrieval in a MySQL database that stores expert question-and-answer pairs, targeting the current subquery. The calculation yields the question portion of each question-answer pair in the database. The BM25 score is used to return the top 10 question-answer pairs with the highest scores. The calculation formula of BM25 score is: where, denotes the total number of problems, denotes the number of problems containing the term , denotes the term , frequency of the term in the problem , frequency of the term in the subquery , denotes the hyperparameter controlling the term frequency saturation and the document length influence; S233. Result fusion: in the mixed retrieval stage, each subquery obtains a set of candidate question and answer pairs from the vector retriever and the keyword retriever, adopts the reciprocal ranking fusion algorithm, calculates the comprehensive score of each question and answer pair according to its ranking in the result list of different retrievers, the earlier the ranking, the greater the contribution to the final score, and obtains the ranking result; In the S233, the comprehensive score The calculation formula of the comprehensive score is represented as: wherein, represents a retriever set, represents a question in a retriever ranking, represents a regulation factor; S234. A rearranger is introduced to perform relevance secondary sorting on the ranking result, that is, the subquery text is spliced with the question part of each question and answer pair, input into the bge-reranker-large model, and a correlation score is calculated, which represents the semantic correlation degree of the model judging the question and the current subquery; S235. After obtaining the correlation score of each candidate question and answer pair calculated by the rearranger, filter out the question and answer pairs with a score lower than 0.5, and finally, the sub-retrieval results corresponding to each subquery are de-duplicated to obtain the final retrieval result.

4. The method of claim 3, wherein, In S3, the following steps are included: S31. Store long-term memory, that is, create and maintain a long-term memory unit for each user, the long-term memory unit is stored in the file system of the system in JSON format and is indexed and retrieved by the user identifier User ID, the data stored in the long-term memory unit is the pre-defined key user information, including: college entrance examination score, provincial ranking, province, gender, subject combination, interest, career planning, intention to study abroad, family economic situation, regional preference of college; S32. Apply the stored long-term memory to memory retrieval, serialize it into a pure text string, and embed it into the system prompt word of the internal large language model of the multi-agent collaboration architecture; In the S32, the following steps are included: S321. When the user conducts any consultation session, retrieve the corresponding long-term memory unit according to the user ID; S322. Serialize the retrieved JSON format long-term memory data into a pure text string, and add a descriptive prefix "The following is the long-term memory information about the user:” in front of the text string; S323. Embed the processed long-term memory text string into the system prompt word of the internal large language model of the multi-agent collaboration architecture; S33. When the user ends a certain session, if new valid interaction information is generated in the session, that is, the session state changes, trigger the memory update process to update the long-term memory; In the S33, the following steps are included: S331. Extract the complete message history list of the session, and filter the message list to remove tool messages and only keep user messages and assistant messages to obtain a filtered message list; S332. Convert the filtered message list into "user-assistant” dialogue text format; S333. Input the user's current long-term memory retrieved in step S321 and the generated dialogue text into the large language model, the large language model analyzes the relationship between the dialogue text and the existing long-term memory, automatically reasons and generates JSON Patch operation instructions for incremental updating of the long-term memory, and finally the system applies the operation instructions to update the long-term memory and save.

5. The method of claim 4, wherein the method is based on multi-agent. In the S4, the following steps are included: S41. Set the role and function of the supervisor agent; In the S41, the supervisor agent is the direct interaction object of the user and the high school counseling system and the core coordinator of the entire collaboration architecture, its responsibilities include: 1) Direct answer: directly answer the user for some simple user questions; 2) Understand the problem and dispatch expert agents: deeply understand the user's complex or multi-aspect consultation questions, accurately identify the user's intention, and intelligently decompose the problem into a series of sub-tasks handled by specific expert agents, then hand over the control to the appropriate expert agent one by one through the "Handoffs” mechanism to execute the task; 3) Integrate results and generate final answer: after each expert agent completes its task and returns the control, the supervisor agent is responsible for integrating the execution results from different expert agents to generate the final answer. S42. The control right transfer and information transmission between agents are realized by the "Handoffs" mechanism. The agent initiating the transfer is configured with a tool in the form of transfer_to_[TargetAgentName](**params), where [TargetAgentName] specifies the name of the target agent, and **params represents a parameter dictionary containing the information required by the target agent to perform its subsequent tasks; S43. Designing expert agents including knowledge retrieval agent, volunteer recommendation agent, admission probability analysis agent, and professional recommendation agent to complete the construction of the multi-agent collaboration architecture; S431. Designing the knowledge retrieval agent; In S431, the knowledge retrieval agent is designed as follows: Responsibilities: Responsible for autonomous retrieval and query of knowledge related to the content from the network and local knowledge base; Required information: Specific query content, transmitted by the supervisor agent through the Handoffs mechanism; Core capabilities and tools: The knowledge retrieval agent is a ReAct-style agent with autonomous reasoning and action capabilities, equipped with all the knowledge retrieval tools described in step S2; Work flow: Autonomous multi-step iteration, in each step, the knowledge retrieval agent determines the next action based on the current progress and long-term memory content, i.e., selects and calls the knowledge retrieval tool, and outputs all the knowledge retrieved by the task completion; S432. Designing the volunteer recommendation agent; In S432, the volunteer recommendation agent is designed as follows: Responsibilities: Recommend appropriate university and professional combination according to the examinee information; Required information: Rank, province, and selected course of the examinee; Core capabilities and tools: The volunteer recommendation agent is a WorkFlow-style agent with pre-defined process reasoning and action, without tools; Work flow: 1) Preliminary screening: Based on the rank, province, and selected course of the examinee, filter out the records with admission score lines within the pre-set floating range from the professional admission score line database; 2) Time-sensitive verification and update: Compare the preliminary screening results with the current university enrollment plan to exclude the specialties that have stopped recruiting this year, forming a preliminary list of volunteer candidates; 3) Personalized refinement: Use a large language model to further filter the preliminary volunteer list based on the user's explicit requirements, long-term memory of user information, and detailed information about the specialty in the current enrollment plan, and finally generate a highly personalized recommended volunteer table; S433. Designing the admission probability analysis agent; In S433, the admission probability analysis agent is designed as follows: Responsibilities: Analyze the probability of being admitted to a target university specified by the examinee, and provide corresponding enrollment recommendations; Required information: Rank, province, and selected course of the examinee, as well as the name of the target university; Core capabilities and tools: The admission probability analysis agent is a WorkFlow-style agent with pre-defined process reasoning and action, without tools; Workflow: 1) Data collection: query the target university's admission score line in the past two years, enrollment plan details, and the latest enrollment plan this year; 2) Trend analysis and comprehensive judgment: use a large language model, combine the dynamic trend of enrollment plans, the fluctuation of admission scores in previous years, and refer to the user information stored in the long-term memory described in step S3 for comprehensive analysis and evaluation, generate probability analysis results about the candidate's admission to the target university, and provide targeted enrollment recommendations; S434. Design a professional recommendation agent; In the S434, the admission probability analysis agent is designed as follows: Responsibilities: recommend learning majors based on the individual needs, interests, and long-term development plans of the candidates, and provide reasons for the recommendations; Required information: none; Core competencies and tools: the professional recommendation agent is a ReAct-style agent with autonomous reasoning and action capabilities, equipped with professional information retrieval tools. The professional recommendation agent has the ability to interact with users in multiple rounds of depth. Workflow: multiple iterations are performed autonomously, and at each step, the professional recommendation agent determines the next action based on the current progress and long-term memory content through internal large language model reasoning, including actively asking users, selecting and calling professional knowledge retrieval tools, and outputting task completion results.

6. A multi-agent based college entrance examination counseling system, characterized in that, A multi-agent-based college entrance examination counseling method for implementing any one of claims 1-5, comprising a multi-agent collaboration architecture (1), a retrieval tool (2), a knowledge base (3), and a memory bank (4); The multi-agent collaboration architecture (1) is connected with the retrieval tool (2) and the memory bank (4), and the retrieval tool (2) is connected with the knowledge base (3).

7. The college entrance examination counseling system based on multiple agents according to claim 6, characterized in that, The multi-agent collaboration architecture (1) includes a supervisor agent and an expert agent; The expert agent includes a knowledge retrieval agent, a volunteer recommendation agent, an admission probability analysis agent, and a professional recommendation agent; The knowledge base (3) includes a local knowledge base and a network knowledge base, and the local knowledge base includes expert knowledge and objective data; The memory bank (4) includes long-term memory and short-term memory.

Images